CN112101660A - Rainfall type landslide displacement prediction model and method based on staged attention mechanism - Google Patents

Rainfall type landslide displacement prediction model and method based on staged attention mechanism Download PDF

Info

Publication number
CN112101660A
CN112101660A CN202010968140.3A CN202010968140A CN112101660A CN 112101660 A CN112101660 A CN 112101660A CN 202010968140 A CN202010968140 A CN 202010968140A CN 112101660 A CN112101660 A CN 112101660A
Authority
CN
China
Prior art keywords
displacement
rainfall
prediction
model
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.)
Granted
Application number
CN202010968140.3A
Other languages
Chinese (zh)
Other versions
CN112101660B (en
Inventor
唐菲菲
唐天俊
王锡斐
朱永亮
王云云
沈诚
宋平
李润杰
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.)
Chongqing Jiaotong University
Original Assignee
Chongqing Jiaotong University
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 Chongqing Jiaotong University filed Critical Chongqing Jiaotong University
Priority to CN202010968140.3A priority Critical patent/CN112101660B/en
Publication of CN112101660A publication Critical patent/CN112101660A/en
Application granted granted Critical
Publication of CN112101660B publication Critical patent/CN112101660B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • 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
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • 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
    • G06N3/045—Combinations of networks
    • 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
    • G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
    • 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
    • G06N3/08—Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Computational Linguistics (AREA)
  • Health & Medical Sciences (AREA)
  • Biophysics (AREA)
  • Artificial Intelligence (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Biomedical Technology (AREA)
  • Business, Economics & Management (AREA)
  • Economics (AREA)
  • Human Resources & Organizations (AREA)
  • Strategic Management (AREA)
  • Development Economics (AREA)
  • Game Theory and Decision Science (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Marketing (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Tourism & Hospitality (AREA)
  • General Business, Economics & Management (AREA)
  • Testing Or Calibration Of Command Recording Devices (AREA)

Abstract

The invention provides a rainfall type landslide displacement prediction model based on a stage attention mechanism, which comprises the following steps: the data preprocessing layer is used for decomposing the rainfall type landslide accumulated displacement into a trend item and a period item according to a time sequence; the data prediction layer is used for predicting a trend item according to the displacement of the adjacent points to obtain a trend item displacement prediction value; the system is used for predicting the period item according to rainfall to obtain a period item displacement predicted value; the system is also used for merging the trend item displacement predicted value and the period item displacement predicted value to obtain an accumulated displacement predicted value of the monitoring point location; the model verification layer is used for optimizing parameters of the prediction model according to the difference value of the accumulated displacement prediction value and the displacement measurement value; a method for predicting rainfall type landslide displacement using the prediction model is also provided. The method can solve the technical problems that the prediction accuracy of an LSTM landslide displacement prediction model on rainfall landslide is not high enough and the long-time sequence prediction reliability is unstable in the prior art.

Description

Rainfall type landslide displacement prediction model and method based on staged attention mechanism
Technical Field
The invention relates to the technical field of geological disaster prediction, in particular to a rainfall type landslide displacement prediction model and method based on a periodic attention mechanism.
Background
The landslide is a natural phenomenon that soil or rock mass on a slope slides downwards along the slope integrally or dispersedly under the action of gravity along a certain weak surface or a weak zone under the influence of factors such as river scouring, underground water activity, rainwater immersion, earthquake, artificial slope cutting and the like. With the development of pattern recognition and machine learning theory, the nonlinear theory is widely applied to landslide displacement prediction, and the landslide displacement prediction theory and method are rapidly developed. In the prior art, CN110470259A provides a landslide displacement dynamic prediction method based on LSTM, which includes, first, constructing a landslide displacement online monitoring system, performing real-time monitoring to obtain complete displacement data in one period, performing outlier elimination processing on the collected displacement data through a 3-algorithm, and performing normalization; then, establishing an LSTM landslide displacement prediction model and training; and finally, inputting the obtained normalized data serving as the input quantity of the model into a landslide displacement prediction model, and processing the input data by the prediction model to realize the prediction of the landslide displacement in the future period.
However, according to the statistical data of the national geological report 2019, about 90% of landslides in China are induced by rainfall, and rainfall type landslides, namely landslides caused by rainfall action, frequently occur in Sichuan, loess areas, southeast coastal areas and the like in China; the occurrence frequency is high and the prediction difficulty is high. Actual monitoring shows that rainfall of rainfall type landslide in early stage of landslide is concentrated and fluctuates greatly, so that rainfall accumulated in early stage of different time sequences is required to be selected during LSTM modeling, and the rainfall is used as a characteristic value to model a landslide displacement prediction model. However, in the above-described technical solution, the reliability of the long-time sequence prediction is unstable, and the accuracy of the prediction of the rainfall type landslide is not high enough.
Disclosure of Invention
Aiming at the defects in the prior art, the invention provides a rainfall type landslide displacement prediction model and a rainfall type landslide displacement prediction method based on a periodic attention mechanism, and aims to solve the technical problems that the rainfall type landslide prediction precision is not high enough and the long-time sequence prediction reliability is unstable in the LSTM landslide displacement prediction model in the prior art.
The technical scheme adopted by the invention is as follows:
in a first aspect, a rainfall type landslide displacement prediction model based on a staged attention mechanism is provided;
in a first implementation, a rainfall-type landslide displacement prediction model based on a episodic attention mechanism includes:
the system comprises a data preprocessing layer, a data prediction layer and a model verification layer;
the data preprocessing layer is used for decomposing the rainfall type landslide accumulated displacement into a trend item and a period item according to a time sequence;
the data prediction layer is used for predicting a trend item according to the displacement of the adjacent points to obtain a trend item displacement prediction value; the system is used for predicting the period item according to rainfall to obtain a period item displacement predicted value; the system is also used for merging the trend item displacement predicted value and the period item displacement predicted value to obtain an accumulated displacement predicted value of the monitoring point location;
and the model verification layer is used for optimizing the parameters of the prediction model according to the difference value of the accumulated displacement prediction value and the displacement measurement value.
With reference to the first implementation manner, in a second implementation manner, the trend term displacement prediction value and the period term displacement prediction value are combined, and the following formula is adopted for calculation:
Xt＝St+Vt (1)
in the above formula, XtTo accumulate the displacement, StAs a trend term, VtIs a period term.
With reference to the first implementation manner, in a third implementation manner, a polynomial regression method is used to predict the trend term, and the following formula is used to calculate:
St＝a1S1+a2S2+a3S3+…+anSn-d
in the above formula, StTo monitor the trend term prediction values of the point locations, S1～SnIs the measured value of the trend term of the adjacent point, a1～anAnd d is a coefficient.
With reference to the first implementation manner, in a fourth implementation manner, a bidirectional long-time and short-time memory neural network based on a periodic attention mechanism is used for predicting the periodic term.
With reference to the fourth implementable manner, in a fifth implementable manner, based on the bidirectional long-short term memory neural network of the stepwise attention mechanism, the bidirectional long-short term memory neural network is used as an encoder, the long-short term memory neural network is used as a decoder, and the stepwise attention mechanism is introduced into the encoder and the decoder.
With reference to the fifth implementation manner, in a sixth implementation manner, a periodic attention mechanism is introduced in the encoder and the decoder, and the attention mechanism layer processing procedure at the time t satisfies the following formula:
Figure BDA0002683096820000031
Figure BDA0002683096820000032
Figure BDA0002683096820000033
in the above formula: score is a function for measuring the importance degree of the predicted value at the t-1 moment to the hidden layer output at the t moment;
Figure BDA0002683096820000034
outputting a vector for the hidden state of the Bi-LSTM at the time t; y ist-1∈RnFor the output value y of the model at the time t-1t-1The vector obtained after the broadcasting mechanism is carried out, and the model training provides an initial condition y0，y0I.e. the initial value of the landslide displacement; ws∈Rn×2n、WT∈Rn×nTraining a weight matrix for an attention training layer; bs∈RnIs a bias vector; alpha is alphatScoring a weight matrix for attention; v is the attention mechanism layer output feature vector.
With reference to the fourth implementable manner, in a seventh implementable manner, the bidirectional long-term and short-term memory neural network based on the episodic attention mechanism includes: the rolling cumulative rainfall formed by the daily rainfall conversion.
With reference to the fourth implementable manner, in the eighth implementable manner, a bidirectional long-and-short-term memory neural network based on a staged attention mechanism is used, and during training, a Huber Loss is used for gradient descent parameter optimization.
In a second aspect, a rainfall type landslide displacement prediction method based on a staged attention mechanism is provided, and a rainfall type landslide displacement prediction model based on the staged attention mechanism is used for predicting rainfall type landslide displacement, and specifically includes the following steps:
acquiring displacement values and daily rainfall of adjacent point positions around the monitoring point position;
inputting the displacement value of the adjacent point and the daily rainfall into a rainfall type landslide displacement prediction model;
the rainfall type landslide displacement prediction model calculates a trend item displacement prediction value through a near point displacement value; calculating to obtain a period item displacement predicted value through daily rainfall;
and the rainfall type landslide displacement prediction model combines the trend item displacement prediction value and the period item displacement prediction value to obtain a rainfall type landslide displacement accumulated displacement prediction value.
In a third aspect, an electronic device is provided, including:
one or more processors;
a storage device for storing one or more programs,
when executed by one or more processors, cause the one or more processors to implement a method for rainfall-type landslide displacement prediction based on a episodic attention mechanism.
In a fourth aspect, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements a method for rainfall-type landslide displacement prediction based on a step-wise attention mechanism.
According to the technical scheme, the beneficial technical effects of the invention are as follows:
1. and decomposing the accumulated displacement of the rainfall type landslide into a trend item and a period item according to the evolution characteristic of the rainfall type landslide. The trend item has strong linear correlation with different time sequence adjacent points, and a polynomial regression method is selected for prediction; the periodic term and the early rainfall with different time sequences have approximately same fluctuation, and the Attention Based Bi-LSTM (bidirectional long-short time memory neural network Based on the periodic Attention mechanism) is selected for prediction, so that higher prediction accuracy can be obtained.
The Attention-Based Bi-LSTM uses Bi-LSTM (bidirectional long-short time memory neural network) as an encoder and uses LSTM (long-short time memory neural network) as a decoder. A stage attention mechanism is introduced into an encoder and a decoder, so that the early stage accumulated displacement of different time spans can be automatically captured, the key information in accumulated rainfall of different time sequences can be better obtained, the key data with large influence on the displacement of the current period item can be decoded and predicted, and the defects that the LSTM network adopts the same parameters and ignores the details of each time sequence prediction are overcome.
3. According to the technical scheme of the embodiment, when the extension Based Bi-LSTM is used for predicting the displacement condition of the rainfall type landslide, the prediction curve is approximately attached to the real value curve, the MAE is 0.088, the MSE is 0.042, and a more accurate prediction effect can be obtained.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below. Throughout the drawings, like elements or portions are generally identified by like reference numerals. In the drawings, elements or portions are not necessarily drawn to scale.
FIG. 1 is a schematic diagram of a prediction model architecture according to embodiment 1 of the present invention;
FIG. 2 is a schematic view of the Attention Based Bi-LSTM structure of embodiment 1 of the present invention;
FIG. 3 is a flowchart of a method of example 2 of the present invention;
FIG. 4 is a diagram showing the predicted results of trend terms in example 2 of the present invention;
FIG. 5 is a graph showing the comparison of the predicted results of the Attention Based Bi-LSTM and LSTM period terms in example 2 of the present invention;
FIG. 6 is a diagram illustrating the accumulated displacement prediction result of the monitoring point location according to the present invention.
Detailed Description
Embodiments of the present invention will be described in detail below with reference to the accompanying drawings. The following examples are only for illustrating the technical solutions of the present invention more clearly, and therefore are only examples, and the protection scope of the present invention is not limited thereby.
It is to be noted that, unless otherwise specified, technical or scientific terms used herein shall have the ordinary meaning as understood by those skilled in the art to which the invention pertains.
Example 1
The invention provides a rainfall type landslide displacement prediction model based on an attention mechanism, which comprises the following steps: the system comprises a data preprocessing layer, a data prediction layer and a model verification layer;
the data preprocessing layer is used for decomposing the rainfall type landslide accumulated displacement into a trend item and a period item according to a time sequence;
the data prediction layer is used for predicting a trend item according to the displacement of the adjacent points to obtain a trend item displacement prediction value; the system is used for predicting the period item according to rainfall to obtain a period item displacement predicted value; the system is also used for merging the trend item displacement predicted value and the period item displacement predicted value to obtain an accumulated displacement predicted value of the monitoring point location;
and the model verification layer is used for verifying the robustness of the prediction model and adjusting the prediction model according to the difference value of the displacement predicted value and the displacement measured value.
The working principle of example 1 is explained in detail below:
in the present embodiment, the rainfall type landslide displacement is predicted using a rainfall type landslide displacement prediction model. The framework of the rainfall type landslide displacement prediction model is shown in fig. 1 and comprises a data preprocessing layer, a data prediction layer and a model verification layer.
The data preprocessing layer is mainly used for time series decomposition of rainfall type landslide accumulated displacement, decomposing the rainfall type landslide accumulated displacement into a trend item and a period item through the time series decomposition. The trend item is mainly influenced by internal geological conditions, geological structures and the like, and the influence is long-term, so that the trend item presents a monotone increasing trend. The period item is mainly influenced by external landslide factors such as rainfall, reservoir water level and the like, and the acting time of the period item on the landslide is short. Decomposing the rainfall type landslide accumulated displacement according to a time sequence according to the following formula:
Xt＝St+Vt (1)
in the above formula (1), XtTo accumulate the displacement, StAs a trend term, VtIs a period term.
The data prediction layer is used for trend item prediction and period item prediction, and different methods are respectively adopted for prediction.
The trend term is predicted by polynomial regression and calculated according to the following formula:
St＝a1S1+a2S2+a3S3+…+anSn-d (2)
in the above formula (2), StIn order to monitor the trend term prediction value of the point location,S1～Snto monitor the trend term actual measurement values of adjacent points in the vicinity of a point location, a1～anAnd d is a coefficient. When training the model, selecting the measured value (S) of the monitoring point location within a certain time periodt) N number of adjacent point actual measurement values (S) in the vicinity of the monitored point1～Sn). From the measured values, the coefficient a can be calculated by the least square method1～anAnd the value of d. When the model is used for prediction, the displacement measured values of adjacent points on the right side of the equation are substituted according to the value as a coefficient, and the displacement predicted value of the trend term on the left side of the equation is calculated. Therefore, the data prediction layer can obtain the trend item displacement prediction value according to the adjacent point displacement measured value.
Early rainfall is the main reason for inducing rainfall type landslide instability, however, landslide displacement evolution has a multi-fractal characteristic, current time sequence landslide displacement increase caused by the early rainfall in multiple time sequences is difficult to determine artificially, early accumulated rainfall and period items are time sequence data, and the LSTM can be used as a tool for processing the time sequence, but the hidden state in the LSTM shares parameters, so that detail information of the time sequence is lost. Specifically, in the stage of the landslide acceleration displacement, the correlation between the rainfall factor (the amount of rainfall) and the period term changes obviously, the correlation between the rainfall factor and the period term in different stages is different, and the correlation between the rainfall factor and the period term gradually increases along with the increase of the acceleration of the landslide accumulated displacement. Therefore, how to identify the correlation change in different stages in the model becomes difficult, in the traditional machine learning algorithm, the correlation change between the characteristic variable and the prediction sequence is difficult to capture, and the model prediction has numerical value mutation after a certain time node. To solve the above problem, in this embodiment, the periodic term is predicted by using Attention-Based Bi-LSTM (bidirectional long-short time memory neural network Based on the periodic Attention mechanism). Performing model training by using rolling accumulated rainfall of different time periods (such as 10 days, 15 days, 30 days, 45 days and 60 days) as sample characteristics; the rolling accumulated rainfall is obtained by accumulating and adding the daily rainfall according to different time period values. In Attention-Based Bi-LSTM, Bi-LSTM (bidirectional long-short memory neural network) is used as an encoder, LSTM (long-short memory neural network) is used as a decoder, and a mechanism of Attention is introduced in the encoder and the decoder, as shown in fig. 2.
The Bi-LSTM is used as an encoder, time sequence data characteristics in two directions can be integrated, front and back time sequence characteristics of a training set can be extracted more effectively, an influence factor sequence and an accumulated displacement sequence are mapped to the same vector space, and the forward LSTM and the backward LSTM are used for predicting the current output of the time sequence data in the two directions. The input processing procedures of the forward LSTM and the backward LSTM at the time t satisfy the following formulas:
Figure BDA0002683096820000071
Figure BDA0002683096820000072
Figure BDA0002683096820000073
in the above equations (7), (8), (9):
Figure BDA0002683096820000074
outputting vectors for the n-th layer hidden state of the forward LSTM at the time t;
Figure BDA0002683096820000075
outputting a vector for the hidden state of the nth layer of the backward LSTM at the time t;
Figure BDA0002683096820000076
inputting a vector for a time sequence of an input layer;
Figure BDA0002683096820000077
is composed of
Figure BDA0002683096820000078
And
Figure BDA0002683096820000079
merging the hidden states of the Bi-LSTM at the moment t in a splicing mode; and | represents a splicing operation.
Attention mechanism is introduced in the encoder and the decoder, because the landslide displacement data belongs to a continuous time sequence, and the future data needs to be predicted by using historical data. Therefore, in this embodiment, according to the characteristics of the rainfall type landslide displacement data, the attention mechanism is improved, the time sequence is introduced into the attention mechanism to form a staged attention mechanism, the encoder outputs the feature vector v at the time t according to the predicted value at the time t-1, the attention distribution probability calculation is performed, and the feature vector and the attention distribution probability are integrated to generate the feature sequence related to the predicted value at the time sequence. The attention mechanism layer satisfies the following formula in the processing process at the moment t:
Figure BDA0002683096820000081
Figure BDA0002683096820000082
Figure BDA0002683096820000083
in the above equations (7), (8), (9): score is a scoring function for measuring the importance degree of the predicted value at the t-1 moment to the hidden layer output at the t moment;
Figure BDA0002683096820000084
outputting a vector for the hidden state of the Bi-LSTM at the time t; y ist-1∈RnFor the output value y of the model at the time t-1t-1The vector obtained after the broadcasting mechanism is carried out, and the model training provides an initial condition y0，y0I.e. the initial value of the landslide displacement; ws∈Rn×2n、WT∈Rn×nTraining a weight matrix for an attention training layer；bs∈RnIs a bias vector; alpha is alphatScoring a weight matrix for attention; v is the attention mechanism layer output feature vector.
A stage attention mechanism is introduced into an encoder and a decoder, so that key information in different time sequences accumulated rainfall can be better captured, the decoding prediction is carried out on key data with large influence on current period item displacement, and the defects that an LSTM network adopts the same parameters and ignores the details of each time sequence prediction are overcome.
The LSTM as a decoder can fully play the advantages of time series processing and effectively record time series dependence information. Inputting the feature sequence after attention mechanism allocation into an LSTM for decoding, wherein the processing process of the LSTM neural network model at the time t meets the following formula:
ft＝σ(Wf·[ht-1,v]+bf) (10)
it＝σ(Wf·[ht-1,v]+bf) (11)
Figure BDA0002683096820000085
Figure BDA0002683096820000086
ot＝tanh(Wo·[ht-1,v]+bo) (14)
ht＝ot⊙tanh(ct) (15)
in the above equations (10) to (15): the method comprises the following steps: f. oft、it、ot、ct、
Figure BDA0002683096820000091
Respectively a forgetting gate, an input gate, an output gate, a candidate memory cell and a memory cell value; h istIs a hidden layer vector; v is the feature vector output by the attention mechanism layer; wf、Wi、Wo∈Rn×2nAre respectively provided withThe weight matrixes are forgetting gates, input gates and output gates; bi、bf、bo∈RnIs a model offset vector; n is the LSTM hidden layer length; σ represents a Sigmoid activation function; an indication of a dot product.
In this embodiment, the number of hidden units in the Attention Based Bi-LSTM model is preferably 64.
In order to improve the robustness of model prediction when training Attention Based Bi-LSTM, daily rainfall is converted into rolling accumulated rainfall to be used as input, and Huber Loss is selected to carry out gradient descent parameter optimization. Huber Loss is a parametric Loss function for the regression problem, and compared with least square linear regression, Huber Loss reduces the punishment degree of outliers and improves the generalization capability of a prediction model. HuberLoss is defined as follows:
Figure BDA0002683096820000092
in the above equation (16): a is an error term, namely the difference between a real value and a predicted value; the Huberloss parameter is set to 1.
The output of the Attention Based Bi-LSTM is the period term displacement prediction value. Thus, the data prediction layer predicts the period item according to the rainfall, and the period item displacement prediction value can be obtained.
Then, a data prediction layer of the prediction model combines the trend item displacement prediction value and the period item displacement prediction value to obtain an accumulated displacement prediction value of the monitoring point location; when the combination is carried out, calculation is carried out according to the formula (1).
And the model verification layer is used for optimizing the parameters of the prediction model according to the difference value of the accumulated displacement prediction value and the displacement measurement value.
In this embodiment, the accumulated displacement of the rainfall type landslide is decomposed into a trend term and a period term according to the evolution characteristic of the rainfall type landslide. The trend item has strong linear correlation with different time sequence adjacent points, and a polynomial regression method is selected for prediction; the periodic item and the early rainfall of different time sequences have approximately same volatility, and the Attention Based Bi-LSTM is selected for prediction, so that higher prediction precision can be obtained, and the reliability of long-time sequence prediction is better.
Example 2
In this embodiment, a rainfall-type landslide displacement prediction method based on a step-by-step attention mechanism is provided, and a prediction model in embodiment 1 is used to predict rainfall-type landslide displacement, as shown in fig. 3, which specifically includes the following steps:
acquiring displacement values and daily rainfall of adjacent point positions around the monitoring point position;
inputting the displacement value of the adjacent point and the daily rainfall into a rainfall type landslide displacement prediction model;
the rainfall type landslide displacement prediction model calculates a trend item displacement prediction value through a near point displacement value; calculating to obtain a period item displacement predicted value through daily rainfall;
and the rainfall type landslide displacement prediction model combines the trend item displacement prediction value and the period item displacement prediction value to obtain an accumulated displacement prediction value of the rainfall type landslide monitoring point position.
The following are exemplified:
in order to evaluate the effect of the prediction accuracy of the model, MSE (mean square error), MAE (mean absolute error) and AE (absolute error) are used in this embodiment, and the calculation formulas are as follows:
Figure BDA0002683096820000101
Figure BDA0002683096820000102
Figure BDA0002683096820000103
in the above formulas (17), (18), and (19): y ismThe actual value is represented by the value of,
Figure BDA0002683096820000104
to representAnd (5) predicting the value.
When the displacement value of a certain monitoring point position in rainfall type landslide needs to be predicted, the method is adopted for predicting. Specifically, the landslide selected in the embodiment is located on the right bank of the Yangtze river, 425km from the upstream to the Chongqing city at the tail of the reservoir, 182km from the downstream to the first three gorges dam of the reservoir, 20km from the northeast to the Fengjie city. The elevation of the rear edge of the landslide is 390 plus 400m, the elevation of the water surface of the front edge of the Yangtze river in the dry period is 175m, the relative height difference is 215 plus 225m, and the gully has no perennial flowing water, generally keeps the water flow for 3-5 days after the precipitation, and is a rain source type gully. According to meteorological data, the average rainfall of the landslide region for many years is 1049.3-1145.1mm, the maximum daily rainfall reaches 199mm, a displacement region is located at the front edge of a bedrock slope or the slope region of the front edge of a terrace, the displacement region is a region with more concentrated slope water flow in the rainfall period, the infiltration of rainfall increases the volume weight of the slope soil body, the displacement of the shallow soil body is stimulated, and rainfall is an external main factor for causing the landslide displacement.
From the meteorological monitoring data, the rainfall of the landslide section is mainly concentrated in 4-8 months, the rainfall reaches the peak value in 7 months, and the rainfall amount is increased. And (3) selecting data of the section of landslide from No. 7/month 13 in 2015 to No. 12/month 20 in 2018 for analysis and modeling, wherein the data comprises monitoring point positions (1), adjacent point positions (3) arranged at the periphery of the monitoring point positions and daily rainfall data. 1285 pieces of original data, 1226 pieces of data with the time sequence of 2015 year 7 month 14 to 2018 year 11 month 20 are adopted in the training set, and 2018 year 11 month 20 to 2018 year 12 month 20 are adopted in the testing set.
Firstly, predicting a trend term, performing polynomial regression modeling by using training set data from No. 7/month 14 in 2015 to No. 11/month 20 in 2018, and predicting by using test set data from No. 11/month 20 in 2018 to No. 12/month 20 in 2018, wherein the prediction results are shown in Table 1 and FIG. 4:
TABLE 1
Figure BDA0002683096820000111
The periodic term is then predicted. When the period term is predicted, the early accumulated rainfall of the landslide region in 10 days, 15 days, 30 days, 45 days and 60 days counted from No. 7/14 in 2015 is respectively calculated and is respectively input into the calculation prediction calculation of LSTM and Attention Based Bi-LSTM as characteristic variables. In this embodiment, the number of hidden units of the Attention Based Bi-LSTM model is set to 64, the number of single LSTM is 4, the number of hidden units is 64, model training is performed with the same data for accuracy comparison, and 7000 times of Adam optimization are performed on the two models respectively. The predicted results for LSTM and Attention Based Bi-LSTM were compared, as shown in Table 2 and FIG. 5:
TABLE 2
Figure BDA0002683096820000112
And finally, combining the trend item predicted value and the period item displacement predicted value to obtain an accumulated displacement predicted value of the monitoring point location, and obtaining the accumulated displacement predicted value of the monitoring point location. Fig. 6 shows a comparison between the predicted accumulated displacement and the measured accumulated displacement.
As can be seen from the above comparative data and comparative figures,
because the single LSTM shares parameters with the hidden layer, the attention distribution probability calculation is not carried out on the influence factors of different time sequences, the model is fitted indifferently by using global data, the attention is dispersed, and the detail information is lost; the prediction result is mutated after 12 months and 9 days, LSTM is well predicted in short term, but the error is increased after the critical point is exceeded.
The extension-Based Bi-LSTM period item prediction curve is approximately fit with the true value curve, and the extension-Based Bi-LSTM keeps better generalization capability in the period item prediction process, and the generalization capability of the extension-Based Bi-LSTM is obviously superior to that of the LSTM.
According to the technical scheme of the embodiment, when the extension Based Bi-LSTM is used for predicting the displacement condition of the rainfall type landslide, the prediction curve is approximately attached to the real value curve, the MAE is 0.088, the MSE is 0.042, the prediction effect with higher precision is obtained, and the reliability of long-time sequence prediction is better.
Example 3
Provided is an electronic device including:
one or more processors;
a storage device for storing one or more programs,
when executed by one or more processors, the one or more programs cause the one or more processors to implement the method for rainfall-type landslide displacement prediction based on a stepwise attention mechanism provided in example 2.
Example 4
There is provided a computer-readable storage medium having stored thereon a computer program which, when executed by a processor, can implement the rainfall-type landslide displacement prediction method based on the stepwise attention mechanism provided in embodiment 2.
Finally, it should be noted that: the above examples are only intended to illustrate the technical solution of the present invention, but not to limit it; while the invention has been described in detail and with reference to the foregoing embodiments, it will be understood by those skilled in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some or all of the technical features may be equivalently replaced; such modifications and substitutions do not depart from the spirit and scope of the present invention, and they should be construed as being included in the following claims and description.

Claims (11)

1. A rainfall type landslide displacement prediction model based on a staged attention mechanism, comprising: the system comprises a data preprocessing layer, a data prediction layer and a model verification layer;
the data preprocessing layer is used for decomposing the rainfall type landslide accumulated displacement into a trend item and a period item according to a time sequence;
the data prediction layer is used for predicting a trend item according to the displacement of the adjacent points to obtain a trend item displacement prediction value; the system is used for predicting the period item according to rainfall to obtain a period item displacement predicted value; the system is also used for merging the trend item displacement predicted value and the period item displacement predicted value to obtain an accumulated displacement predicted value of the monitoring point location;
and the model verification layer is used for optimizing the parameters of the prediction model according to the difference value of the accumulated displacement prediction value and the displacement measurement value.
2. The model of claim 1, wherein the trend term displacement prediction value and the period term displacement prediction value are combined and calculated according to the following formula:
Xt＝St+Vt (1)
in the above formula, XtTo accumulate the displacement, StAs a trend term, VtIs a period term.
3. The model of claim 1, wherein the trend term is predicted by a polynomial regression method, and the model is calculated by the following formula:
St＝a1S1+a2S2+a3S3+…+anSn-d
in the above formula, StTo monitor the trend term prediction values of the point locations, S1～SnIs the measured value of the trend term of the adjacent point, a1～anAnd d is a coefficient.
4. The model of claim 1, wherein the model is based on a stage-wise attention mechanism for predicting displacement of a rainfall-type landslide: and predicting the periodic items by adopting a bidirectional long-time and short-time memory neural network based on a staged attention mechanism.
5. The model of claim 4, wherein the model is based on a stage-wise attention mechanism for predicting displacement of rainfall-type landslide: the bidirectional long-short time memory neural network based on the staged attention mechanism uses the bidirectional long-short time memory neural network as an encoder and uses the long-short time memory neural network as a decoder, and the staged attention mechanism is introduced into the encoder and the decoder.
6. The model of claim 5, wherein the model is based on a stage-wise attention mechanism for predicting displacement of rainfall-type landslide: a stage attention mechanism is introduced into the encoder and the decoder, and the attention mechanism layer satisfies the following formula in the processing process at the time t:
Figure FDA0002683096810000021
Figure FDA0002683096810000022
Figure FDA0002683096810000023
in the above formula: score is a function for measuring the importance degree of the predicted value at the t-1 moment to the hidden layer output at the t moment;
Figure FDA0002683096810000024
outputting a vector for the hidden state of the Bi-LSTM at the time t; y ist-1∈RnFor the output value y of the model at the time t-1t-1The vector obtained after the broadcasting mechanism is carried out, and the model training provides an initial condition y0，y0I.e. the initial value of the landslide displacement; ws∈Rn×2n、WT∈Rn×nTraining a weight matrix for an attention training layer; bs∈RnIs a bias vector; alpha is alphatScoring a weight matrix for attention; v is the attention mechanism layer output feature vector.
7. The model of claim 4, wherein the two-way long-term memory neural network comprises: the rolling cumulative rainfall formed by the daily rainfall conversion.
8. The model of claim 4, wherein the model is based on a stage-wise attention mechanism for predicting displacement of rainfall-type landslide: the bidirectional long-time and short-time memory neural network based on the staged attention mechanism adopts Huber Loss to carry out gradient descent parameter optimization during training.
9. A rainfall-type landslide displacement prediction method based on a staged attention mechanism is characterized in that the prediction model of claim 1 is used for predicting rainfall-type landslide displacement, and the method specifically comprises the following steps:
acquiring displacement values and daily rainfall of adjacent point positions around the monitoring point position;
inputting the displacement value of the adjacent point and the daily rainfall into a rainfall type landslide displacement prediction model;
the rainfall type landslide displacement prediction model calculates a trend item displacement prediction value through the adjacent point displacement value; calculating to obtain a period item displacement predicted value through the daily rainfall;
and the rainfall type landslide displacement prediction model combines the trend item displacement prediction value and the period item displacement prediction value to obtain a rainfall type landslide displacement accumulated displacement prediction value.
10. An electronic device, comprising:
one or more processors;
a storage device for storing one or more programs,
when executed by the one or more processors, cause the one or more processors to implement the method of staged attention mechanism based rainfall landslide displacement prediction of claim 9.
11. A computer-readable storage medium having stored thereon a computer program, wherein the program, when executed by a processor, implements the method for predicting rainfall-type landslide displacement based on the episodic attention mechanism of claim 9.
CN202010968140.3A 2020-09-15 2020-09-15 Rainfall type landslide displacement prediction model and method based on staged attention mechanism Active CN112101660B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010968140.3A CN112101660B (en) 2020-09-15 2020-09-15 Rainfall type landslide displacement prediction model and method based on staged attention mechanism

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010968140.3A CN112101660B (en) 2020-09-15 2020-09-15 Rainfall type landslide displacement prediction model and method based on staged attention mechanism

Publications (2)

Publication Number Publication Date
CN112101660A true CN112101660A (en) 2020-12-18
CN112101660B CN112101660B (en) 2022-06-28

Family

ID=73760131

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010968140.3A Active CN112101660B (en) 2020-09-15 2020-09-15 Rainfall type landslide displacement prediction model and method based on staged attention mechanism

Country Status (1)

Country Link
CN (1) CN112101660B (en)

Cited By (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112945162A (en) * 2021-01-26 2021-06-11 山西大学 Accumulation layer landslide displacement prediction model and prediction method
CN113033091A (en) * 2021-03-24 2021-06-25 桂林电子科技大学 Landslide displacement prediction method
CN113378582A (en) * 2021-07-15 2021-09-10 重庆交通大学 Landslide displacement prediction model and method based on semantic information driving
CN113434989A (en) * 2021-06-28 2021-09-24 山东大学 Pipe network leakage amount prediction method and system based on attention mechanism and LSTM
CN113569479A (en) * 2021-07-27 2021-10-29 天津大学 Long-term multi-step control method, device and storage medium for rock fracture development of stone cave temple
CN114488350A (en) * 2022-01-18 2022-05-13 重庆交通大学 Short-term rainfall forecasting method for mountain city based on radar map information processing
CN114692963A (en) * 2022-03-22 2022-07-01 同济大学 Landslide displacement prediction method based on long-term and short-term memory neural network
CN116206131A (en) * 2023-03-16 2023-06-02 北京百度网讯科技有限公司 Image processing method, training method and device for deep learning model
CN116861286A (en) * 2023-06-21 2023-10-10 四川省人工智能研究院(宜宾) Multi-mode landslide displacement intelligent prediction method based on fuzzy calculation
CN116956046A (en) * 2023-09-07 2023-10-27 西南交通大学 Earthquake landslide hazard analysis method and device based on cyclic neural network
CN117132907A (en) * 2023-05-10 2023-11-28 淮南师范学院 Landslide detection system based on deep learning

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109886492A (en) * 2019-02-26 2019-06-14 浙江鑫升新能源科技有限公司 Photovoltaic power generation power prediction model and its construction method based on Attention LSTM
CN109885977A (en) * 2019-03-25 2019-06-14 成都理工大学 A kind of bank slope Deformation Prediction method
CN110288157A (en) * 2019-06-27 2019-09-27 电子科技大学 A kind of Runoff Forecast method based on attention mechanism and LSTM
CN110334843A (en) * 2019-04-22 2019-10-15 山东大学 A kind of time-varying attention improves be hospitalized medial demand prediction technique and the device of Bi-LSTM
CN110470259A (en) * 2019-07-09 2019-11-19 西安工程大学 Landslide displacement dynamic prediction method based on LSTM
CN110543543A (en) * 2019-09-10 2019-12-06 苏州大学 user movement behavior prediction method and device based on multi-granularity neural network
CN111191841A (en) * 2019-12-30 2020-05-22 润联软件系统(深圳)有限公司 Power load prediction method and device, computer equipment and storage medium
CN111612254A (en) * 2020-05-22 2020-09-01 中国科学院合肥物质科学研究院 Road motor vehicle exhaust emission prediction method based on improved attention bidirectional long-short term memory network

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109886492A (en) * 2019-02-26 2019-06-14 浙江鑫升新能源科技有限公司 Photovoltaic power generation power prediction model and its construction method based on Attention LSTM
CN109885977A (en) * 2019-03-25 2019-06-14 成都理工大学 A kind of bank slope Deformation Prediction method
CN110334843A (en) * 2019-04-22 2019-10-15 山东大学 A kind of time-varying attention improves be hospitalized medial demand prediction technique and the device of Bi-LSTM
CN110288157A (en) * 2019-06-27 2019-09-27 电子科技大学 A kind of Runoff Forecast method based on attention mechanism and LSTM
CN110470259A (en) * 2019-07-09 2019-11-19 西安工程大学 Landslide displacement dynamic prediction method based on LSTM
CN110543543A (en) * 2019-09-10 2019-12-06 苏州大学 user movement behavior prediction method and device based on multi-granularity neural network
CN111191841A (en) * 2019-12-30 2020-05-22 润联软件系统(深圳)有限公司 Power load prediction method and device, computer equipment and storage medium
CN111612254A (en) * 2020-05-22 2020-09-01 中国科学院合肥物质科学研究院 Road motor vehicle exhaust emission prediction method based on improved attention bidirectional long-short term memory network

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
宋丽伟: "基于经验模态分解和LSTM模型的滑坡位移预测", 《人民长江》 *
杨背背 等: "基于时间序列与长短时记忆网络的滑坡位移动态预测模型", 《岩石力学与工程学报》 *

Cited By (18)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112945162B (en) * 2021-01-26 2022-05-31 山西大学 Accumulation layer landslide displacement prediction model and prediction method
CN112945162A (en) * 2021-01-26 2021-06-11 山西大学 Accumulation layer landslide displacement prediction model and prediction method
CN113033091A (en) * 2021-03-24 2021-06-25 桂林电子科技大学 Landslide displacement prediction method
CN113434989A (en) * 2021-06-28 2021-09-24 山东大学 Pipe network leakage amount prediction method and system based on attention mechanism and LSTM
CN113378582A (en) * 2021-07-15 2021-09-10 重庆交通大学 Landslide displacement prediction model and method based on semantic information driving
CN113378582B (en) * 2021-07-15 2022-04-26 重庆交通大学 Landslide displacement prediction model construction and use method based on semantic information driving
CN113569479B (en) * 2021-07-27 2023-11-10 天津大学 Long-term multi-step control method, device and storage medium for rock mass crack development of stone cave temple
CN113569479A (en) * 2021-07-27 2021-10-29 天津大学 Long-term multi-step control method, device and storage medium for rock fracture development of stone cave temple
CN114488350A (en) * 2022-01-18 2022-05-13 重庆交通大学 Short-term rainfall forecasting method for mountain city based on radar map information processing
CN114488350B (en) * 2022-01-18 2024-04-30 重庆交通大学 Mountain city short-time rainfall forecasting method based on radar map information processing
CN114692963A (en) * 2022-03-22 2022-07-01 同济大学 Landslide displacement prediction method based on long-term and short-term memory neural network
CN116206131B (en) * 2023-03-16 2023-09-19 北京百度网讯科技有限公司 Image processing method, training method and device for deep learning model
CN116206131A (en) * 2023-03-16 2023-06-02 北京百度网讯科技有限公司 Image processing method, training method and device for deep learning model
CN117132907A (en) * 2023-05-10 2023-11-28 淮南师范学院 Landslide detection system based on deep learning
CN116861286A (en) * 2023-06-21 2023-10-10 四川省人工智能研究院(宜宾) Multi-mode landslide displacement intelligent prediction method based on fuzzy calculation
CN116861286B (en) * 2023-06-21 2024-01-30 四川省人工智能研究院(宜宾) Multi-mode landslide displacement intelligent prediction method based on fuzzy calculation
CN116956046A (en) * 2023-09-07 2023-10-27 西南交通大学 Earthquake landslide hazard analysis method and device based on cyclic neural network
CN116956046B (en) * 2023-09-07 2023-12-29 西南交通大学 Earthquake landslide hazard analysis method and device based on cyclic neural network

Also Published As

Publication number Publication date
CN112101660B (en) 2022-06-28

Similar Documents

Publication Publication Date Title
CN112101660B (en) Rainfall type landslide displacement prediction model and method based on staged attention mechanism
Khadr Forecasting of meteorological drought using Hidden Markov Model (case study: The upper Blue Nile river basin, Ethiopia)
Song et al. Differences in extremes and uncertainties in future runoff simulations using SWAT and LSTM for SSP scenarios
CN112287608A (en) Dam safety prediction system and method based on fuzzy comprehensive evaluation
Liu et al. Comparative study of three updating procedures for real-time flood forecasting
CN108491974B (en) Flood forecasting method based on ensemble Kalman filtering
Prasetya et al. Rainfall forecasting for the natural disasters preparation using recurrent neural networks
CN112686426A (en) Incoming water quantity early warning method and system based on hydropower station basin key points
Randrianasolo et al. Impact of streamflow data assimilation and length of the verification period on the quality of short-term ensemble hydrologic forecasts
CN115238947A (en) Social and economic exposure degree estimation method for drought, waterlogging and sudden turning event under climate change
Gaetan et al. A hierarchical model for the analysis of spatial rainfall extremes
CN114357670A (en) Power distribution network power consumption data abnormity early warning method based on BLS and self-encoder
Todisco et al. Use of satellite and modeled soil moisture data for predicting event soil loss at plot scale
Chen et al. Combining stochastic weather generation and ensemble weather forecasts for short-term streamflow prediction
CN115983488A (en) Method and device for predicting water inflow of mine
Teja et al. Enhancing the predictability of flood forecasts by combining Numerical Weather Prediction ensembles with multiple hydrological models
Wang et al. Multi-model integrated error correction for streamflow simulation based on Bayesian model averaging and dynamic system response curve
Li et al. A variable-correlation model to characterize asymmetric dependence for postprocessing short-term precipitation forecasts
CN116993030B (en) Reservoir pressure salty taste adjustment method and system under variable conditions
CN116307233B (en) Flood forecasting method based on flood season stage
Long et al. Monthly precipitation modeling using Bayesian non-homogeneous hidden Markov chain
Swain et al. A framework for characterization of extreme floods for dam safety risk assessments
CN115641696A (en) Gridding flood forecasting model construction and real-time correction method based on multi-source information
Wang et al. A hybrid rainfall-runoff model: integrating initial loss and LSTM for improved forecasting
CN106295193A (en) A kind of river based on compressed sensing monthly runoff Forecasting Methodology

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
GR01 Patent grant
GR01 Patent grant