A dual-path forecasting strategy for photovoltaic power with adaptive feature weighting and multi-scale attention

• A dual-path PV forecasting model with adaptive feature weighting and multi-scale attention is proposed. • An error correction enhances signal decomposition for feature extraction (ECVMD) is introduced. • AFWformer and TFMformer are designed to model stable trends and fluctuating signals. • The model on real PV datasets is validated with superior accuracy over baselines. Photovoltaic (PV) power generation is a widely adopted sustainable energy technology. Due to its complex operation nonlinear, and non-stationary characteristics, accurate forecasting of PV power poses challenges. In this paper, a dual-path forecasting strategy for photovoltaic power with adaptive feature weighting and multi-scale attention is developed. Firstly, through an improved variational mode decomposition with an integrated error correction (ECVMD) mechanism, the PV power signal is decomposed into intrinsic mode functions (IMFs) that contain both low-complexity and high-complexity components. Moreover, low-complexity components are forecasted using a lightweight-designed adaptive feature-weighted inspired Informer (AFWformer) model. Meanwhile, high-complexity components are modeled using a Transformer inspired by time–frequency multi-scale (TFMformer). Via the cross-attention mechanism, temporal-domain and frequency-domain features are integrated, and a multi-scale attention module is utilized to capture dynamic features. Besides, the hyperparameters of the model are automatically tuned under different conditions by the Bayesian optimization. The experimental results show that, across three PV plants and diverse weather conditions, the proposed strategy achieves MAE of 0.48–1.36, RMSE of 0.69–1.99, and R 2 of 0.973–0.991, evidencing consistently high accuracy across settings.