An auto-regressive model for battery voltage prediction

Abstract

Accurate modelling of the dynamic behaviour of Lithium-ion (Li-ion) batteries is important in a wide range of scenarios from the determination of appropriate battery-pack size, to battery balancing and state estimation in battery management systems. The prevailing methods used in voltage prediction are the equivalent electrical circuit (EEC) models. EEC models account for the change in the voltage by a series of resistor capacitor networks to mimic the internal resistance of a battery. Thus, given a change in current the EEC models create an appropriate change in the voltage. The downside is that the parameters of the model needs to be fully characterised, across the entire range of usage and life of the battery. This is both time consuming and expensive. In this paper, a linear auto-regressive (AR) process is proposed to account for the short-term dynamic behaviour of the battery cell, allowing for accurate prediction of the voltage given other measurable parameters such as current and temperature. After conducting a sensitivity analysis on the size of the sequence needed to train the AR model, it was found that less than a days worth of raw measurements data is enough to offer a better voltage prediction than a traditional EEC model (the root mean square errors of the two considered voltage estimation approaches were 0.00157 and 0.0133 V, respectively).