Development of Novel Hybrid Intelligent Predictive Models for Dilution Prediction in Underground Sub-level Mining

Abstract

Tenuous dilution estimates in underground mine production scheduling continue to cause significant variations between schedule forecasts and actual production. This arises partly from the inference of dilution from predecessor stopes’ performance, disregarding that these stopes would have undergone multiple intermediate design changes between scheduling and actual mining. The resultant drill and blast-influenced dilution factors gradually lose its robustness over longer planning horizons or when applied to greenfield or brownfield expansions that do not have prior performance data. To overcome this problem, a new methodology is proposed to predict dilution in underground sub-level open stoping (SLOS) using basic geological, geotechnical and stope design attributes available in the early stage of mine planning. The method utilises principal component analysis (PCA), classification and regression tree (CART) algorithm and stepwise selection and elimination (SSE) analysis. First, SSE analysis was conducted to identify the most important independent variables to be used with the CART algorithm (i.e., the SSE-CART model) to provide a predictive model. PCA analysis was then performed, and the new principal components were used to propose a new comparative model (i.e., the PCA-CART model). Low R² values were observed for both models, necessitating the consolidation of dilution categories to increase the models’ prediction bandwidth. The hybrid PCA-CART model outperformed the SSE-CART model with overall F1 score prediction accuracy of 72% and target dilution category prediction accuracy of over 93% against SSE-CART’s 70% and 72%, respectively. Importantly, this study revealed a 13% minimum underestimation of dilution relative to the original design stopes.