摘要
Abstract
Degradation data of battery capacity can be used to predict the battery remaining using life(RUL),but there exist numerous noise data in the battery degradation process caused by factors such as ambient temperature,charge/discharge process and capacity recovery phenomenon.It makes prediction of data-driven life lithium-ion battery challenging.To improve the prediction accuracy and generalization ability of batteries RUL,we proposed a method based on transfer learning and denoising autoencoder-long short term memory(DAE-LSTM).
Firstly,the variational autoencoder-generative adversarial network(VAE-GAN)method was constructed designed.Encoding network was used to estimate the distribution of input data,and generating network and discriminant network were used for data regeneration.It improved the reliability of generated data by VAE method,and solved the problem that GAN method had been difficult to train.Secondly,the DAE-LSTM method was constructed for data denoising and capacity prediction.The DAE can reconstruct the input data and its encoder improved the robustness of the method by adding Gaussian noise.LSTM layer can analyze the temporal characteristics of data for capacity prediction.Due to the small amount of data,the network layer of the overall method was less to avoid overfitting.To reduce the parameters,the same loss function was used in both data denoising and capacity prediction.Finally,the optimal training scheme was determined through the different experiment:The data generated by VAE-GAN was used for method pre-training,then all network layers of the basic method were fine-tuned by actual data.This would effectively improve the prediction accuracy of the method,and ensure the reliability of the prediction results.
Experimental results showed that the proposed method has better predictive performance,and degradation trend of most batteries can be well predicted.MAE and RMSE were controlled within 2.46%and 3.76%respectively,and the lowest was 0.95%and 1.06%.Experimental results with different prediction starting points showed that the prediction were more accurate when the prediction starting point was closer to the failure threshold.This indicates that the method can accurately predict the degradation trend in later stages of battery life.Experimental results with other datasets showed that the proposed method has strong adaptability and generalization ability.It can effectively predict the lithium-ion battery RUL in small data samples.The 90%confidence interval of the prediction results with NASA dataset is narrow,indicating that the method has strong robustness.In addition,we counted the time taken to complete the RUL prediction for different datasets.As the RUL prediction of batteries is offline prediction with low real-time requirement,the training and testing time of the method meets the offline prediction requirement.
The following conclusions can be drawn from the simulation results:(1)The DAE-LSTM method can effectively denoising the degradation data of lithium-ion batteries,and making the prediction result more accurate.(2)VAE-GAN method can generate multiple groups of degradation data conforming to the real degradation to achieve the purpose of data enhancement.(3)Transfer Learning can ensure that the effective information of generated data and real data is fully utilized,so that the prediction model has higher accuracy and better generalization ability.By comparing the prediction results of other literatures,it is proved the proposed method has higher Pre and can be used to predict the RUL of lithium-ion batteries.关键词
锂离子电池/剩余寿命预测/降噪/自编码器/长短时记忆神经网络/迁移学习Key words
Li-ion battery/remaining useful life/denoising autoencoder/long short term memory/transfer learning分类
信息技术与安全科学
