基于注意力机制和多任务LSTM的锂电池容量预测方法

Capacity Prediction of Lithium-ion Batteries Based on Multi-task LSTM with Attention Mechanism

精确的锂电池容量预测可以有效降低电池失效带来的风险与损失。基于神经网络的时间序列预测模型是电池容量预测领域中十分常见的方法。但是,过去的模型预测大多只考虑了未来的目标点,而没有考虑目标点附近信息对预测目标的辅助作用。提出一种基于注意力机制和多任务LSTM的锂电池容量预测方法(MT-LSTM),实现未来不同时刻信息的互补,提高预测的准确性,其中使用硬参数共享方法建立多个任务之间的联系,使用卷积神经网络提取不同抽象水平的特征。通过注意力机制与LSTM模型的结合,有效地提高电池容量预测精度。将所提出的MT-LSTM模型与传统神经网络进行对比,试验结果表明所提模型有更高的预测精度。此外为多任务学习与注意力机制设计了对比试验,验证了这两种方法对电车容量预测精度的积极影响。

Accurate capacity prediction of lithium-ion batteries can effectively reduce the risk and loss caused by battery failure.The time series prediction model based on neural network is a very common method in the field of battery capacity prediction.However,most of the predictions of the past models only considered the future target points,but did not consider the auxiliary role of the information near the target points.A capacity prediction method of lithium-ion batteries based on multi-task LSTM with attention mechanism is proposed to realize the complementation of information at different times in the future and improve the prediction accuracy.The hard parameter sharing method is used to establish the connection among multiple tasks,and the convolutional neural network is used to extract features at different levels of abstraction.Comparing with the traditional neural network,and the experimental results show that the proposed MT-LSTM model has higher prediction accuracy.In addition,comparison experiments are designed for the multi-task learning and the attention mechanism to verify the positive effects of these two methods on the prediction accuracy of battery capacity.