Short-term photovoltaic power forecasting under complex weather conditions using optimized signal decomposition and attention-based learning

Abstract

Photovoltaic (PV) power sequences exhibit significant intermittency and strong non-stationarity under complex weather conditions, particularly during rapid cloud movements that intensify short-term fluctuations, posing substantial challenges for short-term forecasting. To address this, this study proposes a hybrid deep learning forecasting model. First, the Sparrow Search Algorithm (SSA) is employed to adaptively optimize key parameters of Variational Modal Decomposition (VMD), enabling the decomposition of the original power sequences into sub-modes with distinct frequency characteristics. Subsequently, a hybrid forecasting architecture is constructed by serially integrating a Temporal Convolutional Network with an Efficient Channel Attention (TCN-ECA) mechanism and a lightweight Transformer, enabling collaborative modeling of local high-frequency fluctuations and global long-term dependencies. Experimental results from the DKASC photovoltaic power station in Australia demonstrate that the proposed method reduces RMSE and MAE by 31.8% and 21.6%, respectively, compared to suboptimal models on an annual basis. It maintains high prediction accuracy even under cloudy conditions and during summer periods with strong meteorological fluctuations. Further integration with a physically constrained weather classification method based on the clear-sky index and scattering fraction enables systematic evaluation of the model's performance across diverse weather scenarios. By analyzing the relationship between meteorological feature fluctuations and prediction errors, this study reveals the mechanism underlying prediction errors under complex weather conditions. Results indicate that short-term non-stationary fluctuations in irradiance are the key factor amplifying prediction errors under complex weather conditions such as cloudiness.