To establish a comprehensive and systematic high-rise building deformation prediction model, this paper uses Calman filter for deformation data denoising, separation of trend and the error term, and then use GA-BP model and LS-GM (1,1) model to predict the trend, and obtain the trend prediction by combination; secondly, the cumulative the error data is corrected using the Markov chain, further improve the prediction accuracy; finally, the cusp catastrophe theory of stability of high-rise buildings are evaluated to verify the validity of prediction model. The results show that the semi parametric Calman filter has good filtering effect. In the process of forecasting trend, by optimization of the BP neural network, the average prediction accuracy was increased from 4.02% to 2.44%, and the optimization of GM (1,1) model increased the average prediction accuracy from 4.29% to 2.76%, showing that the optimization method in this paper is feasible. Through error correction, the maximum test sample in relative error is only 1.63%, indicating that the error correction model can further improve the prediction accuracy. Catastrophe theory and prediction results were consistent, the high-rise building is in a stable state, and its deformation will continue to be weakend.