Group‐based chaos genetic algorithm and non‐linear ensemble of neural networks for short‐term load forecasting

Abstract

This study presents a non-linear ensemble of partially connected neural networks for short-term load forecasting. Partially connected neural networks are chosen as individual predictors due to their good generalisation capability. A group-based chaos genetic algorithm is developed to generate diverse and effective neural networks. A novel pruning method is employed to develop partially connected neural networks. To further enhance prediction accuracy, an artificial neural network-based non-linear ensemble of partially connected neural network predictors is developed. The proposed non-linear ensemble neural network is evaluated on a PJM market dataset and an ISO New England dataset with promising results of 1.76 and 1.29% error, respectively, demonstrating its capability as a promising predictor.