In North China, especially in Hebei Province, the low temperature and temperature difference in winter pose challenges to the construction of concrete structure of water conservancy projects. Timely monitoring of concrete internal temperature and effective control of temperature cracks are crucial to ensure the safety and stability of the structure. At present, the temperature monitoring of the concrete structure mainly relies on the embedded temperature sensor, but its high cost limits its wide application. Therefore, it is of great significance to construct the temperature estimation model of concrete structure to accurately grasp the temperature change law and prevent temperature cracks. In this paper, we study the winter construction project of Jianqiao Reservoir in Linxi County, Hebei Province, using the measured temperature data and regional temperature data to combine the temporal convolutional neural network (TCN) with the bidirectional long-short-term memory neural network (BiLSTM). Based on this, TimeGAN model and attention mechanism were introduced, and three optimization models, TSTBA, TLTBA and TMTBA, were constructed to further mine the data features. The results show that the introduction of TimeGAN model and attention mechanism significantly improves the model accuracy. Among the three optimized models, TSTBA, TLTBA and TMTBA models lead the performance in turn. Among them, the simulated values of the TSTBA model are highly consistent with the measured values in time and spatial distribution, and the simulated concrete temperature error range is between 0.099 and 0.191, with a high correlation coefficient. The accuracy of the deep neural network model reached 0.920, which effectively estimates the concrete temperature in winter. It provides solid theoretical support for the accurate evaluation of large concrete temperature change prediction and the effective management of concrete performance in water conservancy projects.
