Automated heuristic growing of neural networks for nonlinear time series models

In this paper, we present a method for automatically selecting the optimal architecture of feedforward neural networks to build nonlinear time series models. A heuristic method is used to do an exhaustive search of all possible input/output combinations, while adjusting the lag times and the number of nodes in a fully connected single hidden layer network. Levenberg-Marquardt optimization is performed using the stop-search method of cross-validation. Statistics are maintained for all optimized structures which permits postprocessing based on performance criteria for final model selection. The methodology is applied to a case study for developing multi-variate autoregressive models for the day-ahead forecasting of electricity prices.