Particle peak velocity prediction based on risk-oriented hybrid ensemble learning

In blasting engineering, accurate prediction of peak particle velocity (PPV) is essential to ensuring the safety of surrounding structures. In machine-learning-based PPV prediction, symmetric loss functions (e.g. MSE) are commonly adopted as optimisation objectives. However, these functions treat overestimation and underestimation equally, making them ill-suited for applications with stringent safety requirements, where hazardous underestimation must be avoided. To address this limitation, a risk-oriented hybrid ensemble model is proposed to enhance both safety and reliability while maintaining high prediction accuracy. Three gradient-boosting tree models-LightGBM, XGBoost, and CatBoost-are employed as base learners and integrated using a stacking framework. To obtain near-optimal configurations for the three base learners, Bayesian Optimisation (BO), Grey Wolf Optimiser (GWO), and Particle Swarm Optimisation (PSO) are employed for hyperparameter tuning. Building on the ensemble framework, an asymmetric safety assessment system is proposed. Model performance near the PPV safety threshold is quantified using the asymmetric weighted mean squared error (W-MSE) and the hazardous low-estimation rate (HLR). The results indicate that the integrated model achieves strong overall performance and effectively suppresses hazardous underestimation. The integrated model demonstrates clear advantages in PPV prediction. Moreover, it provides a reusable paradigm for embedding engineering safety constraints into machine learning training and evaluation, thereby offering reliable technical support for safety planning and risk minimisation in blasting projects.