Novel solar forecasting scheme modelled by mixer dual path network and based on sky images

Abstract

The prediction of global horizontal irradiance has become an effective technique to address the intermittence issue of photovoltaic (PV) power generation. This article proposes a novel deep neural network(DNN), named Mixer Dual Path Network (Mixer-DPN), for promising solar forecasting. It shares common features of cloud images and maintains the flexibility to explore new features through dual-path architecture by combining the Mixer layer and Dual Path Network. Therefore, the proposed model can provide more accurate prediction results compared to the classical DNN-based predictors. Moreover, the proposed model shows a faster convergence speed and smaller model size, which makes it suitable for a practical global horizontal irradiance. The merits of the proposed model are verified by testing it with the data from National Renewable Energy Laboratory comparing it with other DNN-based prediction models. Studies have shown that the new model has achieved excellent results in MSE, MAE and other indicators, and the R2 prediction accuracy rate has increased by 14% compared with the baseline model.

Keywords

• Deep learning (DL)
• Solar forecasting
• Dual path network (DPN)
• Convmixer architecture