Abstract: Surface deformation is a common phenomenon in mining area. Estimating the maximal surface subsidence has strong significance to ensuring the safety of ore mining. However the traditional prediction methods have some space for improving accuracy. A new data-driven methodology is proposed based on data-mining algorithms and Extreme Learning Machine (ELM) with data of rock displacement records to predict the maximal surface subsidence. In the proposed methodology, the input parameters are mining thickness, dip-angle, average mining depth, strike length, dip length and overburden lithology. An ELM model including 114 hidden nodes and with sigmoid function as the kernel function is constructed to predict the maximum value of surface subsidence. By case study, ELM performs better than traditional methods including CHAID, Boosted Tree, ANN, BPNN, and SVM in terms of RMSE, MAE, MAPE, maximum residue and rank correlation coefficient. Hence, this framework is valuable for predicting maximum surface subsidence.