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

Area of Science:

• Energy Systems Engineering
• Data Science
• Machine Learning

Background:

• Accurate forecasting of renewable energy generation is vital for grid stability and operational planning.
• Existing methods often struggle with the inherent variability of wind and solar power.

Purpose of the Study:

• To develop and evaluate a comprehensive time-series forecasting framework for wind and solar power.
• To compare the performance of advanced machine learning and deep learning models against statistical methods.

Main Methods:

• A framework integrating lagged variables, rolling statistics, calendar features, and temporal encodings was developed.
• Expanding-window time-series cross-validation was used for robust evaluation.
• Models evaluated include ARIMA, XGBoost, LightGBM, CatBoost, and LSTM on a multi-year dataset.

Main Results:

• Ensemble learning and deep neural models consistently outperformed statistical methods.
• LightGBM and LSTM models showed strong performance for short-term wind and solar forecasting.
• Feature importance analysis highlighted the significance of lagged production and rolling-window statistics.

Conclusions:

• The proposed framework provides a reproducible benchmark for renewable energy forecasting.
• Advanced models offer improved accuracy for operational decision-making in renewable energy integration.
• Forecasting accuracy decreases with longer prediction horizons, indicating increased uncertainty.