Deep Koopman Observer for Lithium-Ion Battery Temperature Estimation

Temperature monitoring is critical for lithium-ion battery (LIB) safety and performance, yet instrumenting every cell in a commercial pack remains impractical due to cost and wiring constraints. Existing sensorless methods rely on either physics-based thermal models requiring extensive parameterization or nonlinear recurrent estimators that cannot integrate sensor feedback when measurements become available. Motivated by these limitations, this paper proposes a Deep Koopman observer that enforces linear latent dynamics, enabling direct compatibility with Kalman filtering. The observer estimates surface temperature from four standard BMS signals and two exponential moving averages of squared current that capture thermal memory at distinct time scales, operating in two modes: fully sensorless for uninstrumented cells, or sensor-fused via a one-state EKF when a thermistor is available. Evaluated under strict cell-to-cell split across twelve drive cycles and five ambient temperatures, the open-loop observer achieves 17% lower error than the strongest reproduced CNN-LSTM baseline without online resistance identification or thermal-model simulation, and the EKF path delivers a further 35% reduction over the open-loop estimate. The evaluation is limited to a single cell chemistry and manufacturing batch; cross-chemistry and aging validation remain for future work.