Toward Precise Long-Term Rockburst Forecasting: A Fusion of SVM and Cutting-Edge Meta-heuristic Algorithms

Rockburst is one of the most hazardous geological disasters in underground engineering due to its complex causes and destructive nature. To address this, there is an imperative for methodologies that can predict rockbursts quickly and effectively to mitigate preemptively the risks and damages. In this study, 259 rockburst instances were analyzed, employing six rockburst feature parameters: maximum tangential stress (σ_θ), uniaxial compressive strength of rock (σ_c), uniaxial tensile strength of rock (σ_t), stress coefficient (σ_θ/σ_t), rock brittleness coefficient (σ_c/σ_t), and elastic energy index (Wet) as inputs. By integrating three novel meta-heuristic algorithms—dingo optimization algorithm (DOA), osprey optimization algorithm (OOA), and rime-ice optimization algorithm (RIME)—with support vector machine (SVM), hybrid models for long-term rockburst trend prediction were constructed. Performance evaluations through fivefold cross-validation revealed that for the no rockbursts, DOA–SVM (Pop = 200) demonstrated superior predictive performance, achieving an accuracy of 0.9808, precision of 0.9231, recall of 1, and an F1-score of 0.96. For moderate rockbursts, OOA–SVM (Pop = 100) emerged as the most effective, registering an accuracy of 0.9808, precision of 0.9545, recall of 1, and an F1-score of 0.9767. For light and severe rockbursts, DOA–SVM, OOA–SVM, and RIME–SVM showcased comparable predictive outcomes. However, these hybrid models outperformed traditional SVM models optimized with conventional algorithms in terms of accuracy across all rockburst hazard levels. Moreover, the hybrid models underwent additional validation with a new dataset of 20 rockburst instances collected globally, confirming their robust efficacy and exceptional generalization capabilities. An ensuing analysis using local interpretable model-agnostic explanations (LIME) on the six key feature parameters revealed a significant positive correlation between σ_θ and Wet with the severity of rockbursts. These results not only affirm the superior optimization performance of the DOA, OOA, and RIME algorithms but also their substantial potential to enhance the predictive accuracy of machine learning models in forecasting long-term rockbursts.