This study evaluates machine learning (ML) and deep learning (DL) approaches for short-term Global Horizontal Irradiance (GHI) forecasting in Sri Lanka’s tropical climate. Seven models, including Support Vector Machine (SVM), Random Forest (RF), XGBoost, Light Gradient Boosting Machine (LightGBM), Artificial Neural Network (ANN), Convolutional Neural Network (CNN), and Long Short-Term Memory (LSTM), were evaluated in different temporal look-back windows at five geographically diverse locations. The results show that traditional ML models consistently outperform DL approaches for operational forecasting windows, with XGBoost achieving the lowest average normalized RMSE of 21.33% and MAE of 13.96% at all locations using a six-hour look-back configuration. Performance varied significantly by location, with Batticaloa showing exceptional results (nRMSE = 16.67%, \(R^{2}\) = 0.93) while Horana presented challenges (nRMSE = 25.97%, \(R^{2}\) = 0.85). Computational efficiency analysis revealed XGBoost required only 4 min for complete hyperparameter tuning versus approximately 30 min for LSTM training. Primary research with 13 solar energy companies revealed that 92% operate without forecasting tools despite substantial growth from 10 MW in 2012 to over 245.94 MW by 2023. This work provides practical implementation guidance for solar forecasting in tropical regions and challenges assumptions about deep learning superiority for operational solar forecasting applications.