PV module power prediction by deep learning on electroluminescence images - Assessing the physics learned by a convolutional neural network

Abstract

Fast and reliable performance monitoring of photovoltaic modules is essential for economic forecasting in large-scale installations. Deep Learning methods, such as convolutional neural networks, have the potential to predict module power directly from electroluminescence images in an automated workflow. However, neural networks must be trained using large numbers of electroluminescence images of defective modules. Due to budgetary or technical limitations, these training images will always be biased, limiting generalization. Here, we demonstrate a transparent method to discriminate the information learned by a convolutional neural network into generally valid physics and bias. Learning of physics is assessed by providing an artificial, unbiased feature list, which is converted into synthetic electroluminescence images. Using these images, we compare the predictions of the neural network trained on the biased dataset to those of a physics-based equivalent. Bias is assessed by a closer look at the deviations between the predictions from the equivalent circuit model and the trained neural network. The assessment of physics knowledge incorporated by a Deep Learning method gives insight into how the method achieves its predictive capacity.