Abstract

To effectively predict and control the peak particle velocity (PPV) induced by tunnel blasting, this study investigates the Qinhuangdao Jiaoshan Tunnel as a case study. This study first proposes an Improved Particle Swarm Optimization (IPSO) algorithm through theoretical derivation. Building upon IPSO, a further enhanced algorithm, termed CIPSO, is developed by integrating a dependency model derived from Copula theory. The CIPSO algorithm is then employed to optimize a Support Vector Regression (SVR) model, establishing the final CIPSO-SVR prediction framework. Copula theory was employed to quantify the correlation between PPV and surface cumulative settlement (S). A regularization term incorporating Kullback-Leibler (KL) divergence was then embedded into the SVR objective function. The Hyperparameters of the CIPSO-SVR model were optimized using fixed-step rolling cross-validation. The model’s predictive performance was rigorously compared against CIPSO-optimized Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) models, as well as against SVR, CNN, and LSTM models optimized by the Grey Wolf Optimizer (GWO) and Moth-Flame Optimization (MFO) algorithms. The results show that the CIPSO-SVR model achieves superior accuracy and robustness on the test set (R2= 0.9569) in predicting PPV compared to the alternative models. Crucially, the model effectively captures the inherent nonlinear relationships of complex engineering problems, even with small-sample data.OPEN ACCESS Received: 18/08/2025 Accepted: 17/10/2025 Published: 23/01/2026

Abstract

Blast-induced ground vibration, quantified by peak particle velocity (PPV), is a crucial factor in mitigating environmental and structural risks in mining and geotechnical engineering. Accurate PPV prediction facilitates safer and more sustainable blasting operations by minimizing adverse impacts and ensuring regulatory compliance. This study presents an advanced predictive framework integrating CatBoost (CB) with nature-inspired optimization algorithms, including the Bat Algorithm (BAT), Sparrow Search Algorithm (SSA), Butterfly Optimization Algorithm (BOA), and Grasshopper Optimization Algorithm (GOA). A comprehensive dataset from the Sarcheshmeh Copper Mine in Iran was utilized to develop and evaluate these models using key performance metrics such as the Index of Agreement (IoA), Nash-Sutcliffe Efficiency (NSE), and the coefficient of determination (R2). The hybrid CB-BOA model outperformed other approaches, achieving the highest accuracy (R2 =0.989) and the lowest prediction errors. SHAP analysis identified Distance (Di) as the most influential variable affecting PPV, while uncertainty analysis confirmed CB-BOA as the most reliable model, featuring the narrowest prediction interval. These findings highlight the effectiveness of hybrid machine learning models in refining PPV predictions, contributing to improved blast design strategies, enhanced structural safety, and reduced environmental impacts in mining and geotechnical engineering.