Automatic Fault Detection and Diagnosis In Photovoltaic Plants Based on Convolutional Neural Networks

This work aims to discover efficient and reliable machine-learning architectures and models for highly automated fault detection, diagnosis and predictive maintenance using the real-time flow of operational data generated at utility-scale photovoltaic plants. Our models consists of Convolutional Neural Networks which are trained on datasets which have been assembled from several years of operational data from about a dozen PV plants located in Europe and Latin America, with an approximate overall generation power of 1 GW, and different types of modules, tracking and terrain. We do minimal data preprocessing to construct our models. We use both 1-D and 2-D convolutional neural networks. For 2-D convolutional neural networks, we transform the time series into images by using Markov Field Transform, Recurrence plots, and Gramian Angular Fields. The performance of our models so far does not show significant improvements over other systems published in the literature, but it must be taken into consideration that the use of real operational data entails a significant challenge, when compared to simulated environments and small plant scenarios. We are improving the architectures and algorithms for higher reliability, a wider catalog of scenarios, and concurrent faults. We are also working on knowledge transfer among PV plants.