Predictive maintenance based on anomaly detection in photovoltaic system using SCADA data and machine learning

Efforts to reduce the increase in average global warming by utilizing renewable energy continue to increase. One of the efforts is to build a solar power plant that is able to generate electricity by converting solar energy into electricity through a photovoltaic (PV) system. The existence of abundant, free, and year-round solar energy makes solar power plant promising in replacing fossil fuels used in thermal power plant. This paper aims to propose a predictive maintenance approach for PV systems using anomaly detection and fault diagnosis. In this study, the daily patterns of irradiance and corresponding AC output from a newly completed solar PV farm are investigated. Given the unlabelled nature of the data, traditional supervised learning methods are unsuitable for anomaly detection in this context. To address this, a long short-term memory autoencoder (LSTM-AE) model is employed to detect anomalies in the time series data. The LSTM-AE model is trained to reconstruct normal operation patterns, and deviations from these reconstructions are flagged as potential anomalies. This approach enables us to identify irregularities in the plant's performance that could indicate system faults or inefficiencies, ultimately providing valuable insights into the maintenance and optimization of solar PV operations. The results show that the anomaly prediction can achieve reasonable accuracy with minimum test errors MSE, RMSE, and MAE of 10.95, 3.30, and 2.76, respectively, and can be applied to the PV system.