Clustering-based Self-imputation of Unlabeled Fault Data in a Fleet of Photovoltaic Generation Systems

Abstract

This work proposes a fault detection and imputation scheme from unlabeled raw data for small-scale urban photovoltaic (PV) systems. Due to its environment-friendly and sustainable characteristics, PV power generation has been spotlighted as a new energy source, and accordingly, its portion of total power generation is increasing. However, direct exposure to the external environment causes frequent failures and it is important to monitor facility maintenance. In the case of small-scale local photovoltaic, there are limitations on the installation cost of environment monitoring sensor. Instead, as the solar radiation is similarly distributed in the same area, the correlation between power data of surrounding facilities is high. Accordingly, this study proposes a framework that detects and corrects anomalies using nearby photovoltaic data to supplement limited environmental data for small-scale PV installations. The proposed framework consists of two phases. In the first step, feature extraction was performed using local solar insolation data and surrounding photovoltaic power generation data, and then classified normal and abnormal state through K-means clustering. In the previous study, the detection of solar failure was mainly done by supervised learning, and the prediction accuracy was compared with the method of detecting local solar failure through prediction. Abnormally detected photovoltaic plant data should be replaced with reliable information for a variety of applications, including future generation and demand scheduling, optimization, and demand response etc. Therefore, in the second stage, the imputation was conducted based on the regression method using neighboring PV data as various explanatory data. When a profile was determined to have an abnormal pattern, imputation for the corresponding data was implemented to use the subset of neighboring PV data clustered as normal. The k-Nearest Neighbor (kNN) method, which was shown to be the best in the conventional solar data restoration, was also tested. For evaluations, 13 rooftop PV data on campus were used. Six abnormal patterns are characterized from observed PV data, and are injected for experimental tests. Correlation analysis shows that adjacent solar irradiance have a higher correlation than environmental data and individual solar power generation data. Based on the clustering-based detection of abnormal data, it was found that the use of adjacent solar power generation profiles improves the accuracy of fault detection and imputation compared to the use of solar radiation data collected over long distances. In addition, the clustering-based detection showed a 6.7% improvement in error rate compared to the prediction-based detection. The anomaly data imputation also confirmed that the correction using the nearby PV data is more accurate than solar irradiance-based kNN approach (scheme?). The results show the effectiveness of self-failure detection and self-imputation in local small-scale PV fleet systems.