Photovoltaic cluster power ultra-short-term cross-seasonal prediction integrating multi-channel information probabilistic diffusion generation and improved offset loss strategy

Accurate photovoltaic (PV) power prediction under complex meteorological conditions remains challenging, particularly given the pronounced seasonal variations that obscure generation patterns. This study presents a novel ultra-short-term prediction framework integrating meteorological volatility analysis with seasonal characteristic modeling. We developed a specialized multi-channel Gram angular summation field (MGASF) transformation matrix to holistically capture meteorological fluctuations, subsequently leveraging denoising diffusion probabilistic model (DDPM) for strategic augmentation of under-represented weather scenarios to enhance similar-day identification. Our hybrid architecture combines multi-channel vision Transformer (VIT) with bidirectional long and short-term memory (BILSTM) networks to synergistically analyze temporal dependencies and spatial patterns in PV similarity recognition. Furthermore, we engineered a seasonal-adaptive prediction system through an improved variable-weight Smooth L1 loss function, establishing an optimized seasonal alignment mechanism that achieves high-precision prediction across varying meteorological conditions with minimal computational overhead. Through rigorous validation using operational data from a utility-scale photovoltaic cluster in Western Inner Mongolia, the proposed method achieved consistent accuracy improvements: 3.02 % reduction in N_RMSE, 1.65 % decrease in N_MAE, and 2.19 % enhancement in R² compared to baseline approaches in PV cluster. These statistically significant enhancements demonstrate our framework’s capability to mitigate seasonal impacts while maintaining prediction reliability in complex meteorological environments.