基于变分深度嵌入-带有梯度惩罚的生成对抗网络的锂离子电池老化特性建模

Ageing Performance Modeling of Li-Ion Batteries Based on Variational Deep Embedding-Wasserstein GAN with Gradient Penalty

锂离子电池老化实验是研究电池老化性能的基本手段,但针对大量电池的老化实验一般很耗时。为了节约时间和测试成本,获得更多电池数据,该文将变分深度嵌入(VaDE)模型与带有梯度惩罚的生成对抗网络(WGANGP)相结合,组成VaDE-WGANGP架构,进而基于该生成模型设计了一种电池老化特性建模与数据生成的方法。该文以一套开放的电池全寿命周期测试数据集为依据展开研究,首先,将电池放电过程中的电压、电流和放电容量这三个外特性作为模型的输入,通过VaDE的编码器将原始数据映射到隐空间,再通过优化获得符合特定规则的分布;然后,通过一定方式对该分布空间进行采样,并将采样所得的隐变量输入解码器中进行数据生成;后续数据测试表明,VaDE-WGANGP在电池外特性数据生成上具有较好的性能,可以实现对电池老化过程中基础外特性的模拟,在数据量不足时也可以为某些数据驱动算法提供有效的扩展数据资源。

Lithium-ion batteryaging tests are primarily used to study battery aging performance.However,aging experiments for a large number of batteries take time and effort.Therefore,combined with Wasserstein GAN and gradient penalty(WGANGP),the Variational deep embedding(VaDE)model is used to form a VaDE-WGANGP architecture.Then,a method of battery aging performance modelling and data generation is designed.This paper aims to generate simulated data onexternal characteristics(such as battery voltage and charge/discharge current during working).The VaDE-WGANGP generation model is effective during different battery SOH intervals,which indicates that simulated battery data can be obtained across the whole battery life span(SOH∈[100%,80%])and coversbattery aging performance.Furthermore,battery external characteristics differ for different battery cells even under the same SOH state.Accordingly,the diversities in battery external characteristics among different battery cells under the same SOH states are considered.An open-source battery test data set,in cludingdata from Toyota Research Institute(TRI)in partnership with MIT and Stanford and working data on LFP/graphite cells,is used.Firstly,the three external characteristics(voltage,current,and discharge capacity during the discharge process)are selected as the input of the model.The latent space by the VaDE encoder can be mapped,and its distribution with specific rules can be obtained through optimization.Then,this distribution space is sampled,and the latent variables obtained by sampling are used as input into the VaDE decoder for data generation.SOH estimation tests show that the VaDE-WGANGP has good generation performance for battery external characteristics,and the primary battery working performance is simulated during the aging process.Besides,effective and extended battery data resources can be provided by this measure for data-driven algorithms,especially in case the primitive data amount is insufficient.The following conclusions can be drawn.(1)Regarding battery data clustering and battery extrinsic characteristics simulation generation,the integrated model of VaDE-WGANGP has better performance than VAE and VaDE.The statistical characteristics of battery extrinsic characteristics in different SOH intervals are accurate.(2)VaDE-WGANGP can generate high-quality simulation data of battery external characteristics.The SOH estimation results prove that the data generated by VaDE-WGANGP is of high quality.This paper provides a novel idea for analyzing the external characteristics of batteries during aging.Given the shortage of original data in data-driven technology,this scheme can also provide an effective method for data extension and is helpful for battery screening.