Self-supervised pre-training in photovoltaic systems via supervisory control and data acquisition data

Abstract

Owing to the availability of sensor data, the operation and maintenance (O&M) of sustainable energy systems have become more intelligent. In particular, data-driven approaches have gained growing interest in supporting intelligent O&M. However, this is not a simple task, as the deficiency of labelled data poses a major challenge. This work proposes a self-supervised pre-training approach for autonomous learning of the Supervisory Control and Data Acquisition (SCADA) data representations for photovoltaic (PV) systems. Specifically, the proposed method first constructs the sample pairs using reasonable assumptions from a large volume of unlabelled SCADA data. Then, it designs a deep Siamese network to extract the representations of the input sample pair and sets the pretext task to measure whether the input pair is similar. The proposed method has been deployed in a PV system with nominal power 2.5 MW located in North China. Experimental results show that the proposed approach achieves accurate similarity assessment for the sample pairs and can potentially support downstream tasks regarding intelligent O&M.