Comparative Analysis of Machine Learning and Deep Learning Models for Solar Energy Forecasting: A Meteorological Data-Driven Approach

Source abstract

Solar energy forecasting plays a crucial role in renewable energy integration and grid stability management. This study presents a comprehensive comparative analysis of machine learning and deep learning models for solar energy production forecasting using meteorological data. The research evaluates eight distinct forecasting approaches, including Long Short-Term Memory (LSTM) networks, XGBoost, linear regression variants (Linear, Ridge, Lasso), and classical time series models (ARIMA, SARIMA, Prophet), on both daily and hourly solar energy datasets. The meteorological dataset incorporates temperature, relative humidity, precipitation, cloud cover, sunshine duration, shortwave radiation, wind speed, and wind direction variables collected from the Kartepe region over seven years (2014-2020). Advanced preprocessing techniques, including feature extraction, lagged variables, moving averages, and robust scaling, were implemented to enhance model performance. The experimental evaluation employed k-fold cross-validation with statistical significance testing and confidence interval analysis to ensure robust model comparison. Results demonstrate that LSTM networks achieve strong performance on hourly data with $R^2 = 0.9622$ and RMSE = 9.92 kWh, effectively capturing complex temporal dependencies. For daily forecasting, Ridge regression exhibits good performance with $R^2 = 0.9997$ and RMSE = 8.53 kWh, demonstrating effective generalization capabilities. XGBoost shows competitive performance on hourly data ($R^2 = 0.9541$, RMSE = 10.94 kWh) while maintaining computational efficiency. Classical time series models, including ARIMA and SARIMA, demonstrate limitations in capturing complex meteorological relationships, particularly in high-frequency datasets. The study reveals that meteorological variables significantly enhance forecasting accuracy compared to approaches relying solely on historical production values. The research provides practical insights for renewable energy system operators and contributes to solar energy forecasting methodologies through detailed performance comparisons across multiple temporal resolutions.