Research And Application Of Deformation Prediction Model For Deep Foundation Pit Based On LSTM

Aiming at the high nonlinearity and ambiguity of deep foundation pit deformation affected by many factors,and the problem with insufficient precision of traditional prediction model dealing with complex nonlinear data,Applying a deep learning model—Long Short-Term Memory(LSTM)network is used to predict the deformation of deep foundation pit in order to improve the precision of deformation prediction and provide a new method for deformation prediction of deep foundation pit.It is found that the selection of optimization algorithm,superparameter selection method,multi-point prediction and other issues need to be deeply analyzed in LSTM network in the deformation prediction of deep foundation pit.The main research contents and conclusions are as follows:(1)The LSTM network is modeled by Stochastic Gradient Descent,Momentum,Nesterov,AdaGrad,RMSProp and Adam optimization algorithms,and the prediction results are compared and analyzed.The example analysis shows that the LSTM network with Adam optimization algorithm has higher prediction accuracy and is more suitable for the deformation prediction of deep foundation pit.(2)In view of the large amount of sample data needed for LSTM network training,and the few monitoring data in the early stage of deep foundation pit construction,the generalization ability of LSTM network is easy to be restricted and its prediction effect is affected.In this paper,by adding the feed-forward neural network layer in front of the LSTM network without increasing the time depth of the LSTM network,the network space depth is increased and the model feature extraction capability is enhanced,in order to enhance the generalization capability of the LSTM network.(3)In LSTM network,hyper-parameters are adjusted manually.This method is inefficient and too human-dependent.Considering that LSTM network training super-parameters are few and the selection method has some rules,the multi-layer grid search method with simple calculation and parallel operation is used to optimize the super-parameters in this paper.(4)The deformation of monitoring points is not isolated and there is a certain connection between adjacent monitoring points.Based on the improved LSTM model,the multi-point prediction model is constructed.The multi-point prediction model can fully mine the internal rules and trends of time series data of different monitoring points by using the self-organization,self-learning and adaptive ability of LSTM network,and make the whole prediction and analysis of the deformation body.Taking a deep foundation pit of Wuhan Metro as an example,several conclusions are drawn:(1)The LSTM network with multi-grid search hyper-parameters is compared with BP neural network,GM(1,1)model,and time series modeling prediction method.The results show that the hyper-parameter optimization based on multi-layer grid search method is more efficient and faster,and the prediction accuracy of LSTM network is higher than the other three models,which can reflect the deformation law of deep foundation pit.(2)Taking the vertical displacement deformation data of deep foundation pit pile as an example,the time series data of 126 periods(all data),60 periods before and 30 periods are selected to construct the sample data,and apply LSTM network and improved LSTM model to train and predict these.The results show that the improved LSTM model can still exert its high-efficiency prediction ability when the monitoring data is relatively small,while the LSTM network model with the decrease of training samples,the generalization ability and the prediction accuracy decreases.(3)Combined with the monitoring data of buildings adjacent to deep foundation pit,based on the improved lstm model,the multi-point prediction and single-point prediction of the building deformation time series data are carried out respectively..The results show that the multi-point prediction model not only has better prediction effect,but also can effectively reduce the total time of model training and prediction.