Loss Minimization for Distributed PV Integration

using Particle Swarm Optimization with Volt-Var

Control
Nathan Kelly Kaveh Dehghanpour Hashem Nehrir
Bonneville Power Administration Electrical and Computer Engineering Electrical and Computer Engineering
Portland, Oregon 97208-3621 Montana State University Montana State University
[email protected] Bozeman, Montana 59717-3780 Bozeman, Montana 59717-3780
[email protected] [email protected]

Abstract—The integration of renewable energy resources into

electric power distribution networks can cause major control
problems, if not planned properly. In this paper, Volt-Var control
(VVC) has been implemented to reduce voltage fluctuation
resulting from high Photovoltaic (PV) penetration. The effect
of VVC on system active losses is also investigated. Additionally,
Particle Swarm Optimization (PSO) has been used to minimize
system losses while maintaining bus voltages within permissible
limits in different test distribution systems.

I. I NTRODUCTION
Fig. 1. Volt-Var droop curve with ANSI C84.1 limits, ±5% of the nominal
The rapid pace of interconnection of Distributed Energy voltage, enforced
Resources (DERs) has been a topic of research for many
years. Photovoltaic (PV) generation system integration has
particularly been of interest due to the continuing falling power grid. Therefore, PV inverters offer a convenient option
cost of PV installation, as well as the dramatic adoption of to regulate voltage as they are capable of providing or absorb-
distributed and centralized PV systems [1]. The growth of PV ing reactive power at a fast time-scale.
systems at the distribution level has faced major hurdles due Indeed, the attractiveness of PV inverters for voltage reg-
to the potential voltage rise that can occur at the Point of ulation has garnered industry collaboration to identify how
Common Coupling (PCC) with the main grid at certain times they may be integrated into the grid [9]. This has led to an
of day [2]. During normal operation conditions, this voltage IEEE working group preparing a new standard, P1547.8 [8], to
deviation should be constrained within ±5% of the nominal address whether local voltage regulation should be a capable
bus voltage according to the ANSI Standard C84.1 [3]. function of new smart inverters. The most recognized method
To keep the voltage rise below the permissible limit defined of voltage regulation is Volt-Var Control (VVC), which is par-
by the ANSI Standard, several techniques have been studied ticularly attractive as it offers a customizable active approach
and, in some cases, deployed. Voltage regulators and On-Load to an inverter’s Var response. Electric Power Research Institute
Tap Changers (OLTCs) can be dispatched to reduce voltage (EPRI) has studied different inverter Var control methods and
levels ahead of high PV power generation time intervals [4]. concluded that VVC can provide adequate voltage support
However, they cannot guarantee to keep the voltage levels for voltage variations on the primary and secondary levels
within bounds throughout the entire feeder [5] [6]. Shunt resulting from changes in PV output power in the system [9].
capacitors can also be used under periods of low sunshine While the theory on which the concept of VVC relies on is
to provide Var support, but, similar to voltage regulators and not necessarily a new one, its relative simplicity makes it an
OLTCs, they do not necessarily have the capability to respond attractive and easily implementable option for modern power
quickly during transient events, such as clouds passing over systems [10]. In its simplest form, VVC consists of a piecewise
the PV panels [7]. Moreover, recently, the IEEE has amended linear curve (or droop curve), as shown in Fig. 1, defined by a
standard 1547 to allow distributed resources to actively support collection of Volt/Var pairs that may, for instance, be chosen by
voltage regulation at the PCC in coordination with the area the designated utility company. This simple droop curve will
Electric Power System (EPS) Operators [8]. Hence, inverter allow the inverter to absorb reactive power when PV terminal
Var control is becoming a simple, attractive, and somewhat voltages are higher than the nominal and inject reactive power
autonomous substitute method for voltage regulation in the when PV terminal voltages are lower than the nominal.

978-1-5386-7703-2/18/$31.00 ©2018 IEEE

Authorized licensed use limited to: Motilal Nehru National Institute of Technology. Downloaded on July 12,2022 at 11:53:55 UTC from IEEE Xplore. Restrictions apply.
 One disadvantage of the VVC droop curve is that PV
terminal voltages can be higher than the nominal value during
the time intervals with higher PV power output, which leads to
higher reactive power absorption. This higher reactive power
absorption by the inverter will, in turn, decrease voltage levels
but will also lead to higher current flow on the lines and
thus higher active system losses. To counteract this effect,
[5] studied the optimal var injections on a radial network.
This study focused on a fast timescale and large voltage
fluctuations when applying a convex relaxation to the power
flow formulation. The authors in [11] modeled DERs as power-
Fig. 2. OpenDSS PVSystem equivalent model diagram
voltage nodes and used Genetic Algorithm (GA) to obtain a
solution for minimizing system losses. An adaptive algorithm
was implemented in [7] on a simple unidirectional radial some cases, real power can be curtailed for more reactive
distribution system to introduce a balance between voltage power availability. However, this paper does not tackle this
regulation and active loss minimization. However, the model issue and will give the priority to real power production in all
was not extended to a larger system. In [12] the authors cases.
addressed the stability of locally controlled droop coefficients The control scheme in this instance was used on a modeled
in VVC. The paper adopts a robust optimization approach, smart inverter connected to a PV System in OpenDSS, as
which depends on high certainty levels in the PV power shown in Fig. 2. OpenDSS is an open source distribution
forecasting. It is demonstrated that optimal droop coefficient system simulator that provides a comprehensive tool for mod-
selection can have superior performance compared to standard eling and evaluating electric utility distribution systems. The
rule-of-thumb selections. software was used to perform a daily power flow study of
In this paper, we have concentrated on VVC and system distribution feeders with variable controls. One of the benefits
energy loss minimization using Particle Swarm Optimization of OpenDSS is that it models control elements separately from
(PSO) [13] for PV inverters over a 24-hour period. PSO was the actual current carrying element. Thus, for example, the
used to minimize the system power losses by finding the capacitor control is separate from the capacitor model and the
optimal droop curve to absorb the least amount of reactive regulator control is separate from the tap-changing transformer
current while ensuring that the voltage levels at the PV that performs the regulation [9].
terminals remain within limits. In order to properly model the The VVC is capable of controlling the power output out of a
behavior of a smart inverter capable of VVC, we have utilized generic generator model. The generic model, in this case, is a
OpenDSS platform [14], developed at EPRI. The behavior of PV system which takes insolation data and converts it first into
VVC is tested on a IEEE 13-bus feeder system with centralized a DC signal which is passed through an input filter capacitor
PV in one case study and distributed PV in another. The PSO- and is then converted from DC to AC with an inverter. This
based VVC method is then successfully extended to a larger model is capable of finding the maximum power point of the
feeder system consisting of 1300 residential, commercial, and PV system quickly and can be considered an adequate PV
light industrial customers via 58 total miles of line length. representation for this study.
Another benefit of the OpenDSS platform is that it supports
II. VVC M ODELING a Component Object Model (COM) interface that allows the
The methodology of the VVC scheme used in this paper user to use the power flow capabilities of the program to
follows a straightforward linear piecewise curve to (Fig. 1) perform analysis and obtain custom solutions with an external
inject or absorb the appropriate reactive power based on the program. In this case MATLAB was used to analyze and
voltage measured at the PV terminal. This curve has Volt/Var loop over solutions, as well as implement a unique PSO code
pair coefficients of (.95 p.u., 1 p.u.) and (1.05 p.u., -1 p.u.) over OpenDSS solution iterations. The steps for the OpenDSS
which implies that when the voltage drops below the nominal solution [9] are as follows:
value to the ANSI C84.1 limit (.95 p.u.), maximum available 1) Initialize node voltages
reactive power (1.0 p.u.) is generated by the inverter and when 2) Obtain compensation currents from loads
the voltage rises above the nominal value to the ANSI C84.1 3) Solve for node voltages
limit (1.05 p.u.), maximum available reactive power (-1.0 p.u.) 4) Repeat steps 2 and 3 until power flow converges
is absorbed. The term “available” is important in this case, as 5) Sample control element inputs
inverter systems have fixed apparent power limits and thus, 6) Take control actions
a fixed reactive power limit. This relationship is given by 7) Repeat steps 2 through 6 until no more control actions
p
Qmax = (S max )2 − (P )2 , where Q defines the inverter are needed at this time step
reactive power in Vars, S is apparent power in Volt-Amps, and This solution algorithm ensures that there is a higher chance
P is the real power, in Watts. This implies that the reactive that power flow convergence occurs and allows for easier
power available will vary as the solar intensity changes. In integration of control schemes in existing models.

Authorized licensed use limited to: Motilal Nehru National Institute of Technology. Downloaded on July 12,2022 at 11:53:55 UTC from IEEE Xplore. Restrictions apply.
 III. PSO-BASED VVC 2700

PSO is a probabilistic search algorithm based on a simplified 2600

social model closely related to swarming theory [15]. PSO 2500

Power (kW)
makes use of a velocity vector, to update the current position 2400

of each particle in the swarm. The position of each particle 2300

is updated based on the knowledge from the swarm, which 2200

enables it to adapt to its environment by returning to promising 2100

regions that were previously discovered [15]. Unlike other 2000

50 55 60 65 70
Hour
evolutionary algorithms, PSO is unique in that all population
members survive from the beginning of a trial until the end Fig. 3. Variable load curve measured at bus 650
[16]. The steps for convergence involve a randomly distributed
initial starting position and velocity, represented by vectors min f (α1 , ..., αm )
α1 ,...,αm
Xi , and Vi , respectively. As variables are updated and the s.t. g(α1 , ..., αm ) ≤ 0
trajectory of each particle shifts to new regions of the search (4)
space, the particles will then cluster around an optima and P F (α1 , ..., αm ) = 0
a final solution is obtained [16]. The methodology for the vmin ≤ vi ≤ vmax ∀i
algorithm is based on [13] and is implemented in MATLAB where, α1 , ..., αm represent the droop parameters of the in-
with OpenDSS. Over the 24-hour interval, openDSS solves the verters, g denotes the constraints on the droop parameters
power-flow iterations and calculates the total system losses. (i.e., minimum/maximum values), P F defines the power flow
The PSO then uses the 24-hour total system losses as the equations which are provided and modeled in OpenDSS, and
fitness function to be minimized. The dynamics of the PSO f is the objective function of the optimization problem (e.g.,
algorithm, with i denoting the positions of the ith particle, can power loss or voltage deviation) to be minimized, and vi is the
be represented as follows: voltage level at the ith node, which is kept within permissible
Vik+1 = ωkVik + c1 r1 (pBest
pBestki − Xik ) bounds ([vmin , vmax ]).
(1)
gBestki − Xik )
+ c2 r2 (gBest IV. N UMERICAL R ESULTS
Xik+1 = Xik + Vik+1 (2) A. Case Study I: 13-Bus System
ωmax − ωmin We will first look at a simple example of VVC on a
ωk = ωmax − k, (3)
kmax IEEE 13-bus distribution feeder example, which is a small
where c1 and c2 are tunable coefficients, r1 and r2 are distribution test feeder operating at 4.16 kV. The system is
uniformly generated random numbers over the interval [0,1], characterized as being relatively heavy loaded, and containing
pBestki and gBestki serve as the personal and global best overhead and underground lines. The peak load supplied
solutions, respectively, kmax denote the maximum number through the substation is about 2.7 MW, as shown in Fig. 3. To
of iterations, and ωk acts as weight parameter with ωmax ensure realistic results, variability has been added to the loads
and ωmin the maximum and minimum values of that weight. that follow a common, hourly, normalized, load shape that is
The parameters were tuned numerically and set to the values provided by EPRI in the OpenDSS example documentation
which showed the best performance. The values were c1 = 1, [14]. A 2-MW, 2.2-MVA PV farm was added to the 13-bus
c2 = 0.1, ωmax = 0.9, ωmin = 0.1, and kmax = 350. The system at bus 675. Insolation data for the PV was synthetic
size of the swarm of particles was set to be 50. data developed with GridPV toolbox and is meant to represent
For each PV inverter the VVC droop curve was determined one day in Livermore, CA with a partly cloudy morning and
by splitting the decision variables into two Volt/Var pairs. clear afternoon [17]. The insolation data was normalized and
One volt/Var pair for values below nominal voltage, and set to 1 minute intervals as displayed in Fig. 4.
one Volt/Var pair for above nominal voltage. Thus, the only The reverse power flow from the PV farm with no VVC
constraint needed was maximum and minimum voltages for caused the voltage at the PV terminal, bus 675, to rise to
above and below nominal. This was set Vmin = 0.93 p.u.
and Vmax = 0.99 p.u. for below nominal and Vmin = 1.01
1.2
p.u. and Vmax = 1.07 p.u. for above nominal. This allows the
1
VVC to be a bit looser in its voltage regulation and will ensure
0.8
legitimate droop curves will be considered. For instance, with
5 PV generators on the IEEE 13-bus system and 2 volt/Var 0.6

pairs for each VVC curve, the vector length for each particle 0.4

in the PSO algorithm is 10. 0.2

VVC can be cast as a constrained optimization problem, 0

0 500 1000 1500
which can be solved using PSO. This optimization problem is Minutes

formulated as follows:
Fig. 4. The normalized PV insolation data

Authorized licensed use limited to: Motilal Nehru National Institute of Technology. Downloaded on July 12,2022 at 11:53:55 UTC from IEEE Xplore. Restrictions apply.
 1.06 -490
PV
-491
1.05

System Losses (kWh)

-492
PV w/ VVC
1.04
Voltage (p.u.)

-493

No PV -494
1.03

-495
1.02
-496

1.01 -497

1
-498
50 55 60 65 70 0 50 100 150 200 250 300 350 400
Hour Iterations

Fig. 5. IEEE 13-bus Voltage comparisons for bus 675 on three different test Fig. 7. Fitness convergence for the PSO algorithm
runs: without PV, with PV, and with PV with VVC
TABLE I
T OTAL SYSTEM ENERGY LOSSES IN K W H

Configuration System Energy Losses

NO PV 807.8 kWh
2 MW Centr. PV 646.6 kWh
2 MW Centr. PV w/ VVC 675 kWh
2 MW Distr. PV 526.5 kWh
2 MW Distr. PV w/ VVC 528.3 kWh
2 MW Distr. PV w/ PSO 490.7 kWh

losses by 7.2% over previous VVC with uniformly set droop

curves and 6.8% over the PV system with no local voltage
regulation. Table I gives a breakdown of the total system losses
Fig. 6. IEEE 13-bus test feeder with distributed PV totaling 2 MW
for various run schemes with the lowest value observed when
the PSO is utilized.
1.055 p.u., as seen in Fig. 5. This rise would fall outside
the ANSI C84.1 limit. For this reason a VVC droop curve B. Case Study II: Larger Feeder System Test
with Volt/Var pairs of (.96 p.u., 1 p.u.) and (1.04 p.u., -1 The methodology was then extended to a larger feeder
p.u.) is implemented on the smart inverter connected to the system provided by the EPRI’s Distributed PV Monitoring
PV system. The resulting voltage with this simple control and Feeder Analysis database [18]. The 12-kV feeder system
scheme is shown again in Fig. 5. From the figure we can shown in Fig. 8, has a peak load of approximately 6 MW that
see that the peak voltage is now reduced to under 1.04 p.u. serves close to 1300 customers. The feeder is studied with 13
and voltage ramps during morning and afternoon hours have distributed PV generators ranging from large 3-phase 1.2 MW
been assuaged. Another benefit of the smart inverter with systems to single-phase rooftop PV generators with 8-10 kW
VVC comes during non-daylight hours. Indeed, we see the capacities, giving a penetration level of around 45%. The goal
inverter providing voltage support during hours when the sun is to deploy the algorithm on an existing feeder under high PV
is not shining (hours 48-53 and 66-72). This allows the utility penetration of various sizes from various locations throughout
company the ability to create control schemes specifically for the distribution network. When run without VVC, the terminal
Var or voltage support at night. This concept, however, may voltage for the 1.2 MW PV system strays above the 5% ANSI
lead to higher system losses if not properly accounted for, limit.
as reactive power absorption to lower terminal voltage will Table II shows the total system losses over the 24-hour
continue if the node bus voltage operates above nominal value. period for different configurations. The PSO algorithm yielded
To further evaluate the performance and capability of VVC, a sufficient droop curve to reduce the overvoltage on the 1.2
a set of distributed PV generators were added to the IEEE MW PV bus (see Fig. 9). In addition, the PSO algorithm was
13-bus test feeder. PV generators were added to buses 633, able to reduce energy losses by 12.2% when compared to a
675, 680, 611, and 646, as seen in Fig. 6. The peak power uniform VVC droop curve for all the 13 PV generators. This
capacity of the collective PV generators was 2 MW, with PVs centralized PSO scheme may be too burdensome especially
at buses 646 and 611 having 200 kW capacity and the other 3
having 600 kW capacity, for a penetration level of around 75%.
The discussed PSO-based VVC algorithm was introduced to TABLE II
provide Volt/Var support through PV inverters. The output of T OTAL SYSTEM ENERGY LOSSES IN MW H
the algorithm produces 10 optimal volt/Var pairs, 2 for each Configuration System Energy Losses
PV generator droop curve. The algorithm was able to converge NO PV 6.1 MWh
to a minimum total system losses of 490.7 kWh within 100 Distr. PV 5.03 MWh
Distr. PV w/ VVC 5.68 MWh
iterations as seen in Fig. 7. By finding the optimal droop curve Distr. PV w/ PSO 4.99 MWh
for each PV system, the algorithm was able to improve system

Authorized licensed use limited to: Motilal Nehru National Institute of Technology. Downloaded on July 12,2022 at 11:53:55 UTC from IEEE Xplore. Restrictions apply.
 Fig. 8. System based on a northeastern U.S. distribution feeder, known as J1

1.06 [2] R. A. Shayani and M. A. G. De Oliveira, “Photovoltaic generation

PV
PV w/ PSO penetration limits in radial distribution systems,” IEEE Transactions on
1.05
Power Systems, vol. 26, no. 3, pp. 1625–1631, 2011.
1.04 [3] A. Standard, “C84. 1-2016,” American National Standard for Electric
Voltage (p.u.)

Power Systems and Equipment-Voltage Ratings (60Hz).

1.03
[4] M. E. Elkhatib, R. El-Shatshat, and M. M. A. Salama, “Novel coordi-
1.02
nated voltage control for smart distribution networks with DG,” IEEE
Transactions on Smart Grid, vol. 2, no. 4, pp. 598–605, 2011.
1.01 [5] M. Farivar, R. Neal, C. Clarke, and S. Low, “Optimal inverter VAR
1
control in distribution systems with high PV penetration,” in IEEE Power
50 55 60 65 70 and Energy Society General Meeting, 2012.
Hour
[6] P. Jahangiri and D. C. Aliprantis, “Distributed Volt/VAr control by PV
inverters,” IEEE Transactions on Power Systems, vol. 28, no. 3, pp.
Fig. 9. Terminal voltage of a large PV generator with and without the
3429–3439, 2013.
proposed algorithm
[7] H. G. Yeh, D. F. Gayme, and S. H. Low, “Adaptive VAR control for
distribution circuits with photovoltaic generators,” IEEE Transactions on
Power Systems, vol. 27, no. 3, pp. 1656–1663, 2012.
for high penetration of PVs, so a distributed-individualized [8] D. L. Bassett, “Update of the status of IEEE 1547.8, expanding on IEEE
Standard 1547,” in Proceedings of the IEEE Power Engineering Society
scheme may be more appropriate, but that is beyond the scope Transmission and Distribution Conference, 2012.
of this paper. This significant decrease in total system losses [9] J. W. Smith, W. Sunderman, R. Dugan, and B. Seal, “Smart inverter
will extend the life of equipment, enhance reliability, and can volt/var control functions for high penetration of PV on distribution
systems,” in 2011 IEEE/PES Power Systems Conference and Exposition,
lead to financial gain for utilities and customers. PSCE 2011, 2011.
[10] S. Civanlar and J. Grainger, “Volt/Var Control on Distribution Systems
V. C ONCLUSIONS with Lateral Branches Using Shunt Capacitors and Voltage Regulators
Part II: The Solution Method,” IEEE Transactions on Power Apparatus
The purpose of this work was to investigate VVC in an effort and Systems, vol. PAS-104, no. 11, pp. 3284–3290, nov 1985.
to reduce overvoltages that stem from power injection from PV [11] T. Niknam, a.M. Ranjbar, and a.R. Shirani, “Impact of distributed
generation on volt/Var control in distribution networks,” 2003 IEEE
generation. A simple VVC droop curve was deployed on an Bologna Power Tech Conference Proceedings,, vol. 3, no. 1, 2003.
IEEE 13-bus test system with a large centralized PV plant, [12] K. Baker, A. Bernstein, E. Dall’Anese, and C. Zhao, “Network-
and a distributed PV system with 5 PV generators. Although Cognizant Voltage Droop Control for Distribution Grids,” pp. 1–8,
2017. [Online]. Available: http://arxiv.org/abs/1702.02969
the VVC curve was successful in reducing overvoltages, the [13] K. Dehghanpour, H. Nehrir, J. Sheppard, and N. Kelly, “Agent-based
absorption of reactive power led to higher system losses. modeling of retail electrical energy markets with demand response,” To
Because of this, a PSO algorithm was developed to find Appear in IEEE Transactions on Smart Grid.
[14] “EPRI: Smart Grid Resource Center: OpenDSS,” accessed: 2017-10-30.
the optimal droop curves for distributed PV generators to [Online]. Available: http://smartgrid.epri.com/SimulationTool.aspx
minimize system losses. The methodology was then extended [15] G. Venter and J. Sobieszczanski-Sobieski, “Particle swarm optimiza-
to a more practical larger system where it led to reduced active tion,” AIAA Journal, vol. 41, no. 8, pp. 1583–1589, 2003.
[16] J. Kennedy, “Particle swarm optimization,” in Encyclopedia of machine
power losses compared to the standard-set VVC droop curves. learning. Springer, 2011, pp. 760–766.
[17] Sandia, “Pv performance modeling collaborative,” accessed: 2017-
R EFERENCES 03-30. [Online]. Available: https://pvpmc.sandia.gov/applications/
gridpv-toolbox/
[1] N. Ameli, M. Pisu, and D. M. Kammen, “Can the U.S. keep the pace? [18] EPRI, “Distributed pv monitoring and feeder analysis,” accessed:
a natural experiment in accelerating the growth of solar electricity,” 2017-03-30. [Online]. Available: http://dpv.epri.com/feeder{\ }j.html
Applied Energy, vol. 191, pp. 163–169, 2017.

Authorized licensed use limited to: Motilal Nehru National Institute of Technology. Downloaded on July 12,2022 at 11:53:55 UTC from IEEE Xplore. Restrictions apply.