Adaptive multi-scale dual-attention network with hyperparameter learning for short-term photovoltaic power forecasting

Abstract Accurate short-term photovoltaic (PV) power forecasting is crucial for reliable operation and efficient dispatch of modern power systems. However, PV generation is strongly affected by meteorological variability and exhibits pronounced intermittency, fluctuations, and heterogeneous temporal characteristics. To address the strong intermittency and complex temporal dynamics of PV generation, this study proposes a hybrid forecasting framework, SSA-DABNet. First, a multi-scale dilated convolutional neural network (MCNN) with parallel branches and different dilation rates is developed to capture short-term fluctuations and long-term trends while enlarging the effective receptive field. Then, a dual-stage attention mechanism is introduced by integrating the squeeze-and-excitation (SE) block with temporal attention, enabling adaptive feature weighting in both channel and temporal dimensions to enhance critical information representation while suppressing redundant features. Finally, the sparrow search algorithm (SSA) is employed to optimize key hyperparameters, including the sliding-window size, SE reduction ratio, and initial learning rate, reducing dependence on manual tuning and improving model adaptability. Experiments conducted on real-world data from a PV power station in Gansu demonstrate that the proposed model achieves stable and accurate forecasting performance, with a coefficient of determination R2 of 0.9404 and a mean absolute error (MAE) of 2.2607. Compared with several benchmark models, the proposed framework exhibits superior performance, providing an effective solution for PV power forecasting and refined operation of renewable energy systems under complex meteorological conditions.