Neural networks ensemble model based on Boosting algorithm for short-term load forecasting

A revised adaptive boosting algorithm for neural networks ensemble model is proposed. In the algorithm, the relevant error criterion is used instead of the absolute error criterion, for it is more closed to the essential of the predict regression model. And at each step of boosting iteration, the new validation subset is obtained from validation sampled subsets, while getting new training subset from training sampled subset. The correspondence between the two subsets is guaranteed. The proposed algorithm is applied to build a neural network ensemble load forecast model using the real data from the California power market of the United States. The numerical simulation results show that the proposed ensemble model can improve the stability of model outputs significantly, and increase the reliability in network structure determination and model selection. With the ensemble model, the better forecasting accuracy is achieved in comparison with the single neural network model.