Electricity price forecasting using evolved neural networks

Evolutionary techniques have capabilities of efficient search space exploration with population models corresponding to the problem. Their ability to capture the non linear dependencies among the system variables has invited economic analysts towards their use in the field of financial time series prediction. Although simple neural networks with sufficient number neuron units in the hidden layer are capable of following dynamics of any deterministic system, the weight search space becomes too complex to be searched using a simple back propagation based training algorithm. This paper presents and evaluates two alternative methods for finding the optimum weights of a neural network to capture the chaotic dynamics of electricity price data. The first method uses evolutionary algorithm to evolve a neural network, and the second method uses Particle Swarm Optimization for NN training. The global search capabilities of these evolutionary methods is used for finding the optimum neural network for forecasting electricity price from the California Power Exchange.