基于模态分解的Transformer-GRU联合电池健康状态估计

Battery health state estimation of combined Transformer-GRU based on modal decomposition

针对锂电池使用过程中因松弛效应导致健康状态(state of health,SOH)呈现非稳定退化并影响SOH预测准确性的问题,提出一种基于变分模态分解(variational mode decomposition,VMD)与粒子群优化(particle swarm optimization,PSO)的变换神经网络(Transformer)和门控循环单元(gate recurrent unit,GRU)的联合方法。首先将锂电池容量信息通过变分模态分解算法分解,为避免分解程度不合理影响预测能力,使用中心频率法判断分解状态作为原数据信息有效解释依据;然后使用粒子群优化算法优化调整后的变换神经网络和门控循环单元结构的超参数,变换神经网络采用线性层代替解码器(decoder)更好适用时序数据,保留编码器(encoder)捕获数据全局特征及内部相关性,提升了单个Transformer及其联合模型预测精度;最后由Transformer和GRU分别对主趋势子序列和高频子序列预测,并将两种模型的预测进行融合以完成对锂离子电池SOH的估算。利用NASA锂电池数据集验证了模型的预测效果,并通过与多层感知机(multi-layer perception,MLP)、循环神经网络(recurrent neural network,RNN)等单一模型和高斯函数-GRU、Transformer-MLP等联合模型进行对比。结果表明本文预测模型无论在精度还是再生现象的拟合程度都优于其他单个模型或者联合模型,预测结果的平均绝对误差和均方根误差维持在0.62%和1.19%以内,决定系数在87.08%之上,验证了所提研究方法的有效性。

The electric vehicle is the inevitable development tendency for the vehicle industry.In addition,lithium-ion batteries act as the energy storage unit of vehicles,which is worthy of being investigated carefully since its life span is related to the accurate estimation of parameters.Unfortunately,the precise estimation of lithium-ion battery parameters,such as state of charge(SOC),state of health(SOH),and so on,is so tough.Therefore,multiple factors will strongly influence the accuracy of conventional estimation methods,especially for estimating non-stable SOH degradation caused by relaxation effects when using lithium-ion batteries.Hence,conducting research on SOH with a new method is significant.In this paper,the research status of common estimation methods for lithium-ion batteries are reviewed first,and the corresponding advantages and disadvantages are analyzed subsequently.Then,the variational mode decomposition(VMD)theory is introduced,including constructing the variational problem,reconstructing the unconstrained variational problem,and multiple iterations.Subsequently,the Transformer model and recurrent neural network(RNN)are brought in,which are useful for disposing of the time-order information.Finally,the particle swarm optimization(PSO)algorithm is described,which is helpful in completing the optimization of neural network super parameters.According to the content mentioned above,a joint method may be proposed based on VMD and PSO for transformer neural networks and gate recurrent units(GRU),improving estimation accuracy.First,VMD technology is used to decompose the capacity information of lithium-ion batteries.The central frequency method is used to determine the decomposition state,which serves as a basis for valid interpretation of the original data information to avoid unreasonable decomposition affecting the prediction ability.Second,the particle swarm optimization algorithm is used to optimize the hyperparameters of the transformed neural network and the gated recurrent unit structure after adjustment.The transformed neural network uses linear layers instead of decoders for better application with time-series data.It retains the encoder to capture global features and internal correlations of the data,thereby improving the prediction accuracy of individual Transformers and their joint models.Finally,the Transformer and GRU are used to predict the main trend and the high-frequency subsequences,respectively.Once the results are merged,the accurate estimation of lithium batteries is accomplished.Simultaneously,the NASA database verifies the model's reliability,and models of multi-layer perception(MLP),RNN,and Gaussian function-GRU are compared with the model proposed in this study.Consequently,the results of predicting accuracy and fitting degree of regeneration phenomenon are better than those of single or joint models,and MAE and RMSE are less than 0.62%and 1.19%,respectively.In addition,the determination coefficient is bigger than 87.08%,indicating the validity of the combined Transformer-GRU model.