A Transfer Learning-Based State of Health Estimation Framework for Supercapacitor Systems Utilizing Tram Field Data

This paper proposes a transfer learning-based state of health (SOH) estimation framework for the onboard super-capacitor systems utilizing the field data collected from a tram. The SOH labels of the supercapacitor systems are obtained from the selected discharging fragments of the field data. In total, 24 health features related to the power, time, current, and voltage of the supercapacitor systems are extracted from the selected charging fragments of the field data. The proposed framework builds on a light gradient boosting machine (LightGBM) model and incorporates two transfer learning strategies: a domain adaptation strategy based on transfer component analysis (TCA) and a fine tuning strategy based on dynamic learning rate adjustment. Specifically, TCA is employed to align the standardized features in the source and target domains while dynamic learning rate adjustment is implemented to accelerate the parameter tuning process. The proposed framework is evaluated using the field data of the four supercapacitor systems in a tram. On average, the SOH estimation errors are less than 1.8%, 1.5%, and 1.5% in terms of the root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE), respectively, which clearly demonstrate the effectiveness of the proposed framework.