A Data-Driven Machine Learning Approach to Solar Energy Generation Forecasting Using Meteorological Variables
TL;DR
Air temperature, wind direction, wind speed, and relative humidity were the top four features affecting solar power generation identified from SHAP analysis, and the results show the potential of using machine learning techniques for accurate solar energy prediction, leading to enhanced grid management, renewable energy integration, and sustainable energy planning.
Abstract
The production of solar power is highly dependent on weather conditions, including temperature, humidity, wind velocity, wind direction, cloud coverage, and solar radiation. These relations must be well understood to develop photovoltaic systems with high efficiency and reliability. A solar power prediction model is proposed in this paper based on meteorological data with more than 30,000 samples gathered at La Trobe University, Victoria, Australia. The purpose is to determine the most significant atmospheric variables and to improve prediction performance by modeling linear and non-linear relationships between solar output and meteorological parameters. Three machine learning (ML) algorithms: Ridge Regression (RR), Elastic Net Regression (ELNET), and Kernel Ridge Regression (KRR) have been employed and compared based on Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and R-Squared evaluation measure. SHapley Additive exPlanations (SHAP) was utilized to explain model outcomes and the significance of the variables. Among all the exemplified approaches, kernel ridge regression provides the best forecasting results. Air temperature, wind direction, wind speed, and relative humidity were the top four important features affecting solar power generation identified from SHAP analysis. The results show the potential of using machine learning techniques for accurate solar energy prediction, leading to enhanced grid management, renewable energy integration, and sustainable energy planning.