5. 3 Operation, Performance and Maintenance of PV Systems
Summary / Abstract:
Underperformance assessment of solar photovoltaic (PV) plants is one of the most important and common studies performed by third-party independent consultants; it requires analysing large amounts of data in order to identify, classify and quantify time periods where the PV plant has failed to produce as much power as possible considering the environmental conditions recorded during such periods. Unfortunately, most tasks required to conduct an underperformance assessment are carried out manually, are time consuming, and highly dependent on the consultant’s expertise and knowledge. On the other hand, digitalisation of information and new technologies, such as Machine Learning (ML), are expected to drive the automation of many tasks that are currently performed by domain experts, and therefore help to improve efficiency and productivity. This paper presents a ML-aided procedure for the underperformance assessment of solar PV plants. The procedure’s goal is to identify, cluster and quantify anomalous periods of time as well as provide relevant information that can help the consultant assign a label to the found clusters of anomalies; furthermore, the procedure can be enhanced by including domain knowledge by means of rule-based criteria. The general procedure is based on the detection of anomalies in the input data and the generation of explainability information that can help to characterize them . Moreover, a data-driven model capable of estimating the expected power of the PV plant under normal conditions  is used as a baseline to assess underperformance for every anomalous data point. Afterwards, explainability information as well as residuals calculated using the expected power model are forwarded to a clustering algorithm that groups the identified anomalies according to their density . Additionally, a rule for identifying total plant unavailability has been introduced into the procedure, this rule will generate a new cluster and override the cluster assignment resulting from the algorithm. Finally, anomalies are grouped according to their assigned cluster and month of the year to produce an underperformance matrix. See Figure 1 for a depiction of the complete underperformance assessment scheme. Results show that the implemented procedure is capable of identifying different underperformance clusters and the use of an expected power model has helped to quantify the amount of energy that has not been produced due to each cluster. Furthermore, the data-driven approach means that no assumptions regarding the data used have been made and there are no restrictions with respect to the input data required. On the other hand, the use of rule-based criteria helps to identify underperformance operation modes that do not occur frequently enough to form a well-defined cluster from a data perspective; although presently only one rule-based criterion has been included in the procedure, this can be easily expanded as long as the criterion can be mathematically defined.