Neural Networks of Combination of Forecasts for Data with Long Memory Pattern

We experience with three neural network models for forecasting electricity peak-load. The goal of the paper is not to compare neural networks with other traditional univariate forecasting models. However, the objective is to better understand the performance of neural networks for the cast when the date exhibits a long memory pattern. The first model is a simple one consisting of the time series values. The second model uses the same inputs as the first model with one additional input representing the combination (average) of six univariate time series forecasting models. The third model consists of all these individual time series forecast; in other words, the neural network is let to combine them in its own way. We obtained a considerably improved performance using information from all univariate forecasting models. The third model significantly outperformed the other two models. The optimum network is identified through extensive sensitivity analysis of the characteristics of the networks. These characteristics included training parameter, number of hidden layers, and testing and training percentages.