Aiming at the problems of strong nonlinearity, difficulty in identifying and eliminating gross errors in concrete dam deformation monitoring data, an approach of gross error identification and a dam safety monitoring model for deformation monitoring data of concrete dams are proposed based on the improved IGGⅢ-ELM method, which combines the advantages of the improved IGGⅢ method with the good robustness to outliers, the extreme learning machine (ELM) method with high efficiency in predicting data sequences and strong ability to handle nonlinear problems and the incremental ELM method with rapid optimal network structure seeking, the IGGⅢ-ELM method is improved by using a four-segment weight function containing two harmonic coefficients, making the first derivative of the weight function smooth everywhere, and enhancing the availability of information in the mutation range of the weight function. In the study case, the processing result of the improved IGGⅢ-ELM method was compared with the IGGⅢ-ELM method, DBSCAN clustering algorithm, Romanovsky criterion and Pauta criterion. The results show that the gross error identification and prediction method of concrete dam deformation monitoring data based on the improved IGGⅢ-ELM method has a gross error identification rate, stronger generalization ability and better prediction effect than the other four methods.