Forecasting battery state of health: Predicting cell degradation using early ageing data only

Retired electric vehicle batteries can potentially be given a second life in energy storage systems before being recycled. Such use realises and enables the transition to a low-carbon, circular economy. However, reliable use of second life batteries depends on accurate forecasting of their state of health and little data on the specific use-case is available. This study considers the little-studied problem of extrapolating forecasts into battery ages not covered by data. I used the open NASA Randomized battery usage dataset to evaluate four models that could plausibly be applied in real-world operations: a linear regression model and three Gaussian process models of varying complexity. The cells in the dataset are cycled in a manner similar to how energy storage systems are used. The simplest Gaussian process model uses a constant kernel, the most complex a search procedure to construct the kernel and find what number of principal components to use. The models train on the first half of battery ageing data, then predict the second half. The more complex models have a best-case normalised root mean square error of 1.5% and are, apart from a set of outlying performances, more robust, yet all models have similar median normalised root mean square errors and the simpler models would be easier to realise in real-world operations.