Machine Learning Prediction of Thermal Losses in MonoPERC Solar Modules: A Novel Clustering Approach for Tropical Climate Applications

Abstract

Thermal losses significantly impact the efficiency of photovoltaic modules, particularly under high-temperature and variable cloud cover conditions in tropical climates. This study presents a novel thermal clustering methodology for predicting thermal losses in Monocrystalline Passivated Emitter and Rear Cell (MonoPERC) solar modules. Seven machine learning algorithms were tested using two methods, a baseline approach and a thermal clustering approach, which allow better energy yield forecasting and a more comprehensive understanding of the behavior of PERC modules. The clustering methodology partitions data into distinct thermal regimes, enabling specialized model training for different temperature operating conditions. K-Nearest Neighbors (KNN) was the best model without clustering, achieving a 0.9612 correlation and a mean prediction error of 7.3 W. With the new thermal clustering method, Multi-Layer Perceptron (MLP) was the top performer, with a 0.9561 correlation and an NMAE of 0.1409. Ensemble methods, such as XGBoost and Random Forest, were also highly effective, while linear methods proved inadequate. Results demonstrate that K-Nearest Neighbors achieved superior baseline performance, while the thermal clustering approach improved prediction accuracy across all algorithms. The Multi-Layer Perceptron emerged as the best performer with the clustering methodology.