Because of the extensive applications of lithium-ion batteries (LIBs) in electric vehicles (EVs), the battery management system (BMS) used to monitor the state and guarantee the operating safety of LIBs has been widely researched. The state of charge (SOC) is one of the most important states of LIBs that is monitored online. However, accurate SOC estimation is challenging because of erratic battery dynamics and SOC variation with current, temperature, operating conditions, etc. In this paper, a method combining a long short-term memory (LSTM) network with an adaptive cubature Kalman filter (ACKF) is proposed. The LSTM network is first utilized to learn the nonlinear relationship between the SOC and measurements, including current, voltage and temperature, and then, the ACKF is applied to smooth the outputs of the LSTM network, thus achieving accurate and stable SOC estimation. The proposed method can simplify the tedious procedure of tuning the parameters of the LSTM network, and it does not need to establish a battery model. Data collected from dynamic stress tests are used as training datasets, while data collected from US06 tests and federal urban driving schedules serve as test datasets to verify the generalization ability of the proposed method. Experimental results reveal that the proposed method can dramatically improve estimation accuracy compared with the solo LSTM method and the combined LSTM-CKF method, and it exhibits excellent generalization ability for different datasets and convergence ability to address initial errors. In particular, the root-mean-square error is less than 2.2%, and the maximum error is less than 4%.

由于锂离子电池（LIB）在电动汽车（EV）中的广泛应用，用于监视状态并保证LIB的运行安全性的电池管理系统（BMS）已得到广泛研究。充电状态（SOC）是可在线监视的LIB的最重要状态之一。但是，由于电池动力学不稳定以及SOC随电流，温度，工作条件等而变化，因此准确的SOC估计具有挑战性。在本文中，将长短期记忆（LSTM）网络与自适应库曼卡尔曼滤波器（ACKF）结合使用的方法）。首先利用LSTM网络了解SOC与测量值（包括电流，电压和温度）之间的非线性关系，然后将ACKF用于使LSTM网络的输出变得平滑，从而实现准确，稳定的SOC估算。所提出的方法可以简化繁琐的LSTM网络参数调整过程，并且不需要建立电池模型。从动态压力测试收集的数据用作训练数据集，而从US06测试和联邦城市驾驶计划收集的数据用作测试数据集，以验证所提出方法的泛化能力。实验结果表明，与单独的LSTM方法和组合的LSTM-CKF方法相比，该方法可以显着提高估计精度，并且对于不同的数据集具有出色的泛化能力，并且能够解决初始误差。特别是，均方根误差小于2.2％，最大误差小于4％。所提出的方法可以简化繁琐的LSTM网络参数调整过程，并且不需要建立电池模型。从动态压力测试收集的数据用作训练数据集，而从US06测试和联邦城市驾驶计划收集的数据用作测试数据集，以验证所提出方法的泛化能力。实验结果表明，与单独的LSTM方法和组合的LSTM-CKF方法相比，该方法可以显着提高估计精度，并且对于不同的数据集具有出色的泛化能力，并且能够解决初始误差。特别是，均方根误差小于2.2％，最大误差小于4％。所提出的方法可以简化繁琐的LSTM网络参数调整过程，并且不需要建立电池模型。从动态压力测试收集的数据用作训练数据集，而从US06测试和联邦城市驾驶计划收集的数据用作测试数据集，以验证所提出方法的泛化能力。实验结果表明，与单独的LSTM方法和组合的LSTM-CKF方法相比，该方法可以显着提高估计精度，并且对于不同的数据集具有出色的泛化能力，并且能够解决初始误差。特别是，均方根误差小于2.2％，最大误差小于4％。