The comparison analysis of neural architectures and learning algorithms for electrical energy losses forecasting

This article is devoted to the comparison analysis of several neural architectures in order to expose the most efficient one towards the task of 0.4-20 kV power distribution networks electrical losses calculation and forecasting. The following artificial neural network architectures were chosen as being the most suitable for the purposes of regression: Multi-Layer Perceptron, Linear Network, Radial Basis Function Network, General Regression Neural Network, Fahlman Cascade Correlation Network.

For training set construction, a data acquired from the parameters of electrical network working regimes modeling, was used. Three populations were corresponding to three different electrical schemes and different levels of regime information availability. In order to increase the quality of neural network training, a «cross-checking» method was used, which means the division of the source training set into several subsets. The least Square Mean Error over the testing subset of the training set was employed as a network-s efficiency criterium.

As a result of the research, it was found that the most efficient neural architectures are Multi-layer Perceptron and Fahlman Cascade Correlation Network. It was also shown, that Perceptron training method which allows getting the least error value is a conjunction of Backpropagation and Adjacent Gradients methods.