A Foundation-Specialist Diffusion Framework for High-Fidelity Wind and Solar Scenario Generation

The transition to power systems with high shares of variable renewable energy demands high-fidelity scenario ensembles capable of accurately capturing the spatiotemporal characteristics of wind and photovoltaic (PV) generation, including multi-scale variability, persistence, ramping behavior, and inter-technology complementarity. However, existing data-driven generative models often face a trade-off among realism, controllability, and computational efficiency. To address this, we propose the Foundation-Adapted Diffusion Framework for Renewable Scenarios (FADFRS). FADFRS employs a foundation and specialist diffusion model architecture. A shared foundation model is first trained on multi-year capacity factor time series to learn generic temporal patterns. Then, lightweight technology-specific adapters are fine-tuned for wind and PV to capture domain-specific dynamics, such as diurnal/seasonal cycles for PV and persistence regimes with extreme ramps for wind. The framework supports conditional generation based on calendar variables and spatial metadata, enabling the creation of spatially coherent multi-site scenarios and the targeted sampling of low-probability, high-impact events (e.g., renewable droughts). Model fidelity is rigorously assessed with a comprehensive suite of diagnostics. This includes established power system metrics (e.g., duration curves, ramp distributions, spectral signatures) as well as advanced probabilistic scores such as the Energy Score, Variogram Score, and FID. Case studies demonstrate that FADFRS consistently outperforms conventional generative baselines in preserving key statistical and dynamical features while maintaining scenario diversity. The work provides a powerful and practical tool for both retrospective analysis and prospective planning of high-renewable power systems.