Exploring Battery Operational Trends in Hybrid-Electric Buses with Interpretable Machine Learning

Abstract

While the advent of Lithium-ion batteries in transportation has provided significant improvements to the performance of vehicle energy storage systems, their long-term performance in application remains hard to predict. The growing accessibility of machine learning tools provides a promising route for developing data-driven approaches to battery prognostics. These approaches can leverage both physicochemical knowledge of degradation mechanisms in individual cells and operational trends for the full energy storage system. One withstanding drawback to developing these models is the current sparsity in open-source battery data spanning long timescales and in real-world applications. Introduced here is an operational Lithium-ion battery (LiFePO4, A123) dataset from a fleet of 174 hybrid-electric buses (BAE Systems HybriDrive), collected during maintenance visits occurring between 2015 and 2019. The dataset contains information relating to the performance of the ESS, like voltage and temperature, but is unique in that it does not contain time-series data typically used to make inferences about the internal state. The dataset was explored using simple visualizations and interpretable machine learning methods to provide insights on battery usage and aging in a real-world transit application. Importantly, this work resulted in the extraction of strong discriminative features for determining declines in battery module performance, that are strongly correlated with time and maintenance replacements.