Sei sulla pagina 1di 34

Optimal Electrification

Planning Incorporating
On- and Off-Grid
Technologies: The Reference
Electrification Model (REM)
This article describes and demonstrates an optimization model and program for
electrification planning that identifies the lowest cost system designs for electricity access.
By P EDRO C ILLER , D OUGLAS E LLMAN , C LAUDIO V ERGARA , A NDRÉS G ONZÁLEZ -G ARCÍA ,
S TEPHEN J. L EE , Graduate Student Member IEEE, C AILINN D ROUIN , M ATTHEW B RUSNAHAN ,
YAEL B OROFSKY, C ARLOS M ATEO , R EJA A MATYA , R AFAEL PALACIOS , ROBERT S TONER ,
F ERNANDO DE C UADRA , AND I GNACIO P ÉREZ -A RRIAGA , Fellow IEEE

ABSTRACT | In many parts of the world, access to basic texts. This paper describes a computer-based optimization
electricity services remains a significant challenge. The status model—named the reference electrification model (REM)—
quo mode of electrification is central grid extension; how- which performs automatic electrification planning and is able
ever, in many areas, off-grid (OG) technologies like minigrids to identify lowest cost system designs to most effectively
(MGs) and standalone (SA) systems are more suitable for provide desired levels of electricity access to populations of
promoting electricity access under cost constraints. Unfortu- any given size. In doing so, REM determines the most suit-
nately, these opportunities are often overlooked due to the able modes of electrification for each individual consumer by
complexities of electrification planning, especially for large specifying whether customers should be electrified via grid
areas. Researchers have designed technoeconomic planning extension, OG MGs, or SA systems. For each system, REM
tools that can be scaled to cut through aspects of this supplies detailed technical designs at the individual customer
complexity and be fit to address different places and con- level. We have used this model in real planning activities in
sub-Saharan Africa and South Asia. The description of REM’s
capabilities is supported by case examples. REM stands apart
Manuscript received October 1, 2018; revised June 3, 2019; accepted from other electrification planning models because of its high
June 3, 2019. Date of publication July 12, 2019; date of current version
September 4, 2019. This work was supported in part by the MIT Tata Center for granularity and its capability to provide concrete plans for a
Technology and Design, in part by the Enel Foundation, in part by Iberdrola, wide range of geographical scales. Because of these benefits,
in part by the World Bank, in part by the Spanish National Plan of Research,
Development and Innovation, in part by General Electric, in part by Shell, in part REM has the potential to help rationalize electrification plan-
by the Shell Foundation, and in part by the German Corporation for International ning and expedite progress toward universal electricity access
Cooperation GmbH (GIZ). (Corresponding author: Fernando de Cuadra.)
P. Ciller, A. González-García, C. Mateo, R. Palacios, and F. de Cuadra are worldwide.
with the Instituto de Investigación Tecnológica, Universidad Pontificia Comillas,
28015 Madrid, Spain (e-mail: cuadra@comillas.edu). KEYWORDS | Electricity access; electrification planning;
D. Ellman, C. Vergara, C. Drouin, and Y. Borofsky were with the Universal geospatial planning model; off-grid (OG) electrification; opti-
Access Lab, Massachusetts Institute of Technology, Cambridge, MA 02139 USA.
S. J. Lee, M. Brusnahan, R. Amatya, and I. Pérez-Arriaga are with the mization planning
Universal Access Lab, Massachusetts Institute of Technology, Cambridge,
MA 02139 USA.
R. Stoner is with MIT Energy Initiative, Massachusetts Institute of Technology, I. I N T R O D U C T I O N
Cambridge, MA 02139 USA.
The lack of electricity access for populations in many
Digital Object Identifier 10.1109/JPROC.2019.2922543 low-income and developing countries is a major contem-

0018-9219 © 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

1872 P ROCEEDINGS OF THE IEEE | Vol. 107, No. 9, September 2019


Ciller et al.: Optimal Electrification Planning Incorporating On- and Off-Grid Technologies

porary issue, with complex social, ethical, environmental, combinations of electrification modes for a given study
economic, and technical dimensions. Despite significant region, including single building SA systems, isolated grids
investments by private and public organizations and wide- with local electricity generation or MGs, and extensions of
spread global progress, current efforts are moving too the existing distribution network (GE). REM performs this
slowly to meet the United Nations’ Sustainable Develop- task with a very high level of spatial granularity, producing
ment Goal 7, which in part targets to ensure universal detailed designs down to the individual consumer level.
access to electricity by 2030 [1]. It is imperative to “think It prescribes network infrastructure layouts, local genera-
big” when considering solutions for this problem if we tion configurations, and storage options. These capabilities
hope to make compelling progress toward achieving these are intended to allow planners to make better-informed
international goals. decisions about electrification modes, budget allocations,
“Thinking big” is not always easy, however. Viewing and bills of materials; ministries and regulators can get
the problem of universal electricity access from a system- quantitative support for policy design; and developers can
level perspective reveals significant complexities related to gain detailed insights into the potential for OG systems
humans, businesses, governments, and other institutions in a region. REM can also facilitate participatory plan-
across a number of cultural, geographic, and legacy con- ning approaches by providing references for least cost
texts. The implications of these considerations are com- electrification designs that can be evaluated by different
pounded by significant technical and economic dimensions stakeholders.
of complexity arising from the dynamic and networked REM considers the specific demand profile of each
nature of large-scale electricity infrastructure planning. customer (incorporating residential, commercial, and
Cost-optimal technoeconomic plans can serve as the foun- industrial loads) and determines the least cost grid/OG
dation for holistic planning approaches that further con- electrification plan by comparing a large number of
sider nontechnical factors. clustering alternatives through a combination of heuris-
The technical complexity of electrification planning can tic optimization, mathematical algorithmic optimization,
be exemplified in part by the range of electrification modes and simulation algorithms. These algorithms account for
available today. In addition to traditional electrification by estimated yearly weather conditions and demand pro-
grid extension, off-grid (OG) minigrids (MGs) and stand- files, targets of quality of electricity supply, the reliabil-
alone (SA) home systems have recently gained momen- ity performance of local distribution feeders, voltage and
tum as effective, alternative ways of providing access to capacity constraints of lines and transformers, catalogs
energy. Challengingly, institutions in developing regions of power system components for grid-extension and OG
today are generally ill-equipped to take full advantage of systems, any existing limits or targets in the use of fossil
these opportunities, as grid extension has represented the fuels or renewables or carbon emissions, and implications
status-quo electrification mode across the world for well of the topology of the terrain: forbidden and penalized
over a century. Rural electrification agencies, energy min- areas, use of prescribed paths such as roads or streets, and
istries, private investors, and entrepreneurs could benefit extra costs due to factors like altitude or the slope of terrain
from the knowledge of what the least cost electrification being considered.
modes and system designs are over their territories of REM has been applied to multiple real electrification
interest, to be used as a basis upon which to add further planning problems, ranging from cases representing small
considerations. areas with hundreds of customers to comprehensive analy-
The scale of the challenge associated with universal ses of entire countries with millions of them. A particular
energy access, the amount of information involved, and configuration of REM, named local REM (LREM), has
the diversity of options for intervention compel the use of been applied to many cases representing villages or small
computer planning models. As a result, research organi- regions to provide detailed electrification designs where all
zations have started to build and provide several of these customers are connected to the same MG.
automated tools, the number and quality of which have This paper is structured as follows. Section II reviews
evolved rapidly in recent years. Utilities, governments, and existing computer-based electrification planning meth-
development organizations have also responded to these ods. An overview of REM is provided in Section III.
opportunities by collecting information and improving the Section IV introduces the input data that REM needs for
availability of digitized and georeferenced data for their a case study. The algorithms of the model are illustrated
jurisdictions. in Sections V–VII. Section VIII describes how the solutions
This paper describes one of these computer models— that the model provides can be processed and presented
the reference electrification model (REM), developed by to the user in detailed reports. We present a case example
the MIT-Comillas Universal Energy Access Laboratory representative of the actual regional electrification plan-
(UEA Lab)1 —which we consider represents the state of the ning studies performed with REM, including descriptions
art in electrification planning. REM performs automated of the input parameters (Section IV), the lookup table
least cost electrification design; it determines cost-optimal of representative MG generation designs (Section V), the
clustering results (Section VI), and the final electrification
1 http://universalaccess.mit.edu plan (Sections VII and VIII). Section IX discusses the appli-

Vol. 107, No. 9, September 2019 | P ROCEEDINGS OF THE IEEE 1873


Ciller et al.: Optimal Electrification Planning Incorporating On- and Off-Grid Technologies

cations and limitations of the model, as well as enhance- Similar procedures are followed by other methods
ments that are still outstanding. Section X describes LREM, described in the literature [9]–[15] as well as software
i.e., the use of REM in a local MG context. The conclu- tools [16]–[18].
sions (Section XI), the acknowledgments, and references Other large-scale approaches compare the cost-
conclude this paper. effectiveness of village OG electricity supply to grid
extension by estimating tradeoffs between the cost of
local generation and grid-extension layout [19], [20].
II. S O F T W A R E S U P P O R T M O D E L S F O R Logiciel d’Aide à la Planification d’Électrification Rurale
E L E C T R I F I C AT I O N P L A N N I N G (LAPER) determines the best electrification mode for a set
of villages or settlements minimizing total investment and
As OG electrification has increasingly entered the main- operation expenses [21]. Challengingly, LAPER requires
stream of electrification planning, there have been substan- its users to provide an initial network that connects all
tial developments addressing the unique challenges that (or most) the settlements and then the model evaluates
both MGs and OG systems present. The work reported sequentially if they are worth disconnecting one line at a
in this paper contributes to two active areas of research: time. LAPER was used for rural electrification planning in
large-area electrification planning and single-system elec- Morocco [22].
trification planning. Large-area approaches deemphasize Network Planner [23] is an open-source model devel-
local accuracy and instead prioritize the identification of oped at Columbia University, which minimizes a cost
trends over large regions, while single-system approaches function to determine the least cost electrification mode
lend themselves to more detailed data collection and of a set of villages and the layout of the correspond-
scrutiny. REM contributes to both bodies of knowledge ing grid-extension network. Network Planner has been
since it is capable of producing very detailed cost-optimal used to provide assistance in electrification planning
solutions at all scales. processes in Nigeria [24], Liberia [25], and Ghana [26].
Parshall et al. [27] describe the heuristic that Network
Planner uses to calculate the grid-extension layout. The
A. Large-Area Planning
researchers use minimum spanning trees (MST) to eval-
Large-area planning models evaluate various electrifica- uate the cost of extending the medium-voltage (MV) grid
tion options for a region and aim to identify the “best” network to some area and to compare with that of elec-
(in some prescribed sense, typically least cost) delivery trifying that area with either diesel-powered MGs or solar
modes for each considered consumer. Given their broad home systems. While the model endeavors to make village-
geographical scope, they frequently have to rely on highly level decisions, the data in [27] are aggregated at the
aggregated or incomplete data about the layout and char- sublocation level (within an average area of 15 km2 ),
acteristics of the existing distribution network and targeted differentiating sublocations only as high or low income.
demand. Furthermore, Network Planner does not account for grid
Several software models have been developed to deal reliability, a factor that can play an important role in grid-
with this problem. Those based on geographic information extension decisions in rural areas.
systems (GIS) typically group consumers into geometric Another tool that follows a similar approach
cells that often represent villages, calculate the levelized is GEOSIM [28], which has been used in several
cost of electricity (LCOE) with grid extension and with OG countries in Africa and Asia, and the model has been
alternatives, and suggest least cost electrification modes. applied to electrification projects for the World Bank
One of the most widely employed computer-based mod- and the European Commission. GEOSIM sorts villages
els that make use of this general approach is the OpeN according to their Indicator for Potential Development,
Source Spatial Electrification Toolkit (OnSSET), which was selects a few of them to be “Development Poles,” and
developed by the KTH Royal Institute of Technology and clusters the villages around them using an algorithm
released as open-source software. OnSSET is affiliated with based on the Huff model [29] before determining the
several projects and studies [2]–[5]. OnSSET makes use of electrification solution for each village.
geospatial information such as proximity to roads and the There are also attempts to deal with this problem using
power grid, population density, and wind, solar, and hydro classical optimization techniques. Zeyringer et al. [30]
potential to determine appropriate electrification solutions formulate a mixed-integer linear programming (MILP)
for each cell. problem and group the consumers into square cells. How-
Another model that relies on this approach is ever, these cells are very large and solar is the only OG
IntiGIS [6], an evolution of the SOLARGIS tool [7] and technology considered. Abdul-Salam and Phimister [31]
further efforts [8]. IntiGIS also operates with cells and apply the hierarchical lexicographic programming to the
minimizes their LCOE to obtain the electrification solution, problem with three different objective functions, although
emphasizing the relevance of GIS technology in rural elec- the number of settlements that are electrified with grid-
trification when gathering data such as wind speed and extension designs is an input in this formulation. Finally,
solar irradiation. Abdul-Salam and Phimister [32] introduce a mixed-integer

1874 P ROCEEDINGS OF THE IEEE | Vol. 107, No. 9, September 2019


Ciller et al.: Optimal Electrification Planning Incorporating On- and Off-Grid Technologies

nonlinear programming (MINLP) formulation and com- 150 000 times from users of more than 190 different
pare the results with those in [27], although this countries [47]. Other models that are worth mentioning
approach assumes that the already-existing network con- are the Distributed Energy Resources Customer Adoption
sists of one single node to reduce the computational Model (DER-CAM) [48], which uses an MILP formulation,
burden. and iHOGA [49], which applies genetic algorithms.
These models work at the village, settlement, or cell Other models are also relevant in the context of local
level, instead of operating at the individual consumer level. OG generation. RETScreen makes use of feasibility analy-
Their network designs are based on geometric consider- sis [50], [51] and Hybrid2 uses simulation [52]–[54].
ations involving distances and demand sizes, instead of Hybrid2 is no longer supported. A comprehensive review
applying power flows and electrical constraints. They do of these software tools is provided in [55].
not optimize the generation designs of OG systems, and The network component of the MG design problem
they do not account for nonserved energy in their decision- obtains the network layout, given the location and demand
making logic. In summary, electrification planning models of the consumers and oftentimes the location of the
have traditionally made overly strong assumptions with generation site in addition. There is plenty of literature
unknown implications regarding the quality of their results about network distribution planning [56], where both
and recommendations. classical optimization techniques [57] and metaheuristic
The most direct approach to overcome this limitation methods [58], [59] have been applied. However, this
is to take the spatial granularity and the specification of problem has not been thoroughly studied from the rural
the demand patterns to the individual consumer level. electrification perspective.
In contrast with the above-mentioned tools, the REM Village power optimization model for renewable
model described in this paper: 1) operates at the individual (ViPOR) [60] designs distribution networks for OG systems
consumer level instead of using aggregate villages or cell with a simulated annealing algorithm [61], although this
representations; 2) calculates network designs considering software is currently unsupported. There are other meth-
electrical constraints and specifications of real equipment; ods for rural electrification network design [62] but they
3) incorporates topography in the optimization process; do not guarantee feasibility with respect to the technical
and 4) obtains generation designs using optimization tech- constraints of electric power systems.
niques and simulation instead of rules based on a priori- Mateo-Domingo et al. [63] describe the reference net-
determined analytical expressions. An early version of work model (RNM), which is a software tool for large-
REM is described in [33] and further developments are scale network design developed at IIT-Comillas in Madrid,
presented in [34]. This paper presents the current level Spain, which can also be used in the design of the network
of development of REM, which is still being enhanced in component of MGs. This model designs the minimum-
several aspects. cost network that meets demand under quality-of-service
specifications, using a user-provided catalog of equipment
B. Local OG Electrification Planning to specify distribution infrastructure down to the individ-
Single-system design models are particularly useful ual consumer level. A key feature in the development of
once a site has been identified for OG electrification, RNM pertains to the high levels of scrutiny it received: its
but they are also an integral part of large-area elec- results were validated by the Spanish distribution utilities;
trification planning processes. Local models have to RNM was then accepted by Spanish regulators as a
solve two quasi-independent problems: the design of the decision-support model to determine appropriate remuner-
local distribution network and that of local generation ation figures for electric power distribution. RNM has been
systems. used for this same purpose in several other countries and
The general methodology for designing a local gener- in many technical studies.
ation system consists of determining candidate designs REM uses RNM as a submodule to calculate network
and evaluating their performance. A variety of meth- designs for MGs and grid extensions, as we describe in this
ods has been described in the literature for the optimal paper. When applied to large-scale electrification planning,
selection of local generation and storage. Upadhyay and REM has to evaluate MG configurations numerous times,
Sharma [35] and Luna-Rubio et al. [36] classify the main each of which requires the design of generation assets
methods of solving this problem. Some of them apply and network layouts using RNM. REM can also be used
classical optimization techniques such as linear program- for smaller scale MG design as LREM (its “local” mode).
ming [37], [38] or metaheuristic algorithms [39]–[44]. This configuration optimizes the generation design of a
The moderate size of the local OG electrification problem given MG, its associated hourly dispatch, and the net-
allows employing multicriteria approaches, involving other work layout. While doing this, LREM considers investment
objective functions that go beyond costs, such as emissions and operation costs plus penalties for nonserved energy.
minimization [45]. Li [64] describes the LREM functionality and applies it to
A few mature models have exclusively focused on the village of Karambi in Rwanda. Brusnahan [65] and
the local supply generation problem. The best known Cotterman [66] also perform MG analyses using LREM in
is HOMER [46]. This tool has been downloaded over villages located in India, Nigeria, and Rwanda.

Vol. 107, No. 9, September 2019 | P ROCEEDINGS OF THE IEEE 1875


Ciller et al.: Optimal Electrification Planning Incorporating On- and Off-Grid Technologies

It can be concluded from this review of large- e) dispatch strategies of local generation and stor-
area and local OG electrification planning models that age, as well as for any applicable demand
REM stands out from previously existing approaches by management schemes;
providing system designs at any geographical scale, work- f) desired reliability of supply.
ing at the individual customer level, employing full rep- When evaluating the cost of the internal network of
resentations of each customer’s hourly demand patterns, an MG or of a grid extension, REM has to design the
respecting the physical laws and constraints of power minimum cost network that meets all prescribed techni-
systems, explicitly modeling reliability targets and costs cal requirements. This network-design process has to be
of unserved energy, accounting for the consideration of done numerous times (many alternative subsystems or
topographical characteristics, and employing optimization clusters), even for problems of moderate size. For this task,
methods to find least cost combinations of electrification REM employs the greenfield network-design software
delivery modes. called RNM, as mentioned previously. RNM employs equip-
ment from a prescribed catalog, designs networks that
meet electrical constraints, and accounts for topological
III. R E M O V E R V I E W
features in the considered region: slope and altitudes of
REM finds the electrification plan that meets a given esti- the terrain, forbidden and penalized zones, and adminis-
mated demand for individual customers in a territory at trative data such as village boundaries (different technical
minimum total cost while satisfying the power system tech- requirements may apply in rural, semiurban, and rural
nical constraints as well as other user-defined constraints areas, for instance).
regarding the reliability of supply, generation mix, admin- When addressing the different procedures embedded in
istrative requirements, and priorities or limits regarding the determination of an electrification plan, REM uses a
modes of electrification (e.g., a preestablished target for mix of established mathematical optimization algorithms
grid extension or a limit to the number of households to and heuristic methods. REM seeks the minimization of a
be supplied with MGs). The demand is defined at very combination of actually incurred costs and social costs.
high levels of spatial and temporal granularity: hourly In fact, incurred costs are related to investment, opera-
demand patterns for each individual customer (building). tion, maintenance, and management activities, which can
In general, the electrification plan consists of a combina- be directly quantified. Social costs measure the loss of
tion of electrification modes: extension of the existing grid, utility or welfare to the end consumers resulting from
OG MGs, and SA systems. Systems prescribed may range poor service quality and potential limitations in electricity
in size from small home solar kits to large and sophisti- utilization associated with their mode of connection.
cated MG and grid-connected systems for schools, clinics, The following outputs are typically obtained when
mines, and other commercial, industrial, or institutional REM is applied to some specific territory.
facilities. 1) The optimal groupings of individual consumers into
When deciding among the alternative electrification electrification cluster so that total system costs are
modes REM may consider the following factors. minimized. These clusters may denote groups of
1) For grid connection: customers to be connected to the same MG systems,
a) layout and technical characteristics of the groups to be connected to the existing grid, or clus-
existing grid and the cost of extending the MV ters of single customers to be supplied with
and low voltage (LV) networks; SA systems.
b) required upstream supply costs of generation 2) The optimal generation mix and network layout for
and networks; each of the OG MGs.
c) reliability of the existing grid; 3) The optimal network layout for each cluster that will
d) catalog of available equipment with technical be connected to the grid.
and economic characteristics and applicable 4) A detailed description of the optimal plan with infor-
grid codes. mation pertinent to decision-making including total
2) For OG supply systems: cost breakdowns, expected reliability data, GIS files
a) costs of local generation, with costs and avail- specifying network layouts, generation and storage
abilities of local resources including solar and specifications, bills of materials, summary charts,
diesel. Biomass, wind, and minihydro will be georeferenced maps of system designs, and reports
included in future versions; in text and spreadsheet formats.
b) costs of extending the local network to every The standard REM process is structured into five
customer with either dc or ac components; sequential stages, which will be explained in detail in
c) customer preferences between SA systems Sections IV–VIII with the help of a case example.
and MGs;
d) extra costs for making MGs “grid-compatible,” A. Data Preparation
so they have the flexibility to connect to the This stage includes partially automated tools used to
main grid at a future date; generate required inputs. For example, a computer vision

1876 P ROCEEDINGS OF THE IEEE | Vol. 107, No. 9, September 2019


Ciller et al.: Optimal Electrification Planning Incorporating On- and Off-Grid Technologies

system that extracts building footprints from satellite The modular structure of REM allows alternative strate-
images is used to identify customers when such infor- gies for these steps. For instance, instead of the bottom-
mation is not readily available and doing so manually is up clustering approach used originally in REM pertaining
too expensive or time-consuming. Other models employ- to (C), a full grid-extension solution can be generated
ing machine-learning concepts classify customers as either initially, with subsequent and systematic disconnection
electrified or nonelectrified given correlated geospatial algorithms to identify OG solutions (a “top-down” cluster-
features. Subsequently, customers may be characterized ing approach) as described in [67].
as belonging to one of several customer types and to be
of a certain size. Each customer can then be assigned a IV. D A T A P R E P A R A T I O N
characteristic demand profile derived from empirical data; A REM study case is characterized by a set of files
demand profiles may also be estimated by aggregating describing the existing and potential electrical demand
typical consumption patterns from constituent electrical in a region, physical characteristics of the region, exist-
appliances. ing infrastructure, available electricity supply technologies,
and information regarding design preferences.
B. Minigrid Generation Design
The different types of files are classified into the follow-
Optimal generation and storage have to be determined ing categories.
for each potential grouping of customers that are eval- 1) Local Information: Information particular to a given
uated, considering its aggregated load profile, the avail- analysis region.
able catalog of system components, and some preselected a) Demand (buildings, types of customers, and
dispatch strategy. In order to save the computation time, demand patterns).
the designs for some number of representative combina- b) Existing network (line segments and their
tions of consumers are solved in advance and stored in a types and transformers).
lookup table so that future designs for other combinations c) Topography and geography (local solar input
of customers can be quickly obtained by interpolation. characterization, elevation, slope of the ter-
Although the lookup table represents only approximate rain, and forbidden and penalized areas).
designs, it is used in the clustering process to speed up the 2) Equipment Catalog: Technical and economic parame-
estimation of local generation cost for each candidate MG. ters of available electrification components for both
REM considers solar panels, batteries, and diesel genera- distribution network extension and OG electricity
tors as available technologies for the generation mix of the supply.
OG systems. a) Networks (types of lines and transformers that
can be used, at the different voltage levels).
C. Clustering b) OG generation (solar panels, diesel generators,
The customers are grouped into a hierarchical struc- batteries, and power electronic equipment).
ture of OG and grid-extension clusters. Both cluster- 3) User Options: User options are specific to the partic-
ing processes consist of bottom-up greedy algorithms ular case study and are structured into eight blocks
(the term “greedy” means here that it is based in irre- of related configurations as follows.
versible decisions seeking immediate and local benefits, a) Workspace definition.
not guaranteeing global optimality) that join customers b) Electrification problem.
into groups if the expected cost of being connected is c) Case-study specification.
lower than the expected cost of being electrified separately. d) Electrification criteria.
Here, robust cost estimations (using partial and uncertain e) Financial, business, and general cost parame-
information) are critical to make sound decisions. ters.
f) Topography/geography.
D. Final Designs g) Network design.
REM explores the structure of clusters obtained in the h) MG generation.
previous stage and calls RNM to evaluate precise network Fig. 1 shows how the parameters of the “User Options”
designs and costs. It then determines the optimal combina- file are organized in blocks, and how the different blocks
tion of SA systems, MGs, and grid extensions. During this contain references to the rest of input data files used
process, proposals for mergers of clusters and proposals for by REM.
grid connections of clusters are accepted or rejected based A detailed description of the contents of different blocks
on design cost. of parameters and the types of data files they refer to
can be found in [68]. Here, we shall explain the data
E. Postprocessing and Reports requirements and typical sources for each one of the major
REM generates several reports containing plots, tables, blocks, with the help of case examples of the application
geospatial data files, and other items, providing all infor- of REM.
mation of interest on the cost-optimal designs that REM Recent experience in the utilization of REM for diverse
has found. situations has shown that at least 80% of the time

Vol. 107, No. 9, September 2019 | P ROCEEDINGS OF THE IEEE 1877


Ciller et al.: Optimal Electrification Planning Incorporating On- and Off-Grid Technologies

OSM provides free and open detailed building and street


annotations using a crowdsourcing-based approach:
millions of participants conduct ground-based surveys and
perform manual labeling on the top of aerial imagery [69].
While OSM rivals proprietary sources in terms of size
and granularity of its map data, the quality of its data
is inconsistent [70]. The availability of OSM’s data is
limited in developing countries and this is especially
true in rural areas. Automatic methods are required to
procure complete building data sets in these regions for
large areas, without performing resource-intensive surveys
Fig. 1. Organization of REM inputs. and manual labeling. Convolutional neural networks
(ConvNets) are the most accurate and effective automatic
and effort required to produce an electrification plan is methods to date for building footprint extraction from
spent in data gathering and preparation. This process satellite and aerial imagery.
includes the following: 1) identifying the existence of data; Here, we describe a basic system for georeferenced
2) identifying and contacting its sources; 3) defining what building footprint extraction based on fully convolu-
is needed; 4) obtaining permission to use proprietary tional networks (FCNs) for semantic segmentation [71].
data or eliminating confidential information; 5) collecting Although FCNs are no longer state of the art for semantic
data; 6) verifying and cleaning it; and 7) making the data segmentation tasks in computer vision, the general method
ready to be used in REM. In some cases, the required is well documented, easy to use, proves to be effective for
data are already available and only needs to be adapted building-footprint extraction, and serves as a benchmark
to the required format for REM. In other cases, specialized for further research. Extended descriptions of each of these
algorithms have had to be developed to obtain the data by topics are given in [72], and a more detailed review of
combining information from different sources. these methods in [68].
The procurement of training data for building extraction
A. Determining Building Locations and Building is often a time and resource-intensive endeavor. The mod-
Electrification Status els described for semantic segmentation require training
data where annotations are captured at the pixel level.
Planners need to know where buildings are located
While the OSM data set provides manual labels that
in order to plan on how to provide them with electric-
may seem promising for model training, its sparsity and
ity. They also need to make determinations on whether
inconsistency in developing countries and its potential
these buildings are already serviced with electric power.
misalignment with satellite imagery decrease its usefulness
In this section, we discuss three steps along our automatic
for our application. Due to these considerations and the
data procurement pipeline for REM: 1) building footprint
fact that our building extraction models may not be highly
extraction from satellite imagery; 2) load localization; and
generalizable across regions, we find it is often beneficial
3) electrification status estimation. This procedure pro-
to procure training data for regions of interest using anno-
duces approximate inputs for REM quickly and with suf-
tation tools for binary image classification and for drawing
ficient quality for many large-scale plans. Ongoing efforts
polygons, as described in [72].
are aimed at improving each of these steps.
Our results [68], [72] demonstrate that contemporary
Choices on whether to pursue such automatic methods
ConvNet models for semantic segmentation are capable
as opposed to manual ones depend on resource constraints
of automatic building footprint extraction from satellite
and study requirements. Automatic methods afford the
imagery and that image quality is highly important. The
ability to scale technoeconomic analyses to large regions
ultimate benefit of using these tools is that they can scale
at low cost; however, they are generally less accurate
massively in the presence of satellite imagery of adequate
than manual methods (e.g., human annotations, ground
quality and with enough training data. Investments in
surveys, etc.). In many cases, it is appropriate to pur-
improved image quality and training data should be made
sue hybrid automatic-manual approaches and iterative
with consideration of alternatives and available resources.
methodologies.
1) Building Footprint Extraction: Building footprint 2) Load Localization: Load localization refers to the
extraction involves the pixelwise classification of process of identifying connection points for individual
satellite and aerial imagery. Building pixels of high- buildings and characterizing the potentially latent load
resolution imagery are classified as either belonging to profile associated with each of them. This is challeng-
buildings or to background. Building footprint extraction is ing because the process of converting building footprint
commonly done using both manual and automatic extractions (which may reflect multiple buildings within
methods. The largest and most notable manual building a single contiguous group of pixels) to individual cus-
labeling endeavor is the OpenStreetMap (OSM) project. tomer points is nonobvious and because information for

1878 P ROCEEDINGS OF THE IEEE | Vol. 107, No. 9, September 2019


Ciller et al.: Optimal Electrification Planning Incorporating On- and Off-Grid Technologies

Fig. 2. GP results corresponding to the survey data zoomed into the capital city, Kampala [72].

determining load profiles is likely not present in satellite and characterization are an important area for contin-
imagery alone. While the problem of load profile assign- ued research.
ment has been largely unstudied and currently requires a
number of assumptions, a few potential solutions exist for 3) Electrification Status Estimation: Electrification status
identifying individual buildings [68], [72]. estimation corresponds to the classification of the iden-
Connection points for individual buildings may be tified buildings as currently electrified or nonelectrified.
obtained by defining polygons that characterize building Information regarding the current electrification status
shapes, producing rasters for building boundaries, or by of buildings is imperative for planning activities. This
sampling points from contiguous groups of pixels denoting information allows planners to avoid designing duplicative
building footprints. The efficacy of different approaches infrastructure and ensure their plans meet the public’s
for identifying connection points on individual buildings needs. Electrification status information also enables the
has not yet been compared for master planning and is the assessment of technology choices; it can inform the use
subject of future work. The choice of approach is likely also of technoeconomic models such as REM that determine
dependent on the method used for characterization of load the attractiveness of OG technologies relative to modes
profile, which itself may be dependent on data availability, of grid extension. Although distribution companies in
generalizability, and other factors. Approaches for estimat- developed parts of the world generally have a wealth
ing load using building polygon or boundary information of digitized infrastructure data, their counterparts
may rely on characterizing relationships between elec- in developing regions are consistently less informed.
tricity consumption and building footprint area, location, Numerous distribution companies in regions with low
and neighborhood building density. If a linear relationship rates of electricity access lack adequate structured
is assumed between electricity consumption and building information regarding their LV distribution lines. Although
footprint area, the connection point sampling approach data collection and digitization efforts have commenced
proposed in [72] may be appropriate. This assumption in some of these regions, requisite data on electrification
would imply that buildings with twice the footprint area status are still largely missing.
would have twice the load and can effectively be modeled Only a few approaches for electrification status esti-
as two separate neighboring buildings. Load localization mation have previously been reported in [68] and [72].

Vol. 107, No. 9, September 2019 | P ROCEEDINGS OF THE IEEE 1879


Ciller et al.: Optimal Electrification Planning Incorporating On- and Off-Grid Technologies

algorithmic constraints, etc.) but population densities and


applicable population boundaries are known, it is possible
to estimate building positions by assuming uniform distrib-
utions of building densities and randomly selecting points.
Fig. 3 shows the difference between manually annotated
building locations and estimated building locations for the
village of Zugu in Nigeria. This procedure is less accurate
in the case of nonuniform buildings distributions, which is
Fig. 3. Left: manually annotated coordinates of all buildings,
more likely to occur when population density and bound-
visualized using Google Earth. Right: randomly sampled building ary information is known at less granular scales.
locations, distributed within the polygon line that delimits the Comparing REM runs using manually annotated build-
boundary of Zugu (Rungan-Gazo) in Nigeria. ing locations with those using building locations that
were randomly sampled according to uniform distribu-
tions within village boundaries has shown similar results
in terms of electrical network design for the metrics of
We have performed experiments with multiple model
interest in planning. For the same number of buildings
types for building level and ∼1-km resolution in the
modeled, there is a small difference in the total length and
estimation of electrification status, comparing results
type of the power lines prescribed, but the costs for the
for models based on logistic regression (LR), Gaussian
generator and transformer are the same. These preliminary
Processes (GPs), and probabilistic graphical models
results suggest that it is still possible to produce useful
(PGMs) for case studies in Uganda. The GP approach
electrification plans in areas with poor quality satellite
demonstrates improved performance metrics relative to LR
imagery if population and village boundary information
due to its ability to capture spatial correlations in electrifi-
is known. In addition, if there are accurate estimations of
cation status. However, we hypothesize that LR approaches
population growth, it may be possible to produce electri-
may be able to better incorporate multimodal data and
fication plans for expected future populations in a similar
generalize to regions outside of those close to direct survey
way. Nevertheless, when it comes to real implementations,
measurements. Ongoing work on a PGM model promises
the design of the local network for each cluster has to
to combine multimodal and multiscale features (such as
be based on highly accurate location information for each
aggregate census statistics, existing grid information, pop-
building to be electrified.
ulation density, nighttime lights values, and satellite image
Population data can typically be obtained from the
features) in addition to capturing spatial correlation [72].
region’s most recent census as long as it may be considered
Fig. 2 shows the example output for a GP classifier
accurate and up to date. It is also possible to obtain
trained with survey data and zoomed into Uganda’s capital
future population estimates from demographic studies.
city, Kampala. Such electrification probability maps can
However, the most valuable information for building loca-
be compared to expert elicitations and used to inform
tion is population density that can be obtained from
master planning efforts. For use with REM, we can use
several geospatial data providers, such as LandScan [73]
probabilistic models such as GPs for electrification status
and the high-resolution settlement layer (HRSL) [74].
estimation by simulating electrification data sets [72].
Estimating household locations based on the freely avail-
4) Combining Information From Different Data Sources: able HRSL information is significantly faster than per-
Sometimes requisite information for areas in need of forming building extraction from raw satellite imagery;
electrification is inaccurate, outdated, or not available. however, it is generally less accurate.
Examples include poor-quality imagery due to blurred
images, clouds, or other artifacts. Even if the quality is B. Demand Profiles for Each Type of
acceptable, readily available imagery may be outdated Building/Load
and not reflective of the situation on the ground. There- One of the most important inputs to an electrification
fore, it is necessary to collect information from differ- model is the expected demand for every customer mod-
ent data sources to corroborate the validity of buildings eled. In the case of REM, customers are categorized into
data and to make estimations. Relevant features include broad archetypes with unique hourly consumption trends.
population and population density, which correlate with The demand of each archetype is modeled either as a single
the number of buildings and its distribution. Since image power utilization pattern or as a linear combination of
processing algorithms can be adjusted in sensitivity to several patterns.
detect a given number of buildings, if the results do The decision to use a linear combination of demand
not match expected population values, it is possible to patterns has two primary reasons as follows.
adjust sensitivities to improve the accuracies of these 1) There is limited information available on how
methods. specific types of customers are expected to use
If buildings cannot be identified accurately using images (or have previously used) electricity. As such, the lin-
(e.g., due to image quality issues, tree or cloud cover, ear manipulation of data can be a useful method of

1880 P ROCEEDINGS OF THE IEEE | Vol. 107, No. 9, September 2019


Ciller et al.: Optimal Electrification Planning Incorporating On- and Off-Grid Technologies

extrapolating reliable load information to different Table 1 Classification of Customers. Source: African Utilities

groups. For example, an assumption can be made


that large households consume three times as much
as small residential customers.
2) To reduce computation time. During the clustering
process, REM calls on a lookup table that includes a
set of typical electrical generation mixes. The model
will then interpolate capacities from the table and
design the future electrification system. To simplify
this process, only a set number of demand pro-
files (i.e., patterns) are allowed. Linearly combining
these patterns allows for more granular demand
projections.
Each demand pattern is characterized by several
demand samples. The samples provide alternative con-
sumption chronologies to account for the uncertainty
in customer usage, for each profile. Each demand
sample specifies two consumption series: critical and
noncritical electricity demands. The series consists
of 8760 values, one for every hour in a year. Demand
profiles identify the minimum amount of electric-
ity needed to satisfy the customer’s load at 100%,
24 h a day. However, REM is not constrained to meet all
the demanded capacity. Instead, it optimizes for cost and,
therefore, may choose to reduce the reliability of the elec-
samples. It should be noted that demand samples are
tricity to avoid significantly scaling system components.
precomputed and act as an input to REM.
User-defined costs of nonserved energy (CNSE), for critical
How people consume electricity varies as temperatures
and noncritical demands, translate these supply failures
change and as sunlight fluctuates. Therefore, climate data
into social costs.
are also factored into expected hourly demands. During
Several methods can be employed to obtain the demand
the analysis of appliance usage, the following factors
profiles that REM needs, as it is described in the following
are considered: 1) the power that each appliance needs;
sections.
2) criticality, i.e., whether the appliance is critical for daily
1) Demand Profile Methodology Based on Appliance use; 3) average number appliances that customers in each
Utilization: One way to obtain demand patterns for a profile are expected to own; 4) probability that a cus-
given profile is by estimating the power of the appliances tomer in the respective demand profile owns the appliance;
expected to be used [33] and by combining them randomly 5) average number of hours the appliance is typically used
to generate a representative, unbiased set of demand for; 6) enabling factor determining that the appliance will
not be used if some condition is not satisfied (e.g., a fan
will not be used if the temperature is lower than 29 ◦ C);
and 7) variability factors in the use of the appliance or in
the household global level of activity. Finally, the profiles
are interpolated across an entire year’s worth of usage.
Fig. 4 shows an example of residential demand including
lighting, television, and fan usage [33]. The input sources
include census data, peer-reviewed studies, surveys, and
demand data captured on the field.

2) Demand Profiles From External Sources: When field


survey data on expected power consumption and appliance
usage are limited, trivial, or not available, commercial
and feeder records can prove useful at estimating nat-
ural demand (i.e., demand of similar customers in sim-
ilar socioeconomic and geographic situations). Some of
the sources used are: 1) commercial information, pro-
viding average energy consumptions per month or year,
Fig. 4. Example of demand samples based on the use of peak load, and installed limiters for different customer
appliances. types; when information about the nature of demand is

Vol. 107, No. 9, September 2019 | P ROCEEDINGS OF THE IEEE 1881


Ciller et al.: Optimal Electrification Planning Incorporating On- and Off-Grid Technologies

available (e.g., classification of households with different


peak demand levels, community, and productive uses),
different hourly load patterns per customer type can be
calculated as an input for REM and 2) feeder infor-
mation when distribution companies can make available
hourly records of the energy supplied by different feeders;
these data are often supplemented with the number (and
archetype) of customers per feeder; these records can then
be disaggregated downstream to the individual house-
hold or commercial level, by assuming similar customer
characteristics. It is especially helpful to have information
Fig. 6. Demand profile samples at the final year of demand
about dedicated feeders for agriculture, commercial, and
growth.
industrial customers. Feeder information is also useful to
determine the reliability of the supply, and how it relates
to the type of customers on the feeder. demand). The data was provided by African utilities, with
For planning purposes, decision-makers may decide to additional consideration of results from a field study devel-
deviate from using the natural or expected demand for oped by the UEA Lab in the Rwandan village of Karambi,
certain types of customers. For instance, they may decide Gicumbi district, and detailed in [64] and [75]. Satel-
to plan for a certain target or tier either above or below the lite imagery alone was not sufficient to assign customer
natural demand. types to consumers. It was necessary to interact with the
A case example has been prepared by the REM team and African utility to determine which loads were community
can be assumed representative of a rural area in many customers or productive customers. The remaining loads
sub-Saharan African countries. The considered area has were assumed residential, and their type (1 or 2) was
65 km × 40 km. All the data, including the layout of the estimated with census data.
existing network, the location of buildings, the catalog of Fig. 5 shows a sample of hourly profiles obtained from
components, and the demand profiles, have been obtained the Karambi field study, which have been used to compute
from actual studies in some of these countries. the hourly demand curve for a whole year considered
Employing the methods described in Section IV-A, it was for this case. It resembles the average use of appliances
determined that the area contains 52 709 nonelectrified reported by the different customer types in the canvased
consumers, most of them rural but some of them periur- village in week and weekend days, also considering
ban. There are 17 different customer types (see Table 1), seasonality in daylight hours (e.g., for the hourly profile
including several residential customers (types 1 and 2), of household lighting) and temperature (e.g., for the use
community customers (types 3–10), and productive cus- of fans).
tomers (types 11–17). Their respective energy demands The objective is to design the cost-optimal electrifica-
have been estimated considering the consumption of sim- tion plan for the demand that is expected to be reached
ilar loads in other rural areas already electrified (natural in 10 years from the present time. The demand pro-
file sample derived, represented in Fig. 6, results from
applying a constant growth rate of 2% per year to the
estimated present demand over the ten years’ period
(22% of accumulated growth). The figure shows the result-
ing profiles of the daily critical and noncritical demands.

C. Existing Distribution Network


REM considers three different electrification modes: SA
systems, MGs, and grid extensions. Unlike the former two
modes of electrification, grid extensions require REM to
both models and interpret already existing infrastructure:
specifically, the location and capacity of connected trans-
mission and distribution lines.
A few simplifications were included when creating REM
and RNM. For instance, both models (REM and RNM)
can apply only one voltage value to each voltage level in
the distribution network (i.e., HV, MV, and LV). Existing
Fig. 5. Sample demand profiles for domestic, community and electrical grids, however, may use several voltage levels in
productive loads in Karambi village [64]. The x-axes are measured in different feeders. Typically, MV lines are set at three-phase
hours and the y-axes are measured in kilowatts.
12.47 kV and LV lines are set at 3-phase 0.4 kV. The use

1882 P ROCEEDINGS OF THE IEEE | Vol. 107, No. 9, September 2019


Ciller et al.: Optimal Electrification Planning Incorporating On- and Off-Grid Technologies

of other voltages or number of phases must be emulated


using equivalent components.
REM also requires its users to define explicitly where the
potential connection points are (i.e., extensions). In our
scenarios, this only includes the MV lines. In general,
the per-unit wholesale cost of energy ($/kWh) could Fig. 8. Examples of catalog components: lines and transformers.
be different for each connection point in an area of
study if there was information to estimate the required
reinforcement infrastructure. The cost of these extra The reliability of feeders can either be based on actual
investments is assumed to be proportional to the amount measurements or estimated through customer surveys.
of energy provided. Fig. 7 shows the location of the 52 709 nonelectrified
To characterize the existing grid, the user can obtain the consumers and the existing MV distribution power grid in
electrical details from available databases or can manually the considered case example. The main grid is assumed
draw it by inspection of images and line diagrams. In REM, to be available at any time with a probability of 90% and
the existing MV network is represented by a set of line the wholesale energy cost at the MV distribution level
segments, each one defined by two coordinates. is 0.10 $/kWh at any connection point. Although REM
REM uses the candidate MV connection points to eval- allows for the hourly specification of reliability and the
uate the network cost (and layout) of a grid-extension specification of different reliability and electricity costs for
solution for all consumer clusters considered. In order to each feeder, the same constant reliability and electricity
do so, REM chooses an appropriate subset of representative cost values were assumed for all supply connection points
MV points for each cluster as potential connection points; in this case study.
it then submits the problem to RNM, which is free to use
any of them to provide an optimized layout. D. Catalog of Components: Networks and
Each MV segment is allocated an energy price in $/kWh, Generation Sites
which could depend on the supply connection point to
The importance of the cost and technical constraints
account for differences in the locational price of deliv-
associated with the physical assets in the distribution net-
ering electricity. REM does not deal directly with the
works requires an accurate technoeconomic characteriza-
upstream transmission network and generation but this
tion of the equipment that can be used in the electrification
feature allows the user to translate different generation
plan.
costs into local wholesale energy costs at MV distribution 1) Catalogs of Network Components:
level. Although the price of delivering electricity is only a) Lines and transformers are described in
considered to compute the cost of energy losses in RNM, terms of electrical parameters (power and
this price is critical in REM decisions to determine cost- impedances) and cost parameters (investment
optimal supply designs. and maintenance).
Each MV segment also exhibits a specific reliability level. b) Separate catalogs can be defined for grid
The reliability level is inputted as a vector of 24 elements extensions and MGs. This is a strategic deci-
that characterizes the hours of a representative day. The sion since the advantages of grid-compatible
reliability level must be within the range [0, 1], where MGs should be weighed against the use of low-
this value indicates the fraction of demand that is sup- cost isolated networks.
plied during that hour. “0” means that no demand can c) REM assumes standard three-phase networks
be supplied, while “1” means that all demand is supplied. with only three voltage levels (HV, MV, and LV).
The data in the catalog have to be verified so that
economies of scale are respected. This is mandatory to
avoid difficulties in REM’s optimization processes. The
components must be ordered by increasing capacity, and
the cost/capacity ratio must decrease or remain constant.
Care must be exercised in curating this catalog since mix-
ing products from different vendors or technologies may
cause inconsistencies.
Fig. 8 shows the portions of a line catalog and a trans-
former catalog, which exhibit the level of detail handled
by REM.
1) Catalogs of Generation-Sites Components:
a) The standard MGs modeled are ac and three-
phase. Solar kits in dc are modeled separately
Fig. 7. Nonelectrified consumers (green dots) and MV power grid and considered only in the final electrification
(black lines).
phase of the algorithm.

Vol. 107, No. 9, September 2019 | P ROCEEDINGS OF THE IEEE 1883


Ciller et al.: Optimal Electrification Planning Incorporating On- and Off-Grid Technologies

costs, and energy costs. To account for the time value


of money, REM discounts future expenditures based
on an appropriate discount rate for each technology.
By allowing different technologies to be discounted with
independent discount rates, REM accounts for the diverse
ownership structures and risk profiles that are possible in
the studies to be performed with the model. For instance,
utility-owned grid-extension projects should have a
Fig. 9. General structure of a generation site in REM. lower discount rate than privately owned solar home
systems.
REM is a “static optimization planning model,” which
b) The local generation site has a flexible determines the minimum cost solution for just a future
structure. Its maximum complexity is shown snapshot situation, i.e., one year in the future. Due to the
in Fig. 9: diesel generator, photovoltaic (PV) wide range of equipment lifetimes used in the electrifica-
panel, battery, charge controller, and inverter. tion space, an annuity for that future year is computed
Five simpler designs are allowed: 1) only a for each technology. This allows to jointly accounting for
diesel generator; 2) only a PV panel; 3) PV + shorter lived products, such as solar home systems, and
diesel; 4) PV + battery; and 5) diesel + battery. assets with longer economic lives, such as lines and trans-
c) For PV panels and batteries, only two options formers in the distribution networks.
for each (defined by their technical characteris- For expenditures occurring on a nonannual basis,
tics and size) are allowed in the current version the expenditure is converted to a yearly annuity A
of REM: the first is more adequate for small
SA systems and the second for large generation
A = C · r/(1 − (1 + r)−L ) (1)
sites.
The above-mentioned requirement about economies of
scale (as in lines and transformers) also applies to diesel where C is the periodic expenditure, r is the discount
generators, charge controllers, and inverters: all compo- rate, and L is the period. Direct monetary costs include
nents must be ordered by increasing capacity, and the investment, operation and maintenance (O & M), manage-
cost/capacity ratio must decrease or remain constant. ment, and energy costs:
Efficiency should increase or remain constant with increas- 1) Investment costs, or capital expenditure (CAPEX),
ing size. are determined directly from the system design and
Fig. 10 shows the portions of a diesel generator catalog associated cost catalog. For grid extension, this is
and a battery catalog. often dominated by the cost/km of the distribution
In the case example, the generation and network cat- network, whereas for MGs, the $/kW for solar PV
alogs have been built by blending data gathered by and $/kWh for battery storage are often the most
members of the MIT/Comillas Universal Energy Access significant components of the total cost of supply.
Lab team from several studies conducted in sub-Saharan As mentioned above, these CAPEXs are converted to
African countries and India. annuities to compare projects in the static optimiza-
E. Cost Drivers and Financial Models tion procedure performed by REM.
2) Each equipment type is assigned an annual
REM considers direct monetary costs and indirect soci- O&M cost based on the local equipment charac-
etal costs in the economic evaluation of electrification teristics and necessary expenditures to maintain
plans. equipment in working condition. For distribution
1) Direct Monetary Costs: The direct monetary lines, this is defined by ($/km)/year, but for trans-
costs include both initial investments and on-going formers, batteries, or diesel generation sets, this
expenditures. The costs are categorized as investment is defined simply as $/year for a given piece of
costs, operations and maintenance costs, management equipment.
3) Annual management costs differ by system type
and size due to the nature of different pieces of
equipment and different business structures. When
modeling the management cost for grid-extension
projects and solar home systems, REM assumes
that economies of scale have been reached, and
the marginal management cost of each additional
customer is uniform. When considering the manage-
Fig. 10. Examples of catalog components: diesel generators and
ment cost associated with MGs, REM assumes that
batteries. each MG will have some fixed management cost,

1884 P ROCEEDINGS OF THE IEEE | Vol. 107, No. 9, September 2019


Ciller et al.: Optimal Electrification Planning Incorporating On- and Off-Grid Technologies

Table 2 Generation Design Samples in the Lookup Table

plus a monotonically decreasing marginal cost per the model delivers reasonable combinations of cost and
additional customer. In this way, the model acknowl- reliability of supply. We often employ this latter method
edges the economies of scale associated with MGs of in our case studies.
increasing size. As an additional feature related to CNSE, REM allows
4) Direct energy and fuel costs are accumulated in the definition of supply limits on demand, in terms of the
grid-extension projects and in MGs with diesel gen- maximum power and time intervals (different for grid-
eration. For grid-extension projects, the cost per extension and MG solutions). This allows modeling the
kilowatt hour of energy is the wholesale electricity following.
price when delivered at MV level (which includes the 1) Power limiters, i.e., devices installed at the house-
true price of energy at wholesale level, plus trans- hold connection point that prevent power con-
mission and HV distribution costs). REM currently sumption to exceed a threshold established by
assumes a flat cost of electricity regardless of the contract.
hour of day or time of year and of the amount of 2) Low-cost supply contracts that cover only partial
energy demanded. REM accepts different values of electricity needs.
this wholesale electricity costs at different connec- A different CNSE penalty is defined for the demand that
tion points as inputs. For MG systems with diesel is not served due to supply limitations, to account for its
generation, the cost of the diesel fuel consumed is equivalent social cost. This penalty is generally set to be
the only energy cost. lower than the critical and noncritical ones since the lack
2) Nonmonetary Costs: The least cost optimization per- of supply is known in advance by contract.
formed by REM also includes nonmonetary costs. The In the considered case example, there is a 0.30-$/kWh
main social cost is the CNSE, which is associated with penalty for not meeting the noncritical demand and
the reliability of supply. REM imposes this penalty on a 0.75-$/kWh penalty for failing to meet critical demand.
per-kilowatt hour basis for every unit of energy demand In other words, these are the social CNSEs for these two
that is not supplied. This penalizing factor ensures that different demand types.
system reliability is properly accounted for while making In the case example presented, the economic life
sure that supply does not become prohibitively expensive span for main distribution network components is set
as direct monetary costs grow rapidly with very high levels to 40 years, whereas for MG network components, it is
of reliability. assumed to be of 20 years. The lifespans of batteries,
A single value of CNSE cannot capture the diversity PV panels, and other components are defined in the corre-
of situations of supply failure, as perceived by different sponding catalogs. The financial discount rate is assumed
customers with different needs and at different times. to be 10% for grid extensions and MGs and 15% for
To approximate to this complex reality, REM distinguishes isolated systems.
between critical and noncritical loads for all customers Furthermore, the case example assumes that the
and applies a different value of CNSE to the curtailment individual connection costs are $65 per consumer for
of critical and noncritical demands. The determination of those serviced by both MGs and grid extensions. The
the appropriate value for CNSE is a difficult task, and it management annual cost is $9 per consumer for those
would theoretically require extensive and well-designed serviced by grid extensions and very large MGs and
surveys or experiments of the customers involved. Fortu- $16 per consumer for medium-sized MGs of about
nately, REM allows users to vary the values of CNSE until 150 consumers (values estimated from MG operators in

Vol. 107, No. 9, September 2019 | P ROCEEDINGS OF THE IEEE 1885


Ciller et al.: Optimal Electrification Planning Incorporating On- and Off-Grid Technologies

East Africa). REM has been set not to accept MGs with five
consumers or fewer. In the clustering algorithm, the annual
management cost for individual lower demand households
(customer type 1) with solar panel and battery has been
set to $60. In the final stage of REM, once the clusters have
been obtained, SA systems for lower income households
are turned into solar kits for the final electrification
plan.
Note that the comparison between solar kits and the
other two delivery modes (i.e., grid extensions and MGs)
is not straightforward. The choice of solar kits cannot
be determined solely by the CNSE. Solar kits, with what
many households might consider acceptable reliability lev-
els for basic energy services, exist at modest prices and
with attractive financing schemes. Solar kits are individ-
ually managed without any external dependencies and the
scarce supplies can be prioritized for individual house-
hold priorities. However, solar kits can only supply appli-
ances with low electricity intensities. In the case example
described, a solar kit is available in the component catalog
with an investment cost of $86.40, which can meet 68% of
the lower demand household’s (customer type 1) demand.
A conventional solar home system using the same MG
technology would meet 95% of the demand with an invest-
ment and operation cost of $181/year, as shown in Table 2
[CAPEX + operational expenditure (OPEX)].

F. Topography and Administrative Divisions


REM and RNM can adjust network costs in accordance
with information about altitude, ground slope, and
forbidden and penalized zones. These adjustments can
be accounted for in the network routing and clustering
algorithms.
The objective function for optimizing the layout and
size of candidate power lines is guided by the net present
value of investment and preventive and corrective main-
Fig. 11. Influence of altitude, slopes, and forbidden zones in the
tenance costs. Apart from other concepts (e.g., the cost network topology. Top image: omits the forbidden zones and the
of the equipment required to improve system reliability, altitude. Bottom image: considers the forbidden zones and the
the cost of ditches, façades and posts, capacitors and altitude [84].

voltage regulators, etc.), the model considers investment


and maintenance costs that depend on the type of power
line. One type of penalty factor is defined for capturing the number of pixels in rows and columns, the coor-
additional costs related to building lines over changing alti- dinates of the southwest corner in a predefined geo-
tudes and a second type is defined for crossing forbidden graphical system, the size of each pixel, and elevation
and penalized zones (i.e., areas where it would be more values.
expensive to construct and maintain lines such as lakes, The National Aeronautics and Space Administration
rivers, and forests). Two types of input files are required (NASA) Shuttle Radar Topography Mission (SRTM) pro-
to calculate these penalty factors: an altitude raster and vides digital elevation data covering all countries. In the
zone-definition polygon files. case example, the topographical data used (altitude raster
The model currently accepts altitude information in the map and forbidden and penalized zones) are from [76].
Environmental Systems Research Institute (ESRI) Amer- The impact of topography and penalized zones in the
ican Standard Code for Information Interchange raster network topology is shown in Fig. 11. Note how the layout
format. The file stores topographic data information in a of the network changes dramatically when considering
numerical matrix where each pixel indicates the altitude forbidden and penalized zones.
of the raster map in a manner that allows rebuilding The penalized area factor depends on the existence of
the slopes and elevation of the terrain. The basic infor- geographical features, such as lakes, wetlands, national
mation to enable reconstructing the raster map includes parks, or other designated zones. RNM and REM use a file

1886 P ROCEEDINGS OF THE IEEE | Vol. 107, No. 9, September 2019


Ciller et al.: Optimal Electrification Planning Incorporating On- and Off-Grid Technologies

that contains a set of polygons. Each polygon is defined V. M I N I G R I D G E N E R A T I O N D E S I G N


by a set of coordinates preceded by a header where the The design of MG generation depends on demand, avail-
user can specify the penalty multipliers for the zone. Large able technology, and local conditions such as the cost of
multipliers (for instance, 100) may be used to completely fuel and the hourly solar irradiance. REM assumes that
avoid penalized areas in network routing decisions, while MGs have centralized generation and operate in islanded
moderate multipliers (for instance, 1.5) may account for mode. The adopted general architecture for any OG system
realistic extra costs. The outline of the countries can also in REM was shown in Fig. 9. REM does not include all the
be used to define the forbidden and penalized zones. This components in every generation design. For instance, if an
is especially relevant if the country has a coastline. MG only has diesel generation and no other generation
In some cases, it is necessary to consider the technology type, converters and inverters are not neces-
administrative boundaries in the clustering process. sary. In the case example, any combination of PV, battery,
For instance, REM can be forced to obtain independent and diesel generation options is permitted; however, diesel
electrification solutions for different administrative is constrained to be only used as a back-up supply option,
regions. This requires delimiting them with polygons, with a utilization limit of no more than 30% of the total
which are used in the model to split the whole problem demand.
into smaller subproblems, each of them processed and
analyzed independently. For this purpose, the model
requires as input data the polygons that define the A. Generation Sizing Algorithm
boundaries of the administrative divisions. The method that REM uses to determine the generation
design of an MG is a variation of the Hooke and Jeeves
algorithm [77]. It starts by picking an initial point in
G. General Electrification Criteria the multidimensional space of the MG design variables
Additional parameters can be used to define overall tar- and then calculates the value of the objective function
gets of the electrification plan, which constrain the primary (i.e., the total cost of investment and operation of the MG)
objective of cost minimization. The most important ones for several points around the initial point before moving
refer to required levels of quality of service and the mix of along the direction of minimum cost. For each candidate
delivery modes (i.e., grid extension, MGs, and SA systems) point in the search space (with as many dimensions as
used to meet the targets. design variables for the MG), REM performs an annual
Reliability of supply is dealt with by setting minimum simulation of the operation of the MG, adopting some
acceptable levels of quality of supply in OG systems and generation dispatch strategy, and calculates the total cost
including the social CNSE in the optimization. Nonserved of that point including investment and operation costs plus
energy costs are further differentiated for critical and a penalty for the nonserved energy.
noncritical demands. This algorithm moves in a 3-D search space where the
By default, REM will search for the least cost electri- dimensions are the diesel generator capacity, the total
fication plan. Ad hoc input “policy penalties” or “cus- capacity of the solar panels, and the battery capacity.
tomer preferences” allow the user to bias the final solution REM sizes the remaining components of the design
toward fewer or more customers served by grid extensions, afterward. The first step of the sizing algorithm is to
MGs, or SA systems. Other parameters can define rela- establish the boundaries of the search space, estimating
tive or absolute budgetary thresholds for the entire plan, the maximum and minimum possible value in each of the
establish a minimum size for the MGs, force connectivity three dimensions.
within the administrative boundary of a village, or force The possible values that the diesel generator could take
grid connections for buildings situated within a short dis- in the search space are provided by the user in the gener-
tance to the existing grid. ation catalog; as a result, the diesel axis is not necessarily
An electrification plan extends over many years. Since a equally spaced. On the other hand, the possible values that
plan includes a variety of technologies and delivery modes, the solar panels and the batteries could take in the search
it is not trivial to compare different plans in terms of total space are given by combining several units of a single solar
cost over a long period. REM addresses this difficulty by panel or battery in a row. This implies that the points along
minimizing the cost of supply for just some chosen future the dimensions of solar and batteries are equally spaced.
year by comparing the equivalent annuity (i.e., assuming Moving from a given diesel generator to one that is
some evolution of costs in time and the different economic immediately larger or smaller could produce significant
lives of the physical assets employed in the plan) for variations of capacity (for example, from 5 to 10 kW).
each plan for that future year and choosing the plan with In order to avoid local minima, the algorithm that REM
the lowest annuity. Strategic input parameters define the uses is based on a master–slave decomposition where the
design horizon (i.e., the future year for which REM obtains master problem controls the diesel axis, and the slave
the optimal plan): demand growth rate, physical compo- problem moves inside a solar-battery plane with a fixed
nent lifetimes, and discount rates that are appropriate for diesel capacity and an initial search point that the master
the considered planning situation. level provides. This nested optimization decomposition,

Vol. 107, No. 9, September 2019 | P ROCEEDINGS OF THE IEEE 1887


Ciller et al.: Optimal Electrification Planning Incorporating On- and Off-Grid Technologies

REM solves this problem by calculating only a few gen-


eration designs for a representative number of candidate
MGs and, if it needs the information of another design,
the model will obtain it using multi-linear interpolation.
Fig. 13 shows a small electrification planning problem that
we will use to illustrate this concept. There are 32 resi-
dential consumers in this problem with the same demand
profile.
Since all the consumers have the same demand,
Fig. 12. Master–slave decomposition.
there are also 32 candidate OG systems with differ-
ent aggregated demands (those with 1, 2, . . . , 31, and
32 consumers). Instead of calculating these 32 generation
which is shown in Fig. 12, has been successfully used in
designs, which would be a feasible strategy for this toy
problems of different nature [78], [79].
example but not for a large-scale problem, REM could
Some components of the cost of the candidate designs
calculate the generation designs related to 1, 2, 5, 15,
depend on operational aspects of the MG considered,
and 32 consumers. If it needs information for a generation
such as fuel, component replacements, and nonserved
design with, for instance, 25 consumers, the model will
energy. These terms are estimated by simulation, using, for
interpolate the values, using the data from the designs with
instance, the “load following” dispatch strategy [80]. For a
15 and 32 consumers.
given generation design, the load following strategy first
In this example, the lookup table would have a single
uses solar energy to meet the demand and, if the battery
axis with five representative points related to residential
is not fully charged, the remaining solar energy is used
consumers. Things, however, are more complicated if we
to charge it. If there is not enough solar energy and the
introduce other types of consumers, such as productive
battery is not fully discharged, the battery is used to meet
loads or bigger households, with different corresponding
the demand. Finally, if there is still demand that cannot
demand profiles. Specifically, adding a new consumer type
be met, the load following strategy either uses the diesel
to the lookup table implies adding one more dimension in
generator or allows some nonserved energy (minimizing
the space of consumers, and the total number of genera-
cost according to the penalties for nonserved energy).
tion designs that REM would need to calculate could be
There are other ways of dealing with the generation
significantly larger.
sizing problem. Instead of using this heuristic approach,
the investment and the operation problems can be formu- C. Dealing With Multiple Types of Consumers
lated as a classic optimization problem that considers the
The number of customer types that REM can process is
generation design and the dispatch at the same time. This
a limitation of the lookup table approach. A case example
formulation has already been implemented in [81] with
with five different consumer types would imply a 5-D
satisfactory results.
lookup table and significantly increase computation times.
B. Lookup Table In order to overcome this limitation, it is important to
realize that generation designs are related to the aggre-
In a large-scale rural electrification planning problem,
gated demand of candidate MGs and not to specific com-
the task of calculating accurate generation designs for
binations of consumer types. This implies that it is worth
all the candidate MGs is computationally infeasible and
associating the axes of the lookup table with “demand
unnecessary. This implies that the iterative application of
patterns” instead of customer types. By doing so, REM can
a generation sizing tool such as HOMER would be impru-
operate with a number of customer types that is higher
dent, and there is a need for finding a new methodology
than the number of axes of the lookup table, as far as the
that balances accuracy with computation time.
demand of any customer type can be expressed as a linear
combination of a set of “basic” demand patterns.
For example, we could have two different types of res-
idential households, large and small, and assume that the
demand for a large household is five times that of a small
one. Then, REM could have one demand pattern related to
the small household profile and the point “5” of the lookup
table could be either five small households or one large one
since the demand is the same in both cases.
Future work is motivated by exploring this idea using
dimensionality reduction techniques that decompose a
large number of demand profiles related to customer types
into a few basic demand patterns associated with axes of
Fig. 13. Electrification problem example. the lookup table. Both the basic demand patterns and the

1888 P ROCEEDINGS OF THE IEEE | Vol. 107, No. 9, September 2019


Ciller et al.: Optimal Electrification Planning Incorporating On- and Off-Grid Technologies

Fig. 14. OG system cost per kilowatt hour of demand served.

linear combinations are specific inputs to REM. Patterns


may be defined directly by the user, using problem-domain
expertise, or computed separately with optimization
algorithms. Fig. 15. Daily sample dispatch of an MG with 10 demand profiles,
where the black line represents the total demand (critical plus
noncritical).
D. Minigrid Design in the Case Example
In the case example, REM creates a lookup table cal-
culating generation designs for 1, 10, 50, 100, 250, 500, The cost per kilowatt hour, as expected, decreases when
5000, and 50 000 sample demand profiles, which are cho- the number of demand profiles increases. This is due to
sen sequentially from those shown in Fig. 6. As mentioned a combination of the economies of scale in management
earlier, when REM needs to estimate the cost of a candi- costs, some economies of scale in the electronic equipment,
date MG with a different number of aggregated demand the profile aggregation, and the existence of diesel genera-
profiles, it will interpolate between the two closest designs tors in the large generation designs. Fig. 14 shows the costs
unless the number of aggregated demand profiles is larger of generation (CAPEX and OPEX) plus nonserved energy
than 50 000. For larger numbers, REM assumes that there for OG systems (network costs will be calculated later for
are no more economies of scale in generation, so the per- each candidate MG).
unit investment and operation costs are the same as in the REM provides the hourly generation dispatch of each
generation design for 50 000 aggregated demand profiles. MG and SA system. The detailed simulation that REM
Table 2 gives the characteristics of the designs for the performs over 1 year (with some optional simplifications
lookup table samples. Note that the management costs to reduce the computational burden) allows the user to
(part of the OPEX shown in Table 2) are calculated assum- analyze days with different patterns of demand coverage,
ing that each sample profile corresponds to a low-demand as shown in Figs. 15 and 16.
residential consumer. Internally, REM uses the actual num- Fig. 15 shows the dispatch for a couple of days in an MG
ber of consumers of a cluster to compute management with 10 demand profiles. Solar panels serve the demand
costs, not the number of sample profiles (since a type of and charge the batteries during the day. At night, batteries
consumer may comprise multiple sample profiles). meet most of the demand.
Generation designs meet almost all the demand, which Fig. 16 shows the dispatch for a couple of days in an
is reasonable given the high penalties for nonserved MG with 5000 demand profiles. This generation design
energy. Diesel generators are part of the generation designs also meets the demand and charges the battery with solar
for the 5000 and 50 000 points of the lookup table. In both panels during the day, but now the demand after midnight
cases, they cover 26% of the total demand, which is below is mostly covered with diesel.
the maximum allowed (30%). The constraint on diesel In addition, REM calculates the statistics for the esti-
utilization is active since the optimum unconstrained value mated annual reliability performance (average amount of
has 74% of diesel production (as shown in Section IX). This demand served for each hour of the day) of the two MGs
case does not reach exactly the 30% limit because: 1) the as discussed in Figs. 15 and 16. Fig. 17 shows this REM
discrete sizes of diesel generators and 2) the shortsighted output, which could be useful to inform developers and
hourly logic of the load-following dispatch is not able to prospective beneficiaries of the expected performance of a
meet medium-term targets accurately. future MG.

Vol. 107, No. 9, September 2019 | P ROCEEDINGS OF THE IEEE 1889


Ciller et al.: Optimal Electrification Planning Incorporating On- and Off-Grid Technologies

Fig. 18. Clustering candidate connections.

tions would be computationally infeasible in large-scale


problems.
REM uses a Delaunay triangulation to obtain the poten-
tial connections among consumers. This procedure has
already been used in clustering algorithms related to dis-
tribution networks [63], [82].
Fig. 18 adds the candidate connections to the example
shown in Fig. 13. In Fig. 18, it is clear that consumers
labeled as 1 and 2 could be electrified together, but a direct
Fig. 16. Daily sample dispatch of an MG with 5000 demand
profiles, where the black line represents the total demand (critical connection between them and consumers 31 and 32 is
plus noncritical). not worth considering. However, by following the logic
explained in Section VI-A, consumers 1 and 32 could be
electrified together if economies of scale justify the gradual
aggregation of more customers until all of them happen to
VI. C L U S T E R I N G
be connected in one large cluster.
The goal of the clustering process is to determine Fig. 19 shows the Delaunay triangulation for the case
which consumers should be electrified together (i.e., with example.
the same system) and what is the most efficient way
to electrify them. Evaluating all the possible combina-
A. OG Clustering Process
The first step of the clustering process (OG cluster-
ing) temporarily assumes that all the consumers will be
electrified individually with OG systems. The algorithm
makes customer-grouping decisions on the basis of two
conflicting driving factors: 1) the savings in generation,
operation, and management costs brought by economies of
scale and customers’ simultaneity factors in larger MGs and

Fig. 17. Demand served each hour of the day for 10 (top) and Fig. 19. Nonelectrified consumers (green) and the corresponding
5000 (bottom) demand profiles. Delaunay triangulation (black).

1890 P ROCEEDINGS OF THE IEEE | Vol. 107, No. 9, September 2019


Ciller et al.: Optimal Electrification Planning Incorporating On- and Off-Grid Technologies

Fig. 20. OG clustering configurations.

2) the increase in network costs associated with grouping


customers together.
REM begins by evaluating the arcs of the Delaunay
triangulation that are more likely to be activated by join- Fig. 22. OG clusters.
ing the corresponding clusters, i.e., from the shortest to
the longest link (the effect of the order of evaluation is
mitigated by running several passes). In each evaluation, described in Section VII when the different alternatives
the model compares the costs of the configurations shown derived from the clustering process are examined and final
in Fig. 20 to determine if the connection should be acti- detailed comparisons are made.
vated. In the figure, triangles represent generation sites. The grid-extension clustering makes use of the arcs of
Configuration 1, where the clusters are separately elec- the Delaunay triangulation that join pairs of two different
trified, has larger generation and management costs. OG clusters. It calculates the cost of several configurations
On the other hand, the network cost of configuration 2 is to determine if it is economically beneficial to join both
larger (the line approximates the incremental network clusters, under the assumption that at least one of them
costs). REM estimates the cost difference between both is going to be connected to the grid. In the first set of
configurations (generation costs are obtained from the configurations, which is shown in Fig. 23, both clusters
lookup table described in Section V) and joins th OG are connected with a line (triangles here represent MV/LV
clustering solution with seven OG clusters for the example transformers; thick lines are MV lines and thin ones are
under consideration, and the Delaunay arcs that have not LV lines). This implies that REM will join both clusters if a
been activated. configuration from this set ends up being the least cost one.
Delaunay arcs are used as potential clustering connec- On the other hand, Fig. 24 shows several configura-
tions; therefore, some of them may be redundant (if they tions with the clusters not connected to one another.
link the same pair of clusters) and just ignored. The In configurations 3 and 4, one of the clusters is indeed
clusters at the end of this step are the OG clusters, and electrified with an OG system (here triangles may be either
they are the starting point of the grid-extension clustering transformers or generation sites). Hence, if a configuration
process. Fig. 22 shows the OG clusters for the case example in Fig. 24 is the least cost one then REM will not connect
(colors of the clusters are randomly assigned). both clusters.
The clusters at the end of this step are named grid-
B. Grid-Extension Clustering Process extension clusters (admittedly, a term that may seem con-
The second step of the clustering process, grid-extension fusing since many of these clusters may end up not being
clustering, starts from the existing network and the OG grid connected). Following the example, Fig. 25 shows
clusters obtained in the first clustering step. Note that REM the corresponding grid-extension clusters at the end of
does not decide the final electrification modes in the clus-
tering process. This is determined in the final design phase,

Fig. 23. Set of alternative configurations that support merging


Fig. 21. OG clustering example. grid-extension clusters together.

Vol. 107, No. 9, September 2019 | P ROCEEDINGS OF THE IEEE 1891


Ciller et al.: Optimal Electrification Planning Incorporating On- and Off-Grid Technologies

Fig. 24. Set of alternative configurations that support keeping


grid-extension clusters separate. Fig. 26. Grid-extension clusters and MV power grid (black lines).

the grid-extension clustering process and the Delaunay extension clusters are about five times larger than OG
arcs that have not been activated. In Fig. 25, none of clusters on average.
the clusters appears as connected to the MV grid, as the The clustering process creates a hierarchical structure
purpose of this grid-extension phase is not to decide on of clusters where the first level contains the grid-extension
grid connection but to get a good set of candidate clusters clusters, the second level contains the OG clusters,
that will be thoroughly analyzed in the final design phase. and the third level contains the individual consumers.
Fig. 26 shows the grid-extension clusters for the case Fig. 28 shows the structure that corresponds to the
example (again, the colors of the clusters are randomly example.
assigned). Note that, in contrast to other electrification This cluster structure is used to determine the electri-
planning methods that are rule-based, REM may find OG fication mode of each consumer and the final solution.
electrification solutions for consumers that are close to However, it is important to note that the origin of each
the network, if the OG solution is less expensive. This cluster—either GE or OG—is completely ignored in the
typically happens when the aggregated demand of these final design phase, and both electrification options (OG
customers is so low that—following a cost minimization and grid extension) are tried and evaluated in detail for all
logic—it does not justify the investment in the minimum the clusters. The only goal of the clustering processes is to
size transformer in the catalog and the corresponding deliver a well-defined, compact, and meaningful structure
wiring cost. of clusters to be thoroughly explored in the final design
Fig. 27 shows the cumulative share of consumers per phase.
cluster size at the end of the two clustering phases. In OG
clusters, about 70% of the consumers are in groups of at C. Other Approaches
most 100 consumers; the reason being that the economies The hierarchical structure of clusters shown in Fig. 28 is
of scale in generation shown in Fig. 14 are weaker beyond created using a bottom-up greedy approach, where individ-
this point and typically not enough to overcome the extra ual consumers form the initial clusters. However, a differ-
network cost required for connection.
Grid-extension clusters (i.e., the clusters generated in
the grid-extension clustering process) are significantly
larger than OG clusters. This is to be expected since the
lower cost of energy from the main grid compensates
for higher costs of connection. Fig. 27 shows that grid-

Fig. 25. Grid-extension clustering example. Fig. 27. Cumulative number of consumers per cluster size.

1892 P ROCEEDINGS OF THE IEEE | Vol. 107, No. 9, September 2019


Ciller et al.: Optimal Electrification Planning Incorporating On- and Off-Grid Technologies

ent strategy, currently under development [67], is to start


with a large grid-extension cluster that contains all the
consumers and then proceed with a bottom-up evaluation
of disconnections.

VII. F I N A L D E S I G N S
In the “final designs” process, REM exploits the hierarchical
structure of clusters in Fig. 28 to determine the least
cost electrification mode for each customer, which belongs
Fig. 29. Final electrification solution.
to three nested clusters: the individual customer, its OG
cluster, and its grid-extension cluster. Since the number of
candidate systems is now significantly lower than it was include a combination of isolated consumers, MGs, and
initially, REM can afford to compute OG and grid-extension grid extensions).
solutions for each cluster regardless of its position on The final step is to compare this temporary optimal
the hierarchical structure. This implies that grid-extension solution with the best electrification option for GE1 as a
designs and MG solutions are calculated for all the clusters, single connected system, either OG or grid extension. The
whether OG or grid-extension clusters. least cost solution becomes the final optimal solution for
Specifically, REM obtains the least cost electrification the GE1 set of consumers.
solution for a cluster by comparing its least cost electri- Fig. 29 shows a possible final electrification solution for
fication mode (MG or grid extension) with the sum of the the example under consideration. In this case, the cluster
optimal electrification solutions of the clusters that are in GE1 is electrified with a grid-extension design, whereas the
the immediately lower level as shown in Fig. 28. Therefore, remaining grid-extension clusters have lower costs when
cost evaluations are propagated bottom-up in the cluster’s electrified with OG systems that are coherent with the
structure. hierarchical structure.
Although this may seem counterintuitive at first (OG In this stage, accurate network designs are calculated
clusters were calculated without considering the grid), it is for the cost comparisons performed to determine the final
sensible to proceed this way. For instance, isolated con- electrification mode of each consumer. REM uses the RNM
sumers might be large factories, so their best electrification to obtain the optimal network layouts and the correspond-
solution could be a grid extension even if there are no other ing costs.
consumers close to them.
In the example in Fig. 28, let us consider the structure A. RNM as a Network Designer
below GE1. First, REM would find the cost of individually
RNM is a flexible model that is able to design a
electrifying the consumers below OG1. Each one could be
quasi-optimal distribution network from scratch, calculat-
provided with an individual generation set or connected to
ing the corresponding costs. RNM can design the entire
the grid. The optimal combination of individual solutions
distribution network starting from the transmission/HV
is defined to be the temporary optimal solution for the
distribution substations, or only the medium and LV com-
OG1 set of consumers, and it is compared with the opti-
ponents, or only the LV network. In this case, RNM needs
mal electrification solution for OG1 as a single connected
as input the location of the transmission substations and
system, either OG or grid extension. The least cost solution
the customers, as well as technoeconomic information
becomes the temporary optimal solution for the OG1 set of
related to the catalog of components (mainly lines and
consumers.
substations/transformers). If RNM is used to design only
The same process is applied to the consumers in OG2,
a part of the distribution network, the locations of the
OG3, and OG4, respectively. The resulting group of least
corresponding upstream substations are required too.
cost solutions becomes the temporary optimal solution for
RNM minimizes cost (CAPEX+OPEX), subject to elec-
the GE1 set of consumers (note that this solution may
trical constraints such as maximum allowed voltage drop
and maximum capacity. The model selects cost-optimal
elements among a defined catalog of components and it
considers the influence of topography when calculating a
network layout. RNM also allows forbidden and penalized
zones, as previously explained for REM.

B. Network Design for Minigrids and Grid


Extensions
In order to obtain the network design for an MG, REM
Fig. 28. Hierarchical structure of clusters in the example. uses RNM twice with different configuration parameters.

Vol. 107, No. 9, September 2019 | P ROCEEDINGS OF THE IEEE 1893


Ciller et al.: Optimal Electrification Planning Incorporating On- and Off-Grid Technologies

C. Dealing With Solar Kits


For low levels of demand, dc solar kits could be pre-
ferred to ac generation systems. Although ac SA systems
can provide more electricity, solar kits are more portable
and less expensive, especially when it comes to O & M
costs. As a result, they may suffice for the small demands
of many households with low levels of affordability. This
implies that solar kits are an option that is worth consider-
ing in rural electrification plans [83].
The CNSE is critical when considering solar kits in
electrification plans since most solar kits are designed to
provide less energy at a very low cost, and for only a few
hours per day. Energy consumption levels from users of
solar kits will typically differ widely from their expected
demand since solar kit users will have to adapt their use
of electricity to the availability of energy both in terms of
peak energy and of total available hours of use. This is a
particular case of supply limitation, and as such, solar kits
should be handled differently from grid-extension or MG
solutions during planning. REM defines a different value
of nonserved energy applicable to solar kits to estimate the
social CNSE due to supply limits.
The solar kits option is not fully compatible with a
bottom-up clustering strategy since they may break the
monotonicity of the function for economies of scale, which
is necessary for the clustering logic previously described.
Because of this, REM only considers solar kits as an electri-
fication option in the final phase, when deciding the cost-
optimal electrification mode of isolated consumers.
Fig. 30. Two examples of MG network layouts. MV line (red line),
LV line (blue line), generation (green triangle), and MV/LV
D. Description of the Final Solution
transformer (purple triangle).
REM explores the structure of clusters and calls RNM
(for precise network designs and costs) in order to deter-
mine the optimal combination of SA systems, MGs, and
REM assumes that all the MGs have an LV generation grid extensions.
system and evaluates two possible networks. The main outputs generated by RNM for each subsystem
1) LV Network: REM assumes that the MG has an (MG or grid extension) are as follows.
LV distribution network. Generation is also con- 1) Detailed network layout in several shapefiles (differ-
nected in LV, so no transformers are needed. ent infrastructure types are represented in separate
2) MV and LV Networks: REM assumes that the MG files).
has an MV network, and MV/LV transformers with 2) Detailed list of components, in HTML files, for MV
LV subnetworks to reach the customers. Gener- and LV networks. They include the list of transform-
ation is assumed to be connected at LV, so an ers, lines, posts, and other elements such as break-
extra MV/LV transformer is needed to feed the MV ers, capacitors, voltage regulators, and maintenance
network. brigades.
The final network design for the MG is the less expensive 3) Cost breakdowns per types of components.
one of the two. As expected, REM selects LV designs for 4) Losses, reliability, and quality analysis.
smaller MGs and MV designs for larger ones, where the The main outputs generated by REM are:
“size” of an MG here must be understood as a combi- 1) Clustering and electrification mode for all the cus-
nation of distance, number of consumers and total load. tomers.
Fig. 30 shows examples of both types of networks, one with 2) Results per system: generation design, cost, and
only LV and the other one with MV and LV. quality parameters.
Note that REM uses RNM only once when calcu-
lating a grid-extension design. RNM computes the MV VIII. P O S T P R O C E S S I N G A N D R E P O R T S
and LV distribution networks of the corresponding grid The plots and figures shown for the case example rep-
extension. resent only a fraction of REM’s capabilities to produce

1894 P ROCEEDINGS OF THE IEEE | Vol. 107, No. 9, September 2019


Ciller et al.: Optimal Electrification Planning Incorporating On- and Off-Grid Technologies

Table 3 Case Study Electrification—Solution Summary

Fig. 31. Case example electrification solution. MV grid extension


(red line), MV existing grid (black line), LV grid extension (blue line),
LV MG (green line), and isolated (purple dot).

useful information for its users. Each stage of the plan-


ning process produces intermediate and final results and The “Final Cost” row is calculated by adding the total
outputs. REM generates text files in different formats, investment and operation cost and the nonserved energy
spreadsheets, and figures. Some of these files are processed cost. Total investment (CAPEX) and operation (OPEX)
further to generate more elaborated reports, via scripts costs include: 1) network costs, which are computed
and tailored applications. Some of these processes are with RNM; 2) connection costs, on a per-customer basis;
exclusively intended for debugging and troubleshooting. 3) generation costs, which are generation investment and
Figures are in MATLAB formats, Shapefiles, and KML files. operation cost in OG systems, and “wholesale energy
GIS files store multiple attributes of the infrastructure costs,” which include also transmission and HV network
represented, which is useful for an interactive analysis of costs in grid extensions; and 4) management costs.
the solutions. Note that the average final cost per consumer provided
Fig. 31 shows the final electrification solution prescribed in Table 3 is higher for grid extensions than MGs and SA
in the case example. systems. This should not lead to the erroneous conclusion
REM’s outputs can be visualized with Google Earth. that it would be better to replace prescribed grid-extension
Fig. 32 shows a prescribed grid-extension design and the designs with OG solutions. REM chooses the most cost-
surrounding area. effective electrification options for each cluster, so any
Table 3 provides a summary of the final electrification incremental change would result in a more expensive
solution, where isolated systems are electrified with a technoeconomic solution on net. In this case, the higher
combination of solar kits for the residential low-demand cost per consumer serviced by grid extensions is due to the
households and ac systems for the remaining loads (note presence of several customer types with different demands.
that the fraction of customers sums up to 1.01 because of
round-off errors).

Fig. 32. Projection onto Google Earth. MV grid extension (red


line), MV existing grid (black line), LV grid extension (blue line), Fig. 33. Total costs per demand served in different systems:
transformer (red triangle), and isolated (yellow dot). investment, operation, and nonserved energy.

Vol. 107, No. 9, September 2019 | P ROCEEDINGS OF THE IEEE 1895


Ciller et al.: Optimal Electrification Planning Incorporating On- and Off-Grid Technologies

those new consumers to the grid. A more accurate model


of reinforcements could change the number of consumers
in grid extensions. On the other hand, the percentage of
OG systems could increase if we considered resources such
as wind, biomass, or minihydro, which are not included in
the current version of REM. Finally, considering MGs with
several generation sites (instead of just one) could also
increase the percentage of OG systems as well.

IX. U T I L I Z A T I O N O F R E M I N
A REGIONAL CONTEXT
REM is a very versatile tool. With REM, a planner can
address electrification cases ranging from the OG supply of
a small village with only tens or a few hundred customers
Fig. 34. System cost per kilowatt hour of demand served:
to the master electrification plan of an entire province,
generation, network, and connection.
state, or country. REM can work at very high levels of
spatial (i.e., customer or building), temporal (i.e., hourly
patterns of demand or solar irradiation), and utilization
Since larger demand loads are typically in grid extensions, (i.e., different customer types) granularity. If necessary to
their cost per consumer is higher, but the cost per kilowatt mitigate computational burdens, REM could also aggre-
hour is lower, as expected. gate individual demands into clusters although so far this
Fig. 33 shows the total system cost per kilowatt hour possibility has not been used in practice.
of demand served for grid extensions and MGs. This cost The flexibility in the level of granularity modeled, the
is calculated by adding the nonserved energy costs to cost-optimization approach (total supply cost minimization
investment and operation costs. Nonserved energy cost per subject to diverse incentives and constraints that include
kilowatt hour is constant for grid extensions because the explicit treatment of reliability or limitations in the use of
grid reliability is uniform. In the case of MGs, nonserved certain fuels or technologies), the engineering design capa-
energy cost is very low due to the high penalties imposed bilities (choice of generation dispatch strategies, physical
to nonserved energy. constraints of power flows, use of actual catalog of existing
Fig. 34 presents a different cost breakdown for grid components, or compliance with grid codes), and the com-
extensions and MGs to show the relative weight of gen- prehensive output capabilities (detailed generation design,
eration, network, and connection costs for different types dispatch, and reliability analysis, lines and transformers
and sizes of systems (management costs and CNSE are not layout, and summary reports with breakdown of costs and
represented here). The individual connection costs are an bill of materials) allows REM to offer an ample range
input and, in this example, they are the same for all grid of modeling capabilities for governments, electrification
extensions and MGs (although REM allows different values agencies, distribution companies, developers, investors,
for each). The energy cost for grid extensions is also an
input, and it has been assumed uniform all over the grid.
Network costs dominate in grid extensions and they show
a large diversity due to the dispersion of consumers (the
case study is in a predominantly rural area). Generation
accounts for the main part of investment and operation
costs in MGs.
Fig. 35 shows the total system cost per kilowatt hour
of demand served for isolated systems. Low-demand
residential consumers (customer type 1) have solar kits
and the remaining loads have ac systems. In the case of
solar kits, “CAPEX” actually represents the total cost of
acquisition of the kit, which includes any support service
that comes with it.
The most dominant electrification option for the case
study is the grid extension, which accounts for 51% of
the consumers. These results, however, should be framed
under the assumption that the wholesale MV energy cost
(0.10 $/kWh) includes the HV distribution and transmis- Fig. 35. Cost per kilowatt hour of demand served of isolated
sion reinforcements that would be necessary to connect systems.

1896 P ROCEEDINGS OF THE IEEE | Vol. 107, No. 9, September 2019


Ciller et al.: Optimal Electrification Planning Incorporating On- and Off-Grid Technologies

donors, regulators, policy-makers, academics, think tanks,


and other stakeholders.
The human planner is the final decision-maker. Although
REM has numerous capabilities, it is still a computer-based
model and is only intended to help in the electrifica-
tion planning process. REM can explore how a cost-
optimal plan would behave if some of our input data and
assumptions were wrong. REM can be adapted to meet
requirements of users working under specific constraints.
The static plan that REM provides—i.e., the optimal power
system design for a given future snapshot scenario—can be
later transformed into an implementable plan consisting of
projects that are developed during the planning horizon.
REM can further support a diversity of regulatory and
policy decisions. This paper describes the capabilities of the
current version of REM. It also elucidates possible future
capabilities with the ongoing enhancements to REM.

A. Sensitivity Analysis
Most of the input data, restrictions, and assumptions
in REM are uncertain and can be inaccurate. However,
with REM, one can explore multiple scenarios based on
variations of the input data set. Additional computation
time is usually immaterial when compared to the effort Fig. 36. Daily sample dispatch with unconstrained diesel usage of
an MG with 5000 demand profiles. Black line: total demand (critical
devoted to gathering and processing data and formulating
plus noncritical).
assumptions. Using REM, the planner can evaluate how
initial plans respond to different “what if” questions to
better understand the effects of input data, restrictions,
assumptions, and scenarios on prescribed cost-optimal share of grid-extension, MG, and SA systems for a specific
plans. scenario; however, a policy-maker may want to reach a
A glimpse of the capabilities of REM for sensitivity certain set of target shares that might be above or below
analysis will be presented here using the same case exam- this optimal mix because of reasons other than cost. Two
ple as described previously. A more detailed account of extreme possibilities are examined here. First, the optimal
these sensitivity analyses with complete information on plan of the case example is compared with one where every
the assumptions and the results can be found in [68]. The customer is connected to the grid. The result is shown
analyses have been organized according to the topic being in Fig. 37, which can be compared with the optimal one
examined. shown in Fig. 31 and has an increased cost of 13.5%.
The opposite situation is one with complete OG electrifi-
1) Restrictions to the Use of Diesel: Many jurisdictions cation with MGs and SA systems, perhaps due to financial
ban or strongly limit the use of diesel for electricity difficulties of the incumbent distribution company or to a
production. This may be motivated for various reasons
including energy dependence and security, complex supply
logistics, cost, theft and security hazards, maintenance
requirements, and environmental concerns.
REM allows the decision-maker to introduce limits to
the use of diesel (e.g., only as a back-up, with an upper
utilization limit) and to examine the sensitivity of the
reference electrification plan to diesel price.
Fig. 36 shows the dispatch for a couple of days in an MG
with 5000 demand profiles in the case example, but now
with no limits to diesel usage (in contrast with the 30%
limitation of the case example). In this scenario, the diesel
generator serves 74% of the demand, only subject to
economic criteria. Note that the new optimal technology
mix does not include batteries.
Fig. 37. Full grid-extension solution. MV grid extension (red line),
2) Preestablished Targets for Grid-Extension or OG Elec- MV existing grid (black line), LV grid extension (blue line), LV MG
trification Shares: REM calculates the optimal (least cost) (green line), and isolated (purple dot).

Vol. 107, No. 9, September 2019 | P ROCEEDINGS OF THE IEEE 1897


Ciller et al.: Optimal Electrification Planning Incorporating On- and Off-Grid Technologies

Fig. 38. Full OG solution. MV grid extension (red line), MV existing Fig. 40. Solution assuming double-demand (for the 17 types of
grid (black line), LV grid extension (blue line), consumers of the case example). MV grid extension (red line), MV
LV MG (green line), and isolated (purple dot). existing grid (black line), LV grid extension (blue line), LV MG
(green line), and isolated (purple dot).

serious deficit of centralized generation.


The solution of this scenario is shown in Fig. 38, cation plans.
which can also be compared with the optimal one shown The correct estimation of future demand growth is
in Fig. 31 and has an increased cost of 8%. equally important. Fig. 40 shows the minimum cost
plan when demand is double than in the reference case
3) Sensitivity to Demand Levels: Errors in the forecast
(17 types of consumers). In this scenario, the percentage of
of demand for different types of customers, or in over-
consumers electrified with grid extensions rises from 51%
all demand growth rates, might lead to very different
(case example) to 73%, favored by the strong economies
planning results. For instance, if there is no information
of scale in the transport capacity of the electricity
at all about the location of each type of demand of the
networks.
future customers in a region, the planner may assume
some average demand for all of them (e.g., one of the 4) Sensitivity to Reliability Levels: REM can consider
lower SE4all tiers or a natural average customer profile reliability either as a cost (the CNSE, with differentiation
extrapolated from close-by grid-connected populations). of critical and less critical loads), as a target or constraint
Fig. 39 shows the electrification solution following this to be met, or both simultaneously.
approach for the demand in the case example, where it The reliability of the main grid is considered an input
has been assumed that all consumers are high-demand to the model. Its impact depends on the values chosen for
residential loads. Again, the solution can be compared with CNSE. When the main grid is very reliable, OG solutions
the one shown in Fig. 31. The differences highlight the only can compete economically in isolated areas far from
importance of collecting information about the nature of the grid. In the case example, the reliability of the main
each load. The experience in using REM indicates that the grid has been set to 90%. Fig. 41 shows the solution when
existence of large communal, commercial, and industrial this reliability is set to 100%. In this scenario, the percent-
customers in a region has a significant impact on electrifi- age of consumers electrified with grid extensions rises from
51% (case example) to 70%.
The discussion about reliability for OG technologies is

Fig. 39. Solution assuming that all the consumers are of type 2
(higher income households). MV grid extension (red line),
Fig. 41. Solution for 100% grid reliability. MV grid extension
MV existing grid (black line), LV grid extension (blue line),
(red line), MV existing grid (black line), LV grid extension (blue line),
LV MG (green line), and isolated (purple dot).
LV MG (green line), and isolated (purple dot).

1898 P ROCEEDINGS OF THE IEEE | Vol. 107, No. 9, September 2019


Ciller et al.: Optimal Electrification Planning Incorporating On- and Off-Grid Technologies

different. REM users may specify a minimum reliability which in principle corresponds to some future date
target level to be achieved by MGs and SA systems, or alter- (e.g., 2030 for Universal Access in the Sustainable Devel-
natively, define the profiles of critical and noncritical loads opment Goals). Once REM delivers an optimal solution
to represent consumer preferences and set values for for the target year, the planner can make use of the
critical and noncritical CNSE that lead to desirable reli- detailed output information of REM to design a temporal
ability levels. Highly reliable OG generation systems are investment trajectory leading from the present moment to
costlier, so the customers may prefer a less reliable system the target date.
tailored to their needs or to their critical demand hours. REM output is detailed down to a number of individual
REM can be used iteratively until a satisfactory trade-off is lots or electrification projects. This level of granularity
attained. allows the planner to decide how to prioritize the different
lots, for instance, according to the annual investment
5) Other Sensitivities: Other factors can substantially
budget available, the existence of community or productive
affect an electrification plan. One key factor is the price of
loads, or the maximization of the number of connections.
wholesale energy, which depends on the generation mix,
the prices of fuels, and the required upstream investments.
A higher price of energy supplied at MV or high-voltage D. Policy and Regulatory Support
distribution level will result in a lower share of grid- The detailed output of REM can help in identifying,
extension solutions. evaluating, and supporting different regulatory and policy
The price of diesel, limits on carbon emissions, and measures.
budgetary constraints are other factors that can be rep- The most significant contribution of REM to mainstream
resented in REM with input parameters, specific purpose regulation of the provision of electricity access is an accu-
penalties, or explicit constraints. rate determination of the cost of service of the differ-
REM can also be used to perform sensitivity analysis ent delivery modes for each considered scenario, and its
on special planning requests by decision-makers, such as: breakdown into cost components at any required level
1) prescribed percentages of one or more of the electri- of detail. Based on this information, policy-makers and
fication modes, as it was illustrated above; 2) choice of regulators can estimate the financing needs of a national
catalog or of grid code in the main network or in MGs; and electrification strategy or electrification plan, which is the
3) adoption of solar kits of prescribed minimum sizes for first step for the corresponding allocation of resources
small residential customers. (not only financial) to their implementation. The planner,
as indicated above, can use sensitivity analysis to search
for alternatives to the initial plan provided by REM, in case
B. Specific Design Requirements
that it does not meet what policymakers want in terms of
Some planners may request REM to adapt to specific cost, mix of delivery modes, fuel use, or any other metric.
design requirements, which, in some occasions, have cre- Establishing the cost of service is a major component in
ated the need for minor enhancements or modifications of determining the viability gap, i.e., the difference between
the original processes. the cost of service and what the beneficiaries of the service
For instance, a policymaker might want to model a case are willing to pay. The quantification of the viability gap
where all customers within the administrative boundary of for different scenarios is necessary to establish a financing
a village or a district be supplied with the same delivery scheme for the plan.
mode, although the least cost solution from REM might Once the cost of service and the viability gap have been
indicate otherwise. This additional constraint implies an estimated, it is possible to define sustainable tariffs and
increase in technoeconomic cost, but it may also facilitate targeted subsidies for different customer types aimed at
implementation and avoid social conflicts. enabling universal electricity access at the national level
REM has been adapted also to “utility in a box” require- within an acceptable time range.
ments. The “utility in a box” approach considers that
all MG generation and storage has a prescribed stan-
E. Future Enhancements to REM
dard design, preinstalled in a “box” or container that
is deployed, operated and maintained according to pre- Despite having numerous present capabilities, enhance-
scribed standard procedures. In this case, REM only sup- ments and improvements to REM are still being devel-
plies OG clusters with a single box, and the clustering oped. These are the major identified shortcomings of the
process groups consumers accordingly, seeking aggregated current model and our focus for ongoing enhancements:
demands that match the capability of the box as close as 1) only solar PV, battery, and thermal generation have
possible—or multiple boxes of the same standard size. been included in the present version of the model; 2) REM
assumes that there is only one generation site in each MG;
3) REM only considers the customers that are assumed
C. Temporal Implementation Strategy to be nonelectrified and also assumes that any existing
REM is a static model that provides a cost-optimal LV lines cannot be extended to connect any nonelectrified
electrification plan for a prescribed future demand level, customers; 4) the design for and the cost of upstream

Vol. 107, No. 9, September 2019 | P ROCEEDINGS OF THE IEEE 1899


Ciller et al.: Optimal Electrification Planning Incorporating On- and Off-Grid Technologies

network and generation reinforcements in REM can be


improved to represent better the impact of electrification
by grid extension on the bulk power system; and 5) a new
version of the clustering algorithm has been developed and
is being tested and compared with the current bottom-
up algorithm; they can complement each other, and it is
possible that each one of them will have advantages for
different kinds of electrification problems.

X. “ L I T T L E ” O R “ L O C A L ” R E M
REM includes extensive MG design capabilities that are
also relevant for the design and planning of a single
MG project. This section describes the alternative use of
the REM model which focuses on the design of a single
MG at the local level and, therefore, does not use all Fig. 42. Network design in which the generation site and streets
the features of REM (no lookup table, clustering, or final are designated.
design phase are needed). This use of REM is called “little”
REM or LREM (see [68]).
The limited size of the problem allows LREM to: financial analysis yields insights into the impact of discount
1) use any number of customer types, each one with its rates, ownership structures, tariffs, connection priorities,
specific demand profile (no need for linear combina- cross-subsidization strategies, government subsidies, and
tions of basic patterns); concessional loans/grants. See [65] and [68] for details.
2) make use of street maps to define corridors for lines
and posts location, a feature of RNM that is often XI. C O N C L U S I O N
not worth using in large-scale runs, and for which It can be concluded that the electrification access problem
detailed information is rarely available. is huge and that sound and efficient planning are critically
Given a fixed cluster of customers and the associated important. The increasing abundance of geospatial data
demand characteristics, LREM utilizes the MG generation and other digital information allows the use of advanced
optimization algorithm described previously to find the computer-based approaches to electrification planning. In
least cost design. This generation provides electricity ser- our view, the model presented in this paper—the REM—
vice to each customer given user-specified constraints such represents the state of the art. REM leverages newly avail-
as reliability level and renewable energy thresholds. The able geospatial information and processing techniques and
final designs are heavily influenced by the generation asset is able to work at very high granularity levels in both time
cost, distribution network cost, geographic characteristics (hourly demand, weather patterns, and supply reliability)
of the cluster, and customer demand patterns. and space (individualized supply to each building).
Given the existing customer locations and demand pat- We claim that this level of detail is not in excess. The
terns, REM produces a least cost network design for the results obtained with REM demonstrate the importance
local MG. Absent a specified location for the generation, of temporal and spatial granularity in obtaining accu-
REM places the generation assets (solar PV, batteries, and rate results. The location of singular loads (i.e., schools,
diesel generation) at the demand-weighted center of the health centers, official buildings, commercial loads, and
village. This location typically results in the least cost net- industrial loads such as trade centers or mines, etc.) and
work design, but this design may not be feasible due to the their consumption patterns have a significant impact on
location of households, roads, or other existing infrastruc- the required technology (grid-extension, MGs, and SA
ture. In cases where a generation site is specified and infor- systems) and geographical layout of the least cost plan.
mation is available regarding the roads, streets, or paths REM is supporting the design of master national elec-
along which the poles can be placed, LREM designs the trification plans and MG initiatives in several developing
LV networks accordingly. An example of this detailed countries. REM can also assess the feasibility of novel
design is shown in Fig. 42. This type of design is often approaches to transform bankrupt incumbent distribu-
necessary due to limitations of the placement of generation tion companies into economically viable business models.
assets and local preferences and constraints on corridors Specifically, REM has been used (or is currently being
for the distribution network. used) to develop National Electrification Plans in Rwanda,
The detailed technical design provided by REM allows Mozambique, and Indonesia, as well as several other
the planner to supplement the design process with a studies in India, Uganda, Nigeria, Kenya, Colombia, and
financial analysis of the revenue sources (customer tariffs, Peru, among others. The research team has been working
potential subsidies, and any other financial arrangements) with planning departments in utilities, ministries, and
necessary for an economically sustainable MG project. This regulatory bodies in some of those countries to develop

1900 P ROCEEDINGS OF THE IEEE | Vol. 107, No. 9, September 2019


Ciller et al.: Optimal Electrification Planning Incorporating On- and Off-Grid Technologies

a user-friendly version of the tool. The targeted users are its deployment to real-world problems, we hope that it can
decision makers and planners in those departments, be used to deliver value for newly emerging applications
with mid to advanced knowledge in distribution power and users worldwide.
systems and geospatial analysis. Other tools will be
made available for specific purposes as satellite imagery Acknowledgments
processing, demand forecasting, implementation monitor- Many people and institutions have contributed to make
ing, decision and policy-making postprocessing tools, and the reference electrification model (REM) a reality. The
scorecards. The final architecture and user specifications Enel Foundation, the MIT Tata Center for Technology and
of the REM platform, including the interaction between Design, and Iberdrola believed in this project when it only
the geospatial database for inputs and outputs, the user was a bunch of ideas and much enthusiasm. The Shell
interface, the interaction with the REM core server, and Foundation has brought REM and the authors’ team to a
the integration with end-customer systems, are presently new dimension in electrification planning, now trying to
under review by a variety of these prospective customers. define the “electric company of the future” for the develop-
They will test and refine the final tool delivered to the ing world. The authors would like to thank P. Sinha, G. Das,
market. and other numerous professionals from Tata Power Delhi
The REM model will be commercialized by Waya Energy for their continuous support in many forms, where a team
Ltd., a startup incorporated by members of the UEA Lab, is developing an industrial–grade interface for the REM
including customer support services, as well as optional model. The authors would like to thank other members of
expert consultancy in planning and regulatory decision their team who have also contributed to or are presently
making. Waya Energy Ltd. will also offer the possibility of working on the development of new capabilities of REM:
developing specific gateways to integrate REM with other P. Dueñas, O. Oladeji, V. Li, T. Cotterman, and M. Barbar.
power systems and geospatial tools in use by the utilities, The authors would also like to thank the many profes-
governments, and international development institutions. sionals in electric utilities and ministries that have worked
The fee for use will be adjusted to sustainably cover the long hours with them providing the data needed to make
model development, customer support, upgrades, and sys- the model useful. The authors have exchanged ideas and
tem costs, and tailored to the needs and sizes of a variety received useful feedback from so many experts that they
of customers. Although REM is still in the initial stages of cannot name them here. The authors thank all of them.

REFERENCES
[1] International Energy Agency. World Energy “Electrification modelling for Nigeria,” Energy Policy, vol. 39, no. 1, pp. 215–227, Jan. 2011. doi:
Outlook 2017. [Online]. Available: Procedia, vol. 93, pp. 108–112, Aug. 2016. doi: 10.1016/j.enpol.2010.09.034.
https://www.iea.org/weo2017/ 10.1016/j.egypro.2016.07.157. [20] D. I. Banks, F. Mocke, E. C. Jonck, E. Labuschagne,
[2] D. Mentis et al., “Lighting the world: The first [11] C. Cader, P. Blechinger, and P. Bertheau, and R. Eberhard, “Electrification planning decision
application of an open source, spatial “Electrification planning with focus on hybrid support tool,” presented at the Domestic Use
electrification tool (OnSSET) on sub-saharan mini-grids—A comprehensive modelling approach Energy Conf., Cape Town, South Africa, Apr. 2000.
Africa,” Environ. Res. Lett., vol. 12, p. 085003, for the global south,” Energy Procedia, vol. 99, [21] R. Fronius and M. Gratton, “Rural electrification
Jul. 2017. doi: 10.1088/1748-9326/aa7b29. pp. 269–276, Nov. 2016. doi: 10.1016/j.egypro. planning software (LAPER),” in Proc. 16th Int.
[3] D. Mentis et al., “The benefits of geospatial 2016.10.116. Conf. Exhib. Electr. Distrib., Amsterdam,
planning in energy access—A case study on [12] E. A. Martinez-Cesena, P. Mancarella, M. Ndiaye, The Netherlands, Jun. 2001.
Ethiopia,” Appl. Geogr., vol. 72, pp. 1–13, and M. Schläpfer, “Using mobile phone data for [22] R. Soler, F. Thomas, N. E. Dhaiby, and M. Bakri,
Jul. 2016. doi: 10.1016/j.apgeog.2016.04.009. electricity infrastructure planning,” Apr. 2015, “Optimizing the place of PV systems in rural
[4] D. Mentis et al., “A GIS-based approach for arXiv:1504.03899. [Online]. Available: electrification planning in Morocco,” in Proc. 3rd
electrification planning—A case study on Nigeria,” https://arxiv.org/abs/1504.03899 World Conf. Photovoltaic Energy Convers., Osaka,
Energy Sustain. Develop., vol. 29, pp. 142–150, [13] S. Szabó, K. Bódis, T. Huld, and M. Moner-Girona, Japan, 2003, pp. 2574–2577.
Dec. 2015. doi: 10.1016/j.esd.2015.09.007. “Sustainable energy planning: Leapfrogging the [23] Columbia University. (2017). Network Planner
[5] F. F. Nerini et al., “A cost comparison of technology energy poverty gap in Africa,” Renew. Sustain. Website. [Online]. Available:
approaches for improving access to electricity Energy Rev., vol. 28, pp. 500–509, Dec. 2013. doi: http://qsel.columbia.edu/network-planner
services,” Energy, vol. 95, pp. 255–265, Jan. 2016. 10.1016/j.rser.2013.08.044. [24] S. Ohiare, “Expanding electricity access to all in
doi: 10.1016/j.energy.2015.11.068. [14] S. Szabó, K. Bódis, T. Huld, and M. Moner-Girona, Nigeria: A spatial planning and cost analysis,”
[6] I. Pinedo-Pascua, “Intigis: Propuesta metodoloǵica “Energy solutions in rural Africa: Mapping Energy Sustain. Soc., vol. 5, p. 8, Dec. 2015. doi:
para la evaluacioń de alternativas de electrification costs of distributed solar and diesel 10.1186/s13705-015-0037-9.
electrificacioń rural basada en SIG,” (in Spanish), generation versus grid extension,” Environ. Res. [25] V. Modi, E. Adkins, J. Carbajal, and S. Shepa,
Ph.D. dissertation, School Agr. Eng., Univ. Lett., vol. 6, p. 034002, Jul. 2011. doi: “Liberia power sector capacity building and
Politecnica Madrid, Madrid, Spain, 2012. 10.1088/1748-9326/6/3/034002. energy master planning,” Columbia Univ.,
[7] C. Monteiro, J. T. Saraiva, and V. Miranda, [15] E. Kaijuka, “GIS and rural electricity planning in New York, NY, USA, Final Rep., 2013. [Online].
“Evaluation of electrification alternatives in Uganda,” J. Cleaner Prod., vol. 15, no. 2, Available: http://sel.columbia.
developing countries-the SOLARGIS tool,” in Proc. pp. 203–217, 2007. doi: 10.1016/j.jclepro. edu/assets/uploads/blog/2013/09/
9th Medit. Electrotechn. Conf. (MELECON), 2005.11.057. LiberiaEnergySectorReform_Phase4Report-
Tel-Aviv, Israel, 1998, pp. 1037–1041. [16] The World Bank Group. (2016). Off-Grid Market Final_2013-08.pdf
[8] J. Amador and J. Domínguez, “Application of Opportunity Tool. [Online]. Available: [26] F. Kemausuor, E. Adkins, I. Adu-Poku,
geographical information systems to rural http://offgrid.energydata.info A. Brew-Hammond, and V. Modi, “Electrification
electrification with renewable energy sources,” [17] Joint Research Centre (European Commission). planning using network planner tool: The case of
Renew. Energy, vol. 30, pp. 1897–1912, Oct. 2005. (2016). RE2nAF. [Online]. Available: Ghana,” Energy Sustain. Develop., vol. 19,
doi: 10.1016/j.renene.2004.12.007. http://re.jrc.ec.europa.eu/re2naf.html pp. 92–101, Apr. 2014. doi:
[9] T. Huld, M. Moner-Girona, and A. Kriston, [18] Centre for Renewable Energy and Energy 10.1016/j.esd.2013.12.009.
“Geospatial analysis of photovoltaic mini-grid Efficiency. (2012). ECOWREX. [Online]. Available: [27] L. Parshall, D. Pillai, S. Mohan, A. Sanoh, and
system performance,” Energies, vol. 10, no. 2, http://www.ecowrex.org/acp-eu V. Modi, “National electricity planning in settings
pp. 218–223, Feb. 2017. doi: [19] U. Deichmann, C. Meisner, S. Murray, and with low pre-existing grid coverage: Development
10.3390/en10020218. D. Wheeler, “The economics of renewable energy of a spatial model and case study of Kenya,”
[10] P. Bertheau, C. Cader, and P. Blechinger, expansion in rural Sub-Saharan Africa,” Energy Energy Policy, vol. 37, no. 6, pp. 2395–2410,

Vol. 107, No. 9, September 2019 | P ROCEEDINGS OF THE IEEE 1901


Ciller et al.: Optimal Electrification Planning Incorporating On- and Off-Grid Technologies

Jun. 2009. doi: 10.1016/j.enpol.2009.01.021. PV-diesel hybrid system for electrification of an [61] T. W. Lambert and D. C. Hittle, “Optimization of
[28] (2010). GEOSIM. [Online]. Available: isolated island-Sandwip in Bangladesh using autonomous village electrification systems by
http://www.geosim.fr genetic algorithm,” Energy Sustain. Develop., simulated annealing,” Sol. Energy, vol. 68, no. 1,
[29] D. L. Huff, “A probabilistic analysis of shopping vol. 13, no. 3, pp. 137–142, Sep. 2009. doi: pp. 121–132, Jan. 2000. doi:
center trade areas,” Land Econ., vol. 39, no. 1, 10.1016/j.esd.2009.07.002. 10.1016/S0038-092X(99)00040-7.
pp. 81–90, 1963. doi: 10.2307/3144521. [45] L. Wang and C. Singh, “Multicriteria design of [62] A. S. Kocaman, W. T. Huh, and V. Modi, “Initial
[30] M. Zeyringer, S. Pachauri, E. Schmid, J. Schmidt, hybrid power generation systems based on a layout of power distribution systems for rural
E. Worrell, and U. B. Morawetz, “Analyzing grid modified particle swarm optimization algorithm,” electrification: A heuristic algorithm for multilevel
extension and stand-alone photovoltaic systems IEEE Trans. Energy Convers., vol. 24, no. 1, network design,” Appl. Energy, vol. 96,
for the cost-effective electrification of Kenya,” pp. 163–172, Mar. 2009. doi: pp. 302–315, Aug. 2012. doi:
Energy Sustain. Develop., vol. 25, pp. 75–86, 10.1109/TEC.2008.2005280. 10.1016/j.apenergy.2012.02.029.
Apr. 2015. doi: 10.1016/j.esd.2015.01.003. [46] T. Lambert, P. Gilman, and P. Lilienthal, [63] C. M. Domingo, T. G. S. Roman,
[31] Y. Abdul-Salam and E. Phimister, “The “Micropower system modeling with HOMER,” in Á. Sanchez-Miralles, J. P. P. Gonzalez, and
politico-economics of electricity planning in Integration of Alternative Sources of Energy. A. C. Martinez, “A reference network model for
developing countries: A case study of Ghana,” Hoboken, NJ, USA: Wiley, 2006, pp. 379–418. large-scale distribution planning with automatic
Energy Policy, vol. 88, pp. 299–309, Jan. 2016. doi: 10.1002/0471755621.ch15. street map generation,” IEEE Trans. Power Syst.,
doi: 10.1016/j.enpol.2015.10.036. [47] (2018). Homer. [Online]. Available: vol. 26, no. 1, pp. 190–197, Feb. 2011. doi:
[32] Y. Abdul-Salam and E. Phimister, “How effective http://www.homerenergy.com 10.1109/TPWRS.2010.2052077.
are heuristic solutions for electricity planning in [48] O. Bailey, C. Creighton, R. Firestone, C. Marnay, [64] V. Li, “The local reference electrification model:
developing countries,” Socio-Econ. Planning Sci., and M. Stadler, “Distributed energy resources in Comprehensive decision-making tool for the
vol. 55, pp. 14–24, Sep. 2016. doi: practice: A case study analysis and validation of design of rural microgrids,” M.S. thesis, Inst. Data
10.1016/j.seps.2016.04.004. LBNL’s customer adoption model,” Lawrence Syst. Soc., Massachusetts Inst. Technol.,
[33] D. Ellman, “The reference electrification model: Berkeley Nat. Lab., Berkeley, CA, USA. Cambridge, MA, USA, 2016.
A computer model for planning rural electricity Tech. Rep., 2003. [Online]. Available: [65] D. M. Brusnahan, “Minigrids for electrification:
access,” M.S. thesis, Dept. Syst. Develop. Eng., https://eetd.lbl.gov/sites/all/files/publications/ Policies to promote industry growth,” M.S. thesis,
Massachusetts Inst. Technol., Cambridge, MA, report-lbnl-52753.pdf Inst. Data, Syst. Soc., Massachusetts Inst.
USA, 2015. [49] Universidad Zaragoza. (2018). iHOGA. [Online]. Technol., Cambridge, MA, USA, 2018.
[34] P. Ciller-Cutillas, “Clustering-related Available: https://ihoga.unizar.es/en/. [66] T. Cotterman, “Enhanced techniques to plan rural
electrical networks using the reference
improvements in the reference electrification [50] A. Adam, N. M. Galal, and M. S. Hamad, “Rural
electrification model,” M.S. thesis, Inst. Data,
model,” M.S. thesis, School Eng., Universidad electrification using a stand-alone photovoltaic
Syst., Soc., Massachusetts Inst. Technol.,
Pontificia Comillas, Madrid, Spain, 2016. system: Case study of Cameroon,” in Proc. Int.
Cambridge, MA, USA, Tech. Rep., 2017.
[35] S. Upadhyay and M. P. Sharma, “A review on Conf. Ind. Eng. Oper. Manage. (IEOM), Dubai, UAE,
[67] O. Oladeji, “Network partitioning algorithms for
configurations, control and sizing methodologies 2015, pp. 1–8.
electricity consumer clustering,” M.S. thesis, Inst.
of hybrid energy systems,” Renew. Sustain. Energy [51] D. Thevenard, G. Leng, and S. Martel, “The
Data, Syst., Soc., Massachusetts Inst. Technol.,
Rev., vol. 38, pp. 47–63, Oct. 2014. doi: RETScreen model for assessing potential PV
Cambridge, MA, USA, Tech. Rep., 2018.
10.1016/j.rser.2014.05.057. projects,” in Proc. Conf. Rec. 28th IEEE Photovoltaic
[68] “Computer-aided electrification planning in
[36] R. Luna-Rubio, M. Trejo-Perea, D. Vargas-Vázquez, Specialists Conf., Anchorage, AK, USA, 2000,
developing countries: The reference electrification
and G. J. Ríos-Moreno, “Optimal sizing of pp. 1626–1629.
model (REM),” Working Paper. MIT & IIT-Comillas
renewable hybrids energy systems: A review of [52] University of Massachusetts Amherst. (1998).
Universal Energy Access Lab., 2018. [Online].
methodologies,” Sol. Energy, vol. 86, no. 4, Hybrid2. [Online]. Available: http://www.umass.
Available: https://www.iit.comillas.edu/
pp. 1077–1088, Apr. 2014. doi: edu/windenergy/research/topics/tools/software/
publicacion/mostrar_publicacion_working_
10.1016/j.solener.2011.10.016. hybrid2
paper.php.en?id=347
[37] F. Huneke, J. Henkel, J. A. B. González, and [53] E. I. Baring-Gould, H. J. Green, and
[69] (2017). OpenStreetMap. [Online]. Available:
G. Erdmann, “Optimisation of hybrid off-grid V. A. P. van Dijk, “Hybrid2—The hybrid power
https://www.openstreetmap.org/
energy systems by linear programming,” Energy system simulation model,” National Renewable [70] J. Yuan, “Automatic building extraction in aerial
Sustain. Soc., vol. 2, p. 7, Apr. 2012. doi: Energy Lab., Golden, CO, USA, Tech. Rep., 1996. scenes using convolutional networks,” Feb. 2016,
10.1186/2192-0567-2-7. [Online]. Available: https://www.nrel.gov/docs/ arXiv:1602.06564. [Online]. Available:
[38] M. Erol-Kantarci, B. Kantarci, and H. T. Mouftah, legosti/old/21272.pdf https://arxiv.org/abs/1602.06564
“Cost-aware smart microgrid network design for a [54] H. J. Green and J. Manwell, “Hybrid2—A versatile [71] J. Long, E. Shelhamer, and T. Darrell, “Fully
sustainable smart grid,” in Proc. IEEE GLOBECOM model of the performance of hybrid power convolutional networks for semantic
Workshops (GC Wkshps), Houston, TX, USA, systems,” Nat. Renew. Energy Lab., Golden, CO, segmentation,” in Proc. IEEE Conf. Comput. Vis.
Dec. 2011, pp. 1178–1182. USA, Tech. Rep., 1995. Pattern Recognit., Boston, MA, USA, Jun. 2015,
[39] Y. A. Katsigiannis, P. S. Georgilakis, and [55] S. Sinha and S. S. Chandel, “Review of software pp. 3431–3440.
E. S. Karapidakis, “Hybrid simulated tools for hybrid renewable energy systems,” [72] S. J. Lee, “Adaptive electricity access planning,”
annealing-tabu search method for optimal sizing Renew. Sustain. Energy Rev., vol. 32, pp. 192–205, M.S. thesis, Dept. Elect., Eng. Comput. Sci.,
of autonomous power systems with renewables,” Apr. 2014. doi: 10.1016/j.rser.2014.01.035. Massachusetts Inst. Technol., Cambridge, MA,
IEEE Trans. Sustain. Energy, vol. 3, no. 3, [56] P. S. Georgilakis and N. D. Hatziargyriou, USA, 2018.
pp. 330–338, Jul. 2012. doi: “A review of power distribution planning in the [73] Oak Ridge National Laboratory. (2018). LandScan.
10.1109/TSTE.2012.2184840. modern power systems era: Models, methods and [Online]. Available: https://landscan.ornl.gov/
[40] Y. A. Katsigiannis, P. S. Georgilakis, and future research,” Elect. Power Syst. Res., vol. 121, [74] Center for International Earth Science Information
E. S. Karapidakis, “Multiobjective genetic pp. 89–100, Apr. 2015. doi: 10.1016/j.epsr. Network, Columbia University. (2018). High
algorithm solution to the optimum economic and 2014.12.010. Resolution Settlement Layer. [Online]. Available:
environmental performance problem of small [57] P. C. Paiva, H. M. Khodr, J. A. Dominguez-Navarro, https://www.ciesin.columbia.edu/data/hrsl/
autonomous hybrid power systems with J. M. Yusta, and A. J. Urdaneta, “Integral planning [75] F. J. Santos-Pérez, “Metodología de ayuda a la
renewables,” IET Renew. Power. Gener., vol. 4, of primary-secondary distribution systems using decisión para la electrificación rural apropiada en
no. 5, pp. 404–419, Sep. 2010. doi: mixed integer linear programming,” IEEE Trans. países en vías de desarrollo,” (in Spanish), Ph.D.
10.1049/iet-rpg.2009.0076. Power Syst., vol. 20, no. 2, pp. 1134–1143, dissertation, School Eng. Universidad Pontificia de
[41] H. Nasiraghdam and S. Jadid, “Optimal hybrid May 2005. doi: 10.1109/TPWRS.2005.846108. Comillas, Madrid, Spain, 2015.
PV/WT/FC sizing and distribution system [58] N. C. Koutsoukis, P. S. Georgilakis, and [76] (2018). DIVA-GIS. [Online]. Available:
reconfiguration using multi-objective artificial bee N. D. Hatziargyriou, “A Tabu search method for http://www.diva-gis.org/
colony (MOABC) algorithm,” Solar Energy, distribution network planning considering [77] R. Hooke and T. A. Jeeves, “Direct search solution
vol. 86, no. 10, pp. 3057–3071, Nov. 2012. doi: distributed generation and uncertainties,” in Proc. of numerical and statistical problems,” J. ACM,
10.1016/j.solener.2012.07.014. Int. Conf. Probabilistic Methods Appl. Power Syst. vol. 8, no. 2, pp. 212–229, Apr. 1961. doi:
[42] Y. Li and B. Zhou, “The application of improved (PMAPS), Durham, U.K., 2014, pp. 1–6. doi: 10.1145/321062.321069.
clonal genetic algorithm in distributed generation 10.1109/PMAPS.2014.6960627. [78] I. P. Nogueira, “Aerodynamic design optimization
planning,” in Proc. Asia–Pacific Power Energy Eng. [59] J. E. Mendoza, M. E. López, S. C. Fingerhuth, based on multi-attribute structured hybrid direct
Conf., Shanghai, China, 2012, pp. 1–4. H. E. Peña, and C. A. Salinas, “Low voltage search. application to industrial problems,” Ph.D.
[43] S. M. Moghaddas-Tafreshi, H. A. Zamani, and S. distribution planning considering micro dissertation, School Eng. Universidad Pontificia
M. Hakimi, “Optimal sizing of distributed distributed generation,” Elect. Power Syst. Res., Comillas, Madrid, Spain, 2017.
resources in micro grid with loss of power supply vol. 103, pp. 233–240, Oct. 2013. doi: [79] X. Liu and Z. Zhang, “A hybrid reliability approach
probability technology by using breeding particle 10.1016/j.epsr.2013.05.020. for structure optimisation based on probability
swarm optimization,” J. Renew. Sustain. Energy, [60] A. K. Rout and M. K. Parida, “Design and analysis and ellipsoidal convex models,” J. Eng. Des.,
vol. 3, p. 043105, Jul. 2011. doi: of SPV-diesel hybrid system for rural vol. 25, nos. 4–6, pp. 238–258, Sep. 2013. doi:
10.1063/1.3610979. electrification in Odisha,” Int. J. Sci. Eng. Res., 10.1080/09544828.2014.961060.
[80] C. D. Barley and C. B. Winn, “Optimal dispatch
[44] B. Bala and S. A. Siddique, “Optimal design of a vol. 4, no. 12, pp. 129–131, Dec. 2013.

1902 P ROCEEDINGS OF THE IEEE | Vol. 107, No. 9, September 2019


Ciller et al.: Optimal Electrification Planning Incorporating On- and Off-Grid Technologies

strategy in remote hybrid power systems,” Sol. pp. 1104–1121, Jan. 2019. doi: in Cameroon Rural Areas. [Online]. Available:
Energy, vol. 58, nos. 4–6, pp. 165–179, Oct. 1996. 10.1016/j.apenergy.2018.09.194. https://www.plugthesun.com/blog/rural-
doi: 10.1016/S0038-092X(96)00087-4. [82] J. Peco, “Modelo de cobertura geográfica de electrification-cameroon/
[81] L. Moretti, M. Astolfi, C. Vergara, E. Macchi, una red de distribución de energía eléctrica,” [84] C. Drouin, “Geospatial cost drivers in
J. I. Pérez-Arriaga, and G. Manzolini, “A design (in Spanish), Ph.D. dissertation, School Eng. computer-aided electrification planning: The case
and dispatch optimization algorithm based on Universidad Pontificia de Comillas, Madrid, Spain, of Rwanda,” M.S. thesis, Dept. Mech. Eng.,
mixed integer linear programming for rural 2001. Massachusetts Inst. Technol., Cambridge, MA,
electrification,” Appl. Energy, vols. 233–234, [83] P. T. Sun. (2018). Solar Kits to Change the Situation USA, 2018.

ABOUT THE AUTHORS

Pedro Ciller received the B.S. and M.S. Andrés González-García received the B.S.
degrees in industrial engineering from Uni- and M.S. degrees in industrial engineer-
versidad Pontificia Comillas, Madrid, Spain, ing and the M.S. degree in energy and
in 2012, the B.S. degree in mathemat- power systems from the ICAI School of
ics from Universidad Complutense, Madrid, Engineering, Universidad Pontificia Comillas,
in 2014, and the M.S. degree in research in Madrid, Spain, in 1995, the M.A. degree in
engineering systems modeling from Univer- international cooperation from Universidad
sidad Pontificia Comillas, in 2016, where he Complutense, Madrid, in 1997, and the Exec-
is currently working toward the Ph.D. degree utive M.B.A. degree from the ICADE Business
in power systems. School, Universidad Pontificia Comillas, in 2000.
In 2015, he joined the Institute for Research in Technology (IIT), He holds an Executive in Driving Government Performance by the
ICAI School of Engineering, Universidad Pontificia Comillas. He is Harvard Kennedy School of Government, Cambridge, MA, USA, and
currently a member of the Universal Energy Access Lab, a joint IESE, Madrid. He is currently an Invited Professor and Researcher
research group of the Massachusetts Institute of Technology, Cam- with the Institute for Research in Technology (IIT), ICAI School of
bridge, MA, USA, and IIT. He has participated in several projects Engineering, an Independent Consultant with the Asian Develop-
related to rural electrification. ment Bank, and a member of the Universal Energy Access Lab,
Massachusetts Institute of Technology, Cambridge, MA, USA, and
IIT. He is the Co-Founder and the Board Member of Waya Energy
Ltd., Cambridge. He specializes in sustainable development, off-
grid and grid electrification planning, business models, regulation,
and energy policy. He has been responsible for the development
of the National Electrification Plan of Rwanda and participated
Douglas Ellman received the A.B. degree
in the Geospatial Electrification Planning of Mozambique and of
in physics from Princeton University,
the regions of Maluku and Papua in Indonesia. He has worked in
Princeton, NJ, USA, in 2009, and the M.S.
Uganda, Kenya, Nigeria, India, Peru, and Colombia.
degree in technology and policy from the
Mr. González-García is the President of the Spanish Federation of
Massachusetts Institute of Technology,
Fair Trade and a member of the Spanish Table for Universal Access
Cambridge, MA, USA, in 2015. He is
to Energy.
currently working toward the Ph.D. degree
in electrical engineering from the University
of Hawai‘i at Mānoa, Honolulu, HI, USA.
He is currently a Research Assistant with the Hawai‘i Advanced Stephen J. Lee (Graduate Student Mem-
Wireless Technologies (HAWT) Institute, Honolulu. His current ber, IEEE) was born in Newton, MA, USA,
research interests include applying optimization and machine in 1991. He received the B.S. degree in
learning to the integration of distributed energy resources and materials science and engineering with a
renewable energy into electric power systems. second major in economics from Johns Hop-
kins University, Baltimore, MD, USA, in 2013,
and the M.S. degree in technology and policy
and the M.S. degree in electrical engineer-
ing and computer science from the Massa-
chusetts Institute of Technology, Cambridge, MA, USA, in 2018.
From 2013 to 2015, he was a Consultant with Deloitte Consulting
Claudio Vergara was born in Santiago,
LLP, Arlington, TX, USA. Since 2018, he has been a Fellow with The
Chile, in 1982. He received the Professional
Energy for Growth Hub, Washington, DC, USA. His current research
degree in electrical engineering from the
interests include machine learning, remote sensing, energy eco-
Universidad de Chile, Santiago, in 2008, and
nomics and policy, and sustainable development.
the Ph.D. degree in sustainable energy sys-
tems from the Universidade do Porto, Porto,
Portugal, in 2015, with extended research
periods at the Massachusetts Institute of Cailinn Drouin was born in Lawrence, KS,
Technology, Cambridge, MA, USA, and Uni- USA, in 1984. She received the B.A. degree
versidad Pontificia Comillas, Madrid, Spain. in French literature and francophone studies
He joined the Massachusetts Institute of Technology as a Post- and the B.S. degree in mechanical engineer-
doctoral Associate and developed design and simulation computer ing from the University of Kansas, Lawrence,
models for distributed energy resources and electricity distribution KS, USA, in 2007 and 2012, respectively,
networks. He is currently the Lead Product Architect at Zola Electric, and the M.S. degree in mechanical engineer-
San Francisco, CA, USA, where he develops ac photovoltaic and ing with a research focus on energy access
battery energy-storage hardware for electricity access applications in developing countries from the Massa-
in developing countries. chusetts Institute of Technology, Cambridge, MA, USA, in 2018.

Vol. 107, No. 9, September 2019 | P ROCEEDINGS OF THE IEEE 1903


Ciller et al.: Optimal Electrification Planning Incorporating On- and Off-Grid Technologies

Matthew Brusnahan received the B.S. Reja Amatya received the B.S. degree
degree in mechanical engineering from the in engineering science from Smith Col-
Georgia Institute of Technology, Atlanta, GA, lege, Northampton, MA, USA, in 2005, and
USA, in 2008, the M.S.E. degree in mechani- the M.S. and Ph.D. degrees in electrical
cal engineering from the University of Michi- engineering from the Massachusetts Insti-
gan, Ann Arbor, MI, USA, in 2011, and the tute of Technology, Cambridge, MA, USA,
M.S. degree in technology and policy from in 2008 and 2012, respectively.
the Massachusetts Institute of Technology, She is currently a Research Scientist with
Cambridge, MA, USA, in 2018. the MIT Energy Initiative, Massachusetts
He was with EnerNOC (EnelX), Boston, MA, USA, consulting on Institute of Technology, and is leading energy projects within
energy management and has recently been responsible for the the initiative with the focus on rural electrification in emerging
master electrification plans in several countries throughout Africa economies. Her current research interests include power systems
and Asia. From 2016 to 2018, he was a Tata Fellow with the Tata involving renewable energy within on and off-grid settings. She has
Center for Technology and Design, Cambridge. He is currently a Co- been involved in universal access projects focused in India, Kenya,
Founder of Waya Energy Ltd., Cambridge, a software company that Rwanda, and Nigeria.
provides electricity distribution planning software to governments
and utilities.

Rafael Palacios was born in Madrid, Spain,


in 1966. He received the M.S. and Ph.D.
degrees in engineering from Universidad
Pontificia Comillas, Madrid, in 1990 and
1998, respectively.
He was with the Massachusetts Institute
Yael Borofsky was born in Philadelphia, PA, of Technology, Cambridge, MA, USA, where
USA, in 1987. She received the B.S. degree he was a Visiting Professor with the Sloan
in human development from Cornell Univer- School of Management from 2001 to 2002,
sity, Ithaca, NY, USA, in 2009, and the M.C.P. the Aeronautics and Astronautics Department from 2009 to 2010,
degree in urban studies and planning and and with the MIT Energy Initiative, Massachusetts Institute of Tech-
the M.S. degree in technology and policy nology, from 2017 to 2018. He is a Senior Associate Professor with
from the Massachusetts Institute of Technol- the Department of Computer Science, and a Researcher with the
ogy, Cambridge, MA, USA, in 2018. She is Institute for Research in Technology (IIT), ICAI School of Engineer-
currently working toward the Ph.D. degree ing, Universidad Pontificia Comillas. He is currently the Head of the
at ETH Zürich, Zürich, Switzerland. programs in telecommunications engineering (undergraduate and
From 2015 to 2016, she was a Researcher with ETH Zürich, master’s), and Coordinator of the master’s degree in cybersecurity.
Zürich, Switzerland, where she has been a Research Scientist since His current research interests include advanced data analysis in
2017. From 2016 to 2017, she was a Researcher and Instructor many different fields, including image processing, geographic infor-
with the MIT Energy Initiative, Massachusetts Institute of Technol- mation systems (GIS), artificial intelligence, and data mining.
ogy. Her current research interests include infrastructure planning
in informal settlements, sociospatial inequality, and energy eco-
nomics and policy. Since 2018, she has been a Fellow with The
Energy for Growth Hub, Washington, DC, USA.

Robert Stoner received the B.S. degree in


engineering physics from Queen’s Univer-
sity, Kingston, ON, Canada, in 1986, and the
Ph.D. degree in condensed matter physics
from Brown University, Providence, RI, USA,
Carlos Mateo received the Ph.D. degree in 1992.
in industrial and computer engineering From 1995 to 2002, he was an Adjunct
from Universidad Pontificia Comillas, Madrid, Member with the Engineering Faculty,
Spain, in 2007. Brown University. He is currently the
He is currently a member of the Insti- Deputy Director for science and technology with the MIT
tute for Research in Technology (IIT), ICAI Energy Initiative, Massachusetts Institute of Technology,
School of Engineering, Universidad Pontificia Cambridge, MA, USA, and the Founding Director of the MIT Tata
Comillas. He has participated in several Center for Technology and Design, Cambridge, MA, USA. He is
European research projects, collaborated currently a member of the MIT Energy Council, the Science and
with the industry and with Spanish institutions, and participated Technology Committee of the Alliance for Sustainable Energy that
in international research projects with prestigious and renowned oversees the U.S. National Renewable Energy Laboratory, the MIT
institutions such as Massachusetts Institute of Technology, Future of Storage Study, the Technical Advisory Board of the India-
Cambridge, MA, USA, NREL, Golden, CO, USA, and the World Bank, based Center for the Study of Science, Technology and Energy
Washington, DC, USA. He has collaborated in the creation of the Policy, and the Rockefeller Foundation funded Global Commission
DSO Observatory for the Joint Research Centre of the European on Universal Electricity Access. He is a Serial Entrepreneur and the
Commission, and has taken part in the SMART-DS project with NREL Inventor and Co-Inventor of numerous computational and ultrafast
and MIT to build large-scale synthetic distribution networks for the optical measurement techniques. For the past decade, his research
U.S. Department of Energy. His current research interests include at MIT has been at the intersection of energy technology, society,
modeling, simulation, and algorithms, especially in the fields of and computation, with an emphasis on power system design,
electricity distribution networks and distributed energy resources. optimization and regulation in developing countries.

1904 P ROCEEDINGS OF THE IEEE | Vol. 107, No. 9, September 2019


Ciller et al.: Optimal Electrification Planning Incorporating On- and Off-Grid Technologies

Fernando de Cuadra was born in Madrid, Ignacio Pérez-Arriaga (Life Fellow, IEEE)
Spain, in 1961. He received the B.S., M.S., received the B.S. and M.S. degrees in indus-
and Ph.D. degrees in industrial engineering trial engineering from the ICAI School of
from the ICAI School of Engineering, Univer- Engineering, Universidad Pontificia Comillas,
sidad Pontificia Comillas, Madrid, Spain, in Madrid, Spain, in 1970, and the M.S. and
1985 and 1990, respectively, and an Exec- Ph.D. degrees in electrical engineering from
utive MBA degree from the ICADE Business the Massachusetts Institute of Technology,
School, Universidad Pontificia Comillas, in Cambridge, MA, USA, in 1978 and 1981,
2010. respectively.
Since 2001, he has been a Full Professor with the ICAI School of He was a Professor with Universidad Pontificia Comillas. He was
Engineering and the Institute for Research in Technology (IIT), Uni- the Founder of the Institute for Research in Technology (IIT), ICAI
versidad Pontificia Comillas. From 2001 to 2010, he was the Dean School of Engineering, Universidad Pontificia Comillas, where he
of the ICAI School of Engineering. In 2016, he joined the Universal was Director from 1984 to 1994. He was a Professor and the
Energy Access Lab, a joint research group of the Massachusetts Director of training with the Florence School of Regulation, Flo-
Institute of Technology, Cambridge, MA, USA, and IIT, where he rence, Italy. Since 2008, he has been a Visiting Professor with the
has mainly been involved in the design and specification of gen- Massachusetts Institute of Technology. He was a Consultant and
eral models, methods, and tools. His current research interests a Lecturer in about 40 countries. His current research interests
include operations research techniques (large-scale simulation and include future trends in power systems, electrification planning
direct-search optimization) applied to diverse sectors, such as the in developing countries, and strategic issues in universal energy
aerospace industry, power systems, and railway systems. access.
Dr. de Cuadra is a member of the Board of Trustees of Universidad Dr. Pérez-Arriaga is a Life Member of the Spanish Royal Academy
Loyola, Andalucía, Spain. of Engineering. He was Commissioner at the Spanish Electricity
Regulatory Commission from 1995 to 2000, an Independent Mem-
ber of the Irish Single Electricity Market Committee from 2007 to
2012, and a member of the Board of Appeal of the EU Agency for
the Coordination of Energy Regulators from 2011 to 2016.

Vol. 107, No. 9, September 2019 | P ROCEEDINGS OF THE IEEE 1905

Potrebbero piacerti anche