Application of optimized grey discrete Verhulst–BP neural network model in settlement prediction of foundation pit

Abstract

Due to the low precision in the prediction of foundation pit settlement of the traditional grey Verhulst model, the optimized discrete grey Verhulst model was selected as the preferred method in settlement prediction. In this work, a combination forecasting model was proposed based on the optimized grey discrete Verhulst model and BP neural network to better predict the foundation pit settlement. For application of the proposed models, the settlement of the foundation pit of a building in Longcheng Industrial Park in Shenzhen, China was predicted. The optimized discrete grey Verhulst model was established on reciprocal transformation of the original data sequence by discretization method. In the modified forecasting model, the predicted result of the optimized grey discrete Verhulst model was used as the input sample value of the BP neural network model and the measured value was used as the target sample value of the neural network model. Furthermore, the neural network was trained to target accuracy and made predict. The maximum number of epochs was 5 × 105. The target error of training is set as 1E−6. The prediction results of these grey models were compared with the prediction results of Kalman filter model. And the two-way verification was carried out to verify that these grey models were suitable for the settlement prediction of the foundation pit. The predicted results of optimized grey discrete Verhulst–BP neural network model display that the average relative errors and mean square errors of the settlement predicted value of two monitoring points CJ12 and CJ23 were 0.0967%, 0.0002 and 0.0795%, 0.00006, respectively. The results revealed that the optimized grey discrete Verhulst–BP neural network model combined the advantages of the two models to achieve complementary advantages, which has higher prediction accuracy and stability. Comparison between the calculated results and the measured ones indicate that the proposed model could satisfactorily describe the settlement monitoring projects.