This paper presents an ECM identification approach for dynamic power profiles of AI training workloads. In AI data centers, clusters of GPUs generate large, rapidly fluctuating load currents. Li ion batteries act as buffers to minimize load fluctuations on the grid side for power quality and backup. Reliable operation then demands accurate capture of battery dynamics to estimate SOC. The proposed method employs electrochemical impedance spectroscopy under varying DC bias levels to parameterize the ECM, explicitly accounting for impedance variations caused by load current magnitude. The model is validated using dynamic current profiles that emulate low frequency fluctuations and transient ripples in AI training. Experimental results show that the model tracks battery terminal voltage with high precision, achieving RMSE ≤ 5.4 mV for 1 Hz and 10 Hz loads. Built on the proposed ECM, an extended Kalman filter based SOC estimator achieves SOC RMSE < 1.24 % under high current profiles that emulate AI training, confirming suitability for real time battery management in AI data centers.
