Time series prediction of the surrounding rock displacement of a soft rock tunnel in the Central Yunnan Water Diversion Project

[Objective] The monitoring value of surrounding rock displacement has the characteristics of complexity and nonlinear dynamic change, and the static one-time learning of previous optimization algorithms combined with a single regression model cannot be practically applied in real scenarios. The regression fitting model uses several displacement monitoring point data to construct a general model of the surrounding rock displacement change, which cannot be applied to predict the future changes in monitoring points. The autocorrelation of the surrounding rock displacement data makes it more practical as a time series prediction problem. However, the generalization performance of a single model is easily disrupted by historical monitoring data, resulting in inaccurate prediction of test applications. In this study, a dynamic prediction method for surrounding rock displacement time series combined with time series monitoring data preprocessing is proposed. [Methods] First, the displacement monitoring data of the tunnel-surrounding rock are preprocessed. The intercepted stability monitoring data are isometrized by cubic spline interpolation, and the monitoring data are decomposed into trend and random term displacement components by variational mode decomposition signal processing. Ada boost integrates 10 long short-term memory networks to construct an integrated optimization model for time series prediction. Then, the weights of the training samples are initialized, the weight coefficients of the base model in the integration are calculated by training the first base model, and the weights of the training samples of the next base model are updated. Finally, the weight coefficients of all base models are obtained. After Adaboost integration optimization, the prediction results are calculated using all base models and their weight coefficients. After training and learning, single-step dynamic prediction is performed, and monitoring changes are updated in real time to model learning. The cumulative displacement prediction results can be obtained by superimposing the trend and random term displacement sequences using the time series decomposition principle. [Results] The displacement components of the rock surrounding the Central Yunnan Water Diversion Project were predicted and superimposed, and three displacement data were obtained. Compared with the traditional time series prediction model, each displacement index exhibited good performance. The complete data of the surrounding rock displacement time series were obtained by the FLAC 3D numerical simulation engineering section, and the application performance of the integrated optimization model was verified. Results showed that the integrated optimization model exhibited good performance in each component and cumulative displacement and was less affected by deformation rate fluctuation than the traditional model. [Conclusions] After preprocessing the time series data, the influencing factors of surrounding rock displacement and deformation are decomposed, and multiple time series prediction models are integrated for single-step dynamic prediction, which improves the shortcomings of previous studies. The correction determination coefficient and symmetrical average absolute percentage error are used as performance indicators to verify that the prediction accuracy achieves the expected goal and is superior to the traditional classical model in solving the time series problem, which promotes the predictability of surrounding rock displacement in practical applications. © 2024 Tsinghua University. All rights reserved.