Comparison and Statistical analysis of various machine learning techniques for daily prediction of solar GHI representing India’s overall solar radiation

Abstract

Abstract Solar energy integration into the grid is a significant challenge because of its varying and unpredictable nature. Therefore, accurate solar energy prediction is vital in ensuring grid stability. To achieve this, the present study uses machine and deep learning methods to estimate the solar global horizontal irradiance. This study aims to predict daily solar GHI for four Indian states (Rajasthan, Madhya Pradesh, Assam, and Meghalaya) with different solar radiation distributions ranging from very high to very low. Four machine-learning techniques (linear regression, support vector machine, ANN and random forest) are used in the present study. Specific sites (Bhadla - Rajasthan, Rewa - Madhya Pradesh, Amguri-Assam, and Shillong-Meghalaya) were chosen in the respective states. The results of the sites represent the overall results for the entire state in this study. The dataset utilized for the study pertains to the selected sites and encompasses the period from January 2019 to November 2022. The study has focused on evaluating the success of machine learning techniques based on seven statistical metrics, including MBE, MAE, MSE, RMSE, Max. Error, R2, and MAPE. The result analysis indicates that all ML techniques' R2, MAPE, and MBE values lie between 0.6108 to 0.9152, 0.0432 to 0.2248, and − 0.2271 to 0.63704 MJ/m2, respectively. The study concludes that all of the machine learning techniques can accurately predict daily solar GHI, with ANN being the best-performing model.