Solar energy has seen an extraordinary growth over the last few years and along with this success expectations of improvements, efficiencies, and a general reduction of the levelized cost of energy (LCoE) are becoming more and more pronounced. At the same time, the extensive digitalization of modern solar photovoltaic (PV) plants has dramatically increased the amount of data that operators have at their disposal; this newly found source of operational data opens the door to a new generation of digital tools that can help to produce added value and reduce costs. One of the tools that is expected to have a major impact on the operation and maintenance (O&M) of PV plants is predictive maintenance.
Under current practices, maintenance is conducted following two major approaches: preventive and corrective. The former relies on guidelines put forward by component manufacturers to define the frequency and type of tasks performed, whereas the latter focuses on unexpected events that prompt immediate maintenance and repair interventions to minimize revenue loss. Unlike preventive and corrective maintenance, a predictive maintenance policy uses monitoring data and artificial intelligence to create a baseline model of plant operation under normal conditions and provide an informed estimate of possible future failure events. Thanks to its failure prediction capabilities, predictive maintenance can help to anticipate and optimize the required preventive and corrective actions, which will result in an increase in the plant’s availability and performance.
Motivated by their unique industry position and multidisciplinary expertise, GreenPowerMonitor, a DNV company have developed a predictive maintenance system for solar inverters that given a set of input data channels, including environmental variables, uses machine learning models to represent an inverter’s normal operation and multivariate analysis to identify anomalous behavior within new streaming data. The developed system has been tested on selected sites where high-quality maintenance logs were available and demonstrated it can detect up to 83% of failures with 75% of True alarms a week ahead of inverter failure. These results are proof that it is possible to predict inverter failures using the developed methodology; however, they required having detailed and high-quality maintenance logs, something that the majority of systems lacks and has prevented further development and validation of the method.
As a solution to the lack of detailed maintenance logs, DNV and GreenPowerMonitor have placed their focus on a more general anomaly detection approach which can be exploited when no maintenance logs for validation of the predictive maintenance system are available. The anomaly detection system relies on the concepts of density and reconstruction error to identify anomalies, and it is complemented with the use of explainability techniques to determine the potential cause for the anomalies. The system has been tested using historical data for multiple inverters and results show that the implemented approach can identify sustained anomalies and determine the data channels responsible for causing the anomalies. Anomaly detection can also play an important role as a support tool for a predictive maintenance system but should not be seen as a substitute.
The added value of predictive maintenance for the PV industry starts with the prediction of a component failure; however, this piece of information can enable new benefits that range from economic gain by preventing plant downtime to better management and planning of staff assignments, spare parts, and replacements, as well as, achieving a better balance between covered and uncovered corrective maintenance activities in contracted services. These potential benefits and added value represent an opportunity that the industry cannot afford to miss, but the absence of detailed maintenance logs and standardization is hindering its development and expansion within the solar industry.
For DNV and GreenPowerMonitor the path towards the future is clear: Owners, operators, and manufacturers must come together and establish procedures and services that enforce the proper and standardized documentation of all relevant events that occur in a PV plant. Moreover, they must exploit the digitalization tools at their disposal in order to facilitate the upkeep and automated access to O&M logs for data analysis routines. However, the industry cannot sit idly by until log availability and standardization are no longer an issue; the deployment of anomaly detection systems can help to close the gap between the current situation and the next generation of predictive maintenance systems. Anomaly detection systems can act as decision-support tools and the alarms generated by them can trigger mandatory validation procedures for plant operators, which will become the building blocks for the much-needed detailed O&M logs. Additionally, operators can make use of the ticketing features included in their monitoring software tools to create and maintain O&M logs without requiring significant changes to the software or data infrastructure.
Finally, the development of a holistic predictive maintenance strategy will require inputs and expertise from many different actors involved in the PV industry, from manufacturers to developers, operators, and researchers. For this reason, it is crucial to encourage collaboration in research and development activities between all sectors in order to deliver the solutions that meet the industry’s needs.
Would you like to read more about it?
Request a copy of our position paper here