CN116821586A - Landslide displacement prediction method based on attention mechanism and bidirectional gating circulating unit - Google Patents

Landslide displacement prediction method based on attention mechanism and bidirectional gating circulating unit Download PDF

Info

Publication number
CN116821586A
CN116821586A CN202310774955.1A CN202310774955A CN116821586A CN 116821586 A CN116821586 A CN 116821586A CN 202310774955 A CN202310774955 A CN 202310774955A CN 116821586 A CN116821586 A CN 116821586A
Authority
CN
China
Prior art keywords
displacement
term
landslide
random
predicted value
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202310774955.1A
Other languages
Chinese (zh)
Inventor
陆鑫
章险锋
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
University of Electronic Science and Technology of China
Original Assignee
University of Electronic Science and Technology of China
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by University of Electronic Science and Technology of China filed Critical University of Electronic Science and Technology of China
Priority to CN202310774955.1A priority Critical patent/CN116821586A/en
Publication of CN116821586A publication Critical patent/CN116821586A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/18—Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/12—Computing arrangements based on biological models using genetic models
    • G06N3/126—Evolutionary algorithms, e.g. genetic algorithms or genetic programming
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
    • G06Q50/10—Services
    • G06Q50/26—Government or public services

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Business, Economics & Management (AREA)
  • Biophysics (AREA)
  • Mathematical Physics (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Software Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Optimization (AREA)
  • Molecular Biology (AREA)
  • Tourism & Hospitality (AREA)
  • Pure & Applied Mathematics (AREA)
  • Computational Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Computing Systems (AREA)
  • Computational Linguistics (AREA)
  • Evolutionary Computation (AREA)
  • Mathematical Analysis (AREA)
  • Physiology (AREA)
  • Algebra (AREA)
  • Databases & Information Systems (AREA)
  • Genetics & Genomics (AREA)
  • Probability & Statistics with Applications (AREA)
  • Development Economics (AREA)
  • Educational Administration (AREA)
  • Operations Research (AREA)
  • Economics (AREA)
  • Human Resources & Organizations (AREA)
  • Marketing (AREA)
  • Primary Health Care (AREA)
  • Strategic Management (AREA)

Abstract

The invention discloses a landslide displacement prediction method based on an attention mechanism and a bidirectional gating circulating unit, and belongs to the fields of landslide displacement prediction and computer deep learning. Based on the idea of time sequence decomposition, the method adopts a genetic algorithm to carry out parameter optimization on a variation modal decomposition method; then decomposing landslide displacement into trend term, periodic term and random term displacement by using an optimized variation modal decomposition method, and simultaneously selecting external inducement and decomposing the external inducement into periodic term components and random term components; obtaining a predicted value of the displacement of the trend term by adopting polynomial least square fitting; obtaining a predicted value of the periodic term displacement by using a bidirectional gating circulating unit model based on an attention mechanism; obtaining a predicted value of random term displacement by adopting a nonlinear autoregressive neural network; and finally, superposing the displacement predicted values to obtain a final landslide displacement predicted value. The method provided by the invention can be used for predicting different displacement components independently, and can be used for effectively improving the accuracy of the landslide displacement predicted value.

Description

Landslide displacement prediction method based on attention mechanism and bidirectional gating circulating unit
Technical Field
The invention belongs to the fields of landslide displacement prediction and computer deep learning, and particularly relates to a landslide displacement prediction method based on an attention mechanism and a bidirectional gating circulating unit.
Background
The geological environment condition of the complicated mountain area is usually fragile, such as mountain area with complicated and various geological structures in the high mountain gorges Gu Demao, and when the mountain is subjected to the action of external induction factors, geological disasters such as collapse, landslide, mud-rock flow and the like easily occur. This can result in serious loss of life and lives and properties for people in the area. Among all natural disasters, landslide is a kind of mountain natural disaster that is liable to occur. It is necessary to predict landslide hazard risk.
Landslide evolution is a dynamically changing process, and when conditions of external induction factors change, the risk of landslide also changes. The risk of landslide geological disasters is completely inadequate if only complex mountainous areas are given. For a monitored area where a landslide geological disaster has been predicted to occur in the near future, it is also necessary to determine the spatial extent of the influence of the landslide geological disaster and the expected occurrence time of the landslide geological disaster risk. Landslide is generally a sliding process from the onset of deformation of the surface of a slope to the final destabilization of the slope. It generally undergoes 3 stages of initial start-up (stage i), slow deformation (stage II) and accelerated deformation (stage iii). The concept of predicting the stage of the landslide according to the landslide displacement change trend and further predicting the possible occurrence time of the landslide is widely accepted. The accurate prediction of landslide displacement variation trend is a precondition for accurate prediction of landslide occurrence time.
The construction of a prediction model by deep learning is a main mathematical method of current prediction. In general, landslide prediction methods can be classified into single-factor and multi-factor prediction methods. For example, the single factor methods such as GM (n, 1) and Verhulst model are applied to landslide prediction to obtain good effects. With the update of information technology and monitoring technology, the use of multi-factor methods is currently preferred. Such as: based on response analysis of the induction factors, building a BP neural network to predict landslide displacement period; based on the time sequence and the PSO-SVR coupling model, establishing a response relation between the periodic term displacement and external factors; based on wavelet transformation and an ELM model, a chaotic time series WA-ELM landslide displacement prediction model is provided. It is generally considered that a "learning" approach that considers multiple factors is more reasonable than a method that considers a single factor when the information is sufficient.
However, at present, the methods mostly adopt static models, that is, the models are trained only once before the models are put into use, and parameters of the models cannot be adjusted along with time changes in the use process. And landslide is a dynamic change process, so that the prediction effect of the model is restricted.
The time sequence prediction of landslide displacement is an important component in a landslide prediction early warning system. The landslide displacement is jointly influenced by the geological condition of the landslide body and external inducement factors, and has the characteristics of dynamics, complexity, nonlinearity and the like. The mountain area is special in topography structure and complex in climate condition, the probability of landslide occurrence is higher, and the difficulty of displacement prediction is higher. At present, although some prediction model methods can be used for landslide surface displacement prediction, most of the prediction model methods are static models, and the trend of landslide self evolution is not considered. For example, the support vector machine model SVM and the neural network model NN belong to static models, and are limited to the unidirectional transmission structural characteristics of the hidden layer network, so that the model cannot consider the trend of landslide self evolution, and the prediction effect of the model is restricted.
Aiming at the influence of external factors such as rainfall and the like on landslide displacement change, most of the current methods utilize a time sequence addition model to decompose displacement into trend term displacement and periodic term displacement. However, they tend to ignore the effect of random term displacement, resulting in displacement prediction accuracy that is not always significantly improved. In order to decompose to obtain random term displacement to improve the accuracy of prediction, methods that can be used include empirical mode decomposition, wavelet analysis, and the like. However, they usually have more than 5 decomposed components, and the physical meaning represented by each component is difficult to be clarified.
The landslide is high in probability of occurrence, difficult to predict, and dynamic evolution characteristics of landslide displacement are difficult to simulate by the traditional time sequence decomposition method. The current landslide displacement prediction method mostly adopts static models, such as a Support Vector Machine (SVM) model and a Neural Network (NN) model, and landslide is a dynamic process, so that the static model prediction effect is not good.
With the wide application of deep learning, the long-short-term memory network can be applied to landslide prediction, so that a dynamic prediction model of a landslide displacement change process is realized. The gating circulation unit (GRU, gated Recurrent Unit) is used as an optimization model of the long-short-term memory network, has good effect on automatic identification and classification, but has relatively few application researches in landslide displacement prediction at present.
Disclosure of Invention
Aiming at the problem that the existing landslide displacement prediction method cannot take the evolution trend of the landslide itself into consideration, the invention provides a landslide displacement prediction method based on an attention mechanism and a bidirectional gating circulation unit. Based on the idea of time sequence decomposition, the method adopts a Genetic Algorithm (GA) to carry out parameter optimization on a variation modal decomposition method (VMD); then decomposing landslide displacement into trend term, periodic term and random term displacement by using an optimized variation modal decomposition method, and simultaneously selecting external inducement and decomposing the external inducement into periodic term components and random term components; obtaining a predicted value of the displacement of the trend term by adopting polynomial least square fitting; obtaining a predicted value of the periodic term displacement by using a bidirectional gating circulating unit model based on an attention mechanism; obtaining a predicted value of random term displacement by adopting a nonlinear autoregressive neural network; and finally, superposing the displacement predicted values to obtain a final displacement predicted value.
The technical scheme adopted by the invention is as follows:
a landslide displacement prediction method based on an attention mechanism and a bidirectional gating circulating unit is characterized by comprising the following steps:
s1, landslide displacement time sequence decomposition:
based on a landslide displacement time sequence addition model shown in a formula (1), performing time sequence decomposition on historical landslide displacement into trend term displacement, periodic term displacement and random term displacement;
Y t ＝T t +P t +R t (2)
wherein Y is t Is the landslide displacement monitoring value at the moment T t For the trend term displacement at time t, P t For the period term displacement at time t, R t Is the random term displacement at time t.
S2, external inducement time sequence decomposition:
adopting a gray correlation model to mine the correlation between the historical displacement components and the influence factors, and selecting two influence factors with the greatest influence on landslide displacement as external inducements by combining the hysteresis characteristics of the influence factors; and carrying out time sequence decomposition on the two external inducers to obtain a periodic term component and a random term component of the two external inducers.
S3, predicting trend item displacement:
and fitting and predicting the displacement of the trend term by using a polynomial least square method to obtain a predicted value of the displacement of the trend term.
S4, predicting the displacement of the period term:
and taking the Bi-GRU neural network as a main network, and inserting an attention module on the basis of the main network to obtain the bidirectional gating circulation unit model based on an attention mechanism.
And inputting the periodic term displacement and the periodic term component into a bidirectional gating circulation unit model based on an attention mechanism to obtain a predicted value of the periodic term displacement.
S5, predicting random term displacement:
and combining the random term displacement and the random term component by adopting a nonlinear autoregressive neural network to obtain a predicted value of the random term displacement.
S6, calculating a landslide displacement predicted value:
and according to the vectorization of the displacement, adding and summing the predicted value of the displacement of the trend item, the predicted value of the displacement of the periodic item and the predicted value of the displacement of the random item to obtain the predicted value of the landslide displacement.
Preferably, in step S4, the attention mechanism-based bi-directional gating cycle unit model includes a back propagation module, a forward propagation module, and an attention module; inputting random term displacement of the T time period into a counter propagation module, inputting an output result of the module into an attention module, and processing the output result by the attention module to obtain a weight vector matrix F; meanwhile, the random term component of the T time period is input into a forward propagation module to obtain a hidden vector h of the T time period t Finally, the hidden vector h of the T time period t And the weight vector matrix F obtained by the attention mechanism module is used as input, and the predicted value of the T+1 time landslide period term displacement is output through a sigmoid function.
Preferably, in the step S1 and the step S2, firstly, a genetic algorithm is used to perform parameter optimization on a secondary penalty factor α and a rising time step parameter τ in a variation modal decomposition method, and then, a variation modal decomposition method after parameter optimization is used to perform time sequence decomposition on historical landslide displacement and external causes;
the parameter optimizing mode is as follows:
B. setting the population individual number as K, the maximum iteration number is greater than 20, the optimizing range of the punishment factor alpha is [0.1, 1000], and the optimizing range of the rising time step parameter tau is [0,1]; wherein, when the landslide displacement is subjected to time series decomposition, the K value is 3, and when the external inducement is subjected to time series decomposition, the K value is 2;
B. selecting individuals by adopting a competitive bidding method according to the fitness;
C. the selected individuals generate offspring through crossover and mutation operations;
D. calculating the fitness of the offspring;
E. merging individuals of the offspring and the previous generation, and sorting according to fitness;
F. if the number of iterations is more than 20 or the optimal fitness of 10 iterations is kept unchanged, ending the algorithm to obtain optimized variation modal decomposition parameters; otherwise, repeating the steps B-F.
Preferably, in the step S5, the functional relationship of the nonlinear autoregressive neural network model is expressed as:
y(t)＝f[y(t-1),…,y(t-d y ),x(t-1),…,x(t-d x )] (3)
wherein f is [.]To activate the function, y (t) represents the predicted value of the random term displacement at time t, d x And d y The delay order of x (t) and y (t), y (t-1), …, y (t-d), respectively y ) Predictive value representing random term displacement at the first d moments, x (t-1), …, x (t-d) x ) Representing the actual value of the combination of the random term displacement and the random term component at the first d moments.
The invention provides a landslide displacement dynamic prediction method aiming at a difficult mountain area, which can improve the accuracy of landslide displacement prediction. Firstly, the method is based on the idea of time sequence decomposition, adopts GA to conduct parameter optimization on VMD, and then uses VMD to decompose landslide displacement into displacement components of trend items, period items and random items. The landslide trend item displacement is expressed as a monotonically increasing curve along with time under the influence of internal factors such as rock-soil characteristics, potential energy and the like, and the trend item displacement can be predicted by adopting a polynomial least square fitting method; adopting a notice mechanism and a Bi-GRU (Bi-directional-Gated Recurrent Unit) neural network to dynamically predict the periodic item displacement in multiple factors; and predicting the random term displacement by adopting a nonlinear autoregressive neural Network (NARX), and finally, superposing each displacement component to obtain a displacement predicted value.
Drawings
FIG. 1 is a flow chart for predicting landslide surface displacement;
FIG. 2 shows the variation of the optimal fitness value;
FIG. 3 shows landslide displacement decomposition results;
FIG. 4 is a time series exploded flow chart of external inducement;
FIG. 5 is a trend term displacement prediction result;
FIG. 6 is an attention mechanism+Bi-GRU model;
FIG. 7 is a graph showing the result of periodic term displacement prediction;
FIG. 8 is a block diagram of a nonlinear autoregressive neural network;
FIG. 9 is a graph of random term displacement prediction results;
fig. 10 shows the result of the landslide accumulated displacement prediction.
Detailed Description
In this embodiment, displacement data of the yaanbao landslide monitoring point provided by the second institute of middle iron is used, and ground surface deformation monitoring data in a period from 1 day of 12 months in 2020 to 31 days of 3 months in 2022 are selected as an example, and the technical scheme of the invention is described in detail with reference to the accompanying drawings of the specification.
1. Landslide displacement decomposition
Under the control of geological conditions of the difficult mountain area, landslide displacement increases along with the increment of time, and the landslide displacement is expressed as an approximate single increasing function which changes along with time, so that the long-term deformation trend of the landslide is reflected. Under the action of periodic external inducement (such as rainfall, etc.), landslide displacement is expressed as periodic fluctuation change. Under the action of a random external inducement (such as earthquake force), landslide displacement is represented as a near white noise sequence. On the basis of the above, the time sequence addition model for obtaining the landslide displacement is as follows:
Y t ＝T t +P t +R t (3)
wherein Y is t Is the landslide displacement monitoring value at the moment T t For the trend term displacement at time t, P t For the period term displacement at time t, R t Is the random term displacement at time t.
The variational modal decomposition method (VMD) is a completely non-recursive, adaptive modal variational and signal decomposition algorithm. The real-valued input signal is decomposed into a plurality of eigenvalue function (intrinsic mode function, IMF) components with specific sparse characteristics through construction of a variation problem and iterative solution, and the method has the advantage that the number of the modal components can be determined in advance. The complex signal can be decomposed by VMD into several eigenmode functions of practical physical significance.
In the eigenmode function of the variation mode decomposition method, the quadratic penalty factor alpha is introduced to convert the eigenmode function into an unconstrained variation model, so that the accuracy of the decomposition result can be improved; in addition, a rise time step parameter τ needs to be set; the secondary penalty factor alpha and the rising time step parameter tau need to be set manually in advance, and the randomness and uncertainty of the manual setting tend to influence the accuracy of the VMD decomposition result. . Therefore, how to accurately and efficiently select parameter combinations is the key to using the VMD algorithm for time series decomposition.
Genetic algorithm is an optimizing algorithm simulating natural selection and genetic evolution of organisms, and generally comprises 3 genetic operators: selection, crossover and mutation. The GA is a population intelligent optimization algorithm, has better global searching capability, can optimize the parameter combination which is difficult to select in the VMD, avoids the interference of subjective factors, and automatically searches the combination of the optimal parameters.
The invention utilizes a genetic algorithm to carry out parameter optimization on the secondary punishment factor alpha and the rising time step parameter tau, and the specific steps are as follows:
1. setting the population individual number K as 3, the optimizing range of the penalty factor alpha as [0.1, 1000], the optimizing range of the ascending time step parameter tau as [0,1], and the maximum iteration number as 20;
2. selecting individuals by adopting a competitive bidding method according to the fitness;
3. the selected individuals generate offspring through crossover and mutation operations;
4. calculating the fitness of the offspring;
5. merging individuals of the offspring and the previous generation, and sorting according to fitness;
6. if the number of iterations is more than 20 or 10, the optimal fitness of the iterations is kept unchanged, ending the algorithm, and obtaining an optimized secondary penalty factor alpha and an ascending time step parameter tau; otherwise, repeating the steps 2-6. The final optimizing results are shown in table 1.
Table 1 GA results of optimizing VMD parameters
The best fitness value change process in the iterative process is shown in fig. 2.
The result obtained by the optimization is brought into a VMD algorithm, the accuracy of the decomposition effect of the data is ensured, the historical landslide displacement is decomposed into trend item displacement, periodic item displacement and random item displacement at each moment of the history in a time sequence, and the decomposition result is shown in figure 3.
2. External incentive selection
There are many external causes of landslide, such as rainfall, air temperature, vegetation coverage, geological structure, geological condition, topography, etc., but the importance of each external cause to landslide disasters is not completely consistent for landslides along the railway in difficult mountainous areas. In order to better analyze landslide external inducement in a difficult mountain area, the main external inducement of the landslide is excavated by adopting a gray correlation model, and the hysteresis characteristics of the external inducement are combined to obtain the sequence of comprehensively evaluating the landslide external inducement from high to low, wherein the sequence comprises the following steps: stratum lithology, daily rainfall, geological structure, gradient, vegetation coverage index, elevation, soil water content, surface coverage type, water system distance, building distance and road distance. And removing the internal factors, and selecting two external factors of the daily rainfall and the soil water content as main factors influencing the landslide period term displacement and the random term displacement.
3. External incentive time series decomposition
Considering that the external inducement has regular periodic information and certain randomness, the two external inducement of rainfall and soil water content are decomposed by a variation modal decomposition method after parameter optimization to obtain periodic term components and random term components of the two external inducement at each moment of history. Setting k=2, the parameter optimizing process is the same as the previous one, and finally obtaining the optimized secondary penalty factor alpha= 65.318 and the rise time step parameter tau=0.53. The 2 components correspond in turn to the periodic and random time series of the external inducements, as shown in fig. 4.
4. Trend term displacement prediction
The trend item displacement reflects the change trend of displacement, is influenced by the characteristics of rock and soil, potential energy, constraint conditions and the like, and shows a monotonically increasing curve along with time. The invention adopts a 3-order polynomial least square method to fit and predict the trend displacement, and the process can be described as follows by a mathematical method:
y＝at 3 +bt 2 +ct+d (2)
wherein y is the displacement of the trend term at the moment t; a. b, c and d are polynomial coefficients and the specific fitting parameters are shown in table 2.
Table 2 trend term displacement fitting parameters
In this embodiment, the trend term displacement prediction result is shown in fig. 5.
5. Periodic term displacement prediction
Under the action of rainfall and soil water content, landslide displacement shows periodic fluctuation change, and periodic term displacement reflects the change of displacement along with periodic external inducement. The invention adopts a bidirectional gating circulation unit model based on an attention mechanism to dynamically predict the periodic item displacement in multiple factors.
The invention adopts Bi-GRU neural network as a backbone network, and captures the characteristics of the displacement vector of the input periodic item from the front direction and the back direction respectively; in addition, in order to obtain more accurate weight values, the invention inserts an attention module on the basis of a backbone network.
Specifically, the bidirectional gating circulation unit model based on the attention mechanism comprises a back propagation module, a forward propagation module and an attention module as shown in fig. 6; inputting random term displacement of the T time period into a counter propagation module, inputting an output result of the module into an attention module, and processing the output result by the attention module to obtain a weight vector matrix F; meanwhile, the random term component of the T time period is input into a forward propagation module to obtain a hidden vector h of the T time period t Finally, the hidden vector h of the T time period t And the weight vector matrix F obtained by the attention mechanism module is used as input, and the predicted value of the landslide period term displacement in the time period T+1 is output through a sigmoid function. In this embodiment, the result of the periodic term displacement prediction is shown in fig. 6.
Landslide displacement prediction result average error rate is 8.3% based on the bidirectional gating circulating unit model based on the attention mechanism; the average error rate of the landslide displacement prediction result of the gate control circulation unit is 12.1%, and the result shows that the landslide displacement prediction error is reduced to a certain extent by the method provided by the invention.
6. Random term displacement prediction
Under the effect of the random external inducement, landslide displacement is represented as a near white noise sequence, and random term displacement reflects the change of displacement along with the random external inducement. The invention predicts the random term displacement by adopting a nonlinear autoregressive neural network.
The nonlinear autoregressive neural network is a dynamic neural network with feedback and memory capabilities, and has obvious advantages compared with the traditional BP neural network. The method has the advantages that the data of the previous stage is reserved and added to the data prediction of the next stage by introducing the delay unit and the output feedback, so that the reserved system information is more complete and has more excellent dynamic properties. The structure of the nonlinear autoregressive neural network is shown in fig. 8.
In FIG. 8, y (t) represents the predicted value of the random term displacement at time t, x (t) represents the actual value obtained by combining the random term displacement at time t and the random term component, and d x And d y Delay orders, ω, of x (t) and y (t), respectively ab Representing weights, ω, between input layer and hidden layer bi Representing weights between the hidden layer and the output layer. The nonlinear autoregressive neural network model can be expressed as a functional relationship:
y(t)＝f[y(t-1),…,y(t-d y ),x(t-1),…,x(t-d x )] (3)
wherein f is [.]To activate the function, y (t-1), …, y (t-d) y ) Predictive value representing random term displacement at the first d moments, x (t-1), …, x (t-d) x ) Representing the actual value of the combination of the random term displacement and the random term component at the first d moments.
The nonlinear autoregressive neural network training expected output is known in the prediction process, so that a Series-Parallel neural network mode can be established, namely the expected output is directly fed back to the input end, thereby reducing the training time and improving the prediction precision.
The random term displacement prediction result is shown in fig. 9.
7. Superimposing the displacement component predictors to obtain a cumulative displacement predictor
According to the time sequence addition principle, the prediction results of the trend term displacement, the period term displacement and the random term displacement are added to obtain the landslide displacement prediction value, and the prediction result is shown in fig. 10.
According to the method, the surface displacement time sequence of the landslide monitoring area is decomposed into trend item displacement, periodic item displacement and random item displacement, and the external inducement time sequence is decomposed into periodic item components and random item components. And the different displacement components are predicted independently, so that the accuracy of the landslide monitoring area earth surface displacement time sequence prediction is improved.

Claims (4)

1. A landslide displacement prediction method based on an attention mechanism and a bidirectional gating circulating unit is characterized by comprising the following steps:
s1, landslide displacement time sequence decomposition:
based on a landslide displacement time sequence addition model shown in a formula (1), performing time sequence decomposition on historical landslide displacement into trend term displacement, periodic term displacement and random term displacement;
Y t ＝T t +P t +R t (1)
wherein Y is t Is the landslide displacement monitoring value at the moment T t For the trend term displacement at time t, P t For the period term displacement at time t, R t The random term displacement at the moment t;
s2, external inducement time sequence decomposition:
adopting a gray correlation model to mine the correlation between the historical displacement components and the influence factors, and selecting two influence factors with the greatest influence on landslide displacement as external inducements by combining the hysteresis characteristics of the influence factors; performing time sequence decomposition on the two external inducements to obtain a periodic term component and a random term component of the two external inducements;
s3, predicting trend item displacement:
fitting and predicting the displacement of the trend term by using a polynomial least square method to obtain a predicted value of the displacement of the trend term;
s4, predicting the displacement of the period term:
taking the Bi-GRU neural network as a main network, inserting an attention module on the basis of the main network, and obtaining a bidirectional gating circulation unit model based on an attention mechanism;
inputting the periodic term displacement and the periodic term component into a bidirectional gating circulation unit model based on an attention mechanism to obtain a predicted value of the periodic term displacement;
s5, predicting random term displacement:
combining the random term displacement and the random term component by adopting a nonlinear autoregressive neural network to obtain a predicted value of the random term displacement;
s6, calculating a landslide displacement predicted value:
and according to the vectorization of the displacement, adding and summing the predicted value of the displacement of the trend item, the predicted value of the displacement of the periodic item and the predicted value of the displacement of the random item to obtain the predicted value of the landslide displacement.
2. The landslide displacement prediction method based on the attention mechanism and the bi-directional gating cycle unit according to claim 1, wherein in step S4, the bi-directional gating cycle unit model based on the attention mechanism comprises a back propagation module, a forward propagation module and an attention module; inputting random term displacement of the T time period into a counter propagation module, inputting an output result of the module into an attention module, and processing the output result by the attention module to obtain a weight vector matrix F; meanwhile, the random term component of the T time period is input into a forward propagation module to obtain a hidden vector h of the T time period t Finally, the hidden vector h of the T time period t And the weight vector matrix F obtained by the attention mechanism module is used as input, and the predicted value of the T+1 time landslide period term displacement is output through a sigmoid function.
3. The landslide displacement prediction method based on the attention mechanism and the bidirectional gating circulation unit as set forth in claim 2, wherein in the steps S1 and S2, the genetic algorithm is used to perform parameter optimization on the secondary penalty factor α and the rise time step parameter τ in the variation mode decomposition method, and then the variation mode decomposition method after parameter optimization is used to perform time sequence decomposition on the historical landslide displacement and the external inducement;
the parameter optimizing mode is as follows:
A. setting the population individual number as K, the maximum iteration number is greater than 20, the optimizing range of the punishment factor alpha is [0.1, 1000], and the optimizing range of the rising time step parameter tau is [0,1]; wherein, when the landslide displacement is subjected to time series decomposition, the K value is 3, and when the external inducement is subjected to time series decomposition, the K value is 2;
B. selecting individuals by adopting a competitive bidding method according to the fitness;
C. the selected individuals generate offspring through crossover and mutation operations;
D. calculating the fitness of the offspring;
E. merging individuals of the offspring and the previous generation, and sorting according to fitness;
F. if the number of iterations is more than 20 or the optimal fitness of 10 iterations is kept unchanged, ending the algorithm to obtain optimized variation modal decomposition parameters; otherwise, repeating the steps B-F.
4. A landslide displacement prediction method based on attention mechanism and bi-directional gating cyclic unit as claimed in claim 2 or 3 wherein in step S5 the functional relationship of the nonlinear autoregressive neural network model is expressed as:
y(t)＝f[y(t-1),…,y(t-d y ),x(t-1),…,x(t-d x )] (3)
wherein f is [.]To activate the function, y (t) represents the predicted value of the random term displacement at time t, d x And d y The delay order of x (t) and y (t), y (t-1), …, y (t-d), respectively y ) Predictive value representing random term displacement at the first d moments, x (t-1), …, x (t-d) x ) Representing the actual value of the combination of the random term displacement and the random term component at the first d moments.
CN202310774955.1A 2023-06-28 2023-06-28 Landslide displacement prediction method based on attention mechanism and bidirectional gating circulating unit Pending CN116821586A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310774955.1A CN116821586A (en) 2023-06-28 2023-06-28 Landslide displacement prediction method based on attention mechanism and bidirectional gating circulating unit

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310774955.1A CN116821586A (en) 2023-06-28 2023-06-28 Landslide displacement prediction method based on attention mechanism and bidirectional gating circulating unit

Publications (1)

Publication Number Publication Date
CN116821586A true CN116821586A (en) 2023-09-29

Family

ID=88128788

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310774955.1A Pending CN116821586A (en) 2023-06-28 2023-06-28 Landslide displacement prediction method based on attention mechanism and bidirectional gating circulating unit

Country Status (1)

Country Link
CN (1) CN116821586A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117132007A (en) * 2023-10-29 2023-11-28 四川轻化工大学 Landslide deformation rate prediction method based on dynamic series PSO-BILSTM

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117132007A (en) * 2023-10-29 2023-11-28 四川轻化工大学 Landslide deformation rate prediction method based on dynamic series PSO-BILSTM
CN117132007B (en) * 2023-10-29 2024-03-08 四川轻化工大学 Landslide deformation rate prediction method based on dynamic series PSO-BILSTM

Similar Documents

Publication Publication Date Title
CN109492814A (en) A kind of Forecast of Urban Traffic Flow prediction technique, system and electronic equipment
Esmaeili et al. Multiple regression, ANN and ANFIS models for prediction of backbreak in the open pit blasting
Huang et al. An integrated neural-fuzzy-genetic-algorithm using hyper-surface membership functions to predict permeability in petroleum reservoirs
Kadkhodaie-Ilkhchi et al. A committee machine with intelligent systems for estimation of total organic carbon content from petrophysical data: An example from Kangan and Dalan reservoirs in South Pars Gas Field, Iran
CN104134351A (en) Short-term traffic flow predicting method
CN108181591A (en) A kind of Forecasting Methodology of the SOC value of battery based on Speed Controlling Based on Improving BP Neural Network
Yi et al. Intelligent prediction of transmission line project cost based on least squares support vector machine optimized by particle swarm optimization
CN116821586A (en) Landslide displacement prediction method based on attention mechanism and bidirectional gating circulating unit
Ahumada et al. Fuzzy logic-based attenuation relationships of strong motion earthquake records
CN113326660B (en) Tunnel surrounding rock extrusion deformation prediction method based on GA-XGboost model
Ghajar et al. An adaptive network-based fuzzy inference system for rock share estimation in forest road construction
CN112884236B (en) Short-term load prediction method and system based on VDM decomposition and LSTM improvement
Sugiharti et al. Predictive evaluation of performance of computer science students of unnes using data mining based on naÏve bayes classifier (NBC) algorithm
CN111709072B (en) Underground vibration amplitude parameter prediction method
CN113610945A (en) Ground stress curve prediction method based on hybrid neural network
CN111058840A (en) Organic carbon content (TOC) evaluation method based on high-order neural network
CN112766603A (en) Traffic flow prediction method, system, computer device and storage medium
CN115130375A (en) Rock burst intensity prediction method
CN112464996A (en) Intelligent power grid intrusion detection method based on LSTM-XGboost
He et al. Principles of machine learning and its potential applications in climate prediction
Robati et al. Inflation rate modeling: Adaptive neuro-fuzzy inference system approach and particle swarm optimization algorithm (ANFIS-PSO)
CN115018193A (en) Time series wind energy data prediction method based on LSTM-GA model
Li et al. Prediction and application of porosity based on support vector regression model optimized by adaptive dragonfly algorithm
Nikafshan Rad et al. Modification of rock mass rating system using soft computing techniques
Xu et al. Stability Analysis of Geotechnical Landslide Based on GA-BP Neural Network Model

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination