Generation of Synthetic Battery Data with Capacity Variation

Abstract

In this article an approach is presented based on the use of measured experimental data from conventional battery packs to generate synthetic operational data for subsequent use in monitoring, predicting and controlling battery pack state of health. Generally speaking, experimentation-based synthetic data is effective in extensive simulation models possessing many varied operating conditions. The results presented in this article focus on proof of concept and are part of a comprehensive study into general capacity estimation and capacity fade prediction in battery packs. Experimental data is derived from scaled operational cycles with multiple charge and discharge pulses applied repetitively on a commercially available battery pack. The resulting synthetically generated data, using Markov chain approaches, has the flexibility of matching user-imposed conditions and can be of any length. Therefore, the focus in this article is the generation of sufficient training data for models built from machine learning techniques, utilizing only a relatively limited amount of actual data. In the context of the overall ongoing study, the behavior of the battery pack is characterized by features and a supervised learning approach is adopted in order to estimate capacity fade during real-time operation without the use of specific capacity tests.