Comparative assessment of statistical, Machine Learning, and Hybrid models for global solar irradiation forecasting in a tropical climate: A case study of Man (Côte d'Ivoire, West Africa)
Abstract
Accurate solar irradiation forecasting is essential for the reliable integration of photovoltaic energy into power systems, particularly in tropical regions where atmospheric conditions exhibit strong temporal variability. However, the nonlinear and multiscale nature of solar radiation dynamics makes forecasting performance highly dependent on the prediction horizon. This study investigates global solar irradiation forecasting in Man, Côte d’Ivoire, using a comparative framework involving statistical (ARMA), machine learning (K-NN, Random Forest, and XGBoost), and deep learning (LSTM and CNN-LSTM) models. Forecasts were developed for one-hour-ahead (H+1), one-day-ahead (D+1), and one-month-ahead (M+1) horizons using meteorological data from NASA's POWER database. Model performance was evaluated using RMSE, nRMSE, MAE, MAPE, and MBE. Feature importance analysis based on Random Forest and the Diebold–Mariano test were used to assess the contribution of input variables and compare model performance. The results reveal a strong scale-dependent forecasting behavior. For the H+1 and D+1 horizons, the CNN-LSTM model achieved the highest predictive accuracy, with nRMSE values of 9.84% and 13.15%, respectively, significantly outperforming the persistence benchmark (p < 0.001). At the M+1 horizon, the K-NN model provided the best performance, with an nRMSE of 4.49%. These findings demonstrate that no single model is universally optimal across temporal scales. Instead, forecasting performance is governed by the interaction among temporal resolution, atmospheric variability, data availability, and model complexity. The proposed scale-aware framework supports reliable solar forecasting and photovoltaic energy integration in tropical and sub-Saharan regions.
