| High slope displacement prediction is an important means to prevent landslide,collapse and other disasters.Aiming at the problem of low accuracy of existing prediction models and inability to effectively reflect the time series information contained in the slope displacement.A new hybrid prediction model,named Gated Deep Recurrent Belief Network(GDRBN),is proposed.Based on the excellent processing capability of recurrent neural network(RNN)for timing information,this paper replaces the hidden layer information of Restricted Boltzmann Machine(RBM)with the timing information contained in gated recurrent unit(GRU),so that the two together form a GRU-RBM unit,stacking several GRU-RBM units can form a GDRBN model.Introduce an adaptive learning rate in the forward process of the traditional deep belief network(DBN)model to speed up the network training efficiency.In the reverse fine-tuning process,the vertical and horizontal error feedback fine-tuning network is introduced to reduce the training error.At the same time,taking the K95+310 section of the main line of the fourth contract section of the Shenshan West Expressway as the research object,combined with the local geological and hydrological conditions and the obtained data,the displacement causes were analyzed,and the influencing factors were used as the model input.Using cubic spline interpolation to process the original displacement data at equal time intervals,taking 5 days as a monitoring period,a total of 55 periodical settlement data sequences were obtained,the first 50 period data were used to build the model,and the last 5 period data were used for model accuracy.A GDRBN-based slope settlement prediction model was established.Finally,the results of the model are compared with the results of the grey GM(1,1)model,BP neural network,RNN model and DBN model.The results of the study are as follows:(1)The influence degree of each influencing factor on slope displacement is as follows:slope foot>excavated layers>slope height>slope length.(2)The average absolute error percentage of the GDRBN model is 1.97%,the correlation coefficient is 0.9995,and the determination coefficient is 0.9990.Compared with GM,BP,RNN,DBN and other prediction models,the accuracy is higher,the error is smaller,and the time series response ability is better.(3)Compared with the GDRBN model with fixed high/low learning rate,the loss function value of the GDRBN model with adaptive learning rate is reduced by 58.82%and 42.62%,respectively.The training times and training time are in the middle.Compared with the GDRBN model that only considers the vertical/horizontal error feedback fine-tuning process,the GDRBN model after fine-tuning with vertical and horizontal error feedback has higher accuracy,correlation coefficient and determination coefficient.The mean absolute error percentages decreased by 13.20%and 25.89%respectively,the correlation coefficients increased by 8.50%and 1.24%respectively,and the determination coefficients increased by 2.12%and 7.28%respectively. |
