Highlights • Roof failure as a destructive disaster in roadways of Tabas longwall mine is studied. • Four neuro-fuzzy models namely PSO-ANFIS, SA-ANFIS, GA-ANFIS, and ANFIS are employed. • The results obtained from PSO-ANFIS model are superior in comparison with other ones. • Proposed model is useful to derive rational judgments for predicting unstable zones.

Abstract The problem of roof failure in underground coal mines is responsible for many fatalities, injuries, downtimes, and delays in production planning. Currently, the support systems in underground roadways are mainly designed based on the miners’ experience or, at worst, on trial and error. Nonetheless, the excessive roof displacements may lead to undesirable instabilities that have adverse effects on the mining operations. The uncontrolled roof failures are the major cause of calamitous consequences in Tabas underground coal mine, northeast of Iran, which brought about many disasters in recent years, from the threat of personnel’s safety to the postponement of coal production. Therefore, this research aims at developing a hybrid neuro-fuzzy model to approximate the unknown nonlinear relationship between the maximum roof displacements ($d_{\max }$) and geomechanical features at Tabas longwall mine. After designing several hybrid models, the Particle Swarm Optimization (PSO) algorithm could significantly improve the performance of the Adaptive Neuro-Fuzzy Inference System (ANFIS). The results of three hybrid neuro-fuzzy models show that the optimization process in PSO is superior in comparison with Genetic Algorithm (GA) and Simulated Annealing (SA). According to the results, the determination coefficients ($R^2$) between the measured and predicted values of $d_{\max }$ for PSO-ANFIS, SA-ANFIS, GA-ANFIS, and ANFIS were respectively obtained as 0.944, 0.907, 0.882, and 0.887. The associated error indicated that the PSO-ANFIS model could yield the best performance when encountered with unseen data. Compared to the ANFIS, the PSO-ANFIS model demonstrated an increase of about 6% in $R^2$, and a decrease of about 34% in the Root Mean Square Error ($\mathrm{RMSE}$). Therefore, our strategy in this research is to predict the $d_{\max }$ at first, and then to establish two milestones as 33% of the $d_{\max }$ for timely installing standing support systems, and 66% of the $d_{\max }$ for announcing an alarm threshold in potentially unstable zones. This may be useful to derive a reasonable judgment for predicting the unstable zones, and implementing preventive measures ahead of time in longwall roadways.