To determine a reasonable size of floor waterproof coal pillar and reduce floor water invasion incidents, support vector machine (SVM) and artificial bee colony algorithm (ABCA) are used to research floor failure depth problem. Prediction accuracy is significantly affected by SVM training parameter choice of penalty factor C and kernel width g . ABCA is used to optimize the selection of training parameters, and a floor failure depth prediction model is established based on SVM. Selecting mining depth, coal seam dip, mining thickness, face plagioclase, anti-destruction capability of floor and whether there is fault fracture zone cutting through the floor as major impact indicators of floor damaged depth, we use 30 sets of measured data for training model and forecasting. The results show that average relative error is 12.5%, average absolute error is 0.986 m, mean square error is 0.005, and squared correlation coefficient is 0.980 . This prediction model has stronger generalization ability and higher prediction accuracy compared with other prediction models.