A deep learning and Krawtchouk polynomials-based approach for fault classification in photovoltaic panels

Abstract

Solar farms play a key role in the global renewable energy transition. They are multi-megawatt installations and they require significant capital investment, making peak operational efficiency essential for a positive return on investment, manual or semi-manual methods for diagnosing faulty solar panels are time consuming and impractical nowadays since solar farms include thousands of solar panels, especially with emerging technologies such as drones and fast computer vision architectures which can process thousands of images in seconds. In this paper, we used Krawtchouk polynomials to extract relevant features such as physical or electrical damage from solar panels. These polynomials’ parameters can be freely chosen to control the number of extracted features and the exact region to extract the most features from, using horizontal and vertical pixel coordinates. Additionally, with the help of an optimization algorithm, we managed to concentrate the feature extraction only on the most important regions containing details such as damage, dirt, or snow. These features were then passed to a convolutional neural network architecture for classification. We trained our model with a weighted categorical cross-entropy loss function and Matthews correlation coefficient (MCC) metric to handle class imbalance. We used multiple metrics to evaluate the proposed method, including F1-score, recall, and precision. The proposed method achieved a MCC score of 99.40%, 99.08% precision, 98.76% recall, and 98.88% F1 score.