Abstract: The operating characteristics of vehicle thermal management systems are jointly affected by driving conditions, temperature settings and occupant comfort. An accurate performance prediction is critical for system optimization and fault diagnosis. This paper employs dual-layer long short-term memory (LSTM) and temporal convolutional network (TCN) for temperature prediction in electric vehicle thermal management systems, analyzes the mechanism of the two algorithms,verifies key modules through ablation experiments, and discusses the influences of data length, time window and training strategies on prediction performance.Using root mean squared error (RMSE), mean squared error (MSE), mean absolute error (MAE), concordance correlation coefficient (CCC) and training time as evaluation metrics, the models are trained and validated based on real-vehicle coolingdown data. The predicted trends of facevent air temperature and breath temperature are compared with measured data. Results show that the dual-layer LSTM model achieves higher accuracy with a CCC value of 0.94 for facevent air temperature. The TCN model exhibits outstanding advantages in training and inference efficiency: its training time is only 2.2% of the LSTM model, and the prediction speed is nearly 15 times higher, while the accuracy still meets engineering requirements. Ablation experiments demonstrate that the dual-layer structure and Dropout enhance the performance of LSTM, and dilated convolution combined with residual connections constitutes the core of efficient TCN modeling. The 30 s time window, 1 650 s data length and optimized training strategies achieve the best model performance. This study provides valuable references for the application of dual-layer LSTM and TCN in vehicle thermal management systems.

摘要： 整车热管理系统运行特性受行驶工况、温度设定与乘员舒适性等多重因素耦合影响， 精准的性能预测对系统优化与故障诊断具有重要意义。文章将双层长短期记忆网络（LSTM） 与时间卷积网络（TCN）用于电动汽车热管理温度预测，对两种算法开展深度剖析，通过消 融试验验证关键模块有效性，并分析数据长度、时间窗口与训练策略对预测效果的作用。以 均方根误差（RMSE）、均方误差（MSE）、平均绝对误差（MAE）、一致性相关系数（CCC） 及训练时间为评价指标，基于实车降温数据完成模型训练与验证，对出风温度与呼吸点温度 趋势进行预测并与实测数据对比。结果表明：双层 LSTM 预测精度更优，出风口温度的 CCC 达 0.94；TCN 在训练与推理效率上优势显著，训练用时仅为 LSTM 的 2.2%，预测速度提升近 15 倍，精度满足工程应用。消融试验显示，双层结构与随机失活（Dropout）可提升 LSTM 性 能，空洞卷积与残差连接是 TCN 高效建模的核心；该研究采用的 30 s 时间窗口、1 650 s 数 据长度与优化训练策略可获得最佳模型表现。该研究可为 LSTM 与 TCN 在整车热管理性能预 测领域的应用提供参考。

关键词: 热管理；双层 LSTM 模型；TCN 模型；神经网络；性能预测