State of Health Forecasting of Heterogeneous Lithium-ion Battery Types and Operation Enabled by Transfer Learning

Due to the global transition to electromobility and the associated increased use of high-performance batteries, research is increasingly focused on estimating and forecasting the state of health (SOH) of lithium-ion batteries. Several data-intensive and well-performing methods for SOH forecasting have been introduced. However, these approaches are only reliable for new battery types, e.g., with a new cell chemistry, if a sufficient amount of training data is given, which is rarely the case. A promising approach is to transfer an established model of another battery type to the new battery type, using only a small amount of data of the new battery type. Such methods in machine learning are known as transfer learning. The usefulness and applicability of transfer learning and its underlying methods have been very successfully demonstrated in various fields, such as computer vision and natural language processing. Heterogeneity in battery systems, such as differences in rated capacity, cell cathode materials, as well as applied stress from use, necessitate transfer learning concepts for data-based battery SOH forecasting models. Hereby, the general electrochemical behavior of lithium-ion batteries, as a major common characteristic, supposedly provides an excellent starting point for a transfer learning approach for SOH forecasting models. In this paper, we present a transfer learning approach for SOH forecasting models using a multilayer perceptron (MLP). We apply and evaluate it on the method presented by von Bülow, Mentz, and Meisen (2021) using five battery datasets. In this regard, we investigate the optimal conditions and settings for the development of transfer learning with respect to suitable data from the target domain, as well as hyperparameters such as learning rate and frozen layers. We show that for the transfer of a SOH forecasting model to a new battery type it is more beneficial to have data of few old batteries, compared to data of many new batteries, especially in the case of superlinear degradation with knee points. Contrarily to computer vision freezing no layers is preferable in 95% of the experimental scenarios.