A combination model for displacement prediction of high arch dams stacking five kinds of temperature factors

The statically indeterminate characteristics of arch dams highlight the temperature deformation effect, making accurate modelling of this effect a key issue in improving the performance of displacement monitoring models. In this paper, causal interpretation ability and prediction accuracy of five kinds of temperature deformation modelling factors, including seasonal harmonic function, segmented average previous air temperature, air temperature hysteresis correction factor, principal components and shape feature clustering-based principal components of measured dam temperatures, are compared. On this basis, a combination prediction model is established using the above five causal models as submodels. The combination process is conducted by three methods of dynamic mutual information coefficient, random forest and support vector machine. Research results of the Jinping-I arch dam show that the shape feature clustering-based temperature principal components can significantly improve the accuracy and adaptability of displacement monitoring models, in which the root mean square error decreases with an average rate of 52%. The combination prediction model can effectively take the advantages of different kinds of temperature deformation modelling factors into account. Compared with the hydraulic-seasonal-time model and the best submodel, prediction accuracy of the support vector machine-based combination model is improved with an average rate of 54% and 28%, respectively.