In a major boost for solar energy planning, a new study has identified an artificial intelligence (AI) model that can predict the sun's seasonal strength with improved accuracy across India's most extreme climates. Researchers from Vellore Institute of Technology, Chennai, found that the Gaussian Process Regression (GPR) machine learning model is the most reliable tool for forecasting seasonal Global Horizontal Irradiance (GHI), an important parameter for assessing the available solar resource for Stand-Alone Photovoltaic (SAPV) systems, like solar power plants.
The GPR model demonstrated exceptional precision, achieving a Root Mean Square Error (RMSE) of just 0.0030 and a Coefficient of Determination of 0.9999, which represents a reduction in error of up to 189.1% compared to other tested models. This work, using data from four climatically distinct locations—Chennai, Jaisalmer, Leh, and Mawsynram—promises to make solar power generation more stable and cost-efficient across the subcontinent.
The core purpose of this research is to enhance the precision of long-term solar resource evaluation and planning across diverse industries by analysing historical climatic fluctuations and their influence on solar irradiance. GHI is the total solar radiation incident on a surface that is perfectly horizontal to the ground. Accurately predicting the GHI seasonal average is essential for estimating the power a solar panel can absorb and for optimising SAPV system planning. SAPV systems, which are not connected to a central power grid, rely entirely on accurate forecasting to meet energy demands throughout the year.
Did you know? The Köppen system is a climate classification method developed by Wladimir Köppen that divides the world's climates into five main groups based on temperature and precipitation: A (tropical), B (dry), C (temperate), D (continental), and E (polar). |
To conduct their investigation, the researchers focused on four locations that represent India's extreme climatic diversity. They were classified using the Köppen system, a climate classification method based on temperature and precipitation. The regions were Chennai, a coastal region; Jaisalmer, a hot desert; Leh, a cold, high-altitude desert; and Mawsynram, a wet mountain region known for its heavy rainfall. They gathered three years of hourly meteorological data from 2017 to 2019 from the National Solar Radiation Database (NSRDB). This raw dataset included 15 key meteorological parameters, such as temperature, relative humidity, and various forms of irradiance.
Three Machine Learning models were selected for comparison: Efficient Linear Regression (ELR), Regression Trees (RT), and Gaussian Process Regression (GPR). ELR is an improved version of traditional linear regression that focuses on capturing relationships between parameters. RT is a decision tree-based approach that excels at capturing non-linear relationships and managing complex datasets. GPR, the eventual winner, is a non-parametric technique that represents data as a distribution over functions, making it highly effective in modelling uncertainty and complex patterns. The models were trained on 2017-2018 data and tested on 2019 data, a temporal split that ensures the models are validated on unseen time periods.
The results confirmed GPR's exceptional performance. The GPR model consistently delivered the lowest prediction errors across all four locations and three seasons (rainy, summer, and winter). While the RT model provided reasonably accurate forecasts, the ELR model was the least reliable, often underestimating peak values and failing to capture daily fluctuations effectively. Diagnostic validation of the GPR model confirmed that its high performance stems from its ability to capture complex and non-linear patterns without overfitting or data leakage.
By leveraging machine learning techniques, the new study can uncover complex patterns in GHI and climatic data that conventional statistical or physical models often overlook. The use of seasonality-based feature selection improves GHI prediction, enabling ML models to enhance SAPV system efficiency by improving forecast accuracy. The research provides a validated tool for solar energy developers and planners. By accurately forecasting seasonal GHI, the GPR model enables SAPV systems to be planned and managed for maximum efficiency and reliability, even in India's most challenging climates.
This article was written with the help of generative AI and edited by an editor at Research Matters.
