A cross-domain information fusion method for non-stationary distributed photovoltaic forecasting

With the increasing penetration of distributed PV systems in distribution networks, their inherent randomness and intermittency pose significant challenges to the safe, stable, and economic operation of power systems. The non-stationarity and uncertainty of PV power generation, particularly in long-term forecasting, severely hinder regulation efficiency in power systems. This paper proposes RCLformer, a novel PV power prediction model that integrates cross-domain information fusion with non-stationary sequence modeling to address these issues. The model constructs a forecasting backbone based on Crossformer and incorporates exogenous variables such as weather conditions, enabling effective cross-domain information fusion and fully exploiting correlations among multi-source data. To capture the rapidly fluctuating characteristics of PV output caused by complex weather patterns, an LSTM module is embedded in the Crossformer embedding layer, enhancing sensitivity to short-term local dynamics and improving the joint extraction of local and global spatiotemporal features. Additionally, a reversible instance normalization module is introduced to improve generalization capability and reduce the impact of trend-related interference on prediction accuracy. Comprehensive multi-task experiments are conducted on three datasets, including two real-world PV operation datasets and one public benchmark dataset, under various forecasting horizons to evaluate robustness in long-term prediction scenarios. Experimental results demonstrate that RCLformer consistently outperforms baseline models, achieving improvements of up to 19.0 % in MSE on real-world PV datasets. These findings substantiate the effectiveness of RCLformer in distributed PV forecasting and highlight its strong potential for real-world engineering applications.