Comparative Analysis of Forecasting Models for Hourly Solar Irradiance in Southern Thailand

Abstract

Accurate forecasting of short-term solar irradiance is essential to optimize energy supply, improve grid stability, and increase the economic viability of photovoltaic systems. This study compares several forecasting models for predicting hourly solar irradiance using real-world data from Southern Thailand. Eight models were evaluated, including six forecasting models (Prophet, LSTM, GRU, XGBoost, Random Forest, and SARIMA with walk-forward validation) and two benchmarks (Naive and Seasonal Naive), using RMSE, MAE, and MAPE metrics. The results demonstrate that ensemble-based machine learning models significantly outperformed traditional statistical and deep learning approaches. Specifically, XGBoost achieved the best performance with RMSE of 79.43 W/m², MAE of 59.63 W/m², and MAPE of 30.07%, followed closely by Random Forest with RMSE of 83.00 W/m² and MAPE of 29.36%. Compared to the best benchmark model, XGBoost reduced RMSE by 50.3% and MAPE by 47.1%. The GRU model outperformed LSTM, demonstrating greater efficiency in capturing short-term temporal dependencies. The findings confirm that ensemble methods effectively capture complex temporal patterns and nonlinear meteorological dependencies, making them highly suitable for real-time solar forecasting applications in tropical climates.