FORECASTING RENEWABLE ENERGY GENERATION: COMPARISON OF ARIMA, XGBOOST, AND A DEEP-LEARNING-INSPIRED PROXY MODEL

Abstract

Accurate short-term forecasting of renewable energy generation is important for power-system operation, balancing, and energy planning. This paper compares three forecasting approaches: the classical ARIMA model, the XGBoost gradient-boosting algorithm, and a simplified window-based proxy model inspired by the logic of LSTM sequence modelling. The study is designed as an undergraduate research paper and uses synthetic hourly data for 2023. The dataset imitates typical daily, seasonal, and stochastic patterns of solar and wind generation. The models were trained on the first eleven months and evaluated on the final thirty-day hold-out period. For solar generation, the proxy model achieved the lowest RMSE (21.50 MW), while XGBoost produced a competitive result (23.44 MW). For wind generation, both approaches showed similar performance, which reflects the stochastic nature of wind power. The study also demonstrates that MAPE may be unstable when generation values are zero or close to zero; therefore, RMSE and MAE are recommended as the main comparison metrics.