Electric Load Forecasting with Genetic Neural Networks

This paper presents an evolution algorithm for optimizing a neural network architecture. The procedure establishes the structure and the training algorithm, as well as searching the minimal topology of the network, eliminating neurons and interconnection weights. The model network, this is, feedforward or feedback, can be selected by the user. This methodology is applied to the real problem of the forecasting in power system load in the city of Malaga (Spain) between the years 1992 and 1993. The results produced by the evolution algorithm are tested with a statistical regression analysis and with other training algorithms of paradigms of neural networks.