Training and Evaluation for Yield-Driven Detection of Losses in PV Systems Utilizing Artificial Neural Networks

Monitoring data of photovoltaic (PV) systems offer a great opportunity for the application of machine learning approaches. Currently, among these approaches, artificial neural networks have become very popular. In this paper, the application of such models will be further investigated and recommendations for data quality and training processes for the specificity of PV system data will be given. In a comprehensive study based on a real PV system, the effect of different input parameters (environmental parameters), insufficient data availability and sensor coverage in the field on the model accuracy was investigated to detect yield losses. Furthermore, multiple training parameters were varied to understand the influence on training quality. The findings were applied using a threshold approach to show deviations in PV yield.