Predicting properties of storage batteries using GPR-based machine learning methods

Yu, Tong

Permalink

https://hdl.handle.net/2142/121381 Copy

Description

Title

Predicting properties of storage batteries using GPR-based machine learning methods

Author(s)

Yu, Tong

Issue Date

2023-07-21

Director of Research (if dissertation) or Advisor (if thesis)

Tartakovsky, Alexandre Miron

Department of Study

Civil & Environmental Eng

Discipline

Civil Engineering

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• Storage Battery
• Machine-learning

Abstract

Accurately predicting the performance and properties of batteries, such as voltages at different states of charge (SOCs) within one cycle and discharge capacities at different cycles, is critical for planning projects that require the application of the batteries. The existing high-fidelity physics-based models are computationally expensive. Also, due to the complexity of the involved physics, certain processes (e.g., battery degradation), cannot be accurately predicted by these models. To overcome these challenges, two machine-learning methods based on the Gaussian process regression (GPR) model are introduced in this thesis. The first model is a multifidelity model to predict the charge-discharge curve within one charge-discharge cycle. In this model, the physics-informed CoKriging (CoPhIK) machine learning method is trained on experimental data collected at the Pacific Northwest National Laboratory (PNNL) for vanadium redox flow batteries (VRFBs). The physics in this model is constrained by the VRFB zero-dimensional physics-based model (0D model). Our results show that a small amount of experimental data is needed for the range of parameters in the 0D model, including current density, flow rate, and initial concentrations, to train the model. To accurately predict the charge-discharge curve, only one initial measurement of voltage in the unknown charge-discharge curve is needed. The model is tested to be robust since the predictions based on this model show good agreement with the experimental results. The second model is an ensemble Gaussian process regression (ensemble GPR) model to predict discharge capacities at different cycles and the cycle lives for commercial lithium iron phosphate/graphite cells under fast-charging conditions before their degradation. In this model, the properties (mean and covariance) of the prior distribution of discharge capacities are calculated from the measured discharge capacities of many batteries. Then, the discharge capacity of the modeled battery is predicted based on the discharge capacity of this battery at the first $N$ cycles as the GPR conditional mean. Uncertainty in the prediction is given by the conditional variance. Our results show that the ensemble GPR model is capable of predicting capacity decay and the life of the battery, i.e., the number of cycles after which the battery capacity drops below the critical value. The significance of the GPR method as a basis for the derived machine learning models is shown for the physics-informed CoKriging model with physics embedded in or in the purely data-driven ensemble GPR model. Both GPR-based models demonstrate great performance in predicting the properties of storage batteries.

Graduation Semester

2023-08

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/121381

Copyright and License Information

Copyright 2023 Tong Yu

Owning Collections