In Situ Adaptive Battery Parameter Estimation Algorithm with Cross-Validation and Observer

Abstract

In this work, a reinforcement learning based algorithm has been developed for online estimation of battery state-of-health. This algorithm utilizes streaming battery data on a second-order equivalent circuit model of a battery cell, with charge reservoir model and fourth order polynomial approximation. The parameters of this model are adapted based on minimization of error between the estimated and measured states of a battery cell. In-situ parameter estimation considerably reduces the amount of information required for accurate state determination, since the algorithm periodically performs training, testing, and validation on the accumulated cell data. Besides, model complexity has been incorporated and online cross-validation has been performed to improve model accuracy. This state-estimation algorithm has been tested on Lithium-ion (Li-ion) cells under the discharge rates of 0.5C, 1C and 2C, at temperatures of 15◦C, 25◦C, and 35◦C.