• Enhanced Transformers with cross-variable attention for PV system dependency capture. • Integrated a linear for trend extraction and a Transformer for complex interactions. • Achieved SOTA on 3 real-world PV datasets (5.6% MSE reduction, 0.7% accuracy gain). Accurate forecasting of photovoltaic (PV) power generation is indispensable to the optimal operation and strategic deployment of PV energy systems, facilitating robust energy management practices and the smooth integration of PV resources into the power grid. Yet, achieving precise PV power predictions remains difficult due to uncertainties stemming from highly dynamic weather conditions and complicated interactions across a range of relevant variables. To address these core challenges, this study introduces PV-Client (Cross-variable Linear Integrated ENhanced Transformer for Photovoltaic power forecasting). PV-Client integrates an Enhanced Transformer module to capture the dependencies among multiple variables of PV systems, alongside a linear module designed to acquire the underlying trend of PV power. Unlike conventional time-series Transformers that rely on cross-time Attention to model temporal dependencies, PV-Client’s Enhanced Transformer module adopts cross-variable Attention to explicitly capture interactions between PV power and weather-related inputs. Additionally, PV-Client incorporates a stabilization module and features optimized internal structures within its constituent modules. Validation experiments carried out on three real-world PV stations demonstrate that PV-Client achieves SOTA performance: it consistently outperforms the strongest baseline at each station—beating the second best model LR by 13.3% (MSE) and 0.9% (accuracy) at Jingang Station, SVR by 10.1% (MSE) and 0.2% (accuracy) at Xinqingnian Station, and LSTM + Attention by 1.4% (MSE) and 1.1% (accuracy) at Hongxing Station. Extensive supplementary experiments further confirm PV-Client’s robustness and generalization.
Gao et al. (Sun,) studied this question.