Prediction of state of health and remaining useful life of lithium-ion battery using graph convolutional network with dual attention mechanisms
Prediction of state of health and remaining useful life of lithium-ion battery using graph convolutional network with dual attention mechanisms
复制标题
DOI：
10.1016/j.ress.2022.108947
复制
发表时间：
2022-11
期刊：
影响因子：
--
通讯作者：
Yupeng Wei;Dazhong Wu
中科院分区：
文献类型：
--
作者：
Yupeng Wei;Dazhong Wu
健康状态和剩余使用寿命的预测对锂离子电池的安全至关重要。现有的健康状态和剩余使用寿命预测方法不能有效地揭示特征之间的相关性。建立相关性可以帮助识别具有高相似性的特征，并将它们聚集在一起以提高预测模型的准确性。此外，现有的方法，如递归神经网络和长期短期记忆，在健康状态和剩余使用寿命预测方面存在局限性，因为它们无法使用时间序列数据中最相关的部分进行预测。为了解决这些问题，引入了一个两阶段优化模型来构造具有最优图熵的无向图。在该图的基础上，建立了具有不同注意力机制的图卷积网络，用于预测电池的健康状态和剩余使用寿命，其中注意力机制使神经网络能够利用时间序列数据中最相关的部分进行预测。实验结果表明，该方法能够准确地预测健康状态和剩余使用寿命，最小均方根误差分别为0.0104和5.8。该方法的性能也优于现有的数据驱动方法，如梯度提升决策树、长短期记忆和高斯过程。
Prediction of state-of-health and remaining useful life is crucial to the safety of lithium-ion batteries. Existing state-of-health and remaining useful life prediction methods are not effective in revealing the correlation among features. Establishing the correlation can help identify features with high similarities and aggregate them to improve the accuracy of predictive models. Moreover, existing methods, such as recurrent neural networks and long short-term memory, have limitations in state-of-health and remaining useful life predictions as they are not capable of using the most relevant part of time-series data to make predictions. To address these issues, a two-stage optimization model is introduced to construct an undirected graph with optimal graph entropy. Based on the graph, the graph convolutional networks with different attention mechanisms are developed to predict the state-of-health and remaining useful life of a battery, where the attention mechanisms enable the neural network to use the most relevant part of time series data to make predictions. Experimental results have shown that the proposed method can accurately predict the state-of-health and remaining useful life with a minimum root-mean-squared-error of 0.0104 and 5.80, respectively. The proposed method also outperforms existing data-driven methods, such as gradient-boosting decision trees, long short-term memory, and Gaussian process.