Flexible, fast and transparent approaches to battery capacity forecasting

Summary:

Battery degradation is a complex phenomenon, leading many to use data-driven techniques to predict future battery health [1]. Linear models are transparent, easy to implement and not computationally demanding. However, simple linear models are often combined with Kalman filters and particle filters when modelling battery degradation [1].
The need for the filtering techniques suggested that a fixed linear model for capacity fade requires flexibility. We proposed to introduce that flexibility by combining automating feature selection and then using piecewise-linear modelling.
The piecewise splitting was performed by finding the peaks in the second derivative of the best correlating feature. By using the feature which correlates with capacity fade, and not time, the process is capable of finding different stages of ageing while being flexible to variable ageing rates.
The piecewise process was compared against a Gaussian process regression (GPR) tool. The data used in this trial was sourced from cells which underwent strenuous lifetimes and all displayed a knee point [2][3]. Full cell profiles were forecasted for each test cell using both techniques.
In trials using over 4,000 predicted capacity profiles randomly drawn from 147 cells available from refs. [2][3], the piecewise model produced very similar RMSE capacity (median 0.97%) to that using GPR (0.95%). That similarity was reflected in the end of life prediction: the piecewise model returned a median end of life error of 1.1% while the GPR approach produced 1.2%. The piecewise approach was a significant improvement on GPR for the 95th percentile of performance.
The process was also found to be robust to reducing the availability of training cells, input features and maximum allowed models in smaller trials with 2,000 predictions.
The simple feature selection and piecewise-linear approaches presented here, when combined, were found to perform comparably with GPR, but with a vastly reduced computational, mathematical and storage requirement.

References:
[1] Yi et al., “Data-driven health … a review,” Renewable and Sustainable Energy Reviews, 113, 2019.
[2] Severson et al., “Data-driven prediction … before capacity degradation,” Nature Energy, 4, 383-391, 2019.
[3] Attia et al., “Closed-loop optimization … machine learning,” Nature, 578, 397-402, 2020.