Search documents

Browse topics

Document details

Fault Detection in Operation and Maintenance of PV Systems
A. Louwen, D. Miorandi, C. Torrero, F. Venturini, D. Moser
Monitoring, PV System, Fault Detection, Machine Learning, Neural Networks
PV Systems – Modelling, Design, Operation and Performance
Subtopic: Operation, Performance and Maintenance of PV Systems
Event: 38th European Photovoltaic Solar Energy Conference and Exhibition
Session: 5CO.9.2
967 - 972
ISBN: 3-936338-78-7
Paper DOI: 10.4229/EUPVSEC20212021-5CO.9.2
0,00 EUR
Document(s): paper, presentation


Concurrently with the strong decline in installed PV system prices, the margin for operation and maintenance (O&M) operators has decreased, as O&M costs are commonly budgeted as a fixed percentage of installed system costs. Hence, there is a need to improve the efficiency of O&M operations. In the PV4.0 project, we are aiming to improve O&M efficiency across the value chain, by implementing Industry 4.0 concepts. One key part of these concepts is to implement an automatic fault detection system, that is used to quickly detect faults in PV plant operation, minimizing electricity generation losses. Furthermore, this fault detection system is to be used to automatically generate maintenance tickets, so corrective maintenance can be performed as soon as possible. In this paper, we present the development of two neural network (NN) based machine learning models. Model A is a four-layer NN, trained to classify individual datapoints as either normal or faulty. Detected faults are subsequently classified as either a hard (zero power output) or soft (power output higher than zero but lower than nominal) failures. Model B is a model that attempts to predict whether the number of failures in the next days exceed a certain threshold. Our results indicate that Model A has a fault detection rate of 79% in our testing dataset, while the performance of Model B is lower, at a fault detection rate of 49% for 1-day-ahead predictions, and only 30% for 2-day-ahead predictions. Further work will focus on improving the failure detection rate for both models. For Model B, we will additionally focus on improving the input dataset, so that better predictions might become possible. Model A will be implemented in a real-time O&M environment to test its performance in real operating conditions.