Enhancing wind and solar energy forecasting through time-series feature engineering and ensemble machine learning

Accurate short- and medium-term forecasting of renewable energy generation is essential for ensuring grid stability, operational planning, and efficient energy management. This study presents a comprehensive time-series forecasting framework for wind and solar power production based on systematic feature engineering and advanced machine learning and deep learning models. The framework integrates lagged production variables, rolling-window statistics, calendar features, and cyclic temporal encodings to capture temporal dependencies and seasonal patterns. Forecasting horizons ranging from 1 h to 24 h are investigated using an expanding-window time-series cross-validation strategy to ensure robust and leakage-free evaluation. Experiments are conducted on a multi-year renewable energy production dataset comprising hourly wind and solar power generation records collected between 2020 and 2025. A broad set of baseline models, including statistical approaches such as ARIMA, machine-learning regressors such XGBoost, LightGBM, and CatBoost, and sequence learning architectures such as LSTM, are implemented and evaluated under identical preprocessing and validation protocols. Model performance is assessed using MAE, RMSE, MAPE, and R² metrics, complemented by bootstrapped confidence intervals and cross-validation stability analysis. Results indicate that ensemble learning and deep neural models consistently outperform conventional statistical methods across all forecasting horizons. For wind power forecasting at the 1 h horizon, LightGBM achieves an RMSE of 1309 MWh and an R² of 0.915, while the LSTM model attains an RMSE of 79,700 MWh on daily-scale aggregation. Performance degrades gradually for longer horizons, highlighting increasing uncertainty in medium-term prediction. Feature ablation experiments demonstrate that lagged production variables and rolling-window statistics contribute most significantly to predictive accuracy, while cyclic temporal encodings improve seasonal representation. Representative case studies further confirm the ability of recurrent models to capture peak generation and rapid ramping events. By providing transparent validation protocols, extensive baseline comparisons, and uncertainty quantification, this study establishes a reproducible benchmark for renewable energy forecasting and supports reliable operational decision-making in large-scale renewable energy integration.