Machine learning approaches for daily solar irradiance forecasting under varying climatic conditions in the Hambantota region

Abstract

The shift toward renewable energy sources has raised the importance of accurate solar irradiance forecasting, especially for optimising photovoltaic energy systems. In regions with variable climates, reliable daily forecasting becomes a challenge due to rapid and unpredictable changes in atmospheric conditions. This research investigates machine learning approaches for improving the accuracy of solar irradiance predictions under different weather conditions, including clear, overcast, partly cloudy, and rainy conditions. A five-year dataset (2020–2024) for Hambantota was collected from the Weather Query Builder database, including global horizontal irradiance (W/m2) and meteorological parameters such as temperature (°C), relative humidity (%), wind speed (km/h), and cloud cover (%). These parameters were used as input variables to train and test multiple machine learning models, including Random Forest (RF), Extreme Gradient Boosting (XGBoost), Convolutional Neural Network (CNN), and a hybrid Convolutional Neural Network–Long Short-Term Memory (CNN-LSTM) architecture. Evaluation metrics included Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE). Results show that the CNN-LSTM consistently achieved the highest accuracy across all weather conditions. Under clear skies, the model achieved a MAE of 2.0 W/m2, a RMSE of 3.3 W/m2, and a MAPE of 0.7%. In overcast conditions, performance remained strong (MAE = 6.65W/m2, RMSE = 13.62 W/m2, MAPE = 4.1%). During rainy days, CNN-LSTM reduced MAPE to 8.5%, compared to 19.7% for CNN, representing a 57% improvement. Error distribution histograms also indicated that the model’s predictions were closely centered and symmetrical, reflecting reduced bias and variance compared to other models. In contrast, RF, XGBoost, and CNN models showed higher deviations, particularly under rainy and partly cloudy conditions. These results confirm the effectiveness of using deep hybrid learning models for handling the complex and nonlinear patterns associated with weatherdependent solar irradiance data. This research contributes toward building better forecasting systems that can increase the stability and efficiency of solar energy integration into power grids, especially in climatically varying regions.