Conditional monitoring in photovoltaic systems by semi-supervised machine learning

With the rapid increase in photovoltaic energy production, there is a need for smart condition monitoring systems ensuring maximum throughput. Complex methods such as drone inspections are costly and labor intensive; hence, condition monitoring by utilizing sensor data is attractive. In order to recognize meaningful patterns from the sensor data, there is a need for expressive machine learning models. However, supervised machine learning, e.g., regression models, suffer from the cumbersome process of annotating data. By utilizing a recent state-of-the-art semi-supervised machine learning based on probabilistic modeling, we were able to perform condition monitoring in a photovoltaic system with high accuracy and only a small fraction of annotated data. The modeling approach utilizes all the unsupervised data by jointly learning a low-dimensional feature representation and a classification model in an end-to-end fashion. By analysis of the feature representation, new internal condition monitoring states can be detected, proving a practical way of updating the model for better monitoring. We present (i) an analysis that compares the proposed model to corresponding purely supervised approaches, (ii) a study on the semi-supervised capabilities of the model, and (iii) an experiment in which we simulated a real-life condition monitoring system.