Core Temperature Estimation and Thermal Management of Lithium-Ion Batteries Using a Nonlinear Adaptive Extended Kalman Filter

The core temperature of a cylindrical lithium-ion cell cannot be measured directly, yet it governs both safety and ageing. This paper presents a nonlinear adaptive extended Kalman filter (AEKF) that estimates core temperature from a single surface thermistor, together with a battery thermal management strategy driven by the resulting estimate. Three features distinguish the formulation from earlier two-node observers. First, heat generation is evaluated at the estimated core temperature rather than treated as an exogenous input; because internal resistance follows an Arrhenius law, this makes the process model genuinely nonlinear in the state and requires an explicit Jacobian, derived here in closed form. The state-feedback term reaches 25.4 % of the dominant conduction term at peak current, and substituting the measured surface temperature into the heat-generation expression incurs an error of up to 1.35 W. Second, the lumped two-node parameters are identified against a radially resolved finite-volume reference model using an excitation containing coolant-flow steps; omitting those steps yields a surface capacitance an order of magnitude too small and an observer that mispredicts the core whenever the pump switches. Third, innovation-based covariance adaptation is made safe for closed-loop use by a Student-t test on the innovation mean that distinguishes sensor degradation from transient model error, and by bounded directional process-noise inflation applied only to the measured state. Validation is by simulation only. Against a structurally different reference plant with deliberate parameter mismatch and a mid-run doubling of thermistor noise, the proposed filter attains a core-temperature RMSE of 0.348 degrees Celsius, against 0.359 for a fixed-gain EKF, 0.429 for a Sage-Husa AEKF and 0.447 for the linear formulation in which heat generation is an input. Under twenty Monte-Carlo realisations with random parameter error the proposed filter and the fixed-gain EKF are statistically indistinguishable. In closed loop, estimate-driven cooling holds the same peak core temperature as a conventional surface-triggered controller while consuming 47.5 % less pump energy.