Capturing lithium-ion battery dynamics with support vector machine-based battery model

• We include recent current history in a data-driven battery model.
• The model captures the voltage dynamics during pulse and relaxation.
• An accurate estimation of the diffusion-related voltage drop is achieved.

During long and high current pulses, diffusion resistance becomes important in lithium-ion batteries. In such diffusion-intense situations, a static support vector machine-based battery model relying on instantaneous current, state-of-charge (SOC), and temperature is not sufficient to capture the time-dependent voltage characteristics. In order to account for the diffusion-related voltage dynamics, we suggest therefore the inclusion of current history in the data-driven battery model by moving averages of the recent current. The voltage estimation performance of six different dynamic battery models with additional current history input is studied during relevant test scenarios. All current history models improve the time-dependent voltage drop estimation compared to the static model, manifesting the beneficial effect of the additional current history input during diffusion-intense situations. The best diffusion resistance estimation results are obtained for the two-step voltage estimation models that incorporate a reciprocal square root of time weighing function for the current of the previous 100 s or an exponential time function with a 20 s time constant (1–8% relative error). Those current history models even improve the overall voltage estimation performance during the studied test scenarios (under 0.25% root-mean-square percentage error).

Figure optionsDownload as PowerPoint slide