Probability interval prediction of shield tunneling thrust based on adaptive variational mode decomposition and deep hybrid networks

Abstract Accurate prediction of shield tunneling thrust has significant importance for ensuring safety and efficiency in tunnel construction. However, traditional methods cannot handle complex nonlinear characteristics and multi-source heterogeneous data. This study proposes a framework for predicting thrust probability intervals by combining dung beetle optimizer enhanced variational mode decomposition (DBOVMD) with deep hybrid networks. The framework first uses DBOVMD to adaptively decompose the original thrust sequence, effectively extracting multi-scale time-frequency features. Subsequently, a BiLSTM-CNN hybrid network is constructed for the decomposed subsequences, comprehensively capturing long-term dependencies and local feature patterns. For residual sequences that are difficult to model deterministically, a Gaussian Process Regression model is introduced for probability interval prediction, enabling the model to quantify uncertainty and provide reliable interval estimates for decision-making. Experimental results demonstrate that the proposed framework exhibits excellent performance on datasets from two distinct engineering cases with different sampling frequencies, achieving 99.83% prediction accuracy for subsequences and 98.64% confidence interval coverage for residual sequences. The research findings provide a theoretical foundation and practical tool for intelligent decision support systems in shield tunneling processes, offering broad engineering application prospects.