Deep learning networks for capacity estimation for monitoring SOH of Li‐ion batteries for electric vehicles

Summary Data‐driven modeling using measurable battery signals tends to provide robust battery capacity estimation without delving deep into electrochemical phenomenon inside the battery. Nowadays, with the advent of artificial intelligence, deep neural networks are playing crucial role in data modeling and analysis. In this article, models of three different families of network architectures such as feed‐forward neural network (FNN), convolutional neural network (CNN), and long short‐term memory neural network (LSTM) are proposed for battery capacity estimation. Measurements from a set of two rechargeable Li‐ion batteries are considered for the model performance evaluation. The battery capacity estimation by different models has been evaluated by considering the effect of certain parameters such as model complexity, sampling rate of battery measurable signals and type of battery measurable signals. With its ability to process time‐series data efficiently by memorizing long‐term dependencies, LSTM outperforms other model architectures in estimating battery capacity more accurately and flexibly with 4.69% and 19.16% decline in average test root mean square error (RMSE) as compared with FNN and CNN, respectively. Simpler architectures of LSTM and FNN are able to perform well as compared with CNN, which needs architecture with certain hidden layers to interpret the battery aging process. Moreover, investigations reveal that sparsely sampled battery signals help all the proposed models to learn the battery dynamics in a better way as compared to densely sampled battery signals which also entails for less complex model learning process. Further, among all battery measurable signals, battery temperature has relatively less weightage in estimating battery capacity. Typical architectures of FNN, CNN, and LSTM with their respective input data formats. Where, Vi, Ii, Ci, and Ti are battery voltage, current, charge capacity, and temperature profiles with time sample points i ε [1,n].