Abstract: Deformation monitoring of deep foundation pits is becoming more and more important in the security control of urban construction. Because of the unavoidable noise in the monitoring data and the prediction residual problem from the single prediction model, it is necessary to improve the prediction accuracy of deep foundation excavation. Taking ZJ52 of a deep foundation pit in Lanzhou as an example, and based on Kalman filtering, the paper used a combined model of ARIMA-NAR neural network to predict and analyze the change trend of the observation point. The results show that the average absolute error, average relative error and residual variance from the established model are 0.43 mm, 0.04 mm and 2.23 mm, respectively. The prediction results are better than those of single ARIMA and NAR neural network models. The prediction results from ARIMA-NAR neural network with Kalman filtering can provide reliable guidance for the security construction of this project.

Key words: deep foundation pit, filtering, combination model, prediction