A proposed neural network model for predicting temperature distribution in aqueducts

【Objective】Temperature fluctuations in aqueducts can induce thermal stress and structural deformation, affecting long-term safety and durability. Accurate prediction of temperature and internal-external surface temperature differences at specific points in an aqueduct is critical for structural monitoring and maintenance. This study proposes a model to address these challenges.【Method】The model was based on the long short-term memory (LSTM) neural network and was applied to a large-span simply supported U-section aqueduct. Measured data from the aqueduct were used to analyze the internal temperature variations and temperature differences between internal and external surfaces at different points in the aqueduct. Using the measured spatiotemporal variations of temperature, we trained the LSTM neural network to predict future temperatures at multiple measurement points, as well as the associated internal-external surface temperature differences.【Result】The internal and surface temperatures of the aqueduct showed noticeable daily and annual variations. Temperatures at the middle of the span and the support sections peaked in August at 45 ℃ and 40 ℃, respectively, while reaching the lowest in January, near 0 ℃ and 5 ℃. The time at which the internal temperature peaked was delayed with increasing depth. A negative temperature gradient occurred when the surface temperatures were low, while a positive gradient was observed when the surface temperatures were high. Compared to the measured data, the coefficient of determination of the LSTM model was close to 1; its mean absolute error was also smaller than those calculated from the CNN and MLP neural networks. The maximum errors of the LSTM model for temperature and internal-external surface temperature difference were 1.681 ℃ and 2.220 ℃/m, respectively.【Conclusion】The LSTM-based model can accurately predict temperature distribution and internal-external surface temperature differences in aqueducts. It can be applied for structural monitoring and safety management of flumes.