Spatiotemporal analysis of deep learning PV forecasting robustness under NWP errors

This study evaluates the robustness of six machine learning and deep sequence models, including PatchTST, GRU, N-HITS, and LightGBM, for photovoltaic power forecasting under numerically weather prediction (NWP) forecast errors. The authors present a physically constrained robustness evaluation framework using virtual PV power to isolate input uncertainty propagation.

• Sequence models demonstrate stronger noise filtering and temporal resilience than the tabular LightGBM baseline under medium to high disturbance regimes.
• Perturbations were heteroscedastic, modulated by clear-sky conditions and Erbs reconstruction to preserve radiation consistency.
• SHAP and Integrated Gradients analyses reveal a feature reallocation tendency where models shift reliance from corrupted future forecasts to stable historical observations and deterministic physical priors.
• A Pareto analysis of accuracy, robustness, and computational latency provides engineering implications for model selection under forecast uncertainty.

The findings translate into practical guidelines for assessing robustness and selecting models that maintain performance when faced with temporally correlated and physically coupled NWP errors.