UTILIZING MACHINE LEARNING ALGORITHMS TO IMPROVE THE ACCURACY OF SOLAR ENERGY YIELD PREDICTIONS BASED ON VARIOUS INPUT VARIABLES

This paper explores a framework for an ML model for solar yield estimation using random forest, gradient boosting, support vector machines, and neural networks There are various factors affecting solar energy production, some of which are weather conditions, solar intensity, temperature, among others, and characteristics of the panel. The trend for development and usage of renewable energy sources continues to gain momentum. The estimation of production rates for solar energy is crucial for the efficiency of energy generation, transmission, and supply. Hypothesis-driven and list-based models of traditional statistics are sometimes rather low in predictive validity because the relationships between these factors may be non-linear and highly interactive. The complex structure of these dependencies indicates that advanced ML algorithms are an appropriate approach to modeling the yield and increasing the utility of the resulting predictions. This work predictively analyzes historical solar generation data and climatic factors and preprocesses the data through normalization and feature engineering. The models are then adjusted and tested using a cross-validation approach in order to minimize the risk of obtaining high accuracy values only on the training data set. Evaluations of performance with mean absolute error, root mean square error, and coefficient of determination are used to measure the efficiency of each model. Machine learning algorithms widely outperform other models when applied to predict the yield of solar energy in the same or related areas. The random forests seem to perform better for prediction because they operate in an ensemble, which allows for modeling interactions between them. Through enhanced prediction precision, the approach proves to be a helpful resource for solar energy managers to attain better correlation between supply and demand, enhance solar energy production, and encourage the use of sustainable energy resources. Extension of this study in future work would involve the use of real-time data coupled with adaptive learning models to improve the prediction resilience.