为实时掌握隧道大体积混凝土内部温度变化，及时制定合理的控温措施，防止隧道大体积混凝土在浇筑过程中由于大范围的水化热反应而产生温度裂缝，基于长短期记忆神经网络LSTM（long short-term memory）和门控循环单元神经网络GRU（gate recurrent unit）2种深度学习算法的多参数、非线性拟合能力，提出隧道大体积混凝土内部温度变化的滚动预测方法，并依托衢州市智慧新城三江中路连通隧道工程建设中采用的双光栅传感器现场混凝土温度实测数据，采用平均绝对误差MAE（mean absolute error）和决定系数（R2）对2类模型的预测结果精度进行检验评价。结果表明： 2种网络模型均能捕捉隧道大体积混凝土内部温度发展规律，准确预测隧道大体积混凝土内部温度变化曲线，且GRU的精度优于LSTM。其中，GRU的MAE为1.34 ℃，比LSTM减小1.07 ℃，同时GRU的R2为0.98，也优于LSTM的R2（0.9）。

The internal temperature change in mass tunnel concrete can be investigated in real time by formulating timely and reasonable temperature control measures that can prevent the temperature cracks caused by the largescale hydration heat reaction during the pouring process of mass tunnel concrete. A rolling prediction method is proposed for the internal temperature change of mass tunnel concrete based on the multiparameter and nonlinear fitting ability of two deep learning algorithms, namely, the long shortterm memory (LSTM) and the gate recurrent unit (GRU). Relying on the onsite concrete temperature measurement data of the doublegrating sensor used to construct the Sanjiang Middle road connecting tunnel project in Quzhou Smart New City, the mean absolute error (MAE) and the 〖WT５《TNR#I》〗R〖WT５《TNR》〗squared (〖WT５《TNR#I》〗R〖WT５《TNR》〗2) values are used herein to test and evaluate the prediction result accuracies of the two models. The results show that the two network models can capture the internal temperature development law of mass tunnel concrete and accurately predict its internal temperature change curve. Furthermore, the GRU accuracy is better than that of LSTM. The MAE of GRU is 1.34 ℃, which is 1.07 ℃ lower than that of LSTM, and the 〖WT５《TNR#I》〗R〖WT５《TNR》〗2 of GRU is 0.98, which is better than that of LSTM (i.e., 0.9).