Application of a New Artificial Neural Network Graduate Free Training
Faculty of Computing, Health and Science
School of Engineering and Mathematics
The performance of a new grid search Artificial Neural Networks (ANN) training method. Without derivative information. Is demonstrated with seasonal peak electric load demand model. The objective is to report the computational concept and the results of the derivative free ANN training that is designed as grid search method. Two 5-5-1 ANN architectures: a logistics and a multiplicative type have been proposed to model seasonal peak electric load and the forecasting efficiency have been compared with the well-known winter's exponential smoothing method. The logistics type ANN architecture forms ridges and the error function possesses high condition number. Accordingly, the logistics type ANN architecture requires more efforts to train compared to the multiplicative one. The multiplicative ANN model performs better than the logistic ANN. The multiplicative ANN retains a mean absolute percentage error (MAPE) 3.25 against 4.10 for the logistic ANN model. The multiplicative ANN outperforms the winter's exponential smoothing method in the validation period. The training method provides robust long-range load forecasting solution.