Recurrent neural network prediction for adjacent building settlement induced by foundation pit excavation

A recurrent neural network (RNN) method for predicting the settlement of adjacent buildings caused by the excavation of foundation pits is introduced. Fully considering the impact of the excavation process on building settlement, the input parameters of the RNN model were analysed and optimised. To improve the predictive performance of the model, data preprocessing and hyperparameter value optimisation were also carried out. A dataset of 204 × 6 typical samples was established based on parameters from the Meituan Shanghai Science and Technology Center Project, with 144 × 6 samples from excavation stages EL-1 to EL-3 used for training. The predicted values of the model were basically consistent with the observed values in terms of their changing trends, which clearly reflected the impact of excavation procedures on building settlement. The root mean square error of the six test results was in the interval [0.39,1.01], indicating that the predicted values were in good agreement with the observed values and the model was robust. Moreover, the maximum prediction error of the maximum cumulative settlement was only 7.3%, which provides a reference for engineering practice.