CausalTransPV: Causal Invariant Representation Learning for Cross-Site Photovoltaic Power Forecasting via Selective Domain Alignment

Cross-site transfer learning is a promising approach to address data scarcity at newly deployed photovoltaic (PV) stations by leveraging knowledge from data-rich source sites. However, existing domain adaptation methods align feature representations without distinguishing physically meaningful causal relationships from site-specific spurious correlations, leading to negative transfer when local environmental conditions differ substantially between stations. This paper proposes CausalTransPV, a causal invariant representation learning framework that integrates explicit temporal causal discovery with selective domain alignment for cross-site PV power forecasting. The framework comprises three synergistic modules: (i) a multi-station temporal causal discovery module that jointly learns shared and station-specific causal graphs through differentiable acyclicity-constrained optimization with a cross-station invariance regularizer; (ii) a causal-guided disentangled encoder that decomposes representations into causal-invariant and site-specific subspaces using the discovered causal graph as a structural prior; and (iii) a causal-subspace transfer and prediction module that performs maximum mean discrepancy (MMD)-based domain alignment exclusively on the causal subspace. Experiments on the Desert Knowledge Australia Solar Centre (DKASC) multi-station dataset under varying target label ratios (0–50%) demonstrate that CausalTransPV achieves relative mean absolute error (MAE) reductions of 6.9–9.9% over the strongest baseline. Ablation studies, causal graph analysis, feature space visualization, and weather-conditioned case studies further validate the contribution of each component. These results suggest that causal-guided selective transfer offers an effective paradigm for reliable PV forecasting under data-scarce cross-site scenarios.