Lithium-ion Battery Capacity Loss Predictions using Domain-Knowledge-Incorporated Capacity Loss Model through Bayesian Inference

Abstract

Lithium-ion batteries (LiBs) are being widely used in applications ranging from mobility to energy storage devices. Therefore, an efficient battery management system (BMS) is required for efficient use of LiB/LiB packs. Physics-informed reduced-order models (PI-ROMs) are being widely used in comparison to computationally expensive full electrochemical models to predict the capacity loss of LiBs and thereby their state of health (SoH), essential for an efficient battery management system. The capacity loss of LiB is often modeled using a standalone or a combination of PI-ROMs. However, statistical validation is sparsely conducted to evaluate the model's interpolation and extrapolation capabilities. This article demonstrates a novel approach of using a Bayesian model (BM) to statistically validate a domain-knowledge-incorporated capacity loss model. The capacity loss model is based on a modified Arrhenius equation incorporating key parameters such as state of charge (SoC), temperature, throughput (time), and discharge rate. For this work, the Python Battery Mathematical Modelling (PyBaMM) framework was employed to generate synthetic capacity loss data for different initial SoCs, temperatures, throughputs (time), and discharge rates to simulate the real-world operating conditions of LiBs in LIB packs. The Bayesian model using synthetic PyBaMM data validates the capacity loss model for the prognostic and diagnostic of LiB. BM predicted the capacity loss of the testing data with RMSE ~7% and R2 ~1. The novelty of the work lies in using trained BM to inversely predict one or more operating parameters of LiB against a target capacity loss, which could be essential for a battery management system. The authors believe that results have substantial implications for the use of BM validated domain-knowledge-incorporated capacity loss model in BMS.