Hybrid Stacking Model with Optuna Optimization for Open Stope Stability Prediction

The stability of underground open stopes is vital for safe mining, and accurate prediction is essential. This study developed a stability prediction framework integrating hybrid stacking ensemble modeling with the Optuna optimization algorithm to enhance traditional stability graphs. The workflow begins with revising the dataset compiled by Mawdesley. The revised dataset was then divided into training and testing subsets, followed by standardized preprocessing applied to both subsets. To reduce overfitting, a five-fold cross-validation strategy was systematically incorporated with the Optuna algorithm for hyperparameter optimization across ten machine learning algorithms. Three global indicators (accuracy, Kappa coefficient, and area under the curve (AUC)) and three local indicators (precision, recall, and F1 score) were employed for performance evaluation. Following performance ranking, the top three algorithms were selected as base learners, while a ridge regression classifier was designated as the meta learner for complementary characteristics and efficiency. The constructed stacking model was subsequently optimized. The optimized stacking ensemble model achieved an accuracy of 0.8689, precision of 0.8542, recall of 0.7519, F1 score of 0.7857, Kappa coefficient of 0.5760, and an AUC value of 0.8944 on the test set. Subsequent application of this model facilitated the generation of an updated stability graph, demonstrating marked improvements over traditional empirical approaches in both prediction accuracy and resistance to subjective bias. Furthermore, a comparative analysis revealed Optuna’s superior efficiency and effectiveness relative to Bayesian optimization, simulated annealing, and random search methodologies. This study provides a more reliable and accurate tool for open stope stability prediction.