Application of machine learning for solar irradiance forecasting in Zambia

Abstract

Zambia has faced persistent energy shortages over the past decade due to heavy reliance on hydropower. Recent droughts, exacerbated by climate variability, have highlighted the vulnerability of this dependence and accelerated the country’s diversification toward solar energy. Effective integration of solar generation into the national grid requires accurate forecasting of global horizontal irradiance (Global Horizontal Irradiance (GHI)) to mitigate intermittency and support efficient energy planning. Despite increasing investment in solar power, limited research exists on forecasting models tailored to Zambia’s climatic conditions. This study develops a data-driven GHI forecasting framework using ten years of local meteorological data, including GHI, temperature, humidity, precipitation, and wind speed, sourced from the Zambia Meteorological Department (ZMD) and Joint Research Center Photovoltaic Geographical Information System (JRC PVGIS). A systematic literature review guided by the Preferred Reporting Items for Systematic Reviews and Meta- Analyses (PRISMA) framework and bibliometric analysis examined global, regional, and local trends in solar irradiance forecasting. Based on the review, three widely applied machine learning models—Long Short-Term Memory (LSTM), Random Forest (RF), and Artificial Neural Networks (ANN)—were selected for evaluation. Feature selection employed Variance Inflation Factor (VIF) analysis and Least Absolute Shrinkage and Selection Operator (LASSO) regression, identifying temperature and humidity as the most relevant predictors. The models were trained and tested using consistent data splits and evaluated using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and R². Results indicate that the ANN model achieved the highest predictive accuracy (MAE = 7.378 W/m², RMSE = 9.584 W/m², R² = 0.845), followed by RF (MAE = 9.845 W/m², RMSE = 12.374 W/m², R² = 0.715) and LSTM (MAE = 0.589 W/m², RMSE = 0.739 W/m², R² = 0.392). These findings demonstrate that reliable short-term GHI forecasting can be achieved using locally relevant predictors and accessible data. The study provides practical value for energy planners, utilities, and policymakers seeking to enhance grid stability, optimize solar dispatch, and improve renewable energy integration in Zambia. It recommends adopting ANN-based forecasting tools and further exploring hybrid and region-specific approaches supported by enhanced local climate data collection.