An Improved Multilayer Perceptron Neural Network Approach for Solar Irradiance Forecasting in Ghana

Abstract

The growing global energy demand, depletion of fossil fuels, and rising environmental concerns necessitate a shift toward sustainable energy sources. Solar energy, particularly in regions like Sub-Saharan Africa, offers significant potential due to its high availability. This study develops a Multilayer Perceptron (MLP) neural network to forecast Plane-of-Array (POA) irradiance in Navrongo, Ghana, using historical meteorological and solar data from 2021-2024. Eight input features, including temporal variables, weather parameters, and direct solar measurements, were used. Data preprocessing and normalization were applied, and training utilized the Levenberg-Marquardt algorithm in MATLAB. The model achieved excellent performance, with an $\mathbf{R}$-value of 0.9993 and MSE of 117.50, outperforming baseline models. External validation using Sunyani data showed good generalization but reduced accuracy, highlighting the need for local calibration. Results demonstrate that incorporating direct irradiance inputs, particularly Global Horizontal Irradiance (GHI) and Direct Normal Irradiance (DNI), significantly improves solar forecasting accuracy in data-scarce regions.