Machine learning-based approach for PV energy forecasting for mono-Si, poly-Si and a-Si Grid-connected PV systems

• RF and DNN are applied to predict PV energy • Production of amorphous, monocrystalline and polycrystalline panels have been compared • Study in semi-desert climate and impact on PV panel performance • Evaluation of the actual accuracy of the grid-connected PV system • RF is more accurate in predicting PV energy of the three technologies • Amorphous has the best accuracy R2 (FR=98.6%, DNN=98.3% and MAPE=9.3%) • Accuracy depends on technology and amorphous is less sensitive The growing global energy demand and the urgent need to reduce greenhouse gas emissions have intensified the search for renewable and sustainable energy sources. Among these, photovoltaic (PV) systems have emerged as a promising solution due to their long lifespan, low maintenance costs, and ability to operate under diverse climatic conditions. However, the intermittent nature of solar energy remains a major challenge for stable integration into electrical grids, especially in semi-desert regions. Despite existing research on PV performance, limited studies have focused on the comparative forecasting of different silicon-based PV technologies using advanced machine learning models in such environments. This study aims to forecast and compare the energy performance of grid-connected monocrystalline silicon, polycrystalline silicon, and amorphous silicon PV systems operating in a semi-desert region of Morocco. Using two years of real measured daily meteorological and energy production data (January 2021 to December 2022), we developed predictive models based on Random Forest and Deep Neural Networks. The models' accuracy was evaluated using multiple error metrics including mean squared error, mean absolute percentage error, mean absolute error, maximum error, and the coefficient of determination. The results demonstrate high predictive accuracy for both models, with amorphous silicon technology showing superior performance, achieving a coefficient of determination of 98.6 % for Random Forest and 98.3 %t for Deep Neural Networks. The MAPEs for amorphous silicon were 8.2 % for Random Forest and 18.7 % for Deep Neural Networks. Monocrystalline silicon achieved 98.5 % and 98.0 % for the coefficient of determination, with MAPEs of 9.3 %t and 20.4 % for Random Forest and Deep Neural Networks, respectively. For polycrystalline silicon, the coefficients of determination were 98.3 % and 98.1 %, with MAPEs of 9.1 % and 24.1 %, respectively. These findings highlight the effectiveness of machine learning models for accurate PV energy forecasting and underline the potential advantages of amorphous silicon technology in semi-desert climates