Robust Estimation of State of Charge in Lithium Iron Phosphate Cells Enabled by Online Parameter Estimation and Deep Neural Networks

Abstract

This paper addresses the state of charge estimation problem in lithium iron phosphate (LFP) battery cells. LFP cells are particularly challenging because their fat open circuit voltage (OCV) curve means OCV-based battery models are weakly observable. This means standard methods for SOC estimation don't easily converge to the true SOC. Additionally, in practice, estimates must be accurate in the face of biased noise on current input, as well as mean-zero noise on measurements. As such, we aim to create an estimator that is accurate when facing these types of noise. We accomplish this with a three-layer estimation technique that uses an adaptive Kalman filter, a Neural Network, and a Kalman Filter to estimate the state of charge. This method achieves an SOC estimation with an RMSE of 2.248%, even in the presence of a 0.2A current measurement bias and 5mA and 5mV random measurement noise. Notably, the proposed approach outperforms state-of-the-art methods like the extended Kalman filter.