CN112347700A - Debris flow occurrence probability and scale forecasting method - Google Patents

Debris flow occurrence probability and scale forecasting method Download PDF

Info

Publication number
CN112347700A
CN112347700A CN202011354891.2A CN202011354891A CN112347700A CN 112347700 A CN112347700 A CN 112347700A CN 202011354891 A CN202011354891 A CN 202011354891A CN 112347700 A CN112347700 A CN 112347700A
Authority
CN
China
Prior art keywords
debris flow
matrix
scale
forecasting
occurrence probability
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.)
Withdrawn
Application number
CN202011354891.2A
Other languages
Chinese (zh)
Inventor
徐根祺
温宗周
李丽敏
张宏伟
马婧
程少康
曹宁
李银兴
任小文
贾亚娟
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.)
Xian Traffic Engineering Institute
Original Assignee
Xian Traffic Engineering Institute
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 Xian Traffic Engineering Institute filed Critical Xian Traffic Engineering Institute
Priority to CN202011354891.2A priority Critical patent/CN112347700A/en
Publication of CN112347700A publication Critical patent/CN112347700A/en
Withdrawn legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • 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
    • 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
    • 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—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/10—Services
    • G06Q50/26—Government or public services
    • G06Q50/265—Personal security, identity or safety
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2111/00—Details relating to CAD techniques
    • G06F2111/08—Probabilistic or stochastic CAD

Landscapes

  • Engineering & Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Economics (AREA)
  • Tourism & Hospitality (AREA)
  • Strategic Management (AREA)
  • Human Resources & Organizations (AREA)
  • Artificial Intelligence (AREA)
  • Health & Medical Sciences (AREA)
  • General Business, Economics & Management (AREA)
  • General Engineering & Computer Science (AREA)
  • Marketing (AREA)
  • General Health & Medical Sciences (AREA)
  • Software Systems (AREA)
  • Development Economics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Quality & Reliability (AREA)
  • Mathematical Physics (AREA)
  • Game Theory and Decision Science (AREA)
  • Biomedical Technology (AREA)
  • Data Mining & Analysis (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Biophysics (AREA)
  • Operations Research (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Computing Systems (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Computer Hardware Design (AREA)
  • Geometry (AREA)
  • Computer Security & Cryptography (AREA)
  • Educational Administration (AREA)
  • Primary Health Care (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention provides a debris flow occurrence probability and scale forecasting method, which solves the problems of data dimension disasters caused by the influence of multiple factors on the current debris flow disasters and insufficient forecasting instantaneity caused by the need of calculating a large number of hidden layer weights in deep learning. The method comprises the following steps: aiming at the occurrence probability of the debris flow disasters and the forecast of the occurrence scale of the debris flow disasters, acquiring corresponding initial influence factor data of the debris flow disasters through field investigation and survey, sorting sample data, and extracting corresponding main influence factor data based on an FMPCE algorithm; respectively constructing a debris flow disaster occurrence probability and scale forecasting model based on the optimal path forest and the matrix random approximate singular value decomposition optimization width learning; and respectively inputting the test sample data into the established debris flow disaster occurrence probability forecasting model and the debris flow disaster occurrence scale forecasting model, and outputting forecasting information of the debris flow occurrence probability and the debris flow disaster occurrence scale.

Description

Debris flow occurrence probability and scale forecasting method
Technical Field
The invention belongs to the technical field of geological disaster forecasting methods, and particularly relates to a debris flow occurrence probability and scale forecasting method.
Background
Debris flow is not only a natural disaster, but also a serious engineering geological disaster. In recent years, serious geological disasters frequently occur, which often cause accidents of house damage, communication facility interruption, road collapse, land damage and even village casualties. And accidents often occur in mountains and valleys with complicated geological structures, gullies, vertical and horizontal and steep terrains, great inconvenience is brought to disaster prevention and reconstruction after disasters, and the scale and the risk of the accidents greatly exceed the tolerable range of people. Therefore, how to forecast the debris flow disasters by using technical means becomes the core of attention of people.
Researchers have proposed various mud-rock flow disaster forecasting methods according to the characteristics of the mud-rock flow. Researchers combine the fuzzy system theory and the artificial neural network to forecast the debris flow disasters, however, the fuzzy system has high calculation complexity, is too subjective when determining the debris flow index weight vector, and although the defect is well overcome by the addition of the artificial neural network, the model is easy to fall into a local minimum value in the training process of the artificial neural network, so that the method has low precision; a scholars applies a logistic regression model to debris flow disaster prediction, analyzes by integrating various indexes, and improves the accuracy to a certain extent, but the logistic regression belongs to a generalized linear regression model and is greatly influenced by multiple collinearity problems; the learners also provide a mud-rock risk forecasting model based on the GA-BP neural network and obtain a good effect, but the analysis of the correlation among indexes is lacked, so that information is possibly overlapped, and dimension disasters are easily caused.
Disclosure of Invention
The invention aims to provide a debris flow occurrence probability and scale forecasting method, and aims to solve the problems of data dimension disasters caused by the influence of multiple factors on the current debris flow disasters and insufficient forecasting instantaneity caused by the fact that a large number of hidden layer weights need to be calculated in deep learning.
In order to achieve the above purpose, the technical scheme provided by the invention is as follows:
a debris flow occurrence probability and scale forecasting method comprises the following steps:
step 1: aiming at forecasting of debris flow occurrence probability and debris flow scale, acquiring initial influence factor data of corresponding debris flow disasters through field investigation and survey, sorting sample data, and extracting corresponding main influence factor data based on FMPCE algorithm;
step 2: respectively constructing a debris flow disaster occurrence probability forecasting model based on an optimal path forest and a debris flow disaster occurrence scale forecasting model based on matrix random approximate singular value decomposition optimization width learning by using the main influence factor data of the debris flow disaster extracted in the step 1;
and step 3: and (3) respectively inputting the test sample data into the debris flow disaster occurrence probability forecasting model and the debris flow disaster occurrence scale forecasting model established in the step (2), and outputting the debris flow occurrence probability and the debris flow scale forecasting information.
The beneficial effect of the invention is that,
(1) the FMPCE algorithm is utilized to extract main influence factors from a large number of initial influence factors of the debris flow disaster, so that the dimension reduction of an input matrix is realized, the dimension disaster possibly caused by huge data volume is successfully avoided, and the complexity of a debris flow disaster forecasting model is greatly reduced.
(2) Aiming at the characteristics of the outburst and the nonlinearity of the debris flow disaster, the debris flow disaster occurrence probability forecasting model is established based on the optimal path forest, so that the probability of the debris flow disaster occurrence can be accurately judged, the training time of the debris flow occurrence probability forecasting model is greatly shortened, and the real-time performance of debris flow occurrence probability forecasting is improved.
(3) Aiming at the characteristics of the paroxysmal and nonlinear characteristics of the debris flow disasters, a debris flow disaster occurrence scale forecasting model is established based on matrix random approximate singular value decomposition optimization width learning, and the model is optimized by adopting matrix random approximate singular value decomposition, so that the scale of the debris flow disasters can be judged more accurately, the off-line training time and the on-line training time of the debris flow scale forecasting model are greatly shortened, and the accuracy and the real-time performance of the debris flow scale forecasting are improved.
In summary, compared with the prior art: according to the method, the debris flow disaster occurrence probability forecast based on the optimal path forest and the debris flow disaster occurrence scale forecast based on matrix random approximate singular value decomposition optimization width learning can eliminate factors which have small influence on debris flow occurrence in input data in a data representation mode, and meanwhile useful information is kept; the method is applied to debris flow disaster forecast, and the probability and scale of debris flow occurrence can be accurately and timely judged.
Drawings
FIG. 1 is a flow chart of a debris flow occurrence probability and scale forecasting method according to the present invention;
FIG. 2 is a schematic diagram of an optimal path forest algorithm involved in the optimal path forest based debris flow forecasting method;
FIG. 3 is a diagram of the prediction result of extracting different dimensionality impact factors by an FMPCE (frequency modulated PCE) involved in a debris flow scale forecasting method based on an optimal path forest;
FIG. 4 is a graph comparing predicted levels with actual levels involved in the example;
FIG. 5 is a diagram of a width learning structure involved in a debris flow scale prediction method for optimizing width learning based on matrix random approximate singular value decomposition;
FIG. 6 is a diagram of the prediction results of extracting different dimensional impact factors from FMPCE involved in the debris flow scale prediction method based on matrix random approximate singular value decomposition optimization width learning;
FIG. 7 is a graph of model predictive fit involved in the example;
fig. 8 is a graph comparing the predicted level and the actual level referred to in the example.
Detailed Description
The invention will be further described with reference to the accompanying drawings and examples, which are to be regarded as illustrative only and not as restrictive. The forecasting method of the debris flow disaster occurrence probability is to analyze the influence factors of the debris flow occurrence and calculate through a model, so that the proposed theoretical model has operability and practicability. Can draw reliable conclusion in practical production application and provide basis for disaster prevention.
The debris flow occurrence probability forecasting model is established by combining a rapid multi-principal-component parallel extraction algorithm FMPCE and an optimal path forest algorithm based on a complete graph, and the debris flow occurrence scale forecasting model is established by optimizing a width learning algorithm based on matrix random approximate singular value decomposition. Extracting main influence factors of the debris flow disasters by utilizing an FMPCE algorithm to replace initial high-dimensional influence factors, so that the problem of dimensional disasters caused by the traditional method can be improved; the optimal path forest algorithm is adopted to predict the debris flow occurrence probability, so that the defect of overlong training time caused by the fact that a large number of hidden layer weights need to be calculated in deep learning can be overcome; the model is optimized by matrix random approximate singular value decomposition, so that the problem of input matrix structure redundancy caused by poor model initialization can be solved; the matrix random approximate singular value decomposition optimization width learning algorithm is adopted to predict the debris flow occurrence scale, so that the defect of overlong training time caused by the fact that a large number of hidden layer weights need to be calculated in deep learning can be overcome, and meanwhile, the online updating capability of the model can be improved.
The implementation of the method for forecasting the occurrence probability and scale of the debris flow can be regarded as mutually independent steps, so that the following description is respectively made in terms of forecasting the occurrence probability and forecasting the occurrence scale of the debris flow.
A. Debris flow forecast based on optimal path forest
Step 1: acquiring initial influence factors of the debris flow disasters through field investigation and survey, and extracting main influence factor data of the debris flow disasters through an FMPCE algorithm;
the specific process of acquiring the initial impact factor of the debris flow disaster in the step 1 and extracting the initial impact factor through FMPCE is as follows:
step 1.1: obtaining an initial influence factor:
in a debris flow incident area, surveying and collecting debris flow disaster influence factor data on the spot according to relevant specifications, and finally determining slope gradient, furrow-bed specific reduction, relative altitude difference, drainage basin area, drainage basin integrity coefficient, drainage basin development degree, segment length ratio of replenishment sections, erosion variation, lithology, vegetation coverage rate, rainfall, soil moisture content, pore water pressure, loose object reserve along ditches, adverse geological phenomena and new structure influence as initial influence factors of debris flow disasters.
Setting the obtained initial influence factors as r groups, wherein each group of samples contains s initial influence factors which are respectively slope s of the hillside1Odds and ends of gully bed2Relative height difference s3Area of drainage basin s4Basin integrity factor s5Degree of development of watershed s6Length ratio of supply segment s7Scouring and sludging amplitude s8Lithology s9Vegetation coverage s10Rainfall s11Water content of soil s12Pore water pressure s13Bulk material reserve s along the trench14Adverse geological phenomena s15And new structural influence s16. Each set of sample data forms a matrix as shown in equation (1):
Figure BDA0002802283200000031
wherein x isij( i 1,2, … r, j 1,2, …, s) represents each influencing factor;
step 1.2: and (3) screening data by using an FMPCE algorithm:
step 1.2.1 the initial impact factor matrix X is normalized according to equation (2):
Figure BDA0002802283200000032
wherein the content of the first and second substances,
Figure BDA0002802283200000045
and sjRespectively representing the mean and variance of the initial impact factors;
step 1.2.2: the linear neural network model is set as follows:
y(k)＝WT(k)x(k) (3)
wherein y (k) e Rr×1Represents the neural network output, W (k) e Rn×rRepresents the weight matrix of the neural network, x (k) epsilon Rn ×1Representing the neural network input, n is the input vector dimension, and r is the dimension of the principal component to be extracted. Make the autocorrelation matrix of the input
Figure BDA0002802283200000041
R is a symmetric positive definite matrix, wherein lambdaiIs a characteristic value of R, uiTo belong to a characteristic value lambdaiI 1,2, …, n, eigenvalue λiAnd > 0, carrying out characteristic value decomposition on R:
R＝UΛUT (4)
wherein U is [ U ]1，u2，…，un]，Λ＝diag{λ1，λ2，…，λnAnd the characteristic value satisfies:
λ1＞λ2＞…＞λr＞…＞λn＞0 (5)
the eigenvectors belonging to these R eigenvalues are the first R principal components of the matrix R, and the space created by these principal components is called the principal subspace. The FMPCE seeks a suitable weight matrix iterative update equation so that the weight matrix can converge to the first R principal components of the matrix R. The algorithm form is as follows:
W(k+1)＝W(k)+ηW(k)[WT(k)W(k)-1-I]+η(RW(k)WTW(k)A2-W(k)AWT(k)RW(k)A) (6)
wherein the matrix A is an r × r diagonal matrix, and the diagonal elements are a1＞a2＞…＞ar> 0, η is the learning rate.
The autocorrelation matrix is estimated by:
Figure BDA0002802283200000042
wherein alpha is a forgetting factor, and satisfies 0 < alpha < 1, and obviously, when k → ∞ is reached, the matrix
Figure BDA0002802283200000043
The autocorrelation matrix is estimated using equation (7), and then the principal component of the input is extracted using equations (3) and (6).
Step 2: constructing a debris flow disaster forecasting model based on the forest of the optimal path by using the influence factors extracted in the step 1;
the concrete method for building the debris flow disaster forecasting model in the step 2 comprises the following steps:
step 2.1: distributing the debris flow disaster influence factor data screened in the step 1 according to the proportion of a training set to a testing set to 8: 2, and taking the data as an input part of a debris flow forecasting model based on the forest with the optimal path;
step 2.2: inputting the given training set in the step 2.1 into a model for training, wherein the training process is as follows:
step 2.2.1: dividing the total sample set Z into training sets Z1And test set Z2By Z1And (5) training. Suppose Z1Is N, wherein each sample aiHas M attributes, respectively ai1、ai2、…、aiM. Constructing a complete graph A consisting of N samples, wherein each node of the complete graph is a training set Z1For all nodes in the complete graph, every two nodes are connected by arcs, and the weight a of each arc isiMExpressed by Euclid distance between nodes, as shown in equation (4).
Figure BDA0002802283200000044
Step 2.2.2: and generating a minimum spanning tree MST according to the complete graph A, and then solving a connecting arc between two different types of nodes in the generated MST according to an equation (4), wherein the two different types of nodes are the root nodes of the tree in the optimal path forest. Node sequence pi formed by path of multiple nodes<s1，s2，…，sk>Wherein(s)i，si+1) Belongs to A and i is more than or equal to 1 and less than or equal to k-1. By pi ·<s，t>Arc of representation<s，t>And a path pi ending with s. For each path in the optimal path forest, according to the path cost function fmaxThe cost can be determined. Path cost function fmaxAs shown in formula (5).
Figure BDA0002802283200000051
fmax(π·<s，t>)＝max{fmax(π)，d(s，t)} (6)
Wherein f ismax(π) represents the maximum distance between all two neighboring nodes lying on a path π, which is a non-trivial path.
Step 2.2.3: solving the maximum value of the distance d (s, t) between the nodes s and t according to the formula (4);
step 2.2.4: solving the cost C(s) of a precursor node s of the node t positioned on the optimal path according to the formula (5);
step 2.2.5: taking the maximum value obtained in the step 2.2.3 and the step 2.2.4 as the cost C (t) of the node t;
step 2.2.6: and solving the type L (R (s)) of the root node on the optimal path of the precursor node, wherein the type of the L (R (s)) is the type of the node t.
Step 2.2.7: and (4) corresponding the node type obtained in the step 2.2.6 with the debris flow occurrence probability to obtain the debris flow occurrence probability output by the model. The debris flow occurrence probability is divided into four categories, namely minimum probability: small, medium, large.
And step 3: and (3) inputting the influence factor data obtained in the step (1) into the debris flow disaster forecasting model established in the step (2), processing and calculating the input influence factor data by the forecasting model, and finally completing forecasting of the occurrence probability of the debris flow disaster through a forecasting device.
The specific process of step 3 is as follows:
and step 3: and inputting the test sample into the trained model, and outputting the debris flow occurrence probability value. The probability corresponds to the debris flow forecast grade, and the debris flow forecast grade is divided into four grades, namely a conventional grade, a prediction grade, an early warning grade and an alarm grade.
Introduction of application examples:
the method is characterized in that test verification is carried out on debris flow data of bridge ditches of a high dam store in Shanyang county of Shanxi Shandong Shanluo City, Shanyang county, Shaanxi province, and 16 influence factors including slope of hillside, ditch bed gradient, relative altitude difference, basin area, basin integrity coefficient, basin development degree, segment length ratio of replenishment, erosion and silt amplitude, lithology, vegetation coverage rate, rainfall, soil moisture content, pore water pressure, loose object reserve along ditches, unfavorable geological phenomena and new structure influence are used as initial influence factors of debris flow disasters. And establishing a debris flow disaster forecasting model by taking the debris flow occurrence probability as an output object. And extracting the main influence factors from 200 sets of initial influence factor data through an FMPCE algorithm.
According to the results in table 1, the finally extracted influence factors are: slope of hillside, specific drop of ditch bed, relative height difference, rainfall, soil moisture content and pore water pressure. Therefore, the above 6 influence factor data are used as the input of the optimal path forest based debris flow probability forecasting model. The FMPCE algorithm successfully reduces the initial impact factors from 16 dimensions to 6 dimensions, and greatly simplifies the complexity of a model data structure.
200 groups of main influence factor data are obtained after extraction by an FMPCE algorithm, and the 200 groups of data are divided into two parts, wherein 180 groups of data are used for model training, and the other 20 groups of data are used for verifying the accuracy of the model. And inputting the 180 groups of data into the model, and obtaining the occurrence probability of the debris flow through an optimal path forest algorithm. The predicted effect of this model is shown in fig. 4. The training set is further expanded, the model is retrained, and the training time of the model for each training set is shown in table 1. When the training set is continuously expanded, the training time is almost unchanged, and the data in the table 1 verifies that the model has faster training time and strong online training capability. Finally, the prediction probability is corresponded to the corresponding grade, the result is shown in fig. 5, and the forecast grade division is shown in table 2.
TABLE 1 training times for different training sets
Figure BDA0002802283200000061
TABLE 2 forecast ratings
Figure BDA0002802283200000062
Therefore, the prediction result and the actual result have good fitting precision, the model training time is shortened, the online training capability of the model is greatly improved, and the prediction performance is good.
B. Debris flow scale forecast based on matrix random approximate singular value decomposition optimization width learning
Step 1: acquiring initial influence factor data of the debris flow disaster through field investigation and survey, and screening main influence factor data of the debris flow disaster through an FMPCE algorithm;
the specific process of acquiring the initial impact factor of the debris flow disaster in the step 1 and extracting the initial impact factor through FMPCE is as follows:
step 1.1: obtaining an initial influence factor:
in a debris flow incident area, surveying and collecting debris flow disaster influence factor data on the spot according to relevant specifications, and finally determining slope gradient, furrow-bed specific reduction, relative altitude difference, drainage basin area, drainage basin integrity coefficient, drainage basin development degree, segment length ratio of replenishment sections, erosion and deposition amplitude, lithology, vegetation coverage rate, rainfall, soil moisture content, pore water pressure, ground sound, infrasound, loose object reserve along ditches, unfavorable geological phenomena and new structure influence as initial influence factors of debris flow disasters.
Setting the obtained initial influence factors as r groups, wherein each group of samples contains s initial influence factors which are respectively slope s of the hillside1Odds and ends of gully bed2Relative height difference s3Area of drainage basin s4Basin integrity factor s5Degree of development of watershed s6Length ratio of supply segment s7Scouring and sludging amplitude s8Lithology s9Vegetation coverage s10Rainfall s11Water content of soil s12Pore water pressure s13Earth sound s14S infrasound of15Bulk material reserve s along the trench16Adverse geological phenomena s17And new structural influence s18. Each set of sample data forms a matrix as shown in equation (1):
Figure BDA0002802283200000071
wherein x isij( i 1,2, … r, j 1,2, …, s) represents each influencing factor;
step 1.2: and (3) screening data by using an FMPCE algorithm:
the initial impact factor matrix X is normalized according to equation (2):
Figure BDA0002802283200000072
wherein the content of the first and second substances,
Figure BDA0002802283200000073
represents normalized data, xminRepresents the minimum value, x, in the datamaxRepresenting the maximum value in the data.
It should be noted that, in the step 1.2, the initial impact factor matrix X may also be normalized by using the step 1.2.1 in step a; similarly, step 1.2.1 in step a may also be normalized by the initial impact factor matrix X in this step 1.2 according to equation (2).
Step 1.3: the FMPCE is used for extracting the main influence factors of the debris flow occurrence scale, and the method comprises the following steps:
for the linear neural network model:
y(k)＝WT(k)x(k) (3)
wherein y (k) e Rr×1Represents the neural network output, W (k) e Rn×rRepresents the weight matrix of the neural network, x (k) epsilon Rn ×1Representing the neural network input, n is the input vector dimension, and r is the dimension of the principal component to be extracted. Make the autocorrelation matrix of the input
Figure BDA0002802283200000074
R is a symmetric positive definite matrix, wherein lambdaiIs a characteristic value of R, uiTo belong to a characteristic value lambdaiI 1,2, …, n, eigenvalue λiAnd > 0, carrying out characteristic value decomposition on R:
R＝UAUT (4)
wherein U is [ U ]1，u2，…，un]，Λ＝diag{λ1，λ2，…，λnAnd the characteristic value satisfies:
λ1＞λ2＞…＞λr＞…＞λn＞0 (5)
the eigenvectors belonging to these R eigenvalues are the first R principal components of the matrix R, and the space created by these principal components is called the principal subspace. The FMPCE seeks a suitable weight matrix iterative update equation so that the weight matrix can converge to the first R principal components of the matrix R. The algorithm form is as follows:
W(k+1)＝W(k)+ηW(k)[WT(k)W(k)-1-I]+η(RW(k)WTW(k)A2-W(k)AWT(k)RW(k)A) (6)
wherein the matrix A is an r × r diagonal matrix, and the diagonal elements are a1＞a2＞…＞ar> 0, η is the learning rate.
The autocorrelation matrix is estimated by:
Figure BDA0002802283200000075
wherein alpha is a forgetting factor, and satisfies 0 < alpha < 1, and obviously, when k → ∞ is reached, the matrix
Figure BDA0002802283200000076
The autocorrelation matrix is estimated using equation (7), and then the principal component of the input is extracted using equations (3) and (6).
Step 2: constructing a debris flow disaster occurrence scale forecasting model based on matrix random approximate singular value decomposition optimization width learning by using the influence factors extracted in the step 1;
the concrete method for building the debris flow disaster forecasting model in the step 2 comprises the following steps:
step 2.1: taking the debris flow disaster influence factor data extracted in the step 1 as an input part of a debris flow disaster occurrence scale forecasting model based on matrix random approximate singular value decomposition optimization width learning;
step 2.2: taking the training set given in the step 2.1 as model input, wherein the expression of the training set is shown as formula (8):
Y＝{(x1，y1)，…，(xk，yk)} (8)
wherein x isi∈Rn，yiE R, i is 1, …, k represents the total number of samples;
step 2.3: inputting a given training set into a forecasting model for training, wherein the specific process comprises the following steps:
step 2.3.1: generating a width learning initial structure, wherein the process is as follows:
let the input data set x have N samples, each sample having a dimension M, Y being the output matrix, and Y being the RN×CFor n feature maps, each map generates k enhanced nodes, represented by equation (9):
Figure BDA0002802283200000081
wherein the content of the first and second substances,
Figure BDA0002802283200000082
and
Figure BDA0002802283200000083
are all randomly generated.
Zn＝[Z1，Z2，…，Zn]Representing the combination of all the feature nodes, the mth enhanced node being defined as
Figure BDA0002802283200000084
Figure BDA0002802283200000085
Hm＝[H1，H2，…，Hm]Combinations of 1 st to mth enhanced nodes are shown. The initial structure of width learning is shown in formula (10):
Figure BDA0002802283200000086
wherein, Wm＝[Zn|Hm]+Y is a connection weight, [ Z ]n|Hm]+Can be obtained from equation (11) (pseudo-inverse ridge regression approximation algorithm):
A+＝limλ→0(λI+AAT)-1ATY (11)
step 2.3.2: adding a new enhanced node into the model, wherein the method comprises the following steps:
let P insert each enhancement node, let Am＝[Zn|Hm]Definition of Am+1＝[Am|Hm+1]Wherein
Figure BDA0002802283200000087
Figure BDA0002802283200000088
The connection weights and biases from the mapped feature node to the newly inserted p enhancement nodes are both randomly generated. According to the RVFLNN dynamic update algorithm, a new input matrix Am can be obtained+1The pseudo-inverse of (c) is shown in formula (12):
Figure BDA0002802283200000089
wherein the content of the first and second substances,
Figure BDA00028022832000000810
Figure BDA00028022832000000811
new weight Wm+1Comprises the following steps:
Figure BDA00028022832000000812
step 2.3.3: adding a new feature mapping node into the model, the method is completely similar to the step 2.3.2, and the finally updated input matrix and weight matrix are respectively shown as the following formula (14) and formula (15):
Figure BDA0002802283200000091
Figure BDA0002802283200000092
step 2.3.4: by XaThe newly added input is represented by the input of the new addition,
Figure BDA0002802283200000093
representing the input to the initial network, which includes n feature mapping nodes and m enhancement nodes. The increment of the feature mapping node and the enhancement node is expressed as:
Figure BDA0002802283200000094
wherein the content of the first and second substances,
Figure BDA0002802283200000095
is added with XaThe latter new set of feature increments is then,
Figure BDA0002802283200000096
are all randomly generated. The updated input matrix is:
Figure BDA0002802283200000097
the corresponding pseudo-inverse is calculated by:
Figure BDA0002802283200000098
wherein the content of the first and second substances,
Figure BDA0002802283200000099
the updated weight is:
Figure BDA00028022832000000910
step 2.3.5: starting from a random initialization network with n groups of mapping feature nodes, the optimization process is as follows:
Figure BDA00028022832000000911
definition of
Figure BDA00028022832000000912
Thus, the
Figure BDA00028022832000000913
The purpose of matrix decomposition is to reduce the number of nodes to simplify the computation. For ZiMatrix random approximation SVD is used for decomposition. From ZiA set of orthogonal bases of column space of (a) constructs a matrix Q such that Zi≈QQ*ZiThen Q is ZiAn approximation of the sub-matrix, QQ*ZiIs ZiA low rank approximation of the constructed subspace. From Q, Z can be easily obtainediAn approximation matrix of (a). Then through ZiApproximate matrix pair Z ofiSingular value decomposition is performed.
Figure BDA00028022832000000914
Figure BDA00028022832000000915
Order to
Figure BDA00028022832000000916
Then
Figure BDA00028022832000000917
Figure BDA0002802283200000101
The model may be defined as:
Figure BDA0002802283200000102
wherein the content of the first and second substances,
Figure BDA0002802283200000103
the optimization results for the insertion of p enhancement nodes are as follows:
Figure BDA0002802283200000104
Figure BDA0002802283200000105
after the width learning is completed, still smaller singular values need to be removed for further optimization. The optimization results are as follows:
Figure BDA0002802283200000106
step 2.3.6: in the formula (26)
Figure BDA0002802283200000107
Then the output Y of the debris flow occurrence scale forecasting model is calculated by the formula (27):
Y＝AFWF (27)
and step 3: and (3) inputting the influence factor data obtained in the step (1) into the debris flow scale forecasting model established in the step (2), processing and calculating the input influence factor data by using the forecasting model, and finally completing forecasting of the occurrence scale of the debris flow disaster by using a forecasting device.
The specific process of step 3 is as follows:
and inputting the test sample into the trained model, and outputting a numerical value corresponding to the occurrence scale of the debris flow. The numerical value corresponds to the debris flow scale grade, and the debris flow scale is divided into 5 grades which are respectively a minimum scale, a small scale, a medium scale, a large scale and a maximum scale.
Introduction of application examples:
the method is characterized in that experiment verification is carried out on debris flow data of bridge ditches of a high dam store in Shanyang county of Shanxi Shandong Shanluo City, 18 influence factors including slope of hillside, ditch bed gradient, relative altitude difference, basin area, basin integrity coefficient, basin development degree, segment length ratio of replenishment, erosion and silt amplitude, lithology, vegetation coverage rate, rainfall, soil moisture content, pore water pressure, ground sound, infrasound, loose object reserve along ditches, unfavorable geological phenomena and new structure influence are used as initial influence factors of debris flow disasters. And establishing a debris flow disaster forecasting model by taking the debris flow scale as an output object. 200 initial impact factor data sets were selected and divided into two parts, 180 for model training and 20 for model testing.
Inputting the 180 groups of data into a model, sequentially setting the number of the influence factors to be extracted to be 1-18, and predicting by using a matrix random approximate singular value decomposition optimized width learning algorithm, wherein the prediction accuracy of the influence factors with different dimensions is shown in fig. 6.
According to the results in fig. 6, 8 main influence factors are finally extracted for predicting the debris flow scale. These 8 influencing factors are: rainfall, lithology, basin development degree, basin integrity coefficient, furrow-bed specific dip, relative elevation difference, infrasound and ground sound. And taking the 8 influence factor data as the input of a debris flow disaster occurrence scale forecasting model based on matrix random approximate singular value decomposition optimization width learning. The FMPCE algorithm successfully reduces the initial impact factors from 18 dimensions to 8 dimensions, and greatly simplifies the complexity of the model data structure.
200 groups of main influence factor data are obtained after extraction by an FMPCE algorithm, and the 200 groups of data are divided into two parts, wherein 180 groups of data are used for model training, and the other 20 groups of data are used for testing the accuracy of the model. Inputting the 180 groups of data into a model, and obtaining the occurrence scale of the debris flow through a matrix random approximate singular value decomposition optimized width learning algorithm. The data fitting effect of this model is shown in fig. 7. The training set is further expanded and the model is retrained, and the training time of the model for each training set is shown in table 3. When the training set is continuously expanded, the training time is almost unchanged, and the data in the table 3 shows that the model has the advantages of fast training time and strong on-line training capability. Finally, the model output values are corresponded to the corresponding grades, the result is shown in fig. 8, and the prediction grade division is shown in table 4.
TABLE 3 training times for different training sets
Figure BDA0002802283200000111
TABLE 4 early warning ranking
Figure BDA0002802283200000112
Therefore, the prediction result and the actual value have good fitting precision, the model training time is shortened, the online training capability of the model is greatly improved, the prediction performance is good, and the problem that the real-time updating speed of the deep neural network is low is solved.

Claims (9)

1. A debris flow occurrence probability and scale forecasting method is characterized by comprising the following steps:
step 1: aiming at forecasting of the debris flow occurrence probability and the debris flow occurrence scale, acquiring initial influence factor data of corresponding debris flow disasters through field investigation and survey, sorting sample data, and extracting corresponding main influence factor data based on an FMPCE algorithm;
step 2: respectively constructing a debris flow disaster occurrence probability forecasting model based on an optimal path forest and a debris flow disaster occurrence scale forecasting model based on matrix random approximate singular value decomposition optimization width learning by using the main influence factor data of the debris flow disaster extracted in the step 1;
and step 3: and (3) respectively inputting the test sample data into the debris flow disaster occurrence probability forecasting model and the debris flow disaster occurrence scale forecasting model established in the step (2), and outputting the debris flow disaster occurrence probability and the debris flow disaster occurrence scale forecasting information.
2. The debris flow occurrence probability and scale forecasting method according to claim 1, wherein in step 1, initial influence factor data of debris flow disaster occurrence is obtained and sample data is sorted according to the forecasting of debris flow occurrence probability, specifically as follows:
surveying and collecting debris flow disaster influence factor data on site according to relevant specifications in a debris flow prone area, and finally determining slope gradient, furrow-bed specific reduction, relative altitude difference, drainage basin area, drainage basin integrity coefficient, drainage basin development degree, segment length ratio of replenishment sections, erosion and deposition variation, lithology, vegetation coverage rate, rainfall, soil moisture content, pore water pressure, loose object reserve along a furrow, adverse geological phenomena and new structure influence as initial influence factors of debris flow disasters;
setting the obtained initial influence factors as r groups, wherein each group of samples contains s initial influence factors which are respectively slope s of the hillside1Odds and ends of gully bed2Relative height difference s3Area of drainage basin s4Basin integrity factor s5Degree of development of watershed s6Length ratio of supply segment s7Scouring and sludging amplitude s8Lithology s9Vegetation coverage s10Rainfall s11Water content of soil s12Pore water pressure s13Bulk material reserve s along the trench14Adverse geological phenomena s15And new structural influence s16Then, each set of sample data constitutes a matrix as shown in formula (1):
Figure FDA0002802283190000011
wherein x isij(i-1, 2, … r, j-1, 2, …, s) represents each influence factor。
3. The debris flow occurrence probability and scale forecasting method according to claim 2, characterized in that, in step 1, main influence factor data is extracted for the debris flow occurrence probability based on an FMPCE algorithm, specifically as follows:
step a 1.1: the initial impact factor matrix X is normalized according to equation (2):
Figure FDA0002802283190000012
wherein the content of the first and second substances,
Figure FDA0002802283190000013
and sjRespectively representing the mean and variance of the initial impact factors;
step a 1.2: the linear neural network model is set as follows:
y(k)＝WT(k)x(k) (3)
wherein y (k) e Rr×1Represents the neural network output, W (k) e Rn×rRepresents the weight matrix of the neural network, x (k) epsilon Rn×1Representing the input of a neural network, wherein n is the dimension of an input vector, and r is the dimension of a principal component to be extracted; make the autocorrelation matrix of the input
Figure FDA0002802283190000024
R is a symmetric positive definite matrix, wherein lambdaiIs a characteristic value of R, uiTo belong to a characteristic value lambdaiI 1,2, n, the eigenvalue λiAnd > 0, carrying out characteristic value decomposition on R:
R＝UΛUT (4)
wherein U is [ U ]1，u2，...，un]，Λ＝diag{λ1，λ2，...，λnAnd the characteristic value satisfies:
λ1＞λ2＞…＞λr＞…＞λn＞0 (5)
the eigenvectors belonging to the R eigenvalues are the first R principal components of the matrix R, the space generated by the principal components is marked as a main subspace, and the FMPCE is used for seeking a proper weight matrix iterative update equation so that the weight matrix can be converged to the first R principal components of the matrix R; the algorithm form is as follows:
W(k+1)＝W(k)+ηW(k)[WT(k)W(k)-1-I]+η(RW(k)WTW(k)A2-W(k)AWT(k)RW(k)A) (6)
wherein the matrix A is an r × r diagonal matrix, and the diagonal elements are a1＞a2＞…＞arGreater than 0, eta is the learning rate;
the autocorrelation matrix is estimated by:
Figure FDA0002802283190000021
wherein alpha is a forgetting factor, and satisfies 0 < alpha < 1, and obviously, when k → ∞ is reached, the matrix
Figure FDA0002802283190000022
The autocorrelation matrix is estimated using equation (7), and then the principal component of the input is extracted using equations (3) and (6).
4. The debris flow occurrence probability and scale forecasting method according to claim 3, wherein the debris flow disaster occurrence probability forecasting model based on the optimal path forest is constructed in the step 2, and the concrete process is as follows:
step a 2.1: distributing the debris flow disaster influence factor data extracted in the step 1 into a training set and a testing set according to a set proportion, and using the training set and the testing set as an input part of a debris flow occurrence probability forecasting model based on an optimal path forest;
step 2.2: inputting the training set determined by allocation into the debris flow occurrence probability forecasting model for training, wherein the training process is as follows:
step a 2.2.1: dividing the total sample set Z into training sets Z1And test set Z2By Z1Training is carried out; suppose Z1Is N, wherein each sample aiHas M attributes, respectively ai1、ai2、...、aiM(ii) a Constructing a complete graph A consisting of N samples, wherein each node of the complete graph is a training set Z1For all nodes in the complete graph, every two nodes are connected by arcs, and the weight a of each arc isiMExpressed by Euclid distance between nodes, as shown in formula (4);
Figure FDA0002802283190000023
step a 2.2.2: generating a Minimum Spanning Tree (MST) according to the complete graph A, and then solving a connecting arc between two different types of nodes in the generated MST according to a formula (4), wherein the two different types of nodes are root nodes of the tree in the optimal path forest; node sequence pi formed by path of multiple nodes<s1，s2，…，sk>Wherein(s)i，si+1) Belongs to A and i is more than or equal to 1 and less than or equal to k-1; by pi ·<s，t>Arc of representation<s，t>A path formed by the path pi with s as an end point; for each path in the optimal path forest, according to the path cost function fmaxCalculating the cost; path cost function fmaxAs shown in formula (5);
Figure FDA0002802283190000031
fmax(π·<s，t>)＝max{fmax(π)，d(s，t)} (6)
wherein f ismax(pi) represents the maximum distance between all pairwise adjacent nodes located on a path pi, which is a non-trivial path;
step a 2.2.3: solving the maximum value of the distance d (s, t) between the nodes s and t according to the formula (4);
step a 2.2.4: solving the cost C(s) of a precursor node s of the node t positioned on the optimal path according to the formula (5);
step a 2.2.5: the cost C (t) of the node t is the maximum value obtained in the step a2.2.3 and the step a 2.2.4;
step a 2.2.6: solving the type L (R (s)) of a root node on the optimal path of the precursor node s, wherein the type L (R (s)) is the type of the node t;
step a 2.2.7: and (6) corresponding the node type obtained in the step (a 2.2.6) with the debris flow occurrence probability to obtain the debris flow occurrence probability output by the model.
5. The debris flow occurrence probability and scale forecasting method according to claim 4, wherein forecasting information of the debris flow occurrence probability is output in the step 3, specifically, the probability value corresponds to a debris flow forecasting grade, and the debris flow forecasting grade is divided into four grades, namely a conventional grade, a forecasting grade, an early warning grade and an alarm grade.
6. The debris flow occurrence probability and scale forecasting method according to claim 1, wherein in step 1, initial impact factor data of debris flow disasters are obtained and sample data are sorted according to the debris flow occurrence scale forecasting, specifically as follows:
surveying and collecting debris flow disaster influence factor data on site according to relevant specifications in a debris flow prone area, and finally determining slope gradient, furrow-bed specific reduction, relative altitude difference, drainage basin area, drainage basin integrity coefficient, drainage basin development degree, segment length ratio of replenishment sections, erosion and deposition variation, lithology, vegetation coverage rate, rainfall, soil moisture content pore water pressure, ground sound, infrasound, loose object reserve along ditches, adverse geological phenomena and new structure influence as initial influence factors of debris flow disasters;
setting the obtained initial influence factors as r groups, wherein each group of samples contains s initial influence factors which are respectively slope s of the hillside1Odds and ends of gully bed2Relative height difference s3Area of drainage basin s4Basin integrity factor s5Degree of development of watershed s6Length ratio of supply segment s7Punching and stamping machineSludge amplitude s8Lithology s9Vegetation coverage s10Rainfall s11Water content of soil s12Pore water pressure s13Earth sound s14S infrasound of15Bulk material reserve s along the trench16Adverse geological phenomena s17And new structural influence s18(ii) a Each set of sample data forms a matrix as shown in equation (1):
Figure FDA0002802283190000032
wherein x isij(i-1, 2, … r, j-1, 2, …, s) represents each influencing factor.
7. The debris flow occurrence probability and scale forecasting method according to claim 6, characterized in that, in the step 1, the main influence factor data is extracted for the occurrence scale of the debris flow disaster based on FMPCE algorithm, and the specific process is as follows:
step b 1.1: and (3) screening data by using an FMPCE algorithm:
the initial impact factor matrix X is normalized according to equation (2):
Figure FDA0002802283190000041
wherein the content of the first and second substances,
Figure FDA0002802283190000042
represents normalized data, xminRepresents the minimum value, x, in the datamaxRepresents the maximum value in the data;
step b 1.2: for the linear neural network model:
y(k)＝WT(k)x(k) (3)
wherein y (k) e Rr×1Represents the neural network output, W (k) e Rn×rRepresents the weight matrix of the neural network, x (k) epsilon Rn×1Representing the input of the neural network, and,n is the dimension of the input vector, r is the dimension of the principal component to be extracted; make the autocorrelation matrix of the input
Figure FDA0002802283190000043
R is a symmetric positive definite matrix, wherein lambdaiIs a characteristic value of R, uiTo belong to a characteristic value lambdaiI 1,2, n, the eigenvalue λiAnd > 0, carrying out characteristic value decomposition on R:
R＝UΛUT (4)
wherein U is [ U ]1，u2，...，un]，Λ＝diag{λ1，λ2，...，λnAnd the characteristic value satisfies:
λ1＞λ2＞…＞λr＞…＞λn＞0 (5)
the eigenvectors belonging to the R eigenvalues are the first R principal components of the matrix R, and the space generated by these principal components is denoted as the principal subspace; FMPCE seeks a proper weight matrix iteration updating equation, so that the weight matrix can be converged to the first R main components of the matrix R; the algorithm form is as follows:
W(k+1)＝W(k)+ηW(k)[WT(k)W(k)-1-I]+η(RW(k)WTW(k)A2-W(k)AWT(k)RW(k)A) (6)
wherein the matrix A is an r × r diagonal matrix, and the diagonal elements are a1＞a2＞…＞arGreater than 0, eta is the learning rate;
the autocorrelation matrix is estimated by:
Figure FDA0002802283190000044
wherein alpha is a forgetting factor, and satisfies 0 < alpha < 1, and obviously, when k → ∞ is reached, the matrix
Figure FDA0002802283190000045
The autocorrelation matrix is estimated using equation (7), and thenThe input principal component is extracted by using the expressions (3) and (6).
8. The debris flow occurrence probability and scale forecasting method according to claim 7, wherein a debris flow occurrence scale forecasting model based on matrix random approximate singular value decomposition optimization width learning is constructed in the step 2, and the concrete process is as follows:
step b 2.1: distributing the debris flow disaster influence factor data extracted in the step 1 into a training set and a testing set according to a set proportion, and using the training set and the testing set as an input part of a debris flow disaster occurrence scale forecasting model for optimizing width learning based on matrix random approximate singular value decomposition;
step b 2.2: and taking the distributed and determined training set as model input, wherein the expression of the training set is shown as the formula (8):
Y＝{(x1，y1)，…，(xk，yk)} (8)
wherein x isi∈Rn，yiE R, i is 1, …, k represents the total number of samples;
step b 2.3: inputting a given training set into a forecasting model for training, wherein the specific process comprises the following steps:
step b 2.3.1: generating a width learning initial structure, wherein the process is as follows:
let the input data set X have N samples, each sample having dimension M, Y being the output matrix, and Y being the RN×CFor n feature maps, each map generates k enhanced nodes, represented by equation (9):
Figure FDA00028022831900000515
wherein the content of the first and second substances,
Figure FDA00028022831900000513
and
Figure FDA00028022831900000514
are all generated randomly;
Zn＝[Z1，Z2，...，Zn]representing the combination of all the feature nodes, the mth enhanced node being defined as
Figure FDA0002802283190000051
Figure FDA0002802283190000052
Hm＝[H1，H2，...，Hm]Represents a combination of 1 st to mth enhanced nodes; the initial structure of width learning is shown in formula (10):
Figure FDA0002802283190000053
wherein, Wm＝[Zn|Hm]+Y is a connection weight, [ Z ]n|Hm]+Can be obtained from formula (11):
A+＝limλ→0(λI+AAT)-1ATY (11)
step b 2.3.2: adding a new enhanced node into the model, wherein the method comprises the following steps:
let P insert each enhancement node, let Am＝[Zn|Hm]Definition of Am+1＝[Am|Hm+1]Wherein
Figure FDA0002802283190000054
Figure FDA0002802283190000055
The connection weight and the bias from the mapping feature node to the newly inserted p enhancement nodes are randomly generated; according to the RVFLNN dynamic update algorithm, a new input matrix A is obtainedm+1The pseudo-inverse of (c) is shown in formula (12):
Figure FDA0002802283190000056
wherein the content of the first and second substances,
Figure FDA0002802283190000057
Figure FDA0002802283190000058
new weight Wm+1Comprises the following steps:
Figure FDA0002802283190000059
step b2.3, 3: adding a new feature mapping node into the model according to the step b2.3.2, and finally updating the input matrix and the weight matrix as shown in the formula (14) and the formula (15) respectively:
Figure FDA00028022831900000510
Figure FDA00028022831900000511
step b 2.3.4: by XaThe newly added input is represented by the input of the new addition,
Figure FDA00028022831900000512
representing the input of an initial network, wherein the input comprises n feature mapping nodes and m enhancement nodes; the increment of the feature mapping node and the enhancement node is expressed as:
Figure FDA0002802283190000061
wherein the content of the first and second substances,
Figure FDA0002802283190000062
is added with XaThe latter new set of feature increments is then,
Figure FDA0002802283190000063
are all generated randomly; the updated input matrix is:
Figure FDA0002802283190000064
the corresponding pseudo-inverse is calculated by:
Figure FDA0002802283190000065
wherein the content of the first and second substances,
Figure FDA0002802283190000066
the updated weight is:
Figure FDA0002802283190000067
step b 2.3.5: starting from a random initialization network with n groups of mapping feature nodes, the optimization process is as follows:
Figure FDA0002802283190000068
definition of
Figure FDA0002802283190000069
Thus, the
Figure FDA00028022831900000610
For ZiDecomposing by matrix random approximation SVD: from ZiColumn (2) ofA set of orthogonal bases of space constructs a matrix Q such that Zi≈QQ*ZiThen Q is ZiAn approximation of the sub-matrix, QQ*ZiIs ZiA low rank approximation of the formed subspace; determining Z from QiThen through ZiApproximate matrix pair Z ofiPerforming singular value decomposition;
Figure FDA00028022831900000611
Figure FDA00028022831900000612
order to
Figure FDA00028022831900000613
Then
Figure FDA00028022831900000614
The model is defined as:
Figure FDA00028022831900000615
wherein the content of the first and second substances,
Figure FDA00028022831900000616
the optimization results for the insertion of p enhancement nodes are as follows:
Figure FDA00028022831900000617
Figure FDA0002802283190000071
and then removing smaller singular values for further optimization, wherein the optimization result is as follows:
Figure FDA0002802283190000072
step b 2.3.6: in the formula (26)
Figure FDA0002802283190000073
Then the output Y of the debris flow occurrence scale forecasting model is calculated by the formula (27):
Y＝AFWF (27)。
9. the debris flow occurrence probability and scale forecasting method according to claim 8, wherein the step 3 of outputting the forecasting information of the debris flow scale is specifically as follows: and the Y value corresponds to the debris flow scale grade, and the debris flow scale is divided into 5 grades which are respectively a minimum scale, a small scale, a medium scale, a large scale and a maximum scale.
CN202011354891.2A 2020-11-27 2020-11-27 Debris flow occurrence probability and scale forecasting method Withdrawn CN112347700A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011354891.2A CN112347700A (en) 2020-11-27 2020-11-27 Debris flow occurrence probability and scale forecasting method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011354891.2A CN112347700A (en) 2020-11-27 2020-11-27 Debris flow occurrence probability and scale forecasting method

Publications (1)

Publication Number Publication Date
CN112347700A true CN112347700A (en) 2021-02-09

Family

ID=74364895

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011354891.2A Withdrawn CN112347700A (en) 2020-11-27 2020-11-27 Debris flow occurrence probability and scale forecasting method

Country Status (1)

Country Link
CN (1) CN112347700A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113379122A (en) * 2021-06-09 2021-09-10 辽宁工程技术大学 Debris flow gully plugging degree prediction method based on probabilistic neural network model

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113379122A (en) * 2021-06-09 2021-09-10 辽宁工程技术大学 Debris flow gully plugging degree prediction method based on probabilistic neural network model
CN113379122B (en) * 2021-06-09 2024-04-23 辽宁工程技术大学 Mud-rock flow ditch blocking degree prediction method based on probabilistic neural network model

Similar Documents

Publication Publication Date Title
Dou et al. Assessment of advanced random forest and decision tree algorithms for modeling rainfall-induced landslide susceptibility in the Izu-Oshima Volcanic Island, Japan
Hong et al. Introducing a novel multi-layer perceptron network based on stochastic gradient descent optimized by a meta-heuristic algorithm for landslide susceptibility mapping
CN110569554B (en) Landslide susceptibility evaluation method based on spatial logistic regression and geographic detector
KR100982448B1 (en) Ground subsidence prediction system and predicting method using the same
CN111310898A (en) Landslide hazard susceptibility prediction method based on RNN
Miao et al. Landslide susceptibility mapping in Three Gorges Reservoir area based on GIS and boosting decision tree model
CN114861277B (en) Long-time-sequence territorial space function and structure simulation method
CN107480808B (en) Method for planning diversion project line in high-altitude mountain area
CN116609860B (en) Hydrologic model real-time correction method and system based on integrated learning algorithm
CN114036841A (en) Landslide incidence prediction method and system based on semi-supervised support vector machine model
Soma et al. Landslide susceptibility map using certainty factor for hazard mitigation in mountainous areas of Ujung-loe watershed in South Sulawesi.
CN112347701A (en) Landslide occurrence probability and scale forecasting method
CN113902580A (en) Historical farmland distribution reconstruction method based on random forest model
Arabameri et al. Application of novel ensemble models and k-fold CV approaches for Land subsidence susceptibility modelling
Lin et al. Applications of cluster analysis and pattern recognition for typhoon hourly rainfall forecast
Abraham et al. Spatio-temporal dynamics of land use land cover changes and future prediction using geospatial techniques
CN112347700A (en) Debris flow occurrence probability and scale forecasting method
CN117540303A (en) Landslide susceptibility assessment method and system based on cross semi-supervised machine learning algorithm
Zhang et al. Landslide Susceptibility Mapping Using Novel Hybrid Model Based on Different Mapping Units
Xu et al. GEO-environmental suitability evaluation of land for urban construction based on a back-propagation neural network and GIS: A case study of Hangzhou
CN115019476A (en) Landslide space-time information monitoring and early warning method based on multi-source information fusion
Tewari et al. Application of LSTMs and HAND in rapid flood inundation mapping
Ye et al. Study on Dynamic Stability Prediction Model of Slope in Eastern Tibet Section of Sichuan-Tibet Highway
Amirabadizadeh et al. Assessment of impacts of future climate change on water resources of the Hulu Langat basin using the swat model
Kong et al. Optimization of random forest model for assessing and predicting geological hazards susceptibility in Lingyun County

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
WW01 Invention patent application withdrawn after publication
WW01 Invention patent application withdrawn after publication

Application publication date: 20210209