Diagnosis of abnormal voltage in individual cells and fault trend prediction in energy storage power stations

Abstract
During the operation of energy storage power stations, individual cells are prone to voltage inconsistency and abnormal voltage fluctuations due to manufacturing variations, inconsistent aging degrees, and changes in operating conditions. To address this issue, this study proposes an individual cells voltage anomaly detection and fault trend prediction method by integrating K-Means-OPTICS clustering, Transformer-based feature extraction, and LSTM-Seq2Seq time-series prediction. First, K-Means and OPTICS clustering algorithms are employed to detect abnormal operating patterns and identify potentially abnormal individual cells. Subsequently, clustering outputs, voltage statistical features, and Shannon entropy features are fed into the Transformer model to achieve abnormal individual cells localization and fault type diagnosis. Finally, the LSTM-Seq2Seq model is utilized to predict future cell voltage difference and minimum cell voltage trends, and early warnings are triggered according to predefined engineering thresholds. Experimental results show that when the voltage of an individual cells becomes abnormal, the Shannon entropy of individual cells No.90 rapidly increases around time point 110, reaching a peak value of 2.5, which is significantly higher than that of normal individual cells. This demonstrates the effectiveness of the proposed model in detecting abnormal voltage fluctuations. Regarding fault prediction, the proposed model accurately tracks the variations of cell voltage difference and minimum cell voltage, providing early warnings of inconsistency faults 140 s in advance and individual cells undervoltage faults 90 s in advance. The proposed method enables accurate localization of abnormal individual cells and prediction of future fault trends, providing an effective safety margin for battery management systems in energy storage power stations.