Machine Learning-Based Forecasting of Hybrid Wind–Solar Energy Generation For Sustainable Grid Integration

Abstract

Integrating renewable energy into power grids requires precise energy forecasting for each area. This study designed a physics-informed machine learning model to predict combined wind and solar power in Punjab, with 60,909 data points (April 2021–August 2025) obtained from five meteorological stations hourly data telemetry. Raw wind speed and solar radiation readings were then corrected for 53 nighttime solar readings (0.21%) which were implausible with a non-zero value and then converted to turbine power output using a turbine power curve, and PV energy output using a PV equation. Five models: Linear Regression, Random Forest, XGBoost, LightGBM, and LSTM models were trained on the leakage-free features (lags, rolling statistics, calendar variables). Random forest and XGBoost had the best base performance (R² = 0.867 and 0.874, respectively) on static test split. A seasonally-aware walk-forward validation used four seasons of training data showed that forecast accuracy was heavily dependent on having at least one season of training data. SHAP analysis was used to interpret the contributions of the features to the model's predictions. It estimates the 2,187 MWh/year renewable energy generation capacity in the region with a solar and wind correlation of near zero (r = -0.0230), thus proving the statistical independence of both energy sources, and their complementarity. These findings provide a region-specific forecasting benchmark for Punjab and methodological evidence on seasonal data requirements for reliable renewable forecasting, offering practical guidance for grid operators and policymakers in northern India.