EU PVSEC Programme Online
EU PVSEC 2021, 6 - 10 September 2021
Presentation: 5CV.3.35 A Machine Learning-Based Anomaly Detection System for Solar Inverters
Type: Visual
Date: Wednesday, 9th September 2020
15:15 - 16:45
Author(s): P. Mercade Ruiz, G. Guerra, L. Landberg
Presenter / Speaker: P. Mercade Ruiz, Greenpowermonitor, Barcelona, Spain
Event: Conference Conference
Session: 5CV.3 Operation, Performance and Maintenance of PV Systems
Type(s) of Access:  Conference Registration
Topic: 5. 3 Operation, Performance and Maintenance of PV Systems
Summary / Abstract: Optimal maintenance planning and scheduling can greatly reduce the levelized cost of energy (LCoE) from solar PV. Predictive maintenance is seen to be more effective than standard preventive strategies at reducing downtime and maintenance costs. Therefore, recent efforts in the Operation and Maintenance (O&M) community have been put into developing predictive maintenance models capable of anticipating the failure of the most critical PV components. Machine Learning-based (ML) models have been a common choice as they can adapt to the various states of operation of PV components. Previous work in [1] and [2] have shown the predictive capabilities of these models for PV inverters, when appropriate amounts of good quality data are available. However, despite the promising results, the implementation of the proposed predictive maintenance models requires detailed maintenance logs, which unfortunately, are not always properly kept. This paper presents an ML-based anomaly detection system that needs no maintenance logs for its implementation and is meant to complement predictive maintenance to help better plan maintenance and reconstruct maintenance logs. The system aims at identifying instantaneous suspicious behaviour of PV inverters using historical data, which, in this work, have been SCADA data collected by GreenPowerMonitor. For this purpose, suspicious behaviour is considered without reference to normal behaviour of PV inverters but instead arises when the current data differ from the previously seen data to the extent that they appear to have been generated by different mechanisms [3]. Normal behaviour is usually implemented into anomaly detection systems using a set of rules that enforce specific relationships among the data channels. However, most of these relationships are based on physical laws that would hold under ideal conditions, but they may not under real-life monitoring conditions and therefore, would conflict with the anomaly detection system on a regular basis. As a result, in the present approach, anomalies are detected without consideration of the normal behaviour when a change of operation happens.