Weather-Feature-Informed Ultra-Short-Term PV Power Forecasting for Hydrogen Production Systems

Abstract

Existing studies incorporating weather information mainly focus on a single forecasting model, making it difficult to separate the contribution of target-time meteorological features from model architecture. This study proposes a unified weather-feature-informed (WFI) framework by introducing target-time meteorological features into extreme gradient boosting (XGBoost) and long short-term memory (LSTM). Observed target-time variables are used as idealized inputs to quantify their contribution; operational deployment requires short-term weather forecasts. Photovoltaic (PV) power 15 min ahead is predicted from the preceding 2 h of power and meteorological data, and test days are classified into three weather types according to fluctuation intensity and solar resource level. The WFI framework improves both baseline models, with larger gains under highly fluctuating conditions. Under Type II and Type III conditions, the mean absolute error (MAE) of WFI-XGBoost decreases by 52.62% and 57.07%, respectively, while that of WFI-LSTM decreases by 33.78% and 41.20%. WFI-XGBoost achieves the lowest MAE and root mean square error (RMSE) under all three weather types. The results demonstrate the cross-model applicability and weather-dependent gains of WFI for short-term operational control in PV hydrogen production systems.