A causal prediction model for the measured temperature field of high arch dams with dual simulation of lag influencing mechanism

Abstract

Mathematical models using measured dam temperature as modelling factors are frequently used in displacement-based structural health monitoring of high arch dams. However, when these models are used to predict future displacement, dam temperature is also unknown, thus it is necessary to predict the temperature field of the dam body in the first step. The main purpose of this paper is to improve the accuracy and rationality of data-driven models in simulating the complex lag influencing mechanism of temperature field in high arch dams. Two strategies are integrated from the perspective of modelling factors and modelling methods. First, based on massive temperature monitoring data, lag influencing times and laws of air temperature and upstream reservoir water level on dam temperature are quantified using cosine similarity and Shapley additive explanation. Based on the main influencing factors and the optimal lag time, dam temperature field zoning and modelling factor optimization methods are proposed. On this basis, a causal prediction model is established for the measured temperature field of high arch dams using the nonlinear autoregressive with exogenous inputs (NARX) neural network, by which the lag influencing mechanism preliminarily simulated through modelling factors can be further optimized. The results of the Jinping-I arch dam indicate that the causal mechanism of arch dam temperature field is nonlinear and hysteresis. The proposed method has achieved the quantification standards of temperature field zoning and modelling factor optimization for high arch dams based on the lag influencing mechanism, and the obtained temperature field zones of the dam body are consistent with the actual situation. Compared with the frequently used support vector machine model, the NARX model can significantly improve the prediction accuracy of dam temperature field, where the root mean square error of all 134 temperature monitoring points on the dam body decreases with an average rate of 76%.