A Time-Driven Approach Leveraging Universally Accessible Features for Photovoltaic Power Forecasting

The rapid expansion of photovoltaic (PV) systems demands accurate hourly power forecasting to maintain grid stability and enable efficient energy management, particularly in regions with limited or unreliable meteorological data. Traditional models rely heavily on weather inputs like irradiance and temperature, which can limit scalability. In contrast, time-based features—such as cyclical encodings of hour and day—capture deterministic solar patterns are universally accessible. This study compares the effectiveness of time-based features versus meteorological inputs for hourly PV power forecasting, aiming to identify robust models that minimize dependency on weather data without compromising accuracy. Using hourly PV generation data from Konstanz, Germany, and augmented with cyclical time encodings and meteorological variables, eight machine learning models—including LSTM, GRU, CNN, ensemble methods (CatBoost, Gradient Boosting), and hybrid models (LSTM + LightGBM, GBR + CatBoost)—were trained and validated. Chronological splits, MinMax scaling, and 5-fold TimeSeriesSplit cross-validation prevented data leakage. The LSTM + LightGBM hybrid achieved the best performance using only time-based inputs (test R² = 0.976, MAE = 0.0053, RMSE = 0.0159), closely approaching weather-inclusive results (R² = 0.9937, MAE = 0.0038, RMSE = 0.0107). Composite scoring ranked this hybrid highest, reducing errors by up to 32% compared to baselines. Sensitivity analysis confirmed optimal sequence lengths of 24 hours, and cross-validation showed strong generalization (average validation R² = 0.884). These results demonstrate that time-based features offer a reliable, low-resource alternative for PV forecasting, supporting solar integration in data-scarce regions and promoting sustainable energy practices.