Short-term photovoltaic power forecasting under extreme weather conditions based on feature mode decomposition and adaptive sparse self-attention

Abstract

Aiming at the strong non-stationarity, pronounced abrupt changes, and error accumulation that characterize photovoltaic power series under extreme weather conditions, this paper proposes a short-term photovoltaic power prediction model, feature mode decomposition (FMD)- long short-term memory (LSTM)- adaptive sparse self-attention (ASSA)-Transformer. First, FMD is employed to perform multiscale decomposition of the original power sequence, and effective modes are selected by combining correlation kurtosis and correlation coefficient, thereby reducing the impact of trend, periodic, and disturbance component aliasing on prediction performance. Second, LSTM is used to extract local temporal features, and an ASSA mechanism is introduced to improve the model’s responsiveness to key abrupt segments such as power dips and rapid recovery. Finally, Transformer is used to model long-term dependencies and improve the stability of multi-step prediction. Based on measured power and meteorological data from a photovoltaic power station collected during 2020–2021, extreme-weather scenarios in four seasons are constructed for validation. The results show that the proposed model outperforms the comparison models in spring, summer, autumn, and winter. Taking spring as an example, the model achieves an root mean square error of 1.725 and an mean absolute error of 1.150, which are 39.13% and 43.49% lower than those of LSTM, respectively, while the R2 reaches 0.963. Decomposition comparison and ablation experiments further demonstrate that FMD and ASSA can effectively improve prediction accuracy and stability.