贫数据条件下锂离子电池容量退化轨迹预测方法

Predicting capacity degradation trajectory for lithium-ion batteries under limited data conditions

在锂离子电池使用过程中,因实际运行条件的限制,通常难以获取大量完整标记的电池数据,对实现电池容量退化轨迹的准确预测构成了显著挑战。为此,本文提出了一种融合容量退化曲线增广和常用神经网络算法的锂离子电池容量退化轨迹预测方法。首先,基于少量完整标记的电池容量退化数据,采用多项式函数和蒙特卡洛方法得到虚拟容量退化曲线,并通过KL散度和欧氏距离进行筛选。之后,构建多层感知机(multi-layer perceptron,MLP)、卷积神经网络(convolutional neural network,CNN)、门控循环单元网络(gated recurrent unit,GRU)和长短期记忆网络(long short-term memory,LSTM)等四类常用神经网络模型,用以映射虚拟容量退化曲线数据至电池实际容量。最后,以虚拟容量退化曲线数据为输入,实际容量为输出,利用少量完整标记电池的数据对模型进行预训练,并利用待预测电池的早期退化数据进行微调,从而实现容量退化轨迹预测。通过77只具有不同放电方案的电池的数据对所提方法进行验证。结果表明,在仅有3只完整标记电池的容量退化数据条件下,所提方法的预测性能不受神经网络类型的影响,四类神经网络均准确预测了其余电池的容量退化轨迹,MAPE和RMSE的均值分别控制在2.3%和31 mAh以下。

In the use of lithium-ion batteries,real-world operating conditions often restrict the availability of extensive,fully labeled data.This limitation presents a significant challenge in accurately predicting battery capacity degradation.To address this issue,this study introduces an approach that combines a capacity degradation curve augmentation algorithm with traditional neural network techniques to predict the trajectory of battery capacity degradation.First,a small set of fully labeled battery capacity degradation data,polynomial functions,and the Monte Carlo method are used to generate virtual capacity degradation curves.These augmented curves are subsequently filtered using KL divergence and Euclidean distance metrics.Next,four widely used neural network models—Multilayer Perceptron,Convolutional Neural Network,Gated Recurrent Unit,and Long Short-Term Memory—are developed to map the virtual degradation curve data to the actual battery capacity.Finally,the models are pretrained with a small amount of fully labeled battery data using the virtual capacity degradation curve data as input and the actual capacity as output.They are then fine-tuned with early degradation data from the target battery to predict its capacity degradation trajectory.The effectiveness of the proposed method is validated using data from 77 lithium-ion batteries subjected to various discharge schemes.The results demonstrate that,even with just three fully labeled battery capacity degradation datasets,the prediction performance of the proposed method remains robust regardless of the type of neural network used.All four neural networks effectively predict the capacity degradation trajectories of the remaining batteries,achieving a mean MAPE and RMSE of less than 2.3%and 31 mAh,respectively.