Predicting Solar Radiation: A Fusion Approach with CatBoost and Random Forest Ensemble Enhance by Explainable AI

Accurate solar radiation estimation is crucial for effectively designing and optimizing solar energy systems. It serves a crucial function in determining the amount of energy that the sun contributes to a specific area. This radiation is crucial for generating energy in photovoltaic plants or solar thermal power plants. However, the occurrence of radiation fluctuates and is greatly influenced by climatic data, leading to an irregular pattern. The fluctuations in solar radiation result in variations in energy generation from these solar power systems. Due to the task’s complexity and the need to predict fluctuations in the performance of photovoltaic generation systems, it is critical to invest in the development of dependable tools that can accurately forecast and assess incident solar radiation. For this study, we have utilized a thorough approach to create a reliable and easily understandable predictive model. The study started by employing the SelectKBest feature selection method to identify the most influential variables from the provided dataset. Following the implementation of the feature selection technique, we proceeded to train a total of eight distinct machine learning models. We selected CatBoost and Random Forest as the top performers among the considered models for further development. In order to improve the system’s predictive capabilities, we developed an ensemble model by merging the two most successful individual models. We utilize a voting technique to combine the results of the CatBoost and Random Forest models, resulting in the ultimate ensemble prediction. The ensemble model gained an impressive accuracy rate of 98.52%, showcasing its exceptional predictive capabilities. We also utilized the Shapash library to enhance the interpretability of the ensemble model we developed. Through the implementation of a comprehensive approach, a robust, highly accurate, and interpretable ensemble model is successfully developed, capitalizing on the strengths of multiple machine learning algorithms.