Thermal consistency-based early warning research for overcurrent faults in battery energy storage systems

Safe operation of large-scale battery energy storage systems (BESS) demands reliable incipient fault prediction. Conventional threshold-based alarms fail to capture weak early fault signatures, leading to delayed alerts and false positives. In this work, a physics-informed early warning framework centered on temperature consistency degradation for overcurrent faults is proposed. The fault evolution pathway is first established as follows: current stress leads to Joule heating, which further leads to temperature consistency degradation, thereby validating temperature consistency as the most sensitive and leading indicator. Guided by this mechanism, a 56-dimensional feature set was constructed using 16 months of real-world BESS operational data. A bidirectional long short-term memory (BiLSTM) model with temporal attention was then developed to dynamically focus on critical latent-phase evolution, enabling a fixed 48-h prediction horizon. Rigorous 5-fold cross-validation yields an event-level accuracy of 87.50%, a recall of 89.2%, and a precision of 87.4%. The indispensable role of temperature features was confirmed via ablation studies: removing these features led to a 22.86% drop in accuracy and a 3.5-fold increase in prediction variance. This interpretable, mechanism-data fused framework provides a practical intelligent solution for enhancing the safety and operational reliability of BESS.