Photovoltaic power prediction system based on multi-stage data processing strategy and improved optimizer

Abstract

Building a reliable forecasting system can quantify future fluctuations in short-term photovoltaic output power, which is essential for optimizing grid configuration and reducing operating costs. However, most of the existing studies only use denoising technology to preprocess data, which results in the elimination of some key information. And the traditional optimizer cannot meet the parameter optimization requirements of the prediction system because of its limited search capacity and search space. Based on the above problems, a combined system based on multi-stage data processing strategy and improved optimizer is proposed, which solves the tradeoff problem between prediction accuracy and stability. Firstly, the multi-stage processing strategy effectively improves the signal-to-noise ratio and preserves more implicit information. Then, the optimal sub-model determination strategy extends the structural framework of model selection and improves the flexibility of the prediction system. Finally, three improved strategies are introduced to improve the optimization ability and convergence speed of the optimizer, which magnifies the advantages of the prediction system. An empirical study using Safi-Morocco data shows that the symmetric mean absolute percentage errors of three photovoltaic modules are 4.777%, 4.755% and 6.033%, respectively, which implies that the system can not only achieve accurate prediction of photovoltaic output power, but also help to balance supply and demand and improve the overall sustainability and stability of the grid-connected power system.