Graph-Attention-Based Core-Temperature Estimation for Large-Format LFP Prismatic Batteries

Abstract Accurate core-temperature estimation is critical for thermal management of large-format lithium iron phosphate (LFP) batteries, yet direct internal sensing is impractical in field-deployed battery management systems. This study presents GAT-TE, a graph-attention-based spatiotemporal estimator that infers the geometric-center core temperature of 230 Ah LFP prismatic cells from externally measurable voltage, current, ambient-temperature, and six surface-temperature measurements. The reported errors are evaluated against thermocouples embedded inside the cells during laboratory cycling; these embedded measurements provide supervised reference targets only and are not used as model inputs during inference. GAT-TE organizes the surface measurement locations and the target core node as a directed surface-to-core graph, applies multi-head graph attention to learn non-uniform spatial contributions from different surface regions, and employs a delta-temperature prediction strategy that anchors the estimate to an observable surface-temperature baseline. Across two independent 230 Ah cells tested at three ambient temperatures under five dynamic operating profiles (DST, NN, UDDS, US06, and real-ship), the method achieves core-temperature MAEs of 0.071 ℃ and 0.082 ℃, with RMSEs of 0.092 ℃ and 0.109 ℃. Systematic ablation, sparse-sensor, and noise-sensitivity analyses confirm that accurate core-temperature estimation depends jointly on temporal encoding, graph-based spatial aggregation, and measurement quality.