Wavelet-Enhanced Machine Learning Models for Hourly Solar Irradiance Forecasting

Abstract

Forecasting solar irradiance accurately is critical for optimizing renewable energy systems. In this study CWT was applied to solar irradiance for time frequency features. The CWT features, meteorological data and lag solar irradiance data were used to train seven ML models. The models were trained with 70% and assessed with 30% of the dataset with five statistical metrics. Results showed that MLP consistently achieved the best predictive accuracy with 12.9494 MAE, 339.81 MSE, 18.43 RMSE, -1.86 MBE and 0.9945 R2, while RF and GBR also performed competitively. Also, kNN with 32.71 MAE, 2115 MSE, 45.99 RMSE, -5.03 MBE and 0.97 R2 exhibited the weakest performance. These results show that CWT coefficient statistical features allow model like MLP, RF and GBR to capture irradiance variability more effectively.