iMOE: prediction of second-life battery degradation trajectory using interpretable mixture of experts

TL;DR

This paper introduces a model to predict battery degradation for second-life use, enabling safer and more efficient energy storage systems.

Contribution

The novel iMOE network predicts battery degradation using field data without needing historical cycling information.

Findings

iMOE achieves 0.95% MAPE in predicting battery degradation trajectories with 0.43 ms inference time.

The model outperforms existing methods by reducing computational time and error rates by 50% and 77%, respectively.

iMOE works effectively with pruned 5MB training data while maintaining prediction accuracy.

Abstract

Retired electric vehicle batteries offer immense potential to support energy infrastructure stability in underdeveloped regions through second-life use, but uncertainties in battery degradation behaviors pose major safety concerns. This work proposes an interpretable mixture of experts (iMOE) network that predicts battery degradation trajectories using partial, field-accessible signals in a single cycling operation. iMOE leverages an adaptive multi-degradation prediction module to classify battery degradation modes using expert weight synthesis learned from battery capacity-voltage and relaxation data. The module produces latent degradation trend embeddings, which are input to a use-dependent recurrent network for long-term degradation trajectory prediction. Validated on three typical use patterns (i.e. consistent operating histories, deeply aged batteries with unknown prior use, and uncertain second-life conditions, including 295 batteries, 93 use conditions, and 84,213 cycles), iMOE achieves an average mean absolute percentage errors (MAPE) of 0.95% with a 0.43 ms inference time for life-long battery degradation trajectory prediction. Compared to state-of-the-art Informer and PatchTST, it reduces computational time and MAPE by 50% and 77%, respectively. Compatible with data sampling in random state of charge regions, iMOE supports a 150-cycle time-horizon degradation trajectory prediction with 1.50% and 6.26% MAPE on average and at maximum, respectively. Notably, iMOE can operate effectively even with pruned 5MB training data while retaining 0.95% MAPE. Broadly, this network offers a deployable, history-free solution for battery degradation trajectory prediction at the time of second-life deployment, redefining how second-life energy storage systems are sensed, evaluated, controlled, and integrated for sustainable energy infrastructures at scale.

The work presents a model that predicts battery degradation using field-accessible data without historical cycling information. It combines a mixture of expert’s network with features from partial charging curves and relaxation voltage to classify degradation modes and predict long-term trajectories under uncertain future conditions.

Figures