Modelling of Li-Ion battery state-of-health with Gaussian processes

Abstract

The problem of lithium-ion cells, which degrade in time on their own and while used, causes a significant decrease in total capacity and an increase in inner resistance. So, it is important to have a way to predict and simulate the remaining usability of batteries. The process and description of cell degradation are very complex and depend on various variables. Classical methods are based, on the one hand, on fitting a somewhat arbitrary parametric function to laboratory data and, on the other hand, on electrochemical modelling of the physics of degradation. Alternative solutions are machine learning ones or nonparametric ones like support-vector machines or the Gaussian process (GP), which we used in this case. Besides using the GP, our approach is based on current knowledge of how to use non-parametric approaches for modeling the electrochemical state of batteries. It also uses two different ways of dealing with GP problems, like maximum likelihood type II (ML-II) methods and the Monte Carlo Markov Chain (MCMC) sampling.

Efficient Gaussian Process Calculations Using Chebyshev Nodes and Fast Fourier Transform

“…In practice, they are useful, for instance, in industrial anomaly detection [29], and crucial for ensuring production quality and continuity. GPs predict battery state parameters [30], which are relevant in electric vehicles and energy management [4]. In bike-sharing systems, they optimize bike allocation, enhancing efficiency [31].…”