锂离子电池的剩余使用寿命（RUL）是电池健康管理的重要参数。电池在实际使用过程中会出现容量再生现象，而且在电池数据采集过程中，通常难以避免噪声干扰，影响数据质量。针对以上问题提出一种基于Transformer结合金字塔卷积网络的电池RUL预测模型，选取容量作为健康因子，利用金字塔卷积网络中不同大小的卷积核提取容量序列的特征信息，利用Transformer中的多头注意力机制进一步学习序列的时序特征。采用加权Huber损失函数，提高模型的鲁棒性；采用Dropout技术，提高模型的泛化能力，防止训练过程中出现过拟合。将所提预测模型在NASA和CALCE数据集上实验，并与其他模型比较。实验结果表明，所提模型的预测精度更高，在NASA和CALCE数据集上的相对误差分别为0.008 6、0.019 3；平均绝对误差分别为0.011 5、0.012 6；均方根误差分别为0.017 3、0.018 9。

The remaining useful life （RUL） of lithium-ion batteries is an important parameter for battery health management. In the actual use process of batteries， the phenomenon of capacity regeneration will occur， and it is difficult to avoid noise interference in the process of battery data acquisition， which affects the quality of data. A battery RUL prediction model based on Transformer combined with pyramid convolutional（PyConv） network is proposed to address the above issues. Capacity is chosen as the health indicator， and feature information of the capacity sequence is extracted by the pyramid convolutional network with convolutional kernels of different sizes. The multi-head attention mechanism in Transformer is used to further learn the temporal features of the sequence. The weighted Huber loss function is used to improve the robustness of the model； Dropout technology is used to improve generalization ability of the model and prevent overfitting during training. The proposed prediction model is tested on the NASA and CALCE datasets and compared with other models. The experimental results show that the proposed model has higher prediction accuracy， on the NASA and CALCE datasets， the relative errors are 0.008 6、0.019 3 respectively； the mean absolute errors are 0.011 5 and 0.012 6 respectively； and root mean square errors are 0.017 3 and 0.018 9 respectively.