PV-Former: A Universal Transformer Framework forFine-Grained Photovoltaic Power Output Forecastingunder Extreme Weather

Forecasting photovoltaic power generation during extreme weather events is particularly difficult, owing to the fundamentalvariability and unpredictability of photovoltaic systems. Traditional forecasting methods often rely on statistical or machinelearning models that struggle to capture the complex, nonlinear dynamics of PV power generation, particularly under extremeweather scenarios. To mitigate the aforementioned limitations, we present PV-Former, a general-purpose transformer-basedframework developed for high-resolution photovoltaic power forecasting. The framework incorporates the Symmetric PoseForecaster, which is composed of three core modules: the Constraint-driven Manifold Optimizer, the Agent-guided EventSegmenter, and the Probabilistic Symmetry Predictor. These modules collectively tackle the intricacies of PV power generationby employing manifold learning to model high-dimensional data structures, event segmentation to identify critical temporalpatterns, and symmetry-aware probabilistic modeling to enhance prediction accuracy. The framework further integratesadvanced strategies, including manifold adaptation refinement and Adaptive Policy-Orchestrated Integration, to improve itsrobustness and adaptability under extreme weather conditions. The experimental evaluation shows that PV-Former consistentlyoutperforms existing state-of-the-art methods, delivering over a 15% improvement in forecasting accuracy and exhibiting robustscalability across a wide range of weather scenarios. These results represent a significant advancement for renewable energysystems, enabling more reliable energy management and improved grid stability in the presence of climate-induced variability.