New categorized machine learning models for daily solar irradiation estimation in southern Morocco's, Zagora city

Abstract

Accurately estimating daily solar irradiation is essential for effectively sizing and simulating solar energy systems. Inaccuracies or discontinuities in solar data can lead to errors in system assessments, potentially resulting in misguided conclusions about their economic feasibility. This study evaluates the performance of various machine learning techniques, including Multilayer Perceptron (MLP), Support Vector Machine Regression (SVMR), and Random Forest (RF), in accurately estimating daily Global and Diffuse Horizontal Irradiations (GHI and DHI), and Direct Normal Irradiation (DNI) in Zagora City, Southern Morocco. The models are categorized based on four classes of data inputs: Temperature-Based Models (GTBMs), Relative Humidity-Based Models (RHBMs) for GHI, Hybrid Parameter-Based Models (HBMs) for GHI, DHI, and DNI, and Clearness Index-Based Models (CIBMs) for DHI and DNI. The results indicate that RF is particularly effective in estimating daily GHI, DHI, and DNI using Hybrid parameters Based Models, achieving validation nRMSEs equal to 8.36 %, 18.55 %, and 10.26 %, respectively. It also performs well in estimating DHI with Clearness index Based Models (nRMSE=20.44 %). The next effective method is MLP, that is the optimal one in estimating daily GHI with Relative Humidity Based Models (nRMSE=11.41 %) and daily DNI using Clearness index Based Models (nRMSE= 12.43 %). Lastly, SVMR is more quitable for estimating daily GHI with Temperature Based Models, with a validation nRMSE of about 9.09 %. The study further demonstrates that empirical models underperform compared to machine learning methods in all included categories.

Keywords

• Solar radiation
• Machine learning models
• Hybrid parameters based models (HBMS)
• Clearness index based models (CIBMS)
• Relative humidity based models (RHBMs)
• Temperature based models (GTBMs)