Search

CN-119202636-B - Dynamic landslide displacement prediction method based on MACSA-BiLSTM

CN119202636BCN 119202636 BCN119202636 BCN 119202636BCN-119202636-B

Abstract

The invention provides a dynamic landslide displacement prediction method based on MACSA-BiLSTM, which belongs to the field of landslide prediction, and comprises the steps of obtaining landslide displacement, setting relevant parameters, preprocessing, decomposing the preprocessed displacement through Ensemble Empirical Mode Decomposition (EEMD) to obtain a periodic term and a trend term, calculating the optimal order of a polynomial fitting trend term according to the trend term, fitting the polynomial according to the optimal order, predicting the trend term, optimizing the super-parameters of a bidirectional long and short time memory neural network (BiLSTM) according to the periodic term by adopting a novel multi-head attention crow search optimization algorithm (MACSA), building the bidirectional long and short time memory neural network according to the optimized super-parameters, obtaining the predicted total displacement by combining the predicted trend term and the preprocessed next predicted periodic displacement, and calculating the residual rate required by the next period prediction.

Inventors

  • LI DEXIN
  • TANG YUFENG
  • ZHANG XIAOLEI

Assignees

  • 郑州大学
  • 黄河水利职业技术学院

Dates

Publication Date
20260512
Application Date
20240903

Claims (8)

  1. 1. The dynamic landslide displacement prediction method based on MACSA-BiLSTM is characterized by comprising the following steps of: Step one, acquiring a landslide displacement data sequence, and setting related parameters for landslide displacement prediction; Step two, carrying out data preprocessing on the landslide displacement data sequence, calculating the residual error rate of the previous prediction period, carrying out ensemble empirical mode decomposition on the preprocessed data, and decomposing the landslide displacement data sequence into a trend item displacement sequence and a period item displacement sequence; Step three, obtaining the optimal order of fitting the trend item through a polynomial, performing polynomial fitting on the trend item displacement sequence through the optimal order, and predicting the trend item displacement of the next period; Step four, optimizing the super parameters of the bidirectional long-short-term memory neural network through a multi-head attention crow search optimization algorithm by combining the periodic term displacement sequences; Step five, according to the optimal super-parameters obtained by optimization, a two-way long short-time memory prediction network is established to predict the displacement of the period item of the next period; Step six, adding the predicted trend item displacement and the predicted periodic item displacement to obtain a predicted total displacement; step seven, adding next monitoring data, repeating the step two to the step six, and predicting the total displacement of the next prediction period every time a cycle is repeated; the multi-head attention crow search optimization algorithm process is as follows: In the initialization stage, adopting logistic-tent mixed chaotic mapping to initialize population positions; In the food hiding stage, calculating the fitness value of each crow i, storing the historical optimal fitness value of each crow i, and taking the corresponding position of the optimal fitness as the food hiding position of each crow i; In the stage of determining the tracked object, selecting M crow with optimal historical optimal fitness value as the tracked object; In the tracking stage, each crow i updates the position after tracking as follows: X t+1 (i,j)= X t (i,j)+r×(c1×rand×X t mem (i,j)+c2×rand×X t mem (w,j)) When A > AP X t+1 (i, j) =lb (j) +rand× (ub (j) -lb (j)) when A≤AP Wherein, X t+1 (i, j) and X t (i, j) are respectively positions of the crow i in the dimension j after updating and before updating, r is a flight step length, c1 and c2 are respectively attention coefficients and c1+c2=1, rand is a random number between (0, 1), X t mem (i, j) and X t mem (w, j) are respectively positions of the crow i at the moment t and the tracked crow in the dimension j, A is a random value in (0, 1), AP is a warning value of the crow when the crow finds other crows, and the value is between (0, 1); In the boundary judging stage, judging whether the position of X t+1 (i, j) in the dimension j is within the search boundary, if so, updating the position of the dimension j according to X t+1 (i, j), otherwise, keeping the position of X t (i, j) unchanged in the dimension j.
  2. 2. The dynamic landslide displacement prediction method based on MACSA-BiLSTM as claimed in claim 1, wherein in the first step, the related parameters of the landslide displacement prediction include a prediction period L, a total number N, MACSA of initialized crow of polynomial fitting maximum order D, MACSA, a tracked crow number M, MACSA-BiLSTM start residual ratio R, and a crow warning value AP.
  3. 3. The dynamic landslide displacement prediction method based on MACSA-BiLSTM as claimed in claim 1, wherein in the second step, the data preprocessing is to calculate the sum of the displacement of each L monitoring data by taking the latest monitoring data as a starting point, convert the original monitoring displacement length sequence into 1/L of the original length, and delete the non-integer part of the data furthest from the current monitoring time if the original length cannot be integer divided by L.
  4. 4. The dynamic landslide displacement prediction method based on MACSA-BiLSTM as claimed in claim 1, wherein in the second step, the residual rate is calculated by the total displacement predicted in the previous cycle and the actual displacement data after preprocessing in the current cycle, namely: C xh =(V xhp - V xh )/ V xh Wherein C xh is the residual rate of the previous cycle, V xhp is the predicted total displacement of the previous cycle, V xh is the actual displacement data of the current cycle after pretreatment, and the initial value of the residual rate in the first cycle is set to 0.
  5. 5. The dynamic prediction method of landslide displacement based on MACSA-BiLSTM as claimed in claim 1, wherein in the third step, the polynomial fitting is performed according to the maximum order D of the polynomial fitting, the root mean square error of the 2 nd order to the D th order fitting is calculated, and the order with the lowest root mean square error is automatically selected as the final order of the polynomial fitting.
  6. 6. The dynamic landslide displacement prediction method based on MACSA-BiLSTM as set forth in claim 1, wherein the logistic-tent mixed chaotic map is initialized, and the mathematical expression is: x (i, j) =lb (j) +X 0 (i,j)×μ×(1- X 0 (i, j))× (ub (j) -lb (j)) when i≤N/2 X (i, j) =lb (j) +x 0 (i, j)/a× (ub (j) -lb (j)) when i > N/2 and X 0 (i, j) < a X (i, j) =lb (j) + (1-X 0 (i, j))/(1-a) × (ub (j) -lb (j)) when i > N/2 and X 0 (i, j) Σa Wherein N is the total number of MACSA initialized crow, X 0 (i, j) is a random value of the ith crow in the dimension j, the size is between (0, 1), mu E [0,4] is called a Logistic parameter, a is a chaos coefficient, ub (j) and lb (j) are the upper limit and the lower limit of a solution domain in the dimension j respectively, and X (i, j) is a Logistic-tent chaos mapping value of the ith crow in the dimension j.
  7. 7. The dynamic prediction method of landslide displacement based on MACSA-BiLSTM as claimed in claim 1, wherein in the fourth step, the super parameters of the bidirectional long-short-term memory neural network are selected to optimize one or more parameters of hidden layer node number, initial learning rate, learning rate reduction factor, regularization coefficient, random inactivation rate and batch processing size, and the MACSA method is adopted to optimize the long-short-term memory neural network only when the absolute value of the last cyclic residual rate is larger than the start residual rate R of MACSA-BiLSTM.
  8. 8. The landslide displacement dynamic prediction method based on MACSA-BiLSTM as claimed in claim 1, wherein in the fifth step, the long-short memory neural network comprises an input layer, a long-short memory layer, a random inactivation layer, an activation layer, a full connection layer and an output layer.

Description

Dynamic landslide displacement prediction method based on MACSA-BiLSTM Technical Field The invention relates to the field of landslide prediction, in particular to a dynamic landslide displacement prediction method based on MACSA-BiLSTM. Background Landslide geological disasters are the most common type of disaster for geological disasters. In the landslide geological disaster prevention and control early warning, landslide displacement is the most important basis, so that landslide displacement prediction has important significance. With the development of big data and artificial intelligence technology, the deep learning method is widely applied to various fields, wherein a bidirectional long and short time memory neural network (BiLSTM) is widely applied due to the advantage in processing time sequence problems. However, if the BiLSTM network model has a plurality of super parameters, if the setting is incorrect, the prediction result is directly affected, so that a proper optimization algorithm is established to optimize the super parameters, and the relationship between the accuracy and the calculation time cost is balanced, and the method has very important practical significance for improving the prediction accuracy and the response speed and increasing the emergency response time after early warning. Disclosure of Invention The invention aims to solve the problems of insufficient precision and low efficiency in the existing landslide displacement prediction model, and provides a dynamic landslide displacement prediction method based on MACSA-BiLSTM, which can predict landslide displacements in different periods in the future according to historical landslide displacement monitoring data, so as to provide scientific basis for landslide prevention and control and early warning. The invention is realized in the following way: A dynamic landslide displacement prediction method based on MACSA-BiLSTM comprises the following steps: Step one, acquiring a landslide displacement data sequence, and setting related parameters for landslide displacement prediction; Step two, carrying out data preprocessing on the landslide displacement data sequence, calculating the residual rate of the previous prediction period, carrying out Ensemble Empirical Mode Decomposition (EEMD) on the preprocessed data, and decomposing the landslide displacement data sequence into a trend item displacement sequence and a period item displacement sequence; Step three, obtaining the optimal order of fitting the trend item through a polynomial, performing polynomial fitting on the trend item displacement sequence through the optimal order, and predicting the trend item displacement of the next period; Step four, optimizing the super parameters of the two-way long and short time memory (BiLSTM) neural network through a multi-head attention crow search (MACSA) optimization algorithm by combining the periodic term displacement sequences; Establishing BiLSTM a prediction network to predict the displacement of the period item of the next period according to the optimal super-parameters obtained by optimization; Step six, adding the predicted trend item displacement and the predicted periodic item displacement to obtain a predicted total displacement; step seven, adding next monitoring data, repeating the step two to the step six, and predicting the total displacement of the next prediction period every time a cycle is repeated; Further, in the first step, the related parameters of landslide displacement prediction include a prediction period L, a polynomial fitting maximum order D, MACSA, initializing total number N, MACSA of the crow, tracking crow number M, MACSA-BiLSTM, starting residual ratio R, and crow warning value AP. In the second step, the data preprocessing is to calculate the sum of the displacement of each L monitoring data by taking the latest monitoring data as a starting point, convert the original monitoring displacement length sequence into 1/L of the original length, and delete the data of the non-integer part which is farthest from the current monitoring time if the original length cannot be integer divided by L. Further, in the second step, the residual rate is calculated by calculating the total displacement predicted in the previous cycle and the actual displacement data after preprocessing in the current cycle, that is: Cxh=(Vxhp- Vxh)/ Vxh Wherein C xh is the residual rate of the previous cycle, V xhp is the predicted total displacement of the previous cycle, V xh is the actual displacement data of the current cycle after pretreatment, and the initial value of the residual rate in the first cycle is set to 0. Further, in the third step, the polynomial fitting is performed by calculating Root Mean Square Error (RMSE) of the 2 nd to the D th order fitting according to the maximum order D of the polynomial fitting, and automatically selecting the order with the lowest RMSE as the final order of the polynomial fitting. Fur