SPI-Net: Temporal scale decoupling and physics-inspired calibration network for non-stationary photovoltaic power forecasting

Photovoltaic power forecasting under non-stationary conditions is challenging because module degradation and soiling cause distribution shifts that undermine conventional deep learning models. To address this, we present SPI-Net, a physics-inspired two-stage framework that decomposes the forecasting task into nominal meteorological response and state-dependent bias correction. A Health State Index and a Cleanliness State Index are introduced as proxy variables to quantify slow, irreversible aging and rapid, reversible soiling effects, respectively. The first stage employs a multi-scale convolutional neural network and long short-term memory baseline predictor using only meteorological features, while the second stage deploys a dual-stream calibrator with bidirectional cross-attention to model the nonlinear interaction between aging and soiling. A dynamic fusion mechanism adaptively integrates outputs from both streams based on real-time system state. Experiments on a multi-year rooftop photovoltaic dataset from the Hong Kong University of Science and Technology show substantial improvements: a 40.7% mean squared error reduction compared to the best baseline method, with particularly strong gains for aged panels (17.4% improvement) and under poor air quality conditions (21.8% improvement). Ablation studies confirm the complementary contributions of dual-stream separation and cross-attention mechanisms. Analysis of the learned dynamic fusion weights reveals emergent state-dependent behavior that provides interpretable cues for condition-based maintenance scheduling.