Data-driven battery state of health diagnostics and prognostics

Thesis

Lithium-ion batteries are increasingly ubiquitous in modern society but the degradation of lithium-ion cells is complex and challenging to predict. Data-driven approaches to estimating and forecasting the state of health of lithium-ion batteries have become increasingly popular in literature due to the growing availability of battery data and the improved flexibility of data-driven approaches relative to physics-based modelling. This thesis begins with a review of health diagnosis and prognosis approaches in battery literature, including a brief introduction to the principles and challenges of using data-driven approaches. The subsequent chapters investigate several of the remaining open questions.

An automated methodology for input feature generation and selection is proposed and thoroughly tested. Gaussian process regression, a well-known non-parametric form of supervised learning, is used to map from the input features to changes in capacity. The inputs are found to be good predictors of degradation and also robust to significantly increased noise and reduced sampling frequencies for the raw cycle data.

Gaussian process regression produces accurate predictions of capacity here. Piecewise, Bayesian linear regression is a faster, more transparent alternative that produces equally accurate end-of-life forecasts. A battery-focussed performance metric is proposed to assess the accuracy of the output probabilistic predictive distributions of both regression models. A proposed adaptation to sparse Gaussian process regression is found to reduce the storage requirements of a Gaussian process battery health model by 98\% without significantly impacting predictive performance.

Lithium-ion cells suffer from cell-to-cell variability in ageing rates, thereby posing a challenge to the control of battery packs and experimental design. An experiment finds that around 12 similarly used cells are required to consistently fit a population-level distribution for an empirical model. By contrast, a Gaussian process regression model requires under 5 cells in a training set to consistently fit the hyperparameters, but around 50 cells are required to produce the capacity predictive performance of other models in this thesis.

Authors: Greenbank, S

• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: statistics; health; forecasting; lithium-ion; battery; machine learning