Performance assessment of an optimization strategy proposed for power systems

TELKOMNIKA Telecommunication, Computing, Electronics and Control
Vol. 18, No. 5, October 2020, pp. 2729~2736
ISSN: 1693-6930, accredited First Grade by Kemenristekdikti, Decree No: 21/E/KPT/2018
DOI: 10.12928/TELKOMNIKA.v18i5.14396  2729
Journal homepage: http://journal.uad.ac.id/index.php/TELKOMNIKA
Performance assessment of an optimization strategy proposed
for power systems
Harold Puin, Cesar Hernandez
Technological Faculty, Universidad Distrital Francisco Jose de Caldas, Colombia
Article Info ABSTRACT
Article history:
Received Oct 24, 2019
Revised May 15, 2020
Accepted Jun 17, 2020
In the present article, the selection process of the topology of an artificial
neural network (ANN) as well as its configuration are exposed. The ANN was
adapted to work with the Newton Raphson (NR) method for the calculation of
power flow and voltage optimization in the PQ nodes of a 10-node power
system represented by the IEEE 1250 standard system. The purpose is to assess
and compare its results with the ones obtained by implementing ant colony and
genetic algorithms in the optimization of the same system. As a result, it is
stated that the voltages in all system nodes surpass 0,99 p.u., thus representing
a 20% increase in the optimal scenario, where the algorithm took 30 seconds,
of which 9 seconds were used in the training and validation processes of
the ANN.
Keywords:
Artificial neural network
Newton Raphson
Optimization
Power system
This is an open access article under the CC BY-SA license.
Corresponding Author:
Cesar Hernandez,
Technological Faculty,
Universidad Distrital Francisco Jose de Caldas,
Calle 68D Bis A Sur # 49F – 70, Bogotá D.C., 111941, Colombia.
Email: cahernandezs@udistrital.edu.co
1. INTRODUCTION
The integration of power generation, transmission and distribution systems with the progress
evidenced in the information and communication technologies (ICT) sector has been encompassed within
the development of the concept known as smart grids. It has allowed to unite in a single management system
the areas of protection coordination, control, measure and economic dispatch of an electric network. Its main
purpose is to achieve integration through the efficient and rational use of energy as well as the increase of
reliability, security and flexibility of electric systems [1-3].
The inherent result of the development of smart grids has been the constant search for
the implementation of algorithms that optimize power systems based on technical and economic criteria as well
as energy quality. Improvement practices are also proposed such as control systems to correct voltage
unbalances in power generation units embedded within the network [4-7], multi-agent systems [8-10] to
compensate the high variability of the energy supply related to alternative sources of generation, and
the forecast of the effects of a massive connection of electric vehicles in the future [11-13]. Other proposals
have focused in the effects of environmental conditions in the projection of the economic dispatch based on
probability functions delivered by the statistical method Latin Hypercube Sampling [14]. Other strategies have
focused on developing algorithms that create usage profiles of the generators linked to a network, according to
the consumption demand [15, 16].
The calculation of the voltage values involves solving a system of simultaneous non-linear equations
in most cases. To achieve this, the Newton-Raphson (NR) and Gauss-Siedel iterative methods [4]. The NR
method consists of an iterative procedure in which non-linear equations are used involving voltage magnitude

 ISSN: 1693-6930
TELKOMNIKA Telecommun Comput El Control, Vol. 18, No. 5, October 2020: 2729 - 2736
2730
and angle variables as well as active and reactive power. Since there are only two power equations, the other
two variables must be calculated [8]. In [9], the NR method is used to calculate photovoltaic parameters and
carry out the modelling process. The Gauss-Seidel method is an iterative method in which the number of
unknown variables is equal to the number of equations to solve. It consists on designing a converging
succession according to a previously defined criteria [11], the convergence values are the solution of the nodal
voltages and powers of the electric system. In [12], the Gauss-Seidel method is used as the main strategy to
accelerate the solution of power flow in a high-performance reconfigurable computer.
It is evident that it is the process to be optimized within the smart grid that determines the type of
algorithm to be implemented, whether it is a linear system or not. Currently, the algorithms of the iterative [17]
(Newton method [18], conjugated gradient, interpolation, etc) and heuristic types [19] (evolutionary
algorithms, genetic algorithms, Nelder Mead, among others) are the most commonly implemented.
Furthermore, there are computational models that can be used in optimization models such as the Artificial
Neural Networks (ANN), a bio-inspired algorithm from 1943 that was relegated to the background due to
the computational capacity that it required at the time. However, with the unstoppable development in
electronics and semi-conductor materials and the manufacture of increasingly powerful processors,
the application of ANN has risen. They can be classified into iterative or heuristic methods depending on their
learning process [20, 21].
Since their reintroduction as a computational model, ANNs have been used to simulate different types
of processes [22-25], due to their swift prediction of variables. Nonetheless, their use as an optimization model
is not yet extended and needs some sort of iterative or heuristic algorithm in order to work. Hence, this article
presents the development process of an artificial neural network combined with the Newton Raphson (NR)
method for the calculation of power flow and the optimization of nodal voltages in power systems with n nodes.
This is subsequently implemented in a 10-node IEEE 1250 standardized power system with the purpose of
guaranteeing voltages over 0.98 p.u. in all system nodes. Finally, the performance will be assessed by
comparing the results obtained with the bee swarm and ant colony algorithms for the same power system.
This helps to determine its potential for implementation in smart grids as an optimization method based on
the electric energy criteria.
2. PROPOSED ALGORITHM
The proposed algorithm is developed in its totality in the MATLAB 2018b numerical computing
software. Initially, the execution of the NR method is carried out in order to obtain the magnitude and angle of
the voltage, based on the impedance matrix and the active and reactive power data. Using the power flow
calculations, the algorithm can proceed to assess the voltage in each PQ node through ANN to determine
whether the nodes are underpowered in comparison to the optimal value used for training. Thus, a capacitive
reactance of 0.1 p.u. is injected if the assessed node does not comply with this optimization parameter. When
the ANN completes the assessment of all power system nodes, the algorithm executes the NR method with
the purpose of determining the new voltages in the system, which are once again assessed by the ANN.
The process is concluded when it is established that all the PQ nodal voltages are equal or above the optimal
reference value.
In this manner, the NR and ANN methods enable the optimization of the power system, by giving
the user a final report in .xlsx format with the optimized values of voltage magnitude and angle for each node
as well as the calculation of generated active-reactive power and the demand. The user is informed on the value
of the capacitive correction required by each node to elevate the voltage to optimal values. Figure 1 shows
the flow diagram of the developed algorithm.
The neural network proposed in the algorith was developed by implementing the base codes of
the MATLAB fitnet function, establishing a structure of three layers: input layer, hidden layer and output layer.
This is illustrated in Figure 2 and the components are explained in this section.
2.1. Input variables
It is the information given to the neural network, for the traning phase as well as the validation and
testing phases. In this specific case, the nodal voltage in p.u. may vary from 0 to 1. The user is informed on
the value of the capacitive correction required by each node to elevate the voltage to optimal values. Figure 1
shows the flow diagram of the developed algorithm.
2.2. Layers
The model has three types of layers: input layer, hidden layer and output layer. The input layer has
one neuron, which individually receives the information of the voltage in each node. The hidden layer has 30
neurons, which is a number determined by considering that the number of neurons present in this layer is

TELKOMNIKA Telecommun Comput El Control 
Performance assessment of an optimization strategy proposed for power systems (Harold Puin)
2731
proportional to the accuracy of the neural network regarding the classification of data but it is also proportional
to the time required to perform such task. In terms of the connections, each neuron in this layer is connected
with a neuron of the input layer depending oh the weights of the connections. It is important to clarify that
there may be more than one hidden layer. The number of hidden layers of a neural network is directly related
with how easy it is to classify the desired output according to the input. Figure 3 shows the output vs input
diagram with typical values of the optimization process of a power system. According to the diagram, for input
variables with values below 0.9, the neural network can generate an output of 0. For input variables over 0.9,
the neural network must generate an output of 0.1. This allows the perfect division of the data through a linear
curve. In case the output states cannot be linearly separated, the number of hidden layers must keep increasing.
The output layer has one neuron, corresponding to the number of output variables with a minimum of 2 states.
2.3. Weights
Weights are coefficients that alter the input value of neurons, starting from the hidden layer,
considering the periodicity in which a specific input value is transmitted compared to a desired output value.
This strengthens the connection with the transmitted neuron at the expense of the connection with other
neurons. Hence, weights also have the functionnality of interconnecting neurons in different layers.
Develop of power flow analysis with NR
RNA training process with 7000 type values
RNA validation porcess with 1500 type
values
RNA test process with 1500 type values
PQ nodes individual RNA evaluation
Power electric system initial data reading
from .xlsx file
The node have
undervoltage?
NO
Generation of the .xlsx report file
YES
Individually inyection of reactive
compensation in the all PQ nodes of the
system with undervoltage
NR algoritm ejecution
Figure 1. Flow diagram of the developed algorithm

 ISSN: 1693-6930
2732
a2.1
a2.2
a2.3
a2.30
a3.1a1.1
Input
layer
Hidden
layer
Output
layer
Nodes Nodes Nodes
Weigth Weigth
Inputs
variables
Output
variables
Figure 2. Structure of the neural network implemented in the optimzzation algorithm
Figure 3. Data classification in a single layer neural network
2.4. Output variables
In (1) is the output of one neuron in the input layer. In (2) is the mathematical model for the output of
a neural network and it represents all possible combinations of the connections between the different layers of
a neural network from the hidden layer up to the output layer. For the ANN discussed in this paper, the output
variable has two states: 0 or 0.1.
𝑎𝑖
𝑘
= 𝑥𝑖 (1)
For k = 1, i > 0, where a is the neuron output and x is the input variable.
𝑎𝑖
(𝑘)
= 𝑓(∑ 𝑎𝑗
( 𝑘−1)𝑛𝑘−1
𝑗=1 ∗ 𝑤𝑗𝑖
( 𝑘−1)
) (2)
for k > 1, i > 0 and j > 0, where a is the neuron output, w are the weights of the neural connections and f(x) is
the limited function.
After establishing the topology of the ANN described in this article, Figure 4 presents the typical
results of the training, validation and testing phases, using 7000, 1500 and 1500 typical data respectively.
The system implements the Bayesian regularization function as a training function. In order to assess its
performance, the geometric mean root is used. A total of 58 iterations were carried out to train the network.
The training, validation and testing processes were stopped because the ANN attained the expected minimum
gradient in 8 seconds.

2733
Figure 4. Typical data in the training, validation and testing processes of the implemented ANN
3. RESULTS
The schematic of the system to be optimized is shown in Figure 5. It is comprised of 10 nodes, from
which two nodes have a PV type, one node has a slack type and seven nodes have a PQ type. The other
characteristics and variables of the electric system are presented in Tables 1 and 2, respectively. In Table 1,
the first two columns represent the interconnection of the system nodes, the electric characteristics of these
connections are established from the third to the fifth column. In Table 2, the first two columns represent
the number of the node and its type. The third column up to the seventh column show the values of the generated
active/reactive power, the demanded active/reactive power and the voltage magnitude in each node.
The performance of the neural network during the training, validation and testing phases for the optimization
process of this specific system is presented in Figure 6. The preliminary result of the nodal voltage after
optimization and the total computation time (ANN training time plus optimization time) are shown in Figure 7.
Figure 5. IEEE 1250 power system

 ISSN: 1693-6930
2734
Table 1. Characteristics of the IEEE 1250 power system
N start N finish R X Bsh/2
1 2 0.08379 0.1772 0.00047
1 6 0.1089 0.2304 0.00061
2 3 0.1843 0.3900 0.0010
2 5 0.1508 0.3191 000085
2 6 0.1675 0.3545 0.00094
3 4 0.2346 0.4963 0.0013
3 5 0.2094 0.4432 0.0011
5 4 0.1843 0.3900 0.0010
4 10 0.1675 0.3545 0.00094
6 7 0.1256 0.2659 0.00071
7 8 0.1005 0.2127 0.00056
7 9 0.0670 0.1418 0.00037
9 8 0.1340 0.2836 0.00075
8 10 0.1089 0.2304 0.00061
Table 2. Known system variables of the IEEE 1250 power system
Node Type P gen Q gen P dem Q dem V mag
1 Slack 0.7 0 0 0 1
2 PQ 0 0 0.15 0.2 0.9433
3 PQ 0 0 0.15 0.2 0.8939
4 PQ 0 0 0.15 0.2 0.8666
5 PV 0 0 0.15 0.2 1
6 PQ 0 0 0.15 0.2 0.9353
7 PQ 0 0 0.15 0.2 0.9335
8 PQ 0 0 0.15 0.2 0.8943
9 PV 0.5 0 0 0 1
10 PQ 0 0 0.15 0.2 0.8373
Figure 6. Training, validaiton and testing of the ANN used in
the optimization of the IEEE 1250 power system
Figure 1. Preliminary results of the nodal voltages and total execution time of the optimization algorithm

2735
4. COMPARATIVE ASSESSMENT
Figure 7 shows the report generated by the algorithm in which the optimization results can be
thoroughly detailed for the 10-node IEEE 1250 standard power system. In Table 3, the node type is coded in
the second column, with 1 for Slack, 2 for PV and 3 for PQ. Table 4 presents the results of the nodal voltage
magnitudes obtained with the genetic algorithm (GA) and the ant colony optimization (ACO). Table 5 shows
the optimization percentage of the voltage magnitude for each node of the IEEE 1250 power system. Each
column represents one of the algorithms discussed in this section.
Table 3. Report of the optimization results for IEEE 1250
Bus i Type node P gen Q gen P dem Q dem Vmag Vang
1 1 0.779 -0.28 0 0 1 0
2 3 0 0.2 0.15 0.2 0.992 -5.96
3 3 0 0.3 0.15 0.2 1.002 -1126
4 3 0 0.3 0.15 0.2 1.008 -12.58
5 2 0 0 0.15 0.2 1 -10.99
6 3 0 0.3 0.15 0.2 1 -4.369
7 3 0 0.3 0.15 0.2 1.002 -5.194
8 3 0 0.3 0.15 0.2 1.004 -7.374
9 2 0,5 0 0 0 1 -2.583
10 3 0 0.3 0.15 0.2 1009 -10.996
Table 4. Results in nodal voltage magnitudes
for ga and aco
Node GA ACO
1 1 1
2 0.9984 0.9566
3 0.9963 0.9109
4 0.9972 0.9038
5 1 1
6 0.9957 0.9722
7 0.9957 0.9735
8 0.9947 0.9513
9 1 1
10 0.9955 0.9075
Table 5. Comparison of the percentages of nodal
voltage increase
Node GA ACO ANN
1 0% 0% 0%
2 5.84% 1.4% 5.16%
3 11.45% 1.9% 12.09%
4 15.07% 4.29% 16.31%
5 0% 0% 0%
6 6.45% 3.94% 6.91%
7 6.55% 4.28% 7.33%
8 11.22% 6.37% 12.26%
9 0% 0% 0%
10 18.89% 8.38% 20.5%
5. ANALYSIS OF THE RESULTS
Figures 4 and 6 exhibit the training, validation and testing process of the neural network. It is
concluded that the time lapses and iterations of convergence are similar under different performance scenarios.
This indicates that the size of the power system to be optimized does not influence the time and quality of
the ANN in detecting subvoltages in the system nodes. Going back to Figures 4 and 6, it can be determined
that the typical parameters for which the training process is stopped are tied to the minimum learning gradient
of the neural network and its performance. This means that the trained neural network guarantees a a high
accuracy in the prediction of values depending on the input variable. This is highlighted by the fact that
the output has two states only distanced by a 0.01. In case that the neural network is halted after completing
the maximum number of iterations, it is recommended to apply once again the learning process since
the network may not successfully predict the expected output values.
Table 3 showcases the report of the algorithm after the optimization process of the IEEE 1250 power
system. This table also shows that, even if the total non-capacitive reactive powers per node do not exceed
the active powers, they are significant enough to affect the power factor associated to the node. This is true not
only for the ANN algoritihm but also for the other algorithms. It is recommended to consider this parameters
for future versions of the algorithms. Furthermore, Table 3 reveals that the reactive power in the PV nodes or
the nodes connected to a generator is increased in order to compensate the capacitive reactive power injected
to the system by the ANN. This results in a normal behavior considering that these nodes must remain with
a voltage equal to 1 p.u. The results of the optimization of the IEEE 1250 power system exhibited by the genetic
algorithm and the ANN are fairly competitive. The minimum difference is 0.46% in the sixth node and
the maximum difference is 1.26% in the last node. It is suggested to assess the optimization time to find
an arguable difference.
6. CONCLUSIONS
The topology of the implemented ANN in the optimization algorithm has shown a stable behaviour
in terms of the parameters used for training, validation and testing, guaranteeing accuracy and swiftness in

 ISSN: 1693-6930
2736
most cases. A neural network that adopts a number of inputs equal to the number nodes in the power system
to be optimized is not viable since it would require a large amount of data to assure a correct training process.
The corresponding computing power would need to surpass the capacity of an average computer. It is
recommended to include the power factor as an optimization parameter keeping in mind that optimization must
be based on the principle of energy quality. The optimization results obtained with both the genetic algorithm
and the implemented ANN are similar.
REFERENCES
[1] Berrío L, Zuluaga C., “Concepts, standards and communication technologies in smart grid,” 2012 IEEE 4th
Colombian Workshop on Circuits and Systems, CWCAS 2012 - Conference Proceedings, 2012.
[2] Heron J. W., Jiang J., Sun H., Gezerlis V., Doukoglou T., “Demand-Response Round-Trip Latency of IoT SmartGrid
Network Topologies,” IEEE Access, vol. 6, pp. 22930-37, 2018.
[3] Peng F. Z., Li Y. W., Tolbert L. M., “Control and protection of power electronics interfaced distributed generation
systems in a customer-driven microgrid,” Conference: Power & Energy Society General Meeting, 2009.
[4] Katiraei F., Iravani R., Hatziargyriou N., Dimeas A., “Microgrids management,” IEEE Power Energy Mag., vol. 6,
no. 3, pp. 54-65, 2008.
[5] Zhang H., Nie Y., Cheng J., Leung V. C. M., Nallanathan A., “Sensing Time Optimization and Power Control for
Energy Efficient Cognitive Small Cell with Imperfect Hybrid Spectrum Sensing,” IEEE Transactions on Wireless
Communications, vol. 16, no. 2, pp. 730-43, 2017.
[6] Karaboga D., Basturk B., “Artificial Bee Colony (ABC) Optimization Algorithm for Solving Constrained
Optimization Problems,” International Fuzzy Systems Association World Congress, 2007.
[7] Jordehi A. R., Jasni J., “Approaches for FACTS optimization problem in power systems,” 2012 IEEE International
Power Engineering and Optimization Conference Melaka, Malaysia, 2012.
[8] Zhao B., Xue M., Zhang X., Wang C., Zhao J., “An MAS based energy management system for a stand-alone
microgrid at high altitude,” Appl Energy, vol. 143, pp. 251-61, 2015.
[9] Chen C., Duan S., “Microgrid economic operation considering plug-in hybrid electric vehicles integration,” J. Mod.
Power. Syst. Clean. Energy., vol. 3, no. 2, pp. 221-31, 2015.
[10] Rizk Y., Awad M., Tunstel E. W., “Decision Making in Multi-Agent Systems: A Survey,” 6th
Internatioonal
Conference, MESAS 2019 Palermo, Italy, 2019.
[11] Mazidi M., Zakariazadeh A., Jadid S., Siano P., “Integrated scheduling of renewable generation and demand response
programs in a microgrid,” Energy Conversion and Management, vol. 86, pp. 1118-27, 2014.
[12] Elsied M., Oukaour A., Gualous H., Hassan R., “Energy management and optimization in microgrid system based
ongreen energy,” Energy, vol. 84, pp. 139-51, 2015.
[13] Mwasilu F., Justo J. J., Kim E. K., Do T. D., Jung J. W., “Electric vehicles and smart grid interaction: A review on
vehicle to grid and renewable energy sources integration,” Renewable and Sustainable Energy Reviews, vol. 34,
pp. 501-16, 2014.
[14] Corchero C., Nunez-Del-Toro C., Paradell P., Del-Rosario-Calaf G., “Integrating ancillary services from demand
side management and distributed generation: An optimal model,” 2018 International Conference on Smart Energy
Systems and Technologies, SEST 2018 - Proceedings. Institute of Electrical and Electronics Engineers Inc.; 2018.
[15] Yang B., Li J., Han Q., He T., Chen C., Guan X., “Distributed Control for Charging Multiple Electric Vehicles with
Overload Limitation,” IEEE Trans Parallel Distrib Syst., vol. 27, no. 12, pp. 3441-54, 2016.
[16] Xu H., Huang H., Khalid R. S., Yu H., “Distributed machine learning based smart-grid energy management with
occupant cognition,” 2016 IEEE International Conference on Smart Grid Communications, 2016.
[17] Carreno Franco E., Toro Ocampo E., Escobar Zuluaga A., Optimización De Sistemas Lineales Usando Métodos De
Punto Interior,” Sci Tech, vol. 1, no. 24, pp. 43-8, 2004.
[18] Anaut D. O., di Mauro G. F., Meschino G., Suárez J. A., “Optimización de Redes Eléctricas Mediante la Aplicación
de Algoritmos Genéticos,” Inf tecnológica, vol. 20, no. 4, pp. 137-48, 2009.
[19] Moreno L. F., Díaz F. J., Peña G. E., Rivera J. C., “Análisis comparativo entre dos algoritmos heurísticos para resolver
el problema de planeación de tareas con restricción de recursos (RCPSP),” Dyna (Medellin, Colombia), vol. 74,
no. 151, pp. 171-83, 2007.
[20] Aprilia E., Meng K., Al Hosani M., Zeineldin H. H., Dong Z. Y., “Unified Power Flow Algorithm for Standalone
AC/DC Hybrid Microgrids,” IEEE Transactions on Smart Grid, vol. 10, no. 1, pp. 639-49, 2019.
[21] Iliya S., Goodyer E., Gow J., Shell J., Gongora M., “Application of Artificial Neural Network and Support Vector
Regression in Cognitive Radio Networks for RF Power Prediction Using Compact Differential Evolution Algorithm,”
2015 Federated Conference on Computer Science and Information Systems (FedCSIS), 2015.
[22] Álvarez D., Hurtado Gómez J., “Optmización basada en confiabilidad por medio de redes neuronales y algoritmos
evolutivos,” Métodos numéricos para cálculo y diseño en Ing Rev Int., vol. 18, no. 4, pp. 573-94, 2002.
[23] Beale M. H., Hagan M. T., Demuth H. B., “Dynamic Neural Networks. Neural Network Toolbox (TM) User’s
Guide,” Natick, MA: The MathWorks, Inc.; 2018.
[24] Abbas N., Nasser Y., Ahmad K. El., “Recent advances on artificial intellixgence and learning techniques in cognitive
radio networks,” EURASIP Journal on Wireless Communications and Networking, vol. 1, pp. 1-20, 2015.
[25] Veitch D., “Wavelet Neural Networks and their application in the study of dynamical systems,” University of York, 2005.

Performance assessment of an optimization strategy proposed for power systems

More Related Content

What's hot (20)

Similar to Performance assessment of an optimization strategy proposed for power systems (20)

More from TELKOMNIKA JOURNAL (20)

Recently uploaded (20)

Performance assessment of an optimization strategy proposed for power systems