STUDY OF ACTIVATION FUNCTIONS IN NEURAL NETWORKS FOR MINIMIZING THE LOSSES IN THE POWER FLOW SYSTEM

Rafael Rangel Szillat¹; Alessandra Freitas Picanco¹

¹ Instituto Federal de Educação, Ciência e Tecnologia da Bahia, Salvador, Brazil.

doi:10.20906/CPS/CILAMCE2017-0076

Resumo

Power Flow Systems (PFS) is a numeric instrument to determine permanent conditions of the electric system. It consists of nonlinear solution methods, that result in interactive algorithms methods to find the solution of this problem. To find the optimal resolution of Load Flow (LF) problems, that would reduce the electrical lost, and also facilitate the management, planning, and operation of the energy system (ES). In this context, several authors have to study techniques to numerical optimization of LF from Newton-Raphson (NR). Because of the number of variables, Artificial Neural Network (ANN) comes as an attractive alternative to solve and optimize FL problems. The present article shows an optimization ANN algorithm of the first order, using Levenberg-Marquardt (LM) and a Novel Levenberg-Marquardt (NLM), to reduce the loss of PFS. The proposed NLM method consists in the application, the actualization of the intern parameters of the function, in the same form of the neural network weights. The codes are tested in IEEE-6, IEEE-30, IEEE-118 e IEEE-300 bus, using 6 different activation functions: Hyperbolic Tangent, Bi-Hyperbolic Tangent, Elliot, Gauss, Sigmoid, and Ackley. The results, that come from this codes, are the actual values of tension and phase of each bus, that results in the loss after the optimization. The values, that come of LM and NLM shows significant reduction loss when compared with NR. For the LM application was a reduction of 7,89 MW (Bi-Hyperbolic Tangent with 0.178ms training time), 4.94 MW (Bi-Hyperbolic Tangent with 0.218ms training time), 3.53 MW (Elliot with 84.3ms training time) and 9.75 MW (Gauss with 309,4ms training time) applied in the IEEE-6, IEEE-30, IEEE-118 and IEEE-300 bus system, respectively. Using NLM, the loss reduction was 9,45 MW in IEEE-6 bus (Hyperbolic Tangent with 86,141ms training time), 6,12 MW in IEEE-30 bus (Bi-Hyperbolic Tangent with 247,206ms training time), 2,24 MW for IEEE-118 bus (Sigmoid with 0,296 µs training time) and 305,38 MW in IEEE-300 bus (Ack

Palavras-chave: Power Flow; Levenberg-Marquardt; Losses; Artificial Neural Network