An Improved Bidirectional Gated Recurrent Unit Method for Accurate State-of-Charge Estimation

Abstract

State-of-Charge (SOC) estimation of lithium-ion batteries have a great significance for ensuring the safety and reliability of battery management systems in electrical vehicle. Deep learning method can hierarchically extract complex feature information from input data by building deep neural networks (DNNs) with multi-layer nonlinear transformations. With the development of graphic processing unit, the training speed of the network is faster than before, and it has been proved to be an effective data-driven method to estimate SOC. In order to further explore the potential of DNNs in SOC estimation, take battery measurements like voltage, current and temperature directly as input and SOC as output, an improved method using the Nesterov Accelerated Gradient (NAG) algorithm based Bidirectional Gated Recurrent Unit (Bi-GRU) network is put forward in this paper. Notably, to address the oscillation problem existing in the traditional gradient descent algorithm, NAG is used to optimize the Bi-GRU. The gradient update direction is corrected by considering the gradient influence of the historical and the current moment, combined with the estimated location of the parameters at the next moment. Compared to state-of-the-art estimation methods, the proposed method enables to capture battery temporal information in both forward and backward directions and get independent context information. Finally, two well-recognized lithium-ion batteries datasets from University of Maryland and McMaster University are applied to verify the validity of the research. Compared with the previous methods, the experimental results demonstrate that the proposed NAG based Bi-GRU method for SOC estimation can improve the precision of the prediction at various ambient temperature.