Photovoltaic power forecasting based on VMD-SSA-Transformer: Multidimensional analysis of dataset length, weather mutation and forecast accuracy

Photovoltaic (PV) power generation is characterized by inherent intermittency and uncertainty, underscoring the critical need for accurate PV power forecasting to enhance grid stability and optimize energy management. This study proposes a novel hybrid forecasting model integrating Variational Mode Decomposition (VMD), Sparrow Search Algorithm (SSA), and a Transformer-LSTM architecture. The proposed model effectively captures the complex interactions between meteorological variables and PV power time series, enabling high-precision forecasts. A sensitivity analysis on training data length indicates that the VMD-SSA-Transformer-LSTM model reduces data requirements by 87.5 % compared to conventional LSTM models, significantly lowering computational costs and improving feasibility for newly deployed rooftop PV systems. Performance evaluation across typical time periods—stable weather, fluctuating weather, and post-fluctuation stabilization—demonstrates exceptional adaptability to dynamic weather patterns. Furthermore, a systematic analysis of forecasting accuracy degradation with increasing prediction horizons was conducted. Finally, comparative experiments across multiple benchmarks—including existing models, diverse geographic locations, varying temporal resolutions, and extreme weather conditions—validate the proposed model's robustness. The normalized RMSE remains below 6.5 % across all scenarios, confirming its superior accuracy and reliability.