Sei sulla pagina 1di 16

The Role of Predictive Models

in Energy Efficiency Optimization


of Industrial Plants and Buildings

Petr Stluka, Jiri Rojicek, Karel Marik


Honeywell Automation and Control Solutions
25 March 2009, Prague
Outline

• Energy Efficiency Applications


- Industrial and building segments
- Energy supply side / demand side
• Decision Making Scenarios
- Process industries: refining
- Cogeneration and buildings
• Energy Demand Forecasting Methods
- Memory-Based Regression

2 Towards e-Environment, Mar 25 2009, Prague


Energy Efficiency

• Terminology
- Energy efficiency = using less energy to provide the same
level of service … e.g. by using more energy efficient
appliances
- Energy conservation = using less energy to achieve a lesser
energy service … e.g. through behavioral change
• Buildings
- Energy efficient appliances (refrigerators, freezers, ovens,
washers, …)
- Monitoring and control of major energy loads (heating,
ventilation, air conditioning, lighting, …)

• Industry
- Optimized energy conversion – fuel switching, running
equipment at its peak efficiency
- Advanced boilers and furnaces (combustion)
- Selection of less energy intensive operating modes
3 Towards e-Environment, Mar 25 2009, Prague
Energy Efficiency Applications
Advanced Industrial
Industrial Process Plants
Control Utilities

Public Utilities
Steam, power

Renewable
Dynamic Energy
Energy
Targeting and Monitoring

Grid
Smart-grid
solutions

Homes & Buildings


Energy Monitoring
Local
Poly-generation HVAC Control
(CHP) Electricity,
heating, Demand-Response
cooling Capable Homes
Optimal Equipment
Scheduling
4 Towards e-Environment, Mar 25 2009, Prague
Energy Efficiency Optimization

Utility Plants Process Plants Buildings


CHP Manufacturing Homes
Business Generate utilities (power, Produce mix of products Depends on the type of
steam, heating, cooling) to meet demands in the building (office, hospital,
objective
to meet customer downstream industries university, army base,
demands (contracts) shopping center, store,
home, etc.)

Attitude Energy directly is the Second largest cost (after Differs by the type of
to energy primary business cost of raw materials) activity. Sometimes energy
objective cost is only a fraction of
other operating costs

Energy Consistently addressed Minimize energy use but Minimize energy use but
efficiency in the plants delivery of products is the keeping occupants
first priority comfort is the first priority

Approach Closed-loop control Dynamic targeting Energy monitoring


Hierarchical optimization Continuous improvement Continuous improvement

Energy Dominated by Dominated by


Demand production schedule occupants needs

5 Towards e-Environment, Mar 25 2009, Prague


Oil Refinery
Operation is mostly Minimizing energy use and environmental Mix of products
determined by the impact by selecting less energy intensive to be produced
specific crude oil and modes on each process unit to meet orders
its properties (e.g. Gas Liquefied petroleum gas
sulfur content) Gas Cooking
Straight Run Straight Run Gas Poly- Di-isobutylene
Isobutylene merization
Gasoline Gasoline Recovery Airplanes

Motor & aviation gasoline


Catalytic Butylene
High octane and
Reforming Butane Hydro-
Naphtha Gasoline Gasoline
Stabilization genation Automobiles
Isomerization Alkylation
Iso-octane
Hydrogen Alternate Isobutane
Crude High octane Farming
Hydro- Alkylate
oil Cracking gasoline and gas Stabilized gasoline
Kerosene Blending
Kerosene

Gas oil Light furnace oil Home heating


Crude Furnace oil or Jet Fuel
Catalytic Cracking
Oil diesel oil Furnace oil
Distillation Heavy
gas oil Diesel oil Railroads
Trucks

Heavy fuel oils


Residue Heavy fuel oils Factories
Atmospheric Heavy fuel oils
or Vacuum Asphalts Ships
Distillation Alternate Light
Hydrotreating
on
Vacuum Solvent lubricating oils Blending Sale
and
Distillation Asphalts Refining Dewaxing Heavy Packaging Asphalts
Asphaltic Roads
residue Heavy gas oil Decolorizing lubricating oils

Vacuum Lubricating distillates Solvent Wax Wax Refined Waxes


Lubricating Refining Plant Waxes
Distillation Lubricating residue Refining
residue Light products to
Alternate recovery and stabilization
Coking Coke
6 Towards e-Environment, Mar 25 2009, Prague
Plant-Wide Optimization Criterion

• Plant profit can be formulated as:

Plant Profit = + ∑( Product


rate
*
Product
value )
- ∑(
Raw Raw
material
usage * material
cost
) Utility consumption can be estimated
potentially for each unit based on the
previously built model that relates
energy consumption with capacity

- ∑( Utility
consumption
*
Utility
value ) utilization, operating modes, etc.

- Maintenance costs
- Penalties for GHG emissions Closely related to on-site energy
generation, types of fuels used,
- other costs overall energy efficiency, etc.

7 Towards e-Environment, Mar 25 2009, Prague


Predictive Model for Crude Distillation Unit

• Correlates energy consumption on CDU unit with production


targets that uniquely define the mode of operation

Predicted energy Fuel energy (preheat, tower, total)


consumption Steam energy (tower, side strippers, total)
Power energy (tower, pump-arounds, total)

Feed Products
Feed flow rate Saturated gas flow rate
Feed temperature Naphtha flow rate
CDU Kerosene flow rate
API density
Diesel flow rate
Assay/blend properties
AGO flow rate
Residue flow rate
Naphtha distillation D86 90% Rec
Disturbances Kerosene distillation D86 10% Rec
Weather Diesel distillation D86 90% Rec
Diesel sulfur wt%

8 Towards e-Environment, Mar 25 2009, Prague


CHP and Buildings

Energy Energy electricity Energy


Suppliers Conversion heat User
electricity cooling
gas steam
CCHP

Real time pricing Variable efficiency Fluctuating demand

• Conversion efficiency • Energy demand primarily


of the CHP system is depends on
optimized - People’s behavior – How
- Gas turbines, steam many occupants are in the
turbines, steam building? What they are
boilers, hot water doing?
boilers, compressor - Weather – Do we need
chillers, absorption cooling or heating? How
chillers much is needed to keep the
adequate comfort level?

9 Towards e-Environment, Mar 25 2009, Prague


Various Daily Profiles …

1 1
Daily Consumption Profiles

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 2 4 6 8 10 12 14 16 18 20 22 24 0 2 4 6 8 10 12 14 16 18 20 22 24

… commerce … two-shift manufacturing

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 2 4 6 8 10 12 14 16 18 20 22 24 0 2 4 6 8 10 12 14 16 18 20 22 24

… casino … administration

10 Towards e-Environment, Mar 25 2009, Prague


Structure of the Predictive Model

Inputs Demand Model Predictions


Data age
Time of day Heating demand
Holiday Cooling demand
Ambient temperature Electricity demand
Wind velocity Steam demand
Humidity …

Accuracy of energy demand


predictions depends heavily on
accuracy of weather forecasts

11 Towards e-Environment, Mar 25 2009, Prague


Energy Demand Modeling Techniques
• Heuristics and benchmarks
- “Rule of thumb”, benchmarking of similar units
• First principle models
- Based on thermodynamics, mass and energy balances
• Time series models
- ARMA with exogenous variables (e.g. outdoor temperature)
• Statistical regression
- Basic regression functions are determined based on the
knowledge how the manufacturing plant is operated
- Resulting “global” models need to be regularly updated
- Linear regression, Partial Least Squares (PLS)
• Local regression
- Regression models are built based on understanding of
variations and correlations in historical data
- Local modeling deals well with non-linear dependencies, and
with segmentation of data in clusters
- Also known as: memory-based regression, locally weighted
regression (non-parametric statistics)

12 Towards e-Environment, Mar 25 2009, Prague


Local Modeling
Local regression models are
built on-the-fly

Y
Y

X2
X1

Y
X2
X2
X1

X1
Y
Current state and its
neighborhood (= past operating
X2
points similar to the current one) X1

… the dependency Y=f(X1,X2) is much simpler


in the local neighborhood than in global context
13 Towards e-Environment, Mar 25 2009, Prague
Local Regression
Points in the neighborhood
are weighted according to - 3σ d2
Kernel function w = exp
2
Forecast – y (load)

Bandwidth
Distance function
Polynomial fit 2
N
Xi* - Xi
d2 = ∑ hi
i=1

Local neighborhood
Query point
x0
Independent variable – x
(time of day)
14 Towards e-Environment, Mar 25 2009, Prague
Forecasting Results

Forecasted load
Forecasted
load maximum

Actual
load Forecasted
load minimum

Past data Forecasts three days ahead

15 Towards e-Environment, Mar 25 2009, Prague


Thank you for your attention!

Potrebbero piacerti anche