A Study on a Hybrid Photovoltaic Power Forecasting Model Based on Measured Weather and Generation Data

Accurate photovoltaic(PV) power forecasting is essential for stable grid operation, yet PV output varies sharply with weather. This study proposes a measured-data-based week-ahead (168-hour) forecasting framework using on-site PV generation and hourly meteorological observations. We construct an hourly dataset of 8,033 records (May 2024–April 2025) and apply cyclic time encodings, lag features, rolling statistics, and interaction features to capture temporal dependency and periodicity. Using timeordered splits, we compare tree-based machine learning models with deep learning baselines. Among single models, XGBoost achieves the best performance (RMSE 70.84 W, R² 0.925). A Ridge-based stacking ensemble further improves accuracy (RMSE 62.21 W, R² 0.943). These results suggest that lightweight stacking improves stability for week-ahead PV forecasting under limited measured data.