Abstract: In order to improve accuracy in identifying mine water inrush source, an IWOA-HKELM water source identification model is proposed. Firstly, HKELM, featuring better learning ability and generalization performance, was constructed on combined basis of Gaussian kernel function and polynomial kernel function. Secondly, IWOA algorithm was proposed considering that WOA was easy to fall into local optimization. Then, three strategies were introduced to reduce probability of premature convergence and obtain better results, including tent mapping, improvement of non-linear factor and setting of reverse elite learning threshold. Finally, water inrush source data of Xinzhuangzi Mine, being taken as simulation data, was put in IWOA-HKELM model for result prediction after dimension reduction. The results show that optimization of HKELM parameters through IWOA can improve the algorithm's overall prediction performance. Prediction results of IWOA-HKELM are completely consistent with actual situation. Compared with other models, the proposed model obviously excels in terms of average classification accuracy with its average mean square error and standard deviation of classification accuracy being significantly reduced.

Key words: water inrush source identification, improved whale optimization algorithm (IWOA), hybrid nuclear extreme learning machine (HKELM), tent mapping, nonlinear factor, reverse elite learning threshold