Lithium-ion batteries have been widely used as energy storage systems and in electric vehicles due to their desirable balance of both energy and power densities as well as continual falling price. Accurate estimation of the state-of-charge (SOC) of a battery pack is important in managing the health and safety of battery packs. This paper proposes a compact radial basis function (RBF) neural model to estimate the state-of-charge (SOC) of lithium battery packs. Firstly, a suitable input set strongly correlated with the package SOC is identified from directly measured voltage, current, and temperature signals by a fast recursive algorithm (FRA). Secondly, a RBF neural model for battery pack SOC estimation is constructed using the FRA strategy to prune redundant hidden layer neurons. Then, the particle swarm optimization (PSO) algorithm is used to optimize the kernel parameters. Finally, a conventional RBF neural network model, an improved RBF neural model using the two stage method, and a least squares support vector machine (LSSVM) model are also used to estimate the battery SOC as a comparative study. Simulation results show that generalization error of SOC estimation using the novel RBF neural network model is less than half of that using other methods. Furthermore, the model training time is much less than the LSSVM method and the improved RBF neural model using the two-stage method.

中文翻译：

---

基于改进的RBF神经网络的锂离子电池组充电状态估计

锂离子电池由于其在能量和功率密度方面的理想平衡以及持续下降的价格而被广泛地用作能量存储系统和电动汽车。准确估计电池组的充电状态（SOC）对于管理电池组的健康和安全性很重要。本文提出了一种紧凑的径向基函数（RBF）神经模型来估计锂电池组的充电状态（SOC）。首先，通过快速递归算法（FRA）从直接测量的电压，电流和温度信号中确定与封装SOC高度相关的合适输入集。其次，使用FRA策略构建用于电池组SOC估计的RBF神经模型，以修剪冗余的隐藏层神经元。然后，粒子群算法（PSO）用于优化内核参数。最后，作为比较研究，还使用常规的RBF神经网络模型，使用两阶段方法的改进的RBF神经模型以及最小二乘支持向量机（LSSVM）模型来估计电池SOC。仿真结果表明，使用新型RBF神经网络模型进行的SOC估计的广义误差小于使用其他方法的误差的一半。此外，模型训练时间比LSSVM方法和使用两阶段方法的改进RBF神经模型要少得多。作为比较研究，还使用最小二乘支持向量机（LSSVM）模型来估计电池SOC。仿真结果表明，使用新型RBF神经网络模型进行的SOC估计的泛化误差小于使用其他方法的泛化误差的一半。此外，模型训练时间比LSSVM方法和使用两阶段方法的改进RBF神经模型要少得多。作为比较研究，还使用最小二乘支持向量机（LSSVM）模型来估计电池SOC。仿真结果表明，使用新型RBF神经网络模型进行的SOC估计的广义误差小于使用其他方法的误差的一半。此外，模型训练时间比LSSVM方法和使用两阶段方法的改进RBF神经模型要少得多。