A composite photovoltaic power prediction optimization model based on nonlinear meteorological factors analysis and hybrid deep learning framework

Key factors influencing photovoltaic (PV) power generation predictions encompass solar radiation, aerosols, sunshine duration, temperature, humidity, wind direction, wind speed, cloud cover, and so on. The various influencing factors exhibit nonlinear correlation correlations, causing high volatility and discreteness in PV power time series. Firstly, to reduce the redundancy of the input for the prediction model and the computational time complexity, while enhancing the robustness and stability of the prediction model, nonlinear correlation search algorithm based on time window extending and time window shrinking strategies have been proposed. Key sequences from nonlinear correlation analysis are used in the next time series prediction model. Afterward, a novel dual-branch architecture that has synthesized the Structured Global Convolution (SGC) and iTransformer branches has been proposed which is called DBSGCformer. This framework enhances the ability to capture long-term dependencies through the combined effects of efficient convolution parameter optimization and variable-oriented multivariate modeling. We perform comprehensive experiments to investigate DBSGCformer’s potential in tackling complex multivariate time series forecasting challenges. Experiments conducted on two PV power datasets and five additional real-world datasets demonstrate that DBSGCformer significantly improves the accuracy of PV power forecasting and exhibits strong generalizability.