Hybrid model for weather forecasting using ensemble of neural networks and mutual information

The chaotic and dynamic nature of weather makes weather forecasting a challenging and controversial task. Various numerical models have been developed and applied for this purpose, however usually they do not provide accurate predictions. Although artificial neural networks have been considerably applied for weather forecasting, they are not able to provide precise results. Consequently some researchers proposed to use ensemble models of neural networks for the prediction task. When considering multiple neural networks, the redundancy caused by having multiple models and also combining the results of different networks are still the main challenges. In this paper we propose a new hybridmodel for weather forecasting, which is based on an ensemble of neural networks. We address the redundancy issue by introducing a modular model in which a feature selection module is first applied to the data. We also, introduce a mutual information approach to tackle the challenge of combining the results of different networks and reducing the redundancy in the hybrid model. The evaluation results presented at the end of paper shows an outperformance of the proposed method compared to the similar methods in the literature.