粒子滤波和GRU神经网络融合的锂电池RUL预测

RUL prediction of lithium battery based on particle filter and GRU neural network fusion

随着锂电池在移动电子设备和电动汽车等领域中得到广泛应用,其剩余使用寿命(remaining useful life,RUL)的精确预测对锂电池的健康管理更具重要意义。本文提出一种基于粒子滤波(particle filter,PF)和门控循环单元(gated recurrent unit,GRU)神经网络融合(PF-GRU)的预测方法预测锂电池的RUL。这种融合方法结合了PF在估计RUL概率分布上的优势以及GRU能够进行时间序列长期预测的能力,获得融合的预测结果。再利用每个预测周期的容量预测结果,采用带移动窗口迭代更新训练数据集的方法对GRU模型进行再训练,提高了GRU的长期预测性能。以上预测步骤迭代进行,直到容量衰减至寿命阈值以下。最后将粒子代表的预测结果外推至寿命阈值,得到电池RUL分布直方图。本文采用美国NASA卓越诊断学中心(PCoE)实验室所提供的锂电池数据对所提方法进行验证,将所提出的融合方法与GRU、PF和无窗口移动融合方法进行RUL预测比较。实验结果表明,本文所提出的融合方法具有良好的性能,在状态和参数估计、RUL预测方面明显优于PF和无窗口移动融合方法,预测精度均高于其他3种预测方法。

As lithium-ion batteries are widely used in mobile electronics and electric vehicles,accurate prediction of their remaining useful life(RUL)is of great importance for health management of lithium batteries.A prediction method based on particle filter(PF)and gated recurrent unit(GRU)neural network fusion method(PF-GRU)is proposed to predict the RUL of lithium batteries.This fusion method combines the advantages of PF in estimating the probability distribution of RUL and the ability of GRU to perform long-term prediction of time series to obtain the fusion prediction result.Using the capacity prediction results of each prediction period,the GRU model is retrained by iteratively updating the training dataset with a sliding window,which improves long-term prediction performance of GRU.The above prediction steps are iteratively performed until the capacity decays below the lifetime threshold.Finally,prediction results represented by particles are extrapolated to the lifetime threshold,and RUL distribution histogram of the battery is obtained.The lithium battery data provided by the NASA prognostics center of excellence(PCoE)laboratory is adopted to verify the proposed method,which compares the proposed fusion method with GRU,PF and fusion method without sliding window for RUL prediction.Experimental results show that the proposed fusion method has good performance,which is obviously better than PF and fusion method without sliding window in terms of state and parameter estimation and RUL prediction accuracy.