Explainable Physics-Informed Transformer Framework for Probabilistic Solar PV Power Forecasting Under Extreme Weather Conditions: A Simulation-Based 20 kW Case Study for Karachi West, Pakistan

This preprint presents an explainable physics-informed Transformer framework for probabilistic short-term solar photovoltaic (PV) power forecasting under extreme weather conditions. A simulation-based 20 kW PV system in Karachi West, Pakistan, is investigated using NASA POWER hourly weather data from 2021–2025 and a pvlib-based digital twin. The proposed framework generates probabilistic forecasts at 1-hour, 2-hour, and 3-hour horizons using q10, q50, and q90 quantiles. It integrates physics-derived PV variables, extreme-weather indicators, conformal calibration, and explainability analysis. On the held-out 2025 test period, the Transformer achieved MAE values of 0.238 kW, 0.266 kW, and 0.318 kW at the 1-hour, 2-hour, and 3-hour horizons, respectively. Calibrated q10–q90 prediction intervals achieved approximately nominal 80% coverage under daylight and extreme-weather conditions. Deterministic benchmarking showed that Histogram Gradient Boosting achieved lower point-forecast errors than the Transformer, and this result is reported transparently. The main contribution of the framework is its combination of probabilistic forecasting, physics-informed modeling, extreme-weather evaluation, uncertainty calibration, and explainability. This study is explicitly simulation-based. The PV power target is generated using a pvlib digital twin and does not represent measured inverter data from a physical 20 kW installation. NASA POWER historical weather data are used rather than archived operational numerical weather prediction forecasts. This preprint is released for research dissemination, reproducibility, academic discussion, and future validation using measured PV plant data. DOI: 10.5281/zenodo.22169867