(362j) Rnn and Transfer Learning for Battery Life Prediction of Electric Vehicles Based on Real-Road Driving Obd Data and Lab Measurements

Due to the strengthening of carbon dioxide emission standards, global electric vehicle (EV) sales in 2021 were 6.6 million units, more than three times higher than in 2019 (220,000 units), showing a dynamic growth. The global market for lithium-ion batteries is expected to increase by 27% annually (2018-2025), reaching $119 billion (about four times compared to 2018) by 2025. Lithium-ion batteries are evaluated as the best technology for storage batteries in that they have high output voltage, high energy density, low self-discharge, long lifespan, and high reliability. However, as the stability is less stable than other batteries, the capacity reduction during overdischarge is very large. In the case of EVs, it is very important to accurately predict the remaining life of an electric vehicle battery because if the capacity is reduced, it can lead to a serious accident such as stopping on the road.

RUL (Remaining Useful Life) prediction studies of lithium-ion batteries using machine learning such as linear regression, Bayesian regression, SVM, Relevance Vector Machine (RVM), Autoregressive Model (AM), and ADNN that integrates autoencoder and DNN were conducted. Since the EV battery test and driving data are managed by the developer themselves and not disclosed, the existing battery remaining life prediction research is mainly based on open DB sources (NASA, Oxford, Stanford, etc.). In the case of the model derived based on the open DB, that does not reflect actual road driving factors and aging factors (abrupt start, sudden stop, complete charging/discharging, etc.). Therefore, there is a problem in that it is not easy to accurately predict the remaining life of a battery installed in an actual EV.

In this study, the LSTM and GRU models, which have strengths in time series analysis among various methods of artificial neural networks, were used. The real-road driving data set collected through OBD was transferred to a pre-trained laboratory data model, and a model for predicting the remaining life of an EV battery was developed through retraining of the model. The actual driving data was collected from the OBD terminal for automatic collection of separate battery information and wireless transmission. The OBD terminal can store vehicle information such as EV cell voltage, temperature, pack voltage, pack current, CCC, CDC, CEC, SOC, and RPM etc..

In the case of LSTM and GRU, more data requirements are required than the parameters of the machine learning methods of previous studies. In addition, in the case of real road driving data, it is difficult to develop a generalized model because it is smaller than the amount of existing laboratory data. By developing a pre-trained model with laboratory data through transfer learning, and reusing the pre-trained deep learning model, the driving data with a small amount of data was supplemented. In addition, since the pre-trained model has already learned the weights, the weights are not newly learned from the beginning, reducing the amount of data required to update the parameters. In order to use it for transfer learning, it is necessary to learn the laboratory data and real road driving data with the same data.

It was confirmed with R2 that the performance of the model that predicted the remaining battery life through transfer learning was better than that of the model that learned only driving data. Through this, it is possible to secure battery performance measurement and management technology for EVs through the development of battery precision measurement technology and battery life estimation algorithm. In addition, it is possible to diagnose the condition of an EV battery through EV battery diagnosis and development of performance indicators, and it is expected that this will help reduce the battery failure rate and prevent accidents.