Dual-Branch Residual latent diffusion with multi-level contrastive learning for robust day-ahead wind power scenario generation

The growing penetration of wind power has intensified challenges to power system stability due to its inherent variability and the rising occurrence of extreme weather. Reliable day-ahead scenario generation is therefore essential, yet conventional methods often fail to represent the non-Gaussian and heavy-tailed characteristics of wind power. To address this issue, PRCLD (Physics-informed Robust Contrastive Latent Diffusion) is proposed as a generative framework that integrates a dual-component residual diffusion mechanism with multi-level contrastive learning. At its core, the Dual-component Residual Diffusion Module integrates a Spectral–Temporal Adaptive Decoupling (STAD) mechanism, which leverages frequency-domain priors to separate low-frequency baseline trends from high-frequency extreme transients. These decoupled components guide two parallel denoising branches to accurately model typical variability and extreme deviations, enabling precise representation of long-tailed uncertainty. To improve robustness under uncertain NWP inputs, PRCLD employs physics-guided multi-level contrastive learning, enforcing embedding consistency across clean and perturbed views and thereby enhancing stability and generalization. Experiments on the GEFCom2014 dataset show that PRCLD significantly outperforms state-of-the-art baselines, reducing CRPS by up to 13.5% relative to a strong diffusion benchmark. A stochastic day-ahead economic dispatch on an enhanced IEEE 33-bus system further verifies its practical value, yielding an expected operational cost within 0.25% of the true benchmark. Overall, PRCLD generates physically coherent and probabilistically calibrated wind power scenarios, effectively capturing both normal variability and extreme ramping events, providing a reliable tool for system scheduling and risk-aware planning with high renewable penetration.