(lET Journal of Energy Technology JET Volume 6 (2013), p.p. 11 - 30 Issue 4, November 2013 http://www.fe.um.si/sl/jet OPTIMIZATION METHOD FOR CONTROL OF VOLTAGE LEVEL AND ACTIVE POWER LOSSES BASED ON OPTIMAL DISTRIBUTED GENERATION PLACEMENT USING ARTIFICIAL NEURAL NETWORKS AND GENETIC ALGORITHMS OPTIMIZACIJSKA METODA ZA NADZOR NAPETOSTNIH NIVOJEV IN IZGUB Z UPOŠTEVANJEM OPTIMALNE IMPLEMENTACIJE RAZPRŠENE PROIZVODNJE S POMOČJO NEVRONSKIH MREŽ IN GENETSKIH ALGORITMOV Marko Vukobratovicw, Predrag Marie, Zeljko Hederic Keywords: distributed generation, Artificial Neural Network, Genetic Algorithm, voltage control Abstract This paper presents a method for reducing active power system losses and voltage level regulation by implementing adequate distributed generation capacity on the appropriate terminal in a distribution system. Active power losses are determined using an Artificial Neural Network (ANN) using simultaneous formulation for the determination process based on voltage level control and injected power. Adequate installed power of distributed generation and the appropriate terminal for distributed generation utilization are selected by means of a genetic m Corresponding author: Marko Vukobratovic, Tel.: +385 31 224 656, Fax: +385 31 224 605, Mailing address: vukobratovic@etfos.hr JET 11 Marko Vukobratovič, Predrag Maric, Željko Hederič JET Vol. 6 (2013) Issue 4 algorithm (GA), performed in a distinct manner that fits the type of decision-making assignment. The training data for Artificial Neural Network (ANN) is obtained by means of load flow simulation performed in DIgSILENT PowerFactory software on a part of the Croatian distribution network. The active power losses and voltage conditions are simulated for various operation scenarios in which the back propagation artificial neural network model has been tested to predict the power losses and voltage levels for each system terminal, and GA is used to determine the optimal terminal for distributed generation placement. Povzetek V članku je predstavljena metoda za zmanjšanje izgub v sistemu in regulacijo napetostnih nivojev z implementacijo razpršenih proizvodnih kapacitet na primernih terminalih distribucijskega sistema. Izgube delovne moči so določene z uporabo Umetne Nevronske Mreže (UNM), kjer je uporabljena sočasna formulacija v procesu odločanja na osnovi nadzora napetostnih nivojev in injiciranih moči. Ustrezne inštalirane moči razpršene proizvodnje in primerni terminali za izkoriščanje razpršene proizvodnje so izbrani na osnovi Genetskih Algoritmov (GA) izvedenih na poseben način, ki ustreza nalogam v procesu odločanja. Podatki za Umetno Nevronsko Mrežo so pridobljeni na osnovi simulacije pretoka energij v programskem paketu ''DIgSILENT PowerFactory'' na delu Hrvaškega distibucijskega omrežja. Simulacije izgub delovne moči in napetostnih razmer so izvedene za različne obratovalne scenarije, v katerih je testiran model ''vzratnega učenja'' umetne nevronske mreže za predvidevanje izgub moči in napetostnih nivojev za vsak sistemski terminal. Genetski algoritem je uporabljen za določitev optimalnega terminala za umestitev razpršene proizvodnje. 1 INTRODUCTION The presence of distributed generation (DG) changes the load characteristics of the distribution network, which gradually becomes an active load network and implies changes in the power flow. Current-voltage conditions are now not only dependent on the current consumption but also on the production from DG. If sized and selected properly, DG can improve electrical conditions, such as improvement of voltage, loss reduction, relieved transmission and distribution congestion, improved utility system reliability and power quality in the distribution network, [1]. In order to determine the impact on the power system of each DG, it is necessary to perform the power flow analysis on a daily or hourly (or even 10-minute) basis. Due to the increased number of small DG, mostly from intermittent sources, it is necessary to implement an advanced management power distribution system to make the distribution network significantly automated. Accordingly, it is necessary to develop mathematical optimization models that can be implemented in the distribution network management system to enable optimal management. According to [2], an automated distribution network has to provide a fast and the accurate solution for power flow and current-voltage conditions control. As an ideal solution, artificial neural networks (ANN) are imposed due to their ability to solve nonlinear problems in a short period of time, and if quality organized and made, they are able to perform real-time calculations necessary for the optimization of the distribution network. ANN have considerable potential in control systems because they can learn and adapt, they can approximate nonlinear functions, they are suited for parallel and distributed processing and model multivariable systems naturally, [3]. Since they are based on human experience and on 12 JET Optimization method for control of voltage level and active power losses based on optimal distributed generation placement using Artificial Neural Networks and Genetic Algorithms logical links between inputs and outputs, they can adopt various learning mechanisms and self-organization or training concepts, pattern recognition, forecasting etc. ANN can be trained to generate control parameters for minimizing power losses and determining the optimal solution for DG implementation in the distribution network. This paper proposes an online real-time power flow optimization and voltage regulation method using ANN and a Genetic Algorithm (GA). ANN are highly robust and provide satisfactory solutions if provided with quality data and can dynamically determine the most appropriate DG solution by means of installed power and position in the system. The GA is used for solving constrained and constrained optimization problems and is based on a natural selection process that mimics biological evolution. The algorithm generates a population of individual solutions that are randomly selected from the population and used as parents for the next generation. Over several generations, the optimal population solution appears. 2 THE PROBLEM FORMULATION Optimization problem can be generally shown with a model of the objective function and associated restrictions: Minf(x,u) So that g(x,u) = 0 (2.1) h(x,u)< 0 Where vector u is a vector of control variables, x is a vector of state variables; scalar f(x) is the objective function, while restrictions are given by the system of equation g(x, u) and inequalities h(x,u). The main goal of the proposed method is to determine the best locations in the distributed system for distributed generation by minimizing different functions related to project goals which are: 1. Reduction of active power losses 2. Voltage profile improvement 2.1 Objective function The main objective function could be described as: F = MinPlosses (Z2) Where Plosses are losses of active power in a system. Minimization of active power losses is an essential requirement in a distribution system for efficient power system operation, [3]. Power losses can be calculated as: JET 13 Marko Vukobratovič, Predrag Maric, Željko Hederič JET Vol. 6 (2013) Issue 4 NB NB I \ / \ Posses=E E A, p Pj+Q, Qj)+B, Q Pj+P, Qj) («) 1=1 j=1 Where: P,Q : real power and reactive power injection at respected terminal Nb : terminal number And Ajj Bjj are represented respectively: ^ _Rijs\n(si-SJ) ^ A_Rijcos (Si-Sj) (24) '' ViVj '' ViVj Rjj : line resistance between terminal; and terminal; V ,8 : voltage and load angle at the selected terminal 2.2 Constraints The objective function of active power loss minimization is not sufficiently suitable without technical restrictions and correct formulation of optimization constraints. Optimal placement of distributed generation and the solution provided with the proposed method must be realistic and should not produce negative impacts on other system aspects. In order to achieve this goal, operational constraints should be properly evaluated and chosen, not only to enable proper operation of the proposed algorithm, but also to support the regular operation of the power system. 2.2.1 Power constraints For the safe operation of the power system, the active power constraints are given by the expression: Pa - Pa - Vt ÎVj • ^sm(^)) (2.5) 7=1 Where: ien : number of nodes in network Pa : active power production in node i Pti : active power consumption in node i 9¡j : angle of mutual admittance ^ of nodes i and j Gij : mutual conductance of nodes i and j 14 JET Optimization method for control of voltage level and active power losses based on optimal distributed generation placement using Artificial Neural Networks and Genetic Algorithms Bij : mutual susceptance of nodes i and j Gn : self-conductance of node i Bu : self-susceptance of node i Reactive power restrictions are given by the expression: Qgi-Qti~Vt (2.6) 7=1 Where: ien : number of nodes in network : reactive power production in node i Qa : reactive power consumption in node i Besides active and reactive power constraints, the apparent power that is transmitted through each branch has to be below the physical limit of the branch transformer in steady-state operation. The constraint of apparent power is given by: Si < Si.max (2.7) Where: St : apparent power in ith branch S^m^ : maximum allowed apparent power in ith branch 2.2.2 Voltage levels constraints Voltage level restrictions are given by the expression: F/-mln< V/< Vi-max (2.8) Where: i^n : number of nodes in network K/-min, Vi-max : voltage limitations Vi : voltage level in node i 2.2.3 Constraints of reactive power production in generator node The generator has the capability curve and the technical operational limits, so the reactive power production is given by the expression: JET 15 Marko Vukobratovič, Predrag Maric, Željko Hederič JET Vol. 6 (2013) Issue 4 Qa-min^ Qa< Qt ie\N pv.N o} (2.9) Where: -max : reactive power production limits in node i N pv : number of PV node No : node of DG The objective function including the reduction of active power losses only could provide the solution without predicting a sufficient amount of reactive power reserves in case of the failure of one or more components in a power system. The appropriate optimization solution has to provide the optimization of voltage levels, voltage reduction, loss of stability risk and the reduction of power losses. Bearing in mind all restrictions and the objective of the optimization, a useful algorithm has to be developed. Because of the complexity and nonlinear interdependence of controlled variables, it is difficult to provide a fast and correct solution using classic (exact) optimization techniques, such as linear programming, the interior point method or mixed integer programming, [5]. ANN can be appropriate for solving such non-linear problems. There are several different types of ANN, including feed-forward neural network, radial basis function (RBF) network, Kohonen self-organizing network, recurrent neural network (RNN), bi-directional RNN, stochastic neural networks, etc. The appropriate neural network has to be properly selected since not every type of neural network will give the best solution for a certain problem. Back-propagation (BP) ANN can be used for the optimization problems since it meets the specific criteria: a flow chart of the problem can be described; there is a relatively easy way to generate a significant or at least necessary number of input and output examples; the problem appears to have considerable complexity but there is a clear solution; outputs may be unambiguous in some extreme cases. The typical back-propagation network has an input layer, an output layer, and at least one hidden layer. The numbers of hidden layers are theoretically infinite but usually one to four layers is adequate to solve any kind of complex problems. Each layer has to be fully connected to the vicinal layer by every neuron, as shown in Figure 1. The relationship between input and output values of multi-layer ANN can be represented as [6]: 3 ARTIFICIAL NEURAL NETWORK DESIGN AND IMPLEMENTATION (3.1) Where: Y : output value 16 JET Optimization method for control of voltage level and active power losses based on optimal distributed generation placement using Artificial Neural Networks and Genetic Algorithms Xj : input value Wj : weighting factor k : threshold value N : layer number f : nonlinear function When the network is created, the process of teaching has to be done in order to organize the neurons. This teaching makes usage of a learning rule, which is the variant of the Delta Rule, [3] The teaching starts with determining the error, which is the difference between the actual outputs and the desired outputs given in the training data. Based on this error, the weighting factor is changed in proportion to the error for the global accuracy. The algorithm for the weighting factor changing based on training data is, [6]: A p Wn = dpPj-o pi)ipi = nSpj iPi (3.2) Where: n : learning rate tPj : j component of pth target output oPj : j component of pth computed output ipi : i component of pth input pattern 8 ■ : error of target and computed output If well trained, an ANN can provide reasonable outputs for a new set of inputs enabling network training on a representative set of inputs with output correction. The training should be done on the largest possible set. Generally, the precision of ANN is increased by the larger training set with more input variables. JET 17 Marko Vukobratovič, Predrag Maric, Željko Hederič JET Vol. 6 (2013) Issue 4 Figure 1: Structure of Artificial Neural Network 3.1 Neural network training For the purpose of ANN training, a training data set has to be generated. Selecting the amount and type of training data is extremely important since the wrong selection could reduce the learning ability of the ANN or even provide an incorrect solution. For better accuracy, all dependent parameters have to be taken into account. The training data for the ANN consists of: DG active power production changed by operation scenarios from 0 kW (no production) to 1.350kW (excessive production) in 10kW increments, injected current from DG production given in kA, and the voltage level on the low-voltage side and the voltage level on the mediumvoltage side, given in per-unit (p.u.) values. Targeted data for the ANN training are total feeder losses for each operation scenario. Accordingly, ANN has four input units and one output unit connected with nine hidden layer units. The training is performed by the Levenberg-Marquardt algorithm for nonlinear least square problems, [7]. Calculations of each operation scenario for the training data generation are performed using DIgSILENT PowerFactory software, a leading power system analysis tool for applications in generation, transmission, distribution and industrial systems, [8]. The ANN is first trained on sample values for one terminal, and later it is tested on all proposed terminals. 18 JET Optimization method for control of voltage level and active power losses based on optimal distributed generation placement using Artificial Neural Networks and Genetic Algorithms The results of each operation scenario are introduced into tables. The power losses in the electrical network can be computed by means of load flow simulation generated in the DIgSILENT PowerFactory software. Quantification and determination of power losses is essential due to the impact on the power system economic operation and the lifetime of the included equipment, [9]. Performance of ANN training is shown in Figure 2. Figure 2: Performance of ANN training For the purpose of electrical network modelling, data is obtained from the Croatian grid company HEP-ODS Elektroslavonija for a part of distribution network with a nominal voltage 35(20)kV and 0.4 kV with 48 terminals, 23 transformers and 25 different low-voltage loads. The distribution network is connected to the transmission network on two sides, but it is never doubly-fed due to operator technical conditions. If fully loaded, the voltage drops under 0.89 p.u. JET 19 Marko Vukobratovic, Predrag Maric, Željko Hederič JET Vol. 6 (2013) Issue 4 Of course, the normal operating conditions for this distribution network are not fully loaded terminals, and it is never doubly-fed, but it is necessary to observe what happens to voltage values. One possible solution for the increase of voltage values is planning for an adequate distributed generation on the convenient terminal in the system. In this case, the continuous electric power production would be as adequate a type as the stable source the network operator could rely on. 4 LOSSES ESTIMATION BY ANN The ANN is modelled in MATLAB, which is a high-level language and interactive environment for numerical computation, visualization, and programming. After the ANN training, the fitting function and associated graph that shows how the results given by the ANN correspond to the control variables and results provided by DIgSILENT PowerFactory could be realized, as shown in Figure 4. The results provided by DIgSILENT PowerFactory power flow calculation are taken as correct real-life values since this software has previously and frequently proven its reliability and precision. 20 JET Optimization method for control of voltage level and active power losses based on optimal distributed generation placement using Artificial Neural Networks and Genetic Algorithms 0.44 0.42 0.4