DWT-Former: Fusing wavelet-based multi-scale features and transformer-based temporal representations for photovoltaic power forecasting

Accurate forecasting of photovoltaic (PV) power generation remains a key challenge due to the difficulty in separately modeling intrinsic signal dynamics and extrinsic influencing factors. Existing approaches often suffer from feature entanglement when processing PV series containing long-term trends, periodic components, and high-frequency noise. Moreover, the homogenized treatment of heterogeneous data sources – such as meteorological variables and historical power records with distinct temporal characteristics – can lead to information dilution, hindering forecasting accuracy. This study proposes DWT-Former, a hybrid deep learning framework designed to address these limitations through a dual-branch architecture. The intrinsic dynamics are handled by a discrete wavelet transform module that decomposes raw PV sequences into trend and seasonal components, enabling multi-scale representation across time, frequency, and scale. Simultaneously, extrinsic drivers are modeled by an enhanced Transformer module that captures long-range dependencies from fused meteorological and historical data. A feature fusion module integrates both branches to form a comprehensive representation of PV generation behavior. Experimental results on multiple real-world PV datasets demonstrate that DWT-Former consistently outperforms existing models in terms of MSE and MAE. The proposed framework effectively combines global dependency modeling with multi-scale signal decomposition, offering a robust and generalizable solution for high-precision PV power forecasting. These findings suggest strong potential for improving grid operation reliability and supporting large-scale solar energy integration.