Combined Prediction Model for High-Speed Railway Bridge Pier Settlement Based on Robust Weighted Total Least-Squares Autoregression and Adaptive Dynamic Cubic Exponential Smoothing

AbstractThe prediction of high-speed railway bridge pier settlement is important for the safety of high-speed railway engineering. At present, a common method in settlement prediction is the curve fitting model in single prediction models. However, it may be difficult to describe the settlement rule of high-speed railway bridge piers using a curve fitting model with limited observation data during time-constrained construction periods. Moreover, relying on only a single prediction model usually does not allow for full exploration of the potential information in the data and poses the problem of poor stability and applicability. To solve this issue, a combined prediction model that uses the optimal nonnegative variable weight combination based on robust weighted total least-squares autoregression (RWTLS-AR) and adaptive dynamic cubic exponential smoothing (ADCES) is proposed to combine the advantages of two single prediction models. The RWTLS-AR model, using a robust weighted total least-squares method, has high prediction accuracy in the case of fewer observation data. At the same time, the adaptive dynamic judgment mechanism is established using the ADCES model to improve stability. The proposed model is applied to the settlement prediction of high-speed railway bridge pier, and three sets of observation data are used for evaluation. A comparison is made with two single prediction models and three other combined prediction models. The results show that the mean absolute error, root-mean-square error, and mean absolute percentage error of the proposed model are respectively 0.092 mm, 0.101, mm and 5.936% in the first set of observations, 0.099 mm, 0.118 mm, and 6.592% in the second set of observations, and 0.177 mm, 0.203 mm, and 15.914% in the third set of observations. This indicates that the proposed model is more accurate and stable than all the aforementioned prediction models.