A new method for short-term photovoltaic power generation forecast based on ensemble model

Abstract

Photovoltaic (PV) power generation is influenced by various factors, including weather conditions, the quality of PV inverters, and the cleanliness of PV modules, with weather conditions having a particularly significant impact on power output. This paper proposes a novel method for PV power generation prediction based on an ensemble forecasting model, aimed at constructing an efficient and stable PV prediction model. Initially, Z-score is employed to filter outliers in the PV data, and Robust STL–bilinear temporal–spectral fusion is introduced for time series feature extraction. Subsequently, an ensemble forecasting model based on bidirectional long short-term memory and extreme gradient boosting is proposed to address the limitations of existing predictive models, which suffer from low robustness and an inability to provide stable forecasts. Furthermore, to mitigate the performance degradation of the prediction model due to manual tuning, a tactics enhanced white shark optimizer is proposed for parameter optimization of the ensemble model. The optimization performance is validated using the IEEE CEC2021 test functions. Finally, the proposed method is tested on PV power generation data from a site in Shandong, China. The results demonstrate that the proposed ensemble forecasting method achieves high accuracy and exhibits strong model stability.