Quality-Aware Photovoltaic Forecasting under Missing Weather Data

haitian ren — Tue, 22 Sep 2026 00:00:00 +0000

Operational photovoltaic forecasts often depend on irradiance and temperature measurements that may be interrupted by sensor or communication faults. This study develops Quality-Aware Dual-Path Conformal Photovoltaic Forecasting (QDPCC-PV), which combines a weather-aware Extra Trees expert with a power-only fallback expert. An availability gate changes their weights according to observed weather inputs, while physical clipping enforces feasible output and condition-specific conformal calibration provides prediction intervals. The method was evaluated chronologically on two measured photovoltaic plants sampled every 15 min at 15-, 60-, and 240-min horizons. Under complete inputs, mean absolute errors were 0.0376, 0.0463, and 0.0684 per unit, corresponding to improvements of 32.8%, 45.7%, and 65.3% over persistence. With 60% random weather loss at the 60-min horizon, QDPCC-PV obtained 0.0492 error, compared with 0.0664 for median imputation. The nominal 90% interval reached 91.8% coverage at 240 min but undercovered at shorter horizons because calibration and test residuals shifted over time. The results show that explicit input-quality routing protects point forecasts from weather-data loss, while adaptive calibration remains necessary for reliable short-horizon intervals.

International Journal of Advanced AI Applications

Quality-Aware Photovoltaic Forecasting under Missing Weather Data