Training artificial neural networks for short-term electricity price forecasting

This paper present a comparative study of training approaches for artificial neural network (ANN) used in forecasting short-term wholesale electricity prices. High probability of volatility in wholesale electricity prices and trends that are generally non-uniform create challenges when forecasting future prices using simple backpropagation feedforward ANN. A number of ANN architectures and training methods have been proposed for a variety of applications, and here we consider three approaches with actual electricity price data. The architectures considered in this study are: the well known feedforward and Elman networks trained with backpropagation, which are compared to feedforward network trained with genetic algorithm. Avoidance of local minima and minimization of computational cost are key performance indicators in ANN training. Number of training iterations needed to achieve target error and the generalization ability are used to compare the methods. This investigation is meant to guide in selecting ANN training method for electricity price forecasting.