Machine Learning-Based Short-Term Wind and Solar Energy Forecasting Framework

Abstract

Wind energy is increasingly important for reducing dependence on fossil fuels, yet its variable and non-stationary nature complicates reliable energy planning. According to the Republic of Türkiye Ministry of Energy and Natural Resources (2026), electricity consumption reached 360.9 TWh in 2025 and is projected to increase to 455.3 TWh in 2030 and 510.5 TWh in 2035. Therefore, accurate short-term renewable energy forecasting is essential for grid planning, balancing, and decision-making and supports United Nations Sustainable Development Goal 7: Affordable and Clean Energy.

This study develops a machine learning-based framework for short-term hourly wind speed and solar radiation forecasting. Hourly meteorological data for winter and summer periods between 2021 and 2025 were obtained from the NASA POWER database for Seferihisar, İzmir. Wind speed at 50 m and total shortwave solar radiation were selected as target variables. For wind forecasting, RF, ET, HGB, XGB, LGBM, CatBoost, SVR, CNN, and LSTM models were evaluated. For solar forecasting, ANN, CNN, LSTM, RF, SVR, and XGB were applied. Each model was tested using both raw inputs and wavelet-based hybrid inputs In the hybrid framework, d1, d2, and d3, which are components derived from the target-related main variables, were added to the input set.

Wind-model inputs included wind speeds at 10 and 50 m, surface temperature, and pressure. Solar-model inputs included solar radiation, clearness index, air temperature, surface pressure, and relative humidity.

The results showed that wavelet-based hybrid models generally improved forecasting performance. W-SVR achieved a 10–20% improvement in short-term wind speed forecasting, reaching r=0.89 for the first 45-hour horizon. For the 442-hour solar forecast, W-ANN improved r from 0.8948 to 0.9203 and R² from 0.6524 to 0.8020 in winter 2021. The highest solar performance was obtained by W-RF in summer 2025, with r=0.9977 and R²=0.9935, while RMSE and MAE decreased by approximately 16%.