Sei sulla pagina 1di 9

This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2018.2856826, IEEE
Transactions on Industrial Informatics
1

Data-Driven Based Optimization for Power


System Var-Voltage Sequential Control
Junbo Zhang, Member, IEEE, Zhihao Chen, Student Member, IEEE, Chuyao He, Zetao Jiang, Lin Guan,
Member, IEEE

 of power losses, minimization of differences between bus


Abstract—This paper proposes a data-driven based optimiza- voltages and their references, maximization of the voltage sta-
tion method for var-voltage sequential control (V2SC). First, bility margin, minimization of the rescheduling cost, and com-
power system var-voltage control characteristics, defined as the binations of the above objectives [3-5]. Constraints include the
function between the reactive power injections and the bus volt-
ages, is approximated by the var-voltage sensitivities (V2S). Then
power flow equation and limitations of reactive power injec-
V2S is estimated online using the noise-assisted ensemble regres- tions. And the solver could be linear programming, nonlinear
sion method. Subsequently, an optimal model is proposed for programming, interior point optimization method (IPOPT) or
V2SC based on the V2S estimation. To avoid frequent back and intelligent algorithms (such as artificial neural networks and
forth control actions, a hysteresis control strategy is employed. genetic algorithms) [6-8].
The performance of the proposed method is statistically validated Once an optimization model is established, the solver can be
in the 8-generator 36-node system and the Nordic32 system with
test data measured from real power systems. implemented in a centralized manner or a sequential manner
[9-10]. The centralized manner provides an ultimate solution,
Index Terms—data-driven method, var-voltage sequential con- the reactive power rescheduling to the pilot buses can be set at
trol, var-voltage sensitivity, optimal voltage control once. However, it relies on the system offline model, which may
be different from the actual system condition especially when
renewable energy resources are integrated. Stochastic optimi-
I. INTRODUCTION zation method could be used to overcome the model bias;
however, its calculation takes a long time and thus it is unsuit-
P OWER system regional voltage control is a measure to
ensure the bus voltage fluctuation remains within a satisfied
range under normal or post-transient operation conditions by
able to online implement [11]. Alternatively, the sequential
manner provides a solution path from the present operation
means of rescheduling the reactive power injections in the condition to a near optimal operation condition via control of
system region [1]. Generally, the objective is achieved by the reactive power injections step by step based on the
maintaining the voltages of the pilot-bus to be within the region var-voltage control characteristics – the function between the
of a reference value, and consequently, the voltages of other injected reactive power and the bus voltages [12]. If the control
buses can also be maintained within the satisfied range [2]. In characteristics can be online estimated, the sequential optimi-
conventional control methods, different pilot-bus voltages are zation can draw the system to the near optimal condition along
independently controlled; therefore, the regional control prob- with the operation trajectory without relying on the offline
lem to a system range can be divided into sub-control ones for physical model, which is more suitable for stochastic system
subareas. However, with the development of power systems, conditions.
pilot buses in a region could be strongly coupled, i.e., a pilot-bus Conventionally, the var-voltage control characteristics are
voltage control may influence voltages of other adjacent buses, approximated by a linear function called the voltage/var sensi-
leading to complex voltage control problems. tivity (V2S), which is calculated by offline system models [13].
To solve the above issues, optimal var-voltage control (V2C) With the development of wide area measurement system
was proposed with a nonlinear objective function and various (WAMS), huge amounts of system operations data have been
constraints [3-8]. The objective function could be minimization collected, providing fundamental data sources for the V2S
estimation [14]. Earlier works attempted to estimate the sensi-
tivities of different power system observations related to the
This work is supported in part by the National Natural Science Foundation operation states [15-18]. References [15] and [16] provided
of China (51607071), the Natural Science Foundation of Guangdong Province,
China (2016A030313477), the Fundamental Research Funds for the Central
weighted least squares estimation and recursive least squares
Universities (2017JQ011), and South China University of Technology. estimation as basic mathematical tools for the sensitivity esti-
Junbo Zhang, Zhihao Chen, Zetao Jiang and Lin Guan are with the School mation; [17] and [18] provided norm-2 penetration algorithm
of Electric Power, South China University of Technology, Guangzhou, P.R.
and noise-assisted algorithm as solutions to data collinearity and
China. (e-mail: epjbzhang@scut.edu.cn, epzhchen@mail.scut.edu.cn,
201620110678@mail.scut.edu.cn, lguan@scut.edu.cn). noise effects. These algorithms are good start points which
Chuyao He is with the Guangzhou Power Supply Bureau Co., Ltd., make an online var-voltage sequential control (V2SC) possible.
Guangzhou, P.R. China. (e-mail: epjbzhang@scut.edu.cn).

1551-3203 (c) 2018 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2018.2856826, IEEE
Transactions on Industrial Informatics
2

In this paper, we combine V2SC and V2S estimation in the controllable system states with stochastic properties, such as
earlier works into a new data-driven based V2SC method. To loads and renewable generations, the voltage values at time t are
implement the proposed method in real power systems, a hys- assumed predictable by time t-1 using time-series approaches
teresis strategy that can prevent frequent back and forth re- [19]. Thus, the V2SC problem is to reschedule reactive power
scheduling actions is also integrated. The proposed method has injections from capacitor and reactor groups at different pilot
the following merits: buses along with time t, so the final optimized vref (t) - v(t) tra-
1) It is a purely data-driven method with capability to draw jectory can be close to the optimal trajectory in theory.
the system to the near optimal condition along with the opera- Assume the function between the actual voltage value v(t)
tion trajectory without relying on the offline physical model, and the reactive power injections is
which is more suitable in modern power systems with stochastic
v (t) = f(q(t)) (1)
renewable energy resources;
2) With consideration of practical issues, it provides similar where f is a multi-input and multi-output nonlinear function and
result to the conventional IPOPT based V2C method, making it q(t) is the reactive power injection vector at the buses. If there
great potentials to be applied in real power systems. are capacitor and reactor groups at each pilot bus, the optimi-
The remainder of the paper is organized as follows. Section II zation model for V2SC at time t can be formulated as follows:
provides theory, framework and implementation procedures for 2
max f (q (t -1)  q (t ))-v ref (t )
the V2SC. Section III discusses key issues in the V2SC, in- q ( t ) 2

cluding implementation of the V2S estimation and the hystere- s.t. q(t ) ≤ q(t ) ≤ q(t ) (2)
sis control strategy. Section IV and Section V validate the
proposed method in the 36-Node system and Nordic32 system, q (t ) ≤ C (t )
respectively. Section VI presents the conclusion. where Δq(t) is the reactive power rescheduled by the capacitor
and reactor groups, q (t ) and q(t ) are the lower and upper
II. DATA-DRIVEN BASED OPTIMAL V2SC
bounds of Δq(t), and C(t) is a constant.
A. Data-Driven Based Optimal Control The above model can be solved when f is available at time t.
A schematic diagram of the data-driven based optimal control However, this requirement is hardly achievable for today’s bulk
is shown in Fig. 1. The optimization path from the current to the power system with many stochastic components. Therefore, to
subsequent point of time is calculated according to the optimal achieve a real-time applicable method, f is approximated with a
problem for the subsequent point of time. Considering that the linear affine function in (3),
system is time-varying and that the optimization step is limited, v (t) = f(q(t)) = φ(t) + Ψ(t)q(t) (3)
in theory, it is not possible to achieve the optimal trajectory
(which is calculated with the actual system model), and only a where φ(t) is a constant vector, and Ψ(t) is the sensitivity matrix
near optimal trajectory will be obtained (the optimized trajec- based on Ψij(t) = ∂vi(t)/∂qj(t). Thus, the objective function in (2)
tory in Fig. 1). Clearly, the gap between the optimized trajectory becomes
and the optimal trajectory in theory depends on the optimization f(q(t-1) + Δq(t)) - vref (t)
algorithm, the optimization step and many practical issues. If
= φ(t-1) + Ψ(t-1)(q(t-1) + Δq(t)) - vref (t), (4)
the whole optimization strategy is well-designed, the final op-
timized trajectory will be close to the optimal trajectory in = v(t-1) - vref (t) +Ψ(t-1)Δq(t)
theory and be far better than the trajectory without optimization. and the optimization model in (2) can be modified as
Objective 2
max v (t -1)-vref (t ) + Ψ(t -1)q(t )
q ( t ) 2
Without optimization
Solution
s.t. q(t ) ≤ q(t ) ≤ q(t ) (5)
path in t1 q(t ) 1 ≤ C (t )
Optimized The above optimization problem is a quadratic programming
problem with v(t-1) and vref (t) already known and Ψ(t-1) to be
estimated online. This problem can be solved using many al-
Goal in t1 Optimal in theory gorithms [20].
t0 t1 t2 t3 t4 Time B. Framework
Fig. 1. Schematic diagram of the O2SL based control. The framework of V2SC is shown in Fig. 2 with three parts in
A. Data-Driven Based Optimal V2SC parallel:
P. 1: Data collection: power system operation data, including
The objective for V2SC in Fig. 1 is the bias between the
power flow, states of the capacitor and reactor groups and
voltage reference vector vref and the actual voltage vector v. The
switching states, are online measured in WAMS.
trajectory is the bias curve vref(t) - v(t) versus time t. For un-

1551-3203 (c) 2018 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2018.2856826, IEEE
Transactions on Industrial Informatics
3

Sequential time t time interval time t+1


Assuming the power system operates in a steady state, by
process
extending (3) for m operating points and assuming Ψ is the same
Power system Yes
Voltage level under small changes of v and q, the following equation holds
voltage
is acceptable? No approximately
monitoring

P.1 Data
P.1 Operation V = φ + ΨQ + ε (10)
data v(t),
collection acquisition q(t) The same with
P.2 Sensitivity procedure
P.2 Sensitivity as in the
estimation
estimation in
(10) ψ(t-1) time  v (1)   q (1)   
interval      
P.3 Sequential
v(t-1), vref(t), P.3 solving
(5) and taking
[t, t+1] V  , Q ,     (11)
 v (m)   q(m) 
upper and lower
voltage
Δq(t-1) bounds of q(t), strategies for  
control and C(t) [Δq(t)]C      
Trigger if
Details in Details in
max{v(t)- where v is a voltage vector of n buses, q is a vector of l reactive
Instructions Section III Section III
vref(t)}
<0.02p.u.
Part A Part B power loads, m is number of observed operation conditions, Ψ
Fig. 2. The Framework of V2SC is the sensitivity matrix with dimensions n×l, and ε is an error
matrix.
P. 2: Sensitivity matrix estimation: with the reactive power The model in (10) is a standard multi-variable regression
injections as the input and the voltage of each bus as the output, model and can be solved using many algorithms, e.g., the
the sensitivity matrix Ψ(t-1) can be online estimated using the weighted regression in [15] or the recursive regression in [16].
method in Section III, part A. If the operation data suffer from collinearity or noise issues,
P. 3: V2SC: by sequentially solving the optimization problem norm-2 penetration based algorithm or the noise-assisted algo-
in (5), Δq(t) is initially obtained; then with the method in Sec- rithm can be employed [17-18].
tion III part B, a modified Δq(t) for the reactive power control is
determined; and finally, the system voltage condition is im- B. Hysteresis Control Strategy
proved once the V2SC is carried out. The solution of (5) is a continuous value Δq which, in general,
is inappropriate for V2SC in real power systems because the var
III. KEY ISSUES IN THE FRAMEWORK compensators, such as the capacitor and reactor groups, are
discretely controlled. Therefore, Δq should be rounded as [Δq]
A. Sensitivity Matrix Estimation according to the discrete capacity of the capacitor and reactor
The incremental power flow equation in a matrix form is groups ([•] is the rounding operator).
shown in (6): However, to track the system operation trajectory, [Δq] may
Δp   J p J pv  Δ   Δ  be positive for time t but negative for time t+1, leading to fre-
Δq  =  J    = J   (6) quent back and forth reschedule actions in the sequential opti-
   q J qv  Δv   Δv  mization process, which thereby shortens the service life of the
where p and q are the injected active and reactive power vectors, switches and var compensators. In this situation, the following
respectively; δ and v are the voltage phase and magnitude hysteresis control strategy can be employed.
vectors, respectively; Δp, Δq, Δδ and Δv are the increments of p, Assume that [Δq(t)] is the suggested rescheduling value, and
q, δ and v, respectively; J is the Jacobin matrix; and Jpδ, Jpv, Jqδ [Δq(t)]C is the executed rescheduling value, where subscript C
and Jqv are the partitioned matrices of J. denotes the execution. The hysteresis control strategy deter-
Generally, Δv is a function of Δp, Δq and Δδ. In the main grid, mines the executed rescheduling value [Δq(t+1)]C at time t+1 by
however, the line reactance is far larger than the line resistance, comparing [Δq(t)]C and [Δq(t+1)] as follows:
leading to small values of Jpv and Jqδ. Therefore, Δp is primarily 1) if [Δq(t)]C and [Δq(t+1)] are all positive, then [Δq(t+1)]C =
determined by Δδ through Jpδ, and Δq is primarily determined min{[Δq(t)], [Δq(t+1)]};
by Δv through Jqv [21]. As a result, a reverse equation of (6) for 2) if [Δq(t)]C and [Δq(t+1)] are all negative, then [Δq(t+1)]C =
Δv and Δq can be derived as max{[Δq(t)], [Δq(t+1)]};
3) if either [Δq(t)]C or [Δq(t+1)] is zero, then [Δq(t+1)]C =
Δv = Jqv-1Δq (7)
[Δq(t+1)];
Alternatively, when V2SC is implemented, Δp is equal to 4) if [Δq(t)]C and [Δq(t+1)] have different directions, then
zero, and (6) can be modified as [Δq(t+1)]C = 0.
q  ( J qv  J q J p1 J pv )v =J R v (8) C. A Comprehensive Implementation Procedure
leading to the reverse form given by A comprehensive implementation procedure for P.2 and P.3
Δv = JR-1Δq (9) at time stamp t is shown as follows:
P.2.1: establish the sensitivity model in (10) using the
Comparing (3), (7) and (9), Jqv is an approximation of JR, and Ψ measured Q(t-1) and V(t-1) in P.1;
is an estimation of Jqv that can be used for V2SC.

1551-3203 (c) 2018 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2018.2856826, IEEE
Transactions on Industrial Informatics
4

P.2.2: estimate Ψ(t-1) by solving the equation (10) with the solve the model. The parameter τ for the weight function in the
noise-assisted algorithm in [18]. And then deliver Ψ(t-1) to NAER is 2, which is tuned with the method in [18]. Due to page
P.3.1. limitations and also because the parameter tuning is not the main
P.3.1: solve (5) for the optimal Δq(t) with v (t-1), vref (t), issue of this work, specific details on the tuning are not provided.
 q(t ) ,  q(t ) , and C(t) measured in P.1, and Ψ(t-1) estimated Readers who are interested in the tuning method can refer [18]
for more information.
in P.2.2;
Without loss of generality, we take the V2S estimated at the
P.3.2: round Δq(t) for [Δq(t)] according to the discrete ca-
first operation point as an example. q(t) is the reactive power
pacity of the capacitor and reactor groups;
vector of the nine loads, and v(t) is the voltage vector of the P-Q
P.3.3: if the hysteresis control strategy is employed, compare
and intermediate nodes. Fig. 4 illustrates examples of the V2S
the optimal result [Δq(t)] with [Δq(t-1)]C to determine the ex-
estimation at buses 9 and 20. For other buses, the results are
ecuted rescheduling value [Δq(t)]C;
similar to those in Fig. 4. Clearly, the estimated V2S to be
P.3.4: execute the rescheduling, and go to the next time stamp
employed in P.3 is similar to the results obtained by the per-
t+1 in Fig. 2.
turbation method, which is calculated with the offline system
model, thereby verifying effectiveness of the V2S estimation.
IV. CASE STUDY IN THE 36-NODE SYSTEM
D. Validation of V2SC
A. Background
In this part, we validate V2SC by comparing its performance
In this part, the 36-node system in [22] is employed to vali- at one operation point to that of IPOPT. The experiment is
date the proposed V2SC (the system configuration is omitted carried out with Monte-Carlo method for 10 trials. Due to page
due to page limitations). The simulation is conducted in Matlab limitations, we only take one trial in the following part as an
with software Power System Analysis Toolbox (PSAT) [23]. example, while the remaining trials have similar results.
The system has eight generators and nine loads. Most gen-
erators are equipped with governor and excitation systems. In
12
the power flow calculation, Generators 2, 4, 5, 7 and 8 are
conventional P-V nodes, Generator 3 is wind power, Generator 10
Power flow curve (p.u.)

6 is a synchronous condenser, Generator 1 is the V-δ node, and 8


the loads are P-Q nodes. There are 6 groups of capacitors and 4
groups of reactors at each load node, with each group having the 6
capacity of 50 Mvar, i.e., 0.5 p.u.. All the capacitor and reactor 4
groups are on the withdrawn mode.
2
B. Data Generation and Simulation of P. 1
0
To simulate power fluctuations caused by renewable energy 0 50 100 150 200 250 300 350
Time (day)
resources and load variations in real power systems, the active
Fig. 3. Active power flow of the generators and loads.
power of the P-V generators, the wind power and the ac-
tive/reactive load consumption in the nine loads are varying 0.015
Voltage/Var sensitivity (p.u.)

with time. The varying data are from real recorded power sys- NAER
Perturbation
tem load data reported in [24] or from real wind farm. Twen- 0.01
ty-four time-series records, with each record covering 365 days,
are employed (6 active power of generators, 9 active loads, and 0.005

9 reactive loads, as shown in Fig. 3), among them the 9 reactive


loads are used for Q(t) and the 28 P-Q bus voltages are used for 0
9 16 18 19 20 21 22 23 29
V(t) in the next subsection. The sampling interval in the original Node number

data is one hour, and in the simulation case, it is interpolated to 5 (a) Estimated voltage/var sensitivities of node 9.

minutes with a cubic spine; afterwards, 40 dB white noise is


0.05
Voltage/Var sensitivity (p.u.)

added to each record to manipulate short term renewable energy NAER


0.04
variations and load behaviors. Next, the data are used in PSAT Perturbation
0.03
for power flow calculation; the outputs are different operation
0.02
cases with manipulated power fluctuations. Afterwards, 45 dB
white noise is added to the power flow data as measurement 0.01

noise, and 105120 operation cases are finally obtained. 0


9 16 18 19 20 21 22 23 29
Node number
C. Sensitivity Estimation and Validation of P. 2 (b) Estimated voltage/var sensitivities of node 20.
With the measured Q(t) and V(t) from P. 1, model (10) is Fig. 4. Estimated V2S and the V2S calculated using the perturbation method.
established for the V2S estimation. The noise-assisted ensemble
regression method (NAER) proposed in [18] is employed to

1551-3203 (c) 2018 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2018.2856826, IEEE
Transactions on Industrial Informatics
5

TABLE II selected. The active power generation and active/reactive load


POWER GENERATION IN THE LIGHT LOAD CONDITION
consumptions are summarized in Tables II and III, respectively.
Generation No. 1 2 3 4
Active power (p.u.) 3.2258 2.6090 1.4760 0.6010 The bus voltages and voltage references are summarized in Fig.
Generation No. 5 6 7 8 5. For generator buses and high voltage side of the transformers,
Active power (p.u.) 2.7360 -0.010 1.1190 1.8070 the voltage references are set to 1.05 p.u.; for the remaining
buses, the voltage references are set to 1.0 p.u.. For P-V or V-δ
TABLE III
LOAD CONSUMPTIONS IN THE LIGHT LOAD CONDITION
nodes, the voltages are the same as the references. For other
Load No. 1 2 3 4 5 nodes, the voltages are in general higher than the references due
Active power (p.u.) 1.9767 2.4747 2.3311 0.4721 0.3283 to overloaded reactive power.
Reactive power (p.u.) 1.2305 1.1965 1.2545 0.3760 0.2284 In the optimal model (5), the objective function is sum of the
Load No. 6 7 8 9 squares of the differences between the 36 node voltages and the
Active power (p.u.) 0.3574 1.3730 1.4966 2.5592
Reactive power (p.u.) 0.2680 1.0849 0.7971 0.1117 voltage references; the constraints  q(t ) and  q(t ) are de-
termined by 1) the capacity limitation of the capacitor/reactor
TABLE IV
groups, from -2 p.u. to 3 p.u. with a minimum step 0.5 p.u.; 2)
COMPARISON OF THE OPTIMIZATION RESULTS
Node Origi- Sensitivity based V2C IPOPT the maximum rescheduling value in the optimization step,
No. nal Step 1 Step 2 Step 3 Step 4 Step 5 Final which is ±1 p.u.; and 3) the constraint C(t) to limit the total
9 - 0.5 -0.5 -1 0.5 0 -0.5 -0.5 rescheduled power in a reasonable range, which is 9 p.u..
16 - -1 -1 0 0 0 -2 0.5 The proposed V2SC is sequentially conducted until the
18 - 0.5 -1 -0.5 0 0 -1 -1
19 - -0.5 0.5 0 0 0 0 0
change of the node voltages between two iterations is less than
20 - -1 -0.5 0 0 0 -1.5 -1 0.001p.u., i.e., ||v(t) - v(t-1)||2 < 0.001 p.u., or the number of the
21 - -1 -1 0 0 0 -2 -2 optimization iterations exceeds 10, i.e., 10 steps of optimization
22 - 1 -1 -1 -1 0 -2 -2 have been completed. In this case, after four sequential opti-
23 - -1 -1 0 0 0 -2 -2 mization steps, ||v(t) - v(t-1)||2 < 0.001 p.u., therefore, the opti-
29 -1 1 0.5 0 0 0.5 -2
mization ends at the fifth step. Δq(t), ||Δq(t)||1 and ||v(t) - vref(t)||2
||ΔQ(t)||1 - 7.5 7.5 3 1.5 0 - 11
Obj. 0.2931 0.1747 0.1157 0.1104 0.1074 0.1074 - 0.1074 at each step is summarized in Table IV, and the final v(t) - vref(t)
Note: “Obj.” represents the value of ||v(t) - vref(t)||2. vector is shown in Fig. 6. The results are compared with those
using IPOPT based centralized optimization (we will use
IPOPT for short in the remaining sections) [9].
Table IV shows V2SC and IPOPT achieves the same objec-
tive result as 0.1074 p.u.. Fig. 6 shows the voltage bias at the
controllable nodes, which indicates the voltage biases for both
methods are significantly smaller than the original situation.
Both methods guarantee max{v(t) - vref(t)} < 0.05 p.u., indi-
cating satisfied voltage control is achieved. Although the total
Δq(t) of V2SC is different from that of IPOPT, the former
performs even better because its value of max{v(t) - vref(t)} at
node 14 is smaller than that of the latter.
Regarding the computation time, it takes several seconds for
the V2SC, whereas considerably more time is required for
Fig. 5. Bus voltages and the expected voltage references.
IPOPT, especially for larger-scale power systems. Therefore,
V2SC is more suitable for online applications.
E. Validation of P. 3
In this part, we validate P.3 with Monte-Carlo method for 10
trials. For each trial, the optimal model (5) and the stop condi-
tion are the same as in the above subsection. The capacity range
of each capacitor/reactor group is -2 to 3p.u., with a minimum
rescheduling step of 0.5 p.u.; the maximum rescheduling step is
±1 p.u., and the constraint C(t) is 9 p.u..
V2SC is implemented with or without the hysteresis control
(here we use H-V2SC to represent the former), and is compared
with the optimal result in theory obtained with IPOPT. The
Fig. 6. v(t) - vref(t) before and after the optimization. sequential control is carried out every 30 minutes, leading to 48
reschedule actions per day. Note that, the above settings are
One operation point among the 105120 ones is randomly reasonable in modern power systems with renewable energy

1551-3203 (c) 2018 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2018.2856826, IEEE
Transactions on Industrial Informatics
6

resources and load variations, where the voltage control should


be carried out along with the operation trajectory.
The results are summarized in Table V. It is concluded that
V2SC and H-V2SC have similar performances in minimizing
the objective function, and both perform very close to the op-
timal result in theory with IPOPT. This is because V2SC and
H-V2SC have the capability to track the system operation tra-
jectory in stochastic operation conditions. With the hysteresis
control strategy, the total rescheduled actions in H-V2SC is far
less than those in V2SC, indicating the effectiveness of the
strategy in preventing unnecessary back and forth actions.
Further, one trial of the ten is randomly selected as an ex-
Fig. 7. Load Curves in the Day-Selected.
ample to illustrate the control behaviors of V2SC and H-V2SC.
Fig. 7 shows the load curves during the sequential control, and
Fig. 8 shows curves of the objective function. It can be figured
out the original objective function varies along with the varying
system operation conditions, and is reduced immediately after
the initial round of rescheduling. Performances of V2SC and
H-V2SC are close to that of IPOPT, indicating the effectiveness
of the proposed V2SC and H-V2SC.
Without loss of generality, the voltage curves of two ran-
domly selected nodes in the load center of the system, 21 and 22,
are examined in Fig. 9. For other nodes, the results are similar to
those in Fig. 9. The similarity of these curves with optimization
indicates V2SC and H-V2SC have similar capacity of tracking
Fig. 8. Objective value of the sequential optimization.
system optimal operation trajectory, and therefore are suitable
for voltage control in stochastic power system conditions.
At last, the executed rescheduling values [Δq(t)]C of V2SC
and H-V2SC are compared in Fig. 10. It is shown that V2SC
takes in general much more control actions than H-V2SC. For
example, at Nodes 9 and 16, V2SC takes 39 control actions,
while H-V2SC takes only 24 and 21 control actions, respec-
tively, indicating the hysteresis control strategy in the H-V2SC
does prevent unnecessary back and forth rescheduling actions.
Another interesting finding is that [Δq(t)]C for H-V2SC is in
general smaller than that for V2SC, indicating H-V2SC is more
preferable as it leaves more control capacity for system excep-
tions in the future, such as transient faults.
(a) Voltage bias curves for node 21.
TABLE V
COMPARISON OF V2SC, H-V2SC AND IPOPT IN P.3
Trial Original V2SC H-V2SC Obj. in theory
No. Obj. Obj. Actions Obj. Actions with IPOPT
1 0.2272 0.1153 171 0.1143 107 0.0990
2 0.2397 0.1326 132 0.1139 118 0.0989
3 0.2375 0.1451 87 0.1161 111 0.0992
4 0.2324 0.1189 186 0.1253 134 0.0984
5 0.2423 0.1169 187 0.1185 112 0.0989
6 0.2247 0.1204 202 0.1173 111 0.0988
7 0.2209 0.1074 117 0.1155 100 0.0986
8 0.2286 0.1140 173 0.1133 99 0.0983
9 0.2423 0.1169 159 0.1249 91 0.0983
10 0.2299 0.1091 142 0.1184 118 0.0981
Note: “Obj.” represents the mean value of ||v(t) - vref(t)||2 along the day trajectory,
(b) Voltage bias curves for node 22.
while “Actions” indicates the number of total rescheduled actions.
Fig. 9. Comparison of voltage bias at nodes with/without optimization.

1551-3203 (c) 2018 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2018.2856826, IEEE
Transactions on Industrial Informatics
7

(a) Δq(t) curves for nodes 9, 16, 18 and 19.

Fig. 11. The Nordic32 System.

P.1 to P.3 are carried out in this system as follows:


P.1: Operation data is generated with the same method as in
the 36-Node system. Renewable energy resources and load
variations are incorporated by injecting load flow fluctuations.
(b) Δq(t) curves for nodes 20, 21, 22 and 23.
After the data generation, there are 105120 operation cases, and
Fig. 10. Comparison of control actions of the sequential optimization methods.
each case includes 20 active power generations, 22 active loads,
F. Section Remarks 22 reactive loads and 74 bus voltages. The 22 reactive power
The proposed H-V2SC is a purely data-driven method suit- from the loads are used for Q(t), while the 74 bus voltages are
able for online voltage control. Compared to IPOPT, it does not used for V(t).
need the system offline models. With the hysteresis control P.2: With the measured Q(t) and V(t) from P. 1, the sensi-
strategy, it can prevent frequent back and forth rescheduling tivity matrix Ψ(t) are estimated using the same identification
actions, yet maintain the same capacity as V2SC to track the method as in the 36-Node system. Due to page limitations and
system optimal operation trajectory. without loss of generality, the V2S estimation results of the 22
reactive power injections to the bus voltages at Nodes 35 and 36
are shown in Fig. 12. The other V2S estimations are similar. It
V. CASE STUDY IN THE NORDIC32 SYSTEM
can be figured out the V2S estimation is almost the same as the
The Nordic32 system in Fig. 11 is a benchmark system true value obtained by the offline model based perturbation
originated from the Swedish and Nordic power system, and is method, which thereby verifies the effectiveness of the sensi-
officially recommended by IEEE Power & Energy Society for tivity estimation.
voltage stability analysis and security assessment [25]. In this P.3: For each day of the 365 days, the proposed H-V2SC and
part, we validate the proposed method in the system with V2SC are carried out from the initial system operation point. All
software DIgSILENT to illustrate the capability of the proposed the simulation/control settings including the voltage references,
method to be applied in large power systems. the optimal model, the stop condition for the sequential opti-
There are 20 generators, 22 loads and 74 buses in the system. mization, the capacity range of each capacitor/reactor group, the
All generators are equipped with step-up transformers and over minimum/maximum rescheduling step, and the constraint C(t),
excitation limiters. In the power flow calculation, all generators are the same as in the 36-Node system.
except Generator 20, which is a V-δ node, are considered as P-V Due to page limitation and without loss of generality, we pick
nodes. The load nodes are P-Q nodes. 12 groups of capacitors the result of one day as an example. The objective functions
with or without optimization are shown in Fig. 13, which illus-
and 8 groups of reactors are implemented at each load node,
trates the effectiveness of the proposed H-V2SC and V2SC for
with the capacity of each group 25 Mvar. In the initial operation
improving the system voltages along with the system operation
condition, all capacitors and reactors are withdrawn.
trajectory. In addition, V2SC and H-V2SC have almost the

1551-3203 (c) 2018 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2018.2856826, IEEE
Transactions on Industrial Informatics
8

same performances in Fig. 13, which is because the executed VI. CONCLUSION
reactive power rescheduling values are almost the same during This paper proposes a data-driven based optimization method
the sequential control process, as illustrated in Fig. 14. It indi- for var-voltage sequential control. Simulation results in the
cates H-V2SC will degenerate to V2SC when the rescheduling
36-Node system and the Nordic32 system verify that the pro-
actions are necessary, which further validates applicability of
posed method has similar optimization performance to the
the method.
conventional IPOPT and has much faster calculation speed in
the field of power system voltage control; it has a capability to
draw the system voltages to the near optimal condition along
with the system operation trajectory without relying on the
offline physical model, and is more suitable in modern power
systems with stochastic renewable energy resources; in addition,
with an integrated hysteresis control strategy, the proposed
method can prevent frequent back and forth rescheduling ac-
tions, thereby can save the service life of switches and var
(a) V2S of the 22 reactive power injections to the bus voltage at Nodes 35.
compensators, and is feasibly applied in practice.

REFERENCES
[1] C. W. Taylor, Power System Voltage Stability. New York: Mcgraw-Hill,
1993.
[2] Voltage Stability Assessment: Concepts, Practices and Tools, IEEE PES
Taskforce Report PES-TR9, SP101PSS, 2002.
[3] B. Zhou, K. W. Chan, T. Yu, H. Wei and J. Tang, “Strength Pareto
Multigroup Search Optimizer for Multiobjective Optimal Reactive Power
Dispatch,” IEEE Trans. Ind. Inf., vol. 10, no. 2, pp. 1012-1022, May
2014
[4] Tabatabaei, Naser Mahdavi, A. J., Bizon and N., & Blaabjerg. Reactive
(b) V2S of the 22 reactive power injections to the bus voltage at Nodes 36. Power Control in AC Power Systems. Springer International Publishing,
Fig. 12. Estimated V2S and the V2S calculated by perturbation method. 2017.
[5] S. M. Mohseni-Bonab and A. Rabiee, “Optimal Reactive Power Dispatch:
A Review, and A New Stochastic Voltage Stability Constrained Mul-
ti-objective Model at the Presence of Uncertain Wind Power Generation,"
IET Gener., Transm. & Distrib., vol. 11, no. 4, pp. 815-829, 3 9 2017.
[6] Z. Yang, A. Bose, H. Zhong, N. Zhang, Q. Xia and C. Kang, “Optimal
Reactive Power Dispatch With Accurately Modeled Discrete Control
Devices: A Successive Linear Approximation Approach,” IEEE Trans.
Power Syst., vol. 32, no. 3, pp. 2435-2444, May 2017.
[7] S. A. Arefifar and Y. A. R. I. Mohamed, “Probabilistic Optimal Reactive
Power Planning in Distribution Systems With Renewable Resources in
Grid-Connected and Islanded Modes,” IEEE Trans. Ind. Electron., vol.
61, no. 11, pp. 5830-5839, Nov. 2014.
[8] E. M. Soler, E. N. Asada and G. R. M. da Costa, “Penalty-Based Non-
linear Solver for Optimal Reactive Power Dispatch With Discrete Con-
trols,” IEEE Trans. Power Syst., vol. 28, no. 3, pp. 2174-2182, Aug.
2013.
Fig. 13. The objective functions with or without optimization. [9] Wei Yan, Juan Yu, D. C. Yu and K. Bhattarai, “A New Optimal Reactive
Power Flow Model in Rectangular Form and Its Solution by Predictor
Corrector Primal Dual Interior Point Method,” IEEE Trans. Power Syst.,
vol. 21, no. 1, pp. 61-67, Feb. 2006.
[10] W. Sheng, K. y. Liu, S. Cheng, X. Meng and W. Dai, “A Trust Region
SQP Method for Coordinated Voltage Control in Smart Distribution
Grid,” IEEE Trans. Smart Grid, vol. 7, no. 1, pp. 381-391, Jan. 2016.
[11] T. Ding, Q. Yang, Y. Yang, C. Li, Z. Bie and F. Blaabjerg, “ A Da-
ta-Driven Stochastic Reactive Power Optimization Considering Uncer-
tainties in Active Distribution Networks and Decomposition Method,”
IEEE Trans. Smart Grid, Early Access.
[12] C. Andalib-Bin-Karim, X. Liang, N. Khan and H. Zhang, “Determine
Q–V Characteristics of Grid-Connected Wind Farms for Voltage Control
Using a Data-Driven Analytics Approach,” IEEE Transactions on Ind.
Appl., vol. 53, no. 5, pp. 4162-4175, Sept.-Oct. 2017.
[13] D. K. Khatod, V. Pant and J. Sharma, “A Novel Approach for Sensitivity
Calculations in the Radial Distribution System,” IEEE Trans. Power
Delivery, vol. 21, no. 4, pp. 2048-2057, Oct. 2006.
[14] D. Ghosh, T. Ghose and D. K. Mohanta, “Communication Feasibility
Analysis for Smart Grid With Phasor Measurement Units,” IEEE Trans.
Fig. 14. Executed rescheduling actions during the sequential optimization. Ind. Inf., vol. 9, no. 3, pp. 1486-1496, Aug. 2013.

1551-3203 (c) 2018 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.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TII.2018.2856826, IEEE
Transactions on Industrial Informatics
9

[15] J. Zhang, C. Y. Chung and Y. Han, “Online Damping Ratio Prediction Zetao Jiang was born in Guangdong, China, in
Using Locally Weighted Linear Regression,” IEEE Trans. Power Syst., 1993. He received his B.Eng. degree from the
vol. 31, no. 3, pp. 1954-1962, May 2016. School of Electric Power, South China University
[16] Y. C. Chen, A. D. Domínguez-García, and P. W. Sauer, “Measure- of Technology. He is currently pursuing his M.Eng.
ment-Based Estimation of Linear Sensitivity Distribution Factors and degree with the same department. His research
Applications,” IEEE Trans. Power Syst., vol. 29, no. 3, pp. 1372 – 1382, interests include optimal energy management,
May 2014. operation control and system planning of
[17] J. Zhang, Z. Wang, X. Zheng, L. Guan and C. Y. Chung, “Locally traditional power systems, multi-energy systems
Weighted Ridge Regression for Power System Online Sensitivity Identi- and smart grid.
fication Considering Data Collinearity,” IEEE Trans. Power Syst., Early
Access.
[18] J. Zhang, C. Y. Chung and L. Guan, “Noise Effect and Noise-Assisted
Ensemble Regression in Power System Online Sensitivity Identification,” Lin Guan (M’14) received the B.Eng. degree and
IEEE Trans. Ind. Inf., vol. 13, no. 5, pp. 2302-2310, Oct. 2017. the Ph.D. degree in electrical engineering from
[19] Box, George E. P., Gwilym M. Jenkins, Gregory C. Reinsel, and Greta M. Huazhong University of Science and Technology,
Ljung. Time Series Analysis: Forecasting and Control. Hoboken, NJ: China, in 1990 and 1995, respectively. She is now a
John Wiley & Sons, 2016. professor and the chair in the School of Electric
[20] Radoslav, Loki. Sequential Quadratic Programming. Optimization Power, South China University of Technology. Her
Theory and Methods. Springer US, 2006:523-560. research interests include power system stability and
[21] A. J. Monticelli, A. Garcia and O. R. Saavedra, “Fast Decoupled Load control, planning and operation, and smart grid
Flow: Hypothesis, Derivations, and Testing,” IEEE Trans. Power Syst.,
vol. 5, no. 4, pp. 1425-1431, Nov 1990.
[22] Z. Wu and X. Zhou, “Power System Analysis Software Package
(PSASP)-An Integrated Power System Analysis Tool”, Power System
Technology, 1998. Proceedings. POWERCON '98. 1998 International
Conference on, Beijing, 1998, pp. 7-11 vol.1.
[23] F. Milano, “An Open Source Power System Analysis Toolbox,” IEEE
Trans. Power Syst., vol. 20, no. 3, pp. 1199-1206, Aug., 2005.
[24] [Online]. Available: http://www.pjm.com/markets-and-operations/ops
-analysis/historical-load-data.aspx
[25] Test Systems for Voltage Stability Analysis and Security Assessment,
IEEE PES Taskforce Report PES-TR19, Aug. 2015, [online] Available:
http://resourcecenter.ieee-pes.org/pes/product/technical-reports/PESTR1
9.

Junbo Zhang (S’10-M’14) received his B.Eng. and


Ph.D. from Tsinghua University in 2008 and 2013,
respectively, followed by a postdoc fellow at the
same university. He studied at The Hong Kong
Polytechnic University from 2009 to 2011 under a
collaborative program between Tsinghua University
and The Hong Kong Polytechnic University. He is
now an associate professor at the School of Electric
Power, South China University of Technology. His
research areas include power system stability and
control, knowledge management in power systems and power system eco-
nomics.

Zhihao Chen (S’17) was born in Guangdong, China,


on September 25, 1995. He received his B.Eng.
degree in electrical engineering from the School of
Electric Power, South China University of Tech-
nology, in 2018. He has studied power system anal-
ysis, statistics, and several power and energy engi-
neering related courses. His research interests in-
clude power system stability and control, system
identification and data analysis in power system.

Chuyao He received the B.Eng. degree and M.Eng.


degree in electrical engineering from the School of
Electric Power in South China University of
Tehnology, Guangzhou, China, in 2014 and in 2017,
respectively. Her current research interest is Voltage
Stability and Data Mining. She is currently an
engineer in the Guangzhou Power Supply of China
Southern Power Grid, Guangzhou.

1551-3203 (c) 2018 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.

Potrebbero piacerti anche