PMWC: A hybrid framework based causal inference and multi-scale feature fusion for day-ahead PV power forecasting

Photovoltaic (PV) generation is affected by both historical trends and short-term climate fluctuations, leading to high uncertainty and posing significant challenges for accurate forecasting. This paper introduces a day-ahead PV forecasting approach grounded in causal inference and multi-scale feature fusion. Given that traditional feature selection methods focus solely on relevance and overlook the varying impact of different lag items, the Peter–Clark Momentary Conditional Independence (PCMCI)-based feature weighting mechanism is employed to quantify the direct causal influences of various lagged features. Moreover, the PV power generation series exhibit long-term dependencies, local dynamics, and periodic structures. To capture these multi-scale characteristics, three models — Bidirectional Long Short-Term Memory (BiLSTM), Temporal Convolutional Network (TCN), and Gramian Angular Field-Based Squeeze-and-Excitation ResNet50 (GAF-SE-ResNet50) — are employed in parallel. However, solely focusing on temporal trend features remains inadequate. To adapt to instantaneous weather fluctuations, preliminary predictions are refined by integrating weather fluctuation features extracted via wavelet decomposition through cross-attention. Experimental results demonstrate that the proposed model achieves superior forecasting accuracy and robustness, reducing RMSE by up to 7.5% and improving R2 by up to 2.4% over the state-of-the-art (SOTA) model across stations. It offers a promising solution for reliable PV output estimation, with potential benefits for intelligent power system scheduling.