CN113033879B - Landslide displacement prediction method based on intuitive fuzzy denominator PSO-LSTM - Google Patents
Landslide displacement prediction method based on intuitive fuzzy denominator PSO-LSTM Download PDFInfo
- Publication number
- CN113033879B CN113033879B CN202110244201.6A CN202110244201A CN113033879B CN 113033879 B CN113033879 B CN 113033879B CN 202110244201 A CN202110244201 A CN 202110244201A CN 113033879 B CN113033879 B CN 113033879B
- Authority
- CN
- China
- Prior art keywords
- population
- sub
- particle
- particles
- fitness
- 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.)
- Active
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
- G06N3/004—Artificial life, i.e. computing arrangements simulating life
- G06N3/006—Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N7/00—Computing arrangements based on specific mathematical models
- G06N7/02—Computing arrangements based on specific mathematical models using fuzzy logic
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
Abstract
The invention discloses a landslide displacement prediction method based on an intuitionistic fuzzy denominator PSO-LSTM model, which comprises the following steps: step 1: establishing an initial LSTM model; step 2: initializing a sub-population; step 3: searching for a sub-seed group concentration degree to reach a threshold value, executing a collision rebound operator, otherwise executing the step 4; step 4: updating the probe population, calculating the fitness and updating the optimal solution; step 5: calculating and developing the Ramahalanobis factor of the particles in the sub-population; step 6: updating the particle speed and position, calculating the particle fitness and updating the optimal solution; step 7: judging whether the development sub-population is converged, and if so, selecting an optimal target area; step 8: returning to the step 3 if the ending condition is not met; step 9: interpreting the optimal solution as LSTM; the LSTM model is trained with training samples. According to the invention, the LSTM prediction model is optimized and is more suitable for landslide prediction through combining a heuristic optimization algorithm with an intuitionistic fuzzy set, so that the prediction precision of the prediction model on the liveness problem is improved.
Description
Technical Field
The invention belongs to the technical field of landslide control, and particularly relates to a landslide displacement prediction method based on an intuitive fuzzy-based PSO-LSTM model.
Background
Currently, many landslide displacement prediction methods exist in landslide control technology, and a Recurrent neural network (Recurrent NeuralNetwork, RNN) is one of the methods, and in the method, a Long Short-term memory neural network (Long Short-term Memory Networks, LSTM) is a deep learning model which is commonly used and has good time sequence prediction capability. The RNN, by associating with a previous unit on the sequence, enables the previous unit to influence the computation of the unit, and thus it can efficiently and orderly process sequence information. The LSTM is further improved on the basis of the RNN, and short-term dependence and long-term dependence of the RNN are effectively avoided by adding the forgetting gate, the input gate and the output gate, so that the LSTM is more suitable for processing time sequences. Therefore, LSTM has the advantages of strong fitting capability, good fault tolerance, strong generalization capability, good feature extraction capability to problems and the like, and quite high prediction accuracy can be obtained under the condition of sufficient sample number.
However, there are the following objective problems in landslide displacement prediction: the collection of landslide data is generally susceptible to objective factors such as environmental self factors (ground water level, rainfall, human activities, earthquakes), equipment factors (equipment accuracy, equipment damage) and the like, so that problems of large scale, nonlinearity, high information redundancy, ambiguity, uncertainty and high noise often exist in the collected sample data, and the problems greatly influence the direct application capability of LSTM on the problems.
In conclusion, the existing landslide displacement prediction technology solves the problems that the accuracy is difficult to improve due to mismatching of a manually designed prediction model caused by factors such as high dimension, blurring, abstraction, uncertainty and the like of data characteristics.
Disclosure of Invention
Aiming at the problems existing in the prior art, the invention aims to provide a landslide displacement prediction method based on an intuitive fuzzy denominator PSO-LSTM model. In order to achieve the above purpose, the present invention provides the following technical solutions:
a landslide displacement prediction method based on an intuitionistic fuzzy-matrix PSO-LSTM model comprises the following steps:
step 1: collecting characteristic data causing landslide displacement, cleaning the data, forming a training sample and a test sample, establishing an initial LSTM model, and setting parameters for an IFMHDPSO algorithm;
step 2: initializing a probe sub-population and a development sub-population, and calculating the fitness of each particle in each sub-population;
step 3: if the cluster concentration of the explored sub-species reaches the threshold value, executing a collision rebound operator, otherwise executing the step 4;
step 4: updating the position and speed of the probe population, calculating the fitness and updating the historical optimal solution;
step 5: calculating and developing the Ramahalanobis factor of each particle in the sub-population;
Step 6: updating the speed and the position of the particles of the development sub-population, calculating the adaptability of each particle in the development sub-population, and updating the historical optimal solution;
step 7: judging whether the development sub-population converges or not, if so, selecting an optimal target area from the current exploration sub-population through an intuitive fuzzy decision, otherwise, executing the step 8;
step 8: judging whether the iteration ending condition is met, if not, adding 1 to the iteration times, returning to the step 3, and if so, entering the step 9;
step 9: interpreting the optimal solution obtained by searching as LSTM; training the LSTM model by using a training sample to obtain a final prediction model.
Further, in the step 1:
the establishing the initial LSTM model comprises the following steps: the first layer is an LSTM model, the last seven layers are full-connection layers FC, wherein the last layer has only one unit, the number of hidden layer units serving as output units is 1, the activation function is fixed to be tanh, the number of super parameters is 14, namely, the search dimension D=14, the first 7 represents the number of hidden units of the corresponding layer, and the last 7 indicates the activation function used by the corresponding layer;
the setting parameters for the IFMHDPSO algorithm comprises the following steps: setting population size N, i.e. exploring the number of particles N of sub-population 1 Developing the particle number n of the sub-population 2 ＝N-n 1 And n is 1 >n 2 The method comprises the steps of carrying out a first treatment on the surface of the Maximum iteration number maxIter; developing the sub-population convergence decision number s epsilon {1,2, …, maxIter }; judgment error epsilon [0,1 ]]The method comprises the steps of carrying out a first treatment on the surface of the Intuitive fuzzy decision of membership degree of each attribute weight
And hesitation->
And->
Two attributes a 1 And a 2 The method comprises the steps of carrying out a first treatment on the surface of the Collision triggering ratio lambda epsilon 0,1]The method comprises the steps of carrying out a first treatment on the surface of the A neighborhood range R; the speed of the probe sub-population is at the upper and lower limit Vmax of the j dimension j And Vmin j The method comprises the steps of carrying out a first treatment on the surface of the Search space upper bound ub, lower bound lb; searching dimension D; initially, the set of all particles i +.>
Iteration number t=1, forbidden set
Further, the step 2 specifically includes the following sub-steps:
step 21: randomly generating n in the search space range of step 1 1 The particles are used as exploring sub-population, and the position x of the ith particle in the exploring sub-population is initialized i Historical optimal solution p of particles i Velocity v of particle i The formula is as follows:
x i,j ＝r(ub j -lb j )+lb j ,i∈{1,2,…,N},j∈{1,2,…,D} (1)
v i,j ＝r(Vmax j -Vmin j )+Vmin j (2)
wherein is a [0,1 ]]Random number on, ub j 、lb j The j-th dimension vector, vmax, representing ub and lb, respectively j And Vmin j The upper and lower limits of the velocity of each particle in the j-th dimension are respectively represented, and the search subThe speed of the population is Vmax at the upper and lower limits of the j dimension respectively j ＝ub j 、Vmin j ＝lb j The method comprises the steps of carrying out a first treatment on the surface of the Initially p i ＝x i ；x i 、p i 、v i Vectors of 14 dimensions respectively; x is x i,j Representing the position of particle i in the j-th dimension; v i,j Representing the velocity of particle i in the j-th dimension;
step 22: calculating the fitness of each particle in the initialized exploring sub-population in the step 21, and selecting the position EBest of the exploring sub-population particle with the minimum fitness;
step 23: randomly generating n in the Ebest neighborhood range obtained in the step 22 2 The particle is used as development sub-population, and the position x of the ith particle in the development sub-population is initialized i Historical optimal solution p of particles i Velocity v of particle i The formula of (2) is as follows:
v i,j ＝r(Vmax j -Vmin j )+Vmin j (4)
wherein is a [0,1 ]]Random number, x on i,j Representing the position of particle i in the j-th dimension, v i,j Representing the velocity of particle i in the j-th dimension, EBest j A value representing the position of the least fitness particle in the search sub-population in the j-th dimension;
step 24: the fitness of each particle of the initialized development sub-population of step 23 is calculated.
Further, in the step 3, judging whether the cluster degree of the explored sub-seeds is larger than a threshold value, if so, executing a collision rebound operator, otherwise, entering the step 4;
the specific judging method comprises the following steps: randomly selecting a particle from the current exploring sub-population, calculating the particle number of the exploring sub-population contained in the particle neighborhood, and marking as neib; the particle i neighborhood is defined as [ x ] i,j -R j ,x i,j +R j ]J e {1,2, …, j, …, D }; if neib>λn 1 Then collision rebound is performedThe operator, specifically, the speed of all particles in the exploring sub-population is updated by using the formula (6), and the positions of the particles are updated by using the formula (5), so that the updated exploring sub-population is obtained;
x i,j ＝max(min(x i,j +v i,j ,ub j ),lb j ) (5)
v i,j ＝max(min(u′ i,j ,Vmax j ),Vmin j ) (6)
the collSet in formula (7) i Representing the collection of particles within the neighborhood of particle i, u' i ＝[u′ i，1 ，u′ i，2 ，...，u′ i，j ，...，u′ i，D ]Is a D-dimensional vector, where u' i,j Is u' i Scalar in the j-th dimension; v in formula (8) i And v k Representing the velocity of particles i and k, m i And m k The mass of particles i and k is two real numbers, which are calculated by the formula (9); fit in (9) i Let minfit denote the fitness of the particle i, and δ denotes the minimum fitness value in the search sub-population, δ being an amount that prevents the denominator from being 0, δ=1.
Further, the step 4 specifically includes the following sub-steps:
step 41, updating the speeds and positions of all particles in the updated exploring sub-population obtained in the step 3 by using the formula (11) and the formula (13) respectively to obtain the updated exploring sub-population;
wherein C is E1 ,C E2 Is two [0,2 ]]Constant C of the upper part E1 ＝C E2 ＝2；
Is a [0.4,1 ]]The random number represents the inertia weight of the exploring sub-population at the t-th iteration; r is (r) 1 ,r 2 Is two [0,1 ]]A random number on the table; speed for arbitrary particle i>
Scalar +. >
Share the same r 1 ,r 2 ；/>
Is->
Is a scalar quantity of the j-th dimension of (c),
represents one of the EB particle positions with the minimum adaptability in the exploring sub-population at the t-th iteration, k is randomly selected from the EB particles, EB is preset in advance, and EB epsilon {1,2, …, n 1 }，EB＝3；/>
Is->
J-th dimensional scalar,/->
Representing the optimal position found by the particles i through the t iteration; />
Is->
Scalar on the j-th dimension, +.>
Is->
Scalar in the j-th dimension;
step 42: calculating the fitness of each particle in the updated exploring sub-population obtained in the step 41;
step 43: updating the historical optimal solutions of all particles of the probe population;
the method specifically comprises the following steps: traversing the updated search sub-population particles i obtained in step 41, if
Make->
Otherwise->
Performing elite retention policy on the first EB optimal particles, i.e. if the particles are more adaptive after updating than when not updated +.>
Is the position corresponding to one of the best EB particles.
Further, the step 5 includes the following steps:
step 51: judging the set corresponding to all particles of the current development sub-population i If the two are empty sets, executing the step, then executing the step 6, otherwise executing the step 52;
the specific operation is as follows: traversing each particle in the development sub-population; randomly selecting two particles k and one particle l from the development sub-population; if it is 
Make->
set k ＝set k U-shaped U i; no->
set l ＝set l ∪i；
Step 52: calculating and developing a Ramahalanobis factor eta of each particle in the sub-population;
the method specifically comprises the following steps: (1) Calculate all satisfaction
The lamac factor of the particles of (a);
wherein the method comprises the steps of
Indicating the fitness of the particle k at the t-th iteration; />
Represents the lamac factor of particle i at the t-th iteration;
(2) To satisfy the following
Particles i, eta thereof i Equal to the mean of all η's greater than 0 at the t-th iteration, the formula:
wherein avg represents an averaging function;
step 53: judging
If the number of the particles is less than 2, executing the step, then executing the step 6, otherwise executing the step 54;
the method specifically comprises the following steps: the set corresponding to all particles of the current development sub-population is first set i Setting the sample as an empty set; then traversing each particle in the development sub-population, and randomly selecting two particles k and one particle l from the particles; if it is
Make->
set k ＝set k U-shaped U i; otherwise->
set l ＝set l U-shaped U i; executing the step 6;
step 54: selection of each particle of a current development sub-population by a tournament mechanism
The specific operation is as follows: firstly, the set corresponding to all particles of the current development sub-population is set i Set to empty and then traverse the current setDeveloping each particle in the sub-population; from the slave
Randomly selecting two particles k and l from the particles of (a); if it is 
Make->
set k ＝set k U-shaped U i; otherwise->
set l ＝set l U-shaped U i; executing the step 6;
further, the step 6 specifically includes the following sub-steps:
step 61: updating the speed and the position of the particle i by using the formulas (17) and (19) for each particle in the development sub-population;
wherein C is M1 、C M2 Is two [0,2 ]]Constant of C M1 ＝C M2 ＝2；
Is one of [0.2,0.8 ]]The random number on the random number is used for the random number,representing inertial weights of developing sub-populations at the t-th iteration; r is (r) 1 、r 2 Is two [0,1 ]]A random number on the table; speed for arbitrary particle i>
Scalar +.>
Share the same r 1 ,r 2 ；/>
Is MBest t Jth dimensional scalar, MBest t Representing the position of the particle with the smallest adaptability in the development sub-population at the t-th iteration; />
Is->
Scalar in the j-th dimension; />
Is->
Scalar in the j-th dimension;
step 62: calculating the fitness of each particle in the updated development sub-population obtained in the step 61;
step 63: updating and developing historical optimal solutions of all particles of the sub-population;
the method specifically comprises the following steps: traversing the updated particles i of the development sub-population obtained in step 62, if fitness
Make->
Otherwise->
Further, the step 7 specifically includes the following steps:
judging whether the development sub-population converges or not, specifically: if fit (MBest) t-n )-fit(MBest t )<E, considering that the development sub-population is converged and using the position MBest of the optimal particles of the development sub-population obtained in the step 6 t Replacing the position of the particle with the greatest adaptability of the probe population; otherwise, executing the step 8;
the method comprises the following steps of:
step 81: normalizing the fitness of all particles obtained in the step 4, and specifically adopting logarithmic normalization shown in a formula 20:
in the formula (20), nrzfit i The fitness of the particle i after logarithmic normalization is represented; the EWorstFit represents the maximum fitness of the particles of the exploring sub-population, and is obtained by executing a max function on the fitness of all the particles of the exploring sub-population; EBestFit represents the minimum fitness of the particles in the exploring sub-population, and is obtained by executing a min function on the fitness of all the particles in the exploring sub-population; fit i Representing the fitness of the particles i in the exploring sub-population; e is a natural constant;
step 82: converting the fitness normalized in the step 81 into an intuitive fuzzy set of fitness attributes of the corresponding particles, namely calculating and exploring the properties a of the sub-population particles i 1 Degree of membership on
Non-membership->
And hesitation->
Exp in the above formula represents an exponential function based on e;
step 83: calculating the number of particles of other exploring sub-populations in the neighborhood of each particle in the exploring sub-population, and marking the number of particles in the neighborhood of the particle i as agg i The method comprises the steps of carrying out a first treatment on the surface of the The neighborhood R is obtained by the step 4, and the range of the neighborhood of any particle i in the j-th dimension is expressed as [ agg ] i,j -R j ,agg i,j +R j ]；
Step 84: converting the particle number in the neighborhood of each particle i obtained in step 83 into an intuitive fuzzy set of the aggregation degree attribute, namely calculating the attribute a of the particle i 2 Membership degree of (C)
Non-membership->
And hesitation->
Max agg in the above i Indicating that the largest agg is selected from the particles of the exploring sub-population, the agg being calculated in step 83;
step 85: the two properties of particle i are combined into a comprehensive intuitional fuzzy set, denoted as A i The method comprises the steps of carrying out a first treatment on the surface of the The calculation formula is shown as follows;
in the formula (28)
And->
Respectively representing the membership and the hesitation of the attribute J; weight +.A weight of property J is synthesized in equation (28) using symmetric weight coefficients>
K in the formulas (29), (30) represents participationThe number of attributes for decision making, K is 2; in formula (30), particle i is in attribute a J Membership degree->
And non-membership->
Calculated from steps 82 and 84;
step 86: calculating approximation degree xi of particles i and B according to ideal solution B and negative ideal solution G i The method comprises the steps of carrying out a first treatment on the surface of the The calculation method comprises the following steps:
in the formula, ideal solution B= { mu B ＝1,π B ＝0,γ B =0 }, non-ideal solution g= { μ } G ＝0,π G ＝0,γ G =1 } is two fixed constant parameters; in the above formula D (X, Y) is used to calculate the distance between the intuitionistic ambiguity set X and Y, here using hamming distance calculation, the formula is as follows:
Step 87: and (3) performing zeta corresponding to the particle i in the exploring sub-population obtained in the step (86) i Sequencing from small to large; checking the corresponding x in order i Whether in the forbidden set, if so, check the next x i Otherwise let target=x i Step 88 is entered; if the examination is over and none of the particles satisfies the condition, then let target=x 1 Where x is 1 Representing particles of the exploring sub-population ranked first;
step 88: storing the target obtained in the step 87 into a forbidden set; i.e. forbidset=forbidset;
step 89: by position x of target target Velocity v target History optimal solution p target Respectively replacing and developing positions of optimal particles of sub-populationsSetting x, speed v and history optimal solution p.
Further, the step 9 specifically includes the following sub-steps:
step 91: selecting particles with the smallest fitness from the exploring sub-population updated in the step 4 and the developing sub-population updated in the step 6 as optimal particles Best, and taking the position x of Best best The value in each dimension is converted into the corresponding super parameter in the LSTM model;
step 92: training the model by using the training sample obtained in the step 1 to obtain a trained LSTM model;
step 93: substituting the characteristic data of the landslide displacement to be tested into a trained LSTM model to obtain the predicted landslide displacement distance.
Further, in the steps 2, 4 and 6, the calculating the fitness of the particles means:
position x of particle i i The value in each dimension is converted into the corresponding super parameter in the LSTM model; training the model by using the training sample obtained in the step 1; the test samples are then substituted and finally the sum of the errors of all test samples is multiplied by 100 as the fitness of the particle i.
The method has the beneficial effects that the LSTM model structure is adaptively adjusted, so that the prediction accuracy of the prediction model on the liveness problem is improved. The problems are overcome together by a heuristic optimization algorithm in combination with an intuitive fuzzy set (Intuitionistic Fuzzy Sets, IFS) theory, and the LSTM predictive model is optimized and made more suitable for landslide prediction. Firstly, the heuristic algorithm is a global optimization algorithm for simulating animal natural community behaviors, and has the advantages of simplicity, high calculation speed, no need of objective function information and the like. It has been widely used to solve various black box problems. Secondly, the intuitionistic Fuzzy set theory realizes an extension of the Zadeh Fuzzy Sets (FS) and can describe three states of support, objection and neutral, so that the IFS has stronger analysis capability in a Fuzzy environment and can describe Fuzzy objects in the real world more finely and better fit with the actual situation compared with the FS. Therefore, the IFS may convert abstract, fuzzy, complex environmental information in loess landslide into clear, identifiable environmental factors. Therefore, the optimization capacity of the heuristic algorithm is improved, the model quality is improved, and the prediction accuracy is improved.
Drawings
FIG. 1 is a flow chart of an algorithm of the present invention;
FIG. 2 is a flow chart for calculating fitness;
FIG. 3 is an LSTM initial model enumerated by an embodiment of the present invention;
FIG. 4 is a graph of real prediction results of a comparative experiment according to an embodiment of the present invention;
FIG. 5 is a graph of the predicted results of the IFMHDPSO-LSTM model provided by an embodiment of the invention.
Detailed Description
The objects, technical solutions and advantages of the present invention will become more apparent by the following detailed description of the present invention with reference to the accompanying drawings. It should be understood that the description is only illustrative and not limiting on the scope of the invention.
The experimental platform used in this embodiment is a 64-bit win10 system with matlab2018a installed, a processor bit Intel Core i7-6700 processor, and memory 8G. The LSTM model is implemented by the third party library keras, version 2.0.8.
The invention discloses a landslide displacement prediction method (IFMHDPSO-LSTM) based on an intuitive fuzzy-matrix PSO-LSTM model, which specifically comprises the following steps:
step 1: characteristic data that causes landslide displacement are collected and the data are cleaned. Wherein:
features include landslide displacement distance, rainfall, air temperature, humidity and soil moisture.
The cleansing data includes a uniform format/unit, removal of invalid data, removal of unreasonable data, removal of outlier noise points, and padding data.
Step 2: training samples and test samples are constructed.
Taking all the characteristics in the step 1 as the input of the model, and taking the displacement characteristics (landslide displacement distance) as the output of the model; for any sample, the real output value corresponding to the time T is the input value of the displacement characteristic at the time T+stepN, and stepN is the step length; the data from time T to time T + stepN are synthesized and input as one sample to LSTM. In this example, stepn=5, t units are days, and a total of 900 days of landslide displacement data are collected as a total sample.
And dividing the total sample according to the ratio of 7:3 to obtain a training sample and a test sample.
Step 3: an initial LSTM model is built.
In this embodiment, eight layers of initial models are provided, as shown in fig. 2, where the first layer is an LSTM model, the second seven layers are full-connection layers FC, where the last layer has only one unit, and the number of hidden layer units as output units is 1, and the activation function is fixed to be tanh. The initial model is constructed in this way in this embodiment to prove that the method can still guarantee the result when the super parameters are more. The solution vector, the information to be represented in each dimension, and the optimizing boundary are designed according to the super parameters of the LSTM. In this embodiment, the number of super-parameters is 14, i.e., the search dimension d=14. Wherein the first 7 represent the number of hidden units of the corresponding layer (layer 1-layer 7), the search range is 16,17, …, 32. The last 7 indicate the activation functions used for the corresponding layers (layer 1-layer 7), with search ranges {0,1,2} representing sigmoid, tanh, relu functions, respectively.
The IFMHDPSO-LSTM of the present invention is an algorithm that searches in continuous space, so the search space is adjusted to: front 7 dimensions [15.5,32.4], back 7 dimensions [ -0.4,2.4]. The lower and upper bounds lb and ub are therefore respectively: lb= [15.5,15.5,15.5,15.5,15.5,15.5,15.5, -0.4, -0.4, -0.4, -0.4, -0.4, -0.4, -0.4]; ub= [32.4,32.4,32.4,32.4,32.4,32.4,32.4,2.4,2.4,2.4,2.4,2.4,2.4,2.4].
The first 7 dimensions of the search space correspond in turn to the number of hidden units per layer of the model of fig. 2. The 7-dimensional last corresponds to the activation function of each layer of the model of fig. 2 in turn;
in this embodiment, including in the comparative experiment, the loss functions of the LSTM model are Mean Square Error (MSE) functions, and the back propagation algorithms are adam algorithms.
Step (a) 4: parameters are set for the IFMHDPSO algorithm.
The parameters to be set include: population size N, i.e. exploring the number of sub-population particles N 1 Development ofParticle count n of sub-population 2 ＝N-n 1 And n is 1 >n 2 The method comprises the steps of carrying out a first treatment on the surface of the Maximum iteration number maxIter; developing the sub-population convergence decision number s epsilon {1,2, …, maxIter }; judgment error epsilon [0,1 ]]The method comprises the steps of carrying out a first treatment on the surface of the Intuitive fuzzy decision of membership degree of each attribute weight
And hesitation degree
And->
The method shares two attributes: a, a 1 And a 2 The method comprises the steps of carrying out a first treatment on the surface of the Collision triggering ratio lambda epsilon 0,1 ]The method comprises the steps of carrying out a first treatment on the surface of the A neighborhood range R; the speed of the probe sub-population is at the upper and lower limit Vmax of the j dimension j And Vmin j The method comprises the steps of carrying out a first treatment on the surface of the Search space upper bound ub, lower bound lb; the dimension D is searched.
In this embodiment, n=40, N 1 ＝24,n 2 ＝16，maxIter＝40，s＝2,ε＝0.001，
λ=0.8, per-dimension vector in neighborhood R +.>
Vmax j ＝R j 、Vmin j ＝-R j Ub, lb, D are obtained in step 3, where ub j And lb j The j-th dimension vectors of ub and lb are represented, respectively.
Initially, the set of all particles i
Iteration number t=1, forbidden set +.>
Step 5: initializing a probe sub-population and a development sub-population, and calculating fitness of each particle in various populations. The method specifically comprises the following substeps:
step 5-1: randomly generating n within the search space of step 4 1 The particles are used as exploring sub-population, and the position x of the ith particle in the exploring sub-population is initialized i Historical optimal solution p of particles i Velocity v of particle i The formula is as follows:
x i,j ＝r(ub j -lb j )+lb j ,i∈{1,2,…,N},j∈{1,2,…,D} (1)
v i,j ＝r(Vmax j -Vmin j )+Vmin j (2)
wherein is a [0,1 ]]Random number on, where ub j And lb j The j-th dimension vectors of ub and lb are represented, respectively. Vmax j And Vmin j The upper and lower limits of the velocity of each particle in the j-th dimension are shown, respectively, and the velocity of the search sub-population in this embodiment is Vmax in the j-th dimension j ＝ub j 、Vmin j ＝lb j . Initially p i ＝x i The method comprises the steps of carrying out a first treatment on the surface of the D is obtained in the step 3; x is x i 、p i 、v i Vectors of 14 dimensions respectively; x is x i,j Representing the position of particle i in the j-th dimension; v i,j Representing the velocity of particle i in the j-th dimension;
step 5-2: and (3) calculating the fitness of each particle in the initialized exploring sub-population in the step (5-1), and selecting the position EBest of the particle in the exploring sub-population with the minimum fitness.
The method specifically comprises the following steps: traversing the search sub-population to locate the position x of particle i i The value in each dimension is converted into the corresponding super parameter in the LSTM model; training the model by using the training sample obtained in the step 2; then substituting the test sample; finally the error sum of all test samples is multiplied by 100 as the fitness of particle i. Here multiplied by 100 to improve the differentiation.
Step 5-3: randomly generating n in the Ebest neighborhood range obtained in the step 5-2 2 The particle is used as development sub-population, and the position x of the ith particle in the development sub-population is initialized i Historical optimal solution p of particles i Velocity v of particle i The formula of (2) is as follows:
v i,j ＝r(Vmax j -Vmin j )+Vmin j (4)
wherein is a [0,1 ]]Random number, x on i,j Representing the position of particle i in the j-th dimension, v i,j Representing the velocity of particle i in the j-th dimension, EBest j The value of the position of the minimum particle with the lowest fitness in the exploring sub-population in the j-th dimension is represented, and the range of the neighborhood R is obtained in the step 4.
Step 5-4: and calculating the fitness of each particle of the initialized development sub-population in the step 5-3.
The method specifically comprises the following steps: traversing the development sub-population to locate the position x of particle i i The value in each dimension is converted into the corresponding super parameter in the LSTM model; training the model by using the training sample obtained in the step 2; then substituting the test sample; finally the error sum of all test samples is multiplied by 100 as the fitness of particle i. Here multiplied by 100 is to improve the differentiation.
Step 6: judging whether the cluster concentration of the explored sub-species is larger than a threshold value, if so, entering a step 7, otherwise, entering a step 8.
The specific judging method comprises the following steps: randomly selecting one particle from the current exploring sub-population, calculating the particle number of the exploring sub-population contained in the particle neighborhood, and marking the particle number as neib. The particle i neighborhood is defined as [ x ] i,j -R j ,x i,j +R j ]J ε {1,2, …, j, …, D }. If neib>λn 1 Step 7 is entered, otherwise step 8 is entered, wherein λ is as defined in step 4.
Step 7: a collision rebound operator is performed. Specifically, the speeds of all particles in the search sub-population are updated by using the formula (6), and the positions of the particles are updated by using the formula (5), so that an updated search sub-population is obtained.
x i,j ＝max(min(x i,j +v i,j ,ub j ),lb j ) (5)
v i,j ＝max(min(u′ i,j ,Vmax j ),Vmin j ) (6)
The collSet in formula (7) i Representing the collection of particles within the neighborhood of particle i, u' i ＝[u′ i,1 ,u′ i,2 ,…,u′ i,j ,…,u′ i,D ]Is a D-dimensional vector, where u' i,j Is u' i Scalar in the j-th dimension. V in formula (8) i And v k Representing the velocity of particles i and k, m i And m k The mass of particle i and particle k is calculated from equation (9), and is two real numbers. Fit in (9) i Let minfit denote the fitness of the particle i, and δ denotes the minimum fitness value in the search sub-population, which is an amount that prevents the denominator from being 0, typically δ=1.
Step 8: updating the exploring sub-population. The method specifically comprises the following substeps:
and 8-1, updating the speeds and positions of all particles in the updated exploratory sub-population obtained in the step 7 by using the formula (11) and the formula (13) respectively to obtain the updated exploratory sub-population.
Wherein C is E1 ,C E2 Is two [0,2 ]]Constant of above, C in this example E1 ＝C E2 ＝2。
Is a [0.4,1 ]]The random number on the model represents the inertia weight of the exploring sub-population at the t-th iteration. r is (r) 1 ,r 2 Is two [0,1 ]]Random numbers on the same. Speed for arbitrary particle i>
Scalar +.>
Share the same r 1 ,r 2 。/>
Is->
J-th dimensional scalar,/->
Represents one of the EB particle positions with the minimum adaptability in the exploring sub-population at the t-th iteration, k is randomly selected from the EB particles, EB is preset in advance, and EB epsilon {1,2, …, n 1 Eb=3 in this example. />
Is->
J-th dimensional scalar,/->
The optimal position found by the particles i through the t-th iteration (i.e., the position corresponding to the particle i when the fitness of the particle i is minimum in the t-th iteration) is represented. / >
Is->
Scalar on the j-th dimension, +.>
Is->
Scalar in the j-th dimension.
Step 8-2: and (3) calculating the fitness of each particle in the updated exploring sub-population obtained in the step 8-1.
The method specifically comprises the following steps: traversing the search sub-population to locate the position x of particle i i The value in each dimension is converted into the corresponding super parameter in the LSTM model; training the model by using the training sample obtained in the step 2; then substituting the test sample, and multiplying the total error by 100 to obtain the fitness of the particle i.
Step 8-3: and updating the historical optimal solutions of all particles of the probe population.
The method specifically comprises the following steps: traversing the updated particle i of the search sub-population obtained in step 8-1 if
Make->
Otherwise->
Performing elite retention policy on the first EB optimal particles, i.e. if the particles are more adaptive after updating than when not updated +.>
Is the position corresponding to one of the best EB particles.
Step 9: calculating the Ramahalanobis factor of each particle in the development sub-population and selecting for each particle of the development sub-population
This step includes the following sub-steps and all operations are performed only on the development sub-population.
Step 9-1: judging the set corresponding to all particles of the current development sub-population i If they are empty sets, the step is executed, then step 10 is executed, otherwise 9-2 is executed.
The specific operation is as follows: traversing each particle in the development sub-population; randomly selecting two particles k and one particle l from the development sub-population; if it is
Make->
set k ＝set k U-shaped U i; no->
set l ＝set l ∪i。
Step 9-2: the lamac factor η is calculated for each particle in the development sub-population. The method specifically comprises the following steps: (1) Calculate all satisfaction
Is a lamac factor of the particles of (a).
Wherein the method comprises the steps of
Indicating the fitness of particle k at the t-th iteration. />
The lamac factor of particle i at the t-th iteration is represented.
(2) To satisfy the following
Particles i, eta thereof i Equal to the mean of all η's greater than 0 at the t-th iteration. Namely the following formula:
where avg represents the averaging function.
Step 9-3: judging
If the number of particles is less than 2, then the step is executed, then step 10 is executed, otherwise 9-4 is executed.
The method specifically comprises the following steps: the set corresponding to all particles of the current development sub-population is first set i Put it as empty set. Then traversing each particle in the development sub-population, and randomly selecting two particles k and one particle l from the particles; if it is
Make->
set k ＝set k U-shaped U i; otherwise->
set l ＝set l U-shaped U i; executing step 10;
step 9-4: developing sub-populations for current use by tournament mechanismsEach particle selection
The specific operation is as follows: firstly, the set corresponding to all particles of the current development sub-population is set i Setting an empty set, and then traversing each particle in the current development sub-population; from the slave
Randomly selecting two particles k and l from the particles of (a); if it is
Make->
set k ＝set k U-shaped U i; otherwise->
set l ＝set l U-shaped U i; executing step 10;
step 10: the speed and location of developing sub-population particles is updated. The method specifically comprises the following steps:
step 10-1: for each particle in the development sub-population, the velocity and position of particle i are updated by equations (17) and (19).
Wherein C is M1 、C M2 Is two [0,2 ]]Constant of above, C in this example M1 ＝C M2 ＝2。
Is one of [0.2,0.8 ]]And the random number on the model represents the inertia weight of the development sub-population at the t-th iteration. r is (r) 1 、r 2 Is two [0,1 ]]Random numbers on the same. Speed for arbitrary particle i>
Scalar +.>
Share the same r 1 ,r 2 。/>
Is MBest t Jth dimensional scalar, MBest t Representing the position of the particle with the smallest adaptability in the development sub-population at the t-th iteration; here, for developing all particles of a sub-population, their maximum and minimum velocities Vmax j And Vmax j Defined by step 4. />
Is->
Scalar in the j-th dimension. />
Is->
Scalar in the j-th dimension.
Step 10-2: and calculating the fitness of each particle in the updated development sub-population obtained in the step 10-1.
The method specifically comprises the following steps:traversing the development sub-population to locate the position x of particle i i The value in each dimension is converted into the corresponding super parameter in the LSTM model; training the model by using the training sample obtained in the step 2; then substituting the test sample, and multiplying the total error by 100 to obtain the fitness of the particle i.
Step 10-3: and updating and developing historical optimal solutions of all particles of the sub-population.
The method specifically comprises the following steps: traversing the updated particles i of the development sub-population obtained in step 10-2 if
Make->
Otherwise->
Step 11: and judging whether the development sub-population obtained in the step 10 is converged.
If fit (MBest) t-n )-fit(MBest t )<E, considering that the development sub-population is converged, and using the position MBest of the optimal particles of the development sub-population obtained in the step 10-1 t Replacing the position of the particle with the greatest adaptability of the probe population; otherwise, step 14 is performed. s, ε, maxIter are available from step 4.
Step 12: the optimal target area is selected from the current exploring sub-population through intuitive fuzzy decision, and the method specifically comprises the following substeps, which are only executed on the exploring sub-population particles if no special explanation exists.
Step 12-1: and normalizing the fitness of all the particles obtained in the step 8. Preferably, the normalization method employs logarithmic normalization as shown in equation 20.
In the formula (20), nrzfit i The fitness of particle i after logarithmic normalization is shown. EWorstFit represents the exploring sub-population grains The maximum fitness of the sub-population is obtained by performing a max function on the fitness of all particles of the exploring sub-population. EBestFit represents the minimum fitness of particles in the exploring sub-population, obtained by performing a min function on the fitness of all particles in the exploring sub-population. fit i Indicating the fitness of the particles i in the search sub-population. e is a natural constant.
Step 12-2: converting the fitness after normalization in the step 12-1 into an intuitive fuzzy set of fitness attributes of the corresponding particles, namely calculating and exploring the properties a of the sub-population particles i 1 Degree of membership on
Non-membership->
And hesitation->
Exp in the above formula represents an exponential function based on e.
Step 12-3: calculating the number of particles of other exploring sub-populations in the neighborhood of each particle in the exploring sub-population, and marking the number of particles in the neighborhood of the particle i as agg i . The neighborhood R is obtained by the step 4, and the range of the neighborhood of any particle i in the j-th dimension is expressed as [ agg ] i,j -R j ,agg i,j +R j ]。
Step 12-4: counting the number of particles in the vicinity of each particle i obtained in step 12-3Conversion to an intuitive fuzzy set of concentration properties, i.e. computing the properties a of particle i 2 Membership degree of (C)
Non-membership->
And hesitation->
Max agg in the above i Indicating that the largest agg is selected from the particles exploring the sub-population, the agg being calculated in step 12-3.
Step 12-5: two attributes of particle i (attribute a 1 Attribute a 2 ) Is synthesized into a comprehensive intuitionistic fuzzy set, which is marked as A i . The calculation formula is shown as follows.
In the formula (28)
And->
The membership and hesitation of the attribute J are respectively represented, and the weight of each attribute is represented by the result of step 4. Weight +.A weight of property J is synthesized in equation (28) using symmetric weight coefficients>
In the formulas (29) and (30), K represents the number of attributes involved in decision, and K is 2 in step 4. In formula (30), particle i is in attribute a J Membership degree->
And non-membership degree
Calculated from step 12-2 and step 12-4.
Step 12-6: calculating approximation degree xi of particles i and B according to ideal solution B and negative ideal solution G i 。
The calculation method is as follows.
In the formula, ideal solution B= { mu B ＝1,π B ＝0,γ B =0 }, non-ideal solution g= { μ } G ＝0,π G ＝0,γ G =1 } are two fixed constant parameters. In the above formula D (X, Y) is used forThe distance between the intuitionistic ambiguity sets X and Y is calculated, here using the hamming distance calculation, as follows.
Step 12-7: and (3) performing zeta corresponding to the particle i in the exploring sub-population obtained in the step 12-6 i Ordering from small to large. Checking the corresponding x in order i Whether in the forbidden set, if so, check the next x i Otherwise let target=x i Step 12-8 is entered. If the check is over, none of the particles satisfies the condition, i.e. all particles are in the forbidden set, then target=x 1 Where x is 1 Representing the particles of the first population of explored sub-population.
Check x i Method of whether in forbidden set: traversing all forbidden locations in the forbidden set forbidSet, using forbid k Represents the position where k is forbidden, if x i In forbid k In the neighborhood, i.e. x i ∈[forbid k -R,forbid k +R]Then consider to be in the forbidden region if x i The particle is not considered to be in the forbidden set if it is not in all forbidden regions.
Step 12-8: and storing the target obtained in the step 12-7 into a forbidden set. I.e. forbidset=forbidset;
step 13: by position x of target target Velocity v target History optimal solution p target And respectively replacing the position x, the speed v and the historical optimal solution p of the optimal particles of the development sub-population.
The replacement action of the step is to exchange information contained in the two sub-populations, so that the optimizing efficiency of the algorithm is improved.
Step 14: judging whether the iteration times t is greater than maxIter, if not, enabling t=t+1, and returning to the step 6; if so, step 15 is entered.
Step 15: and obtaining a final prediction model, training and predicting. The method specifically comprises the following substeps:
step 15-1: selecting particles with minimum fitness from the exploring sub-population updated in the step 8 and the developing sub-population updated in the step 10 as optimal particles Best, and taking the position x of Best best The values in each dimension are converted into corresponding superparameters in the LSTM model.
Step 15-2: and (3) training the model by using the training sample obtained in the step (2) to obtain a trained LSTM model.
Step 15-3: substituting the characteristic data of the landslide displacement to be tested into a trained LSTM model to obtain the predicted landslide displacement distance.
The method of the present invention ends.
In order to verify the effect of the method, a simulation experiment is carried out on the same experiment platform by using a common LSTM model of manual parameter adjustment, wherein a predefined model is shown as a figure 3, and the corresponding super parameter is [32,32,32,32,32,32,32,1,1,1,1,1,1,1], which indicates that the number of hidden layer units from layer 1 to layer 7 is 32; the activation functions of the 1 st layer to the 7 th layer are tanh; the hidden unit of the output layer is 1, and the activation function is a tanh function. The data preprocessing steps are as step 1 and step 2 of the method. And carrying the obtained training sample into the LSTM model for training, and carrying the test sample into the trained test model. The output result is compared with the true value, and the experimental result is shown in fig. 4. Substituting the test sample obtained in the step 2 into the trained LSTM model obtained in the step 15-2, and comparing the output result with a true value to obtain a test result figure 5. In fig. 5, the dashed line represents the predicted result, the realization represents the real result, and the closer the two curves are, the better the result.
As can be seen by comparing fig. 4 and 5, the predicted result (dotted line) and the true value (solid line) of fig. 5 are more similar, whereas the predicted result and the real data in fig. 4 differ considerably. The method is characterized in that the traditional method for manually defining the super-parameters cannot effectively process landslide problems with large scale, nonlinearity, high information redundancy, ambiguity, uncertainty and high noise, and an algorithm overcomes the uncertainty problems and adaptively adjusts the super-parameters of the neural network through an intuitionistic ambiguity theory. The final experimental result proves that the method for predicting landslide displacement based on the intuitive fuzzy-based PSO-LSTM model (IFMHDPSO-LSTM for short) can better adapt to various characteristics of landslide data, the model quality and the prediction precision are obviously improved, and the method is superior to the existing LSTM prediction model needing manual adjustment.
Claims (4)
1. A landslide displacement prediction method based on an intuitionistic fuzzy-matrix PSO-LSTM model is characterized by comprising the following steps:
step 1: collecting characteristic data causing landslide displacement and cleaning the data to form a training sample and a test sample, establishing an initial LSTM model, and setting parameters for the landslide displacement prediction method based on the intuitive fuzzy denominator PSO-LSTM model; the characteristics comprise landslide displacement distance, rainfall, air temperature, humidity and soil moisture;
Step 2: initializing a probe sub-population and a development sub-population, and calculating the fitness of each particle in each sub-population; the method specifically comprises the following substeps:
step 21: randomly generating n in the search space range of step 1 1 The particles are used as exploring sub-population, and the position x of the ith particle in the exploring sub-population is initialized i Historical optimal solution p of particles i Velocity v of particle i The formula is as follows:
x i，j ＝r(ub j -lb j )+lb j ，i∈{1，2，...，N}，j∈{1，2，...，D} (1)
v i，j ＝r(Vmax j -Vmin j )+Vmin j (2)
wherein r is a [0,1 ]]Random number on, ub j 、lb j The j-th dimension vector, vmax, representing ub and lb, respectively j And Vmin j The upper and lower limits of the speed of each particle in the j-th dimension are respectively represented, and the speed of the exploring sub-population is respectively Vmax in the upper and lower limits of the j-th dimension j ＝ub j 、Vmin j ＝lb j The method comprises the steps of carrying out a first treatment on the surface of the Initially p i ＝x i ；x i 、p i 、v i Vectors of 14 dimensions respectively; x is x i，j Representing the position of particle i in the j-th dimension; v i，j Representing the velocity of particle i in the j-th dimension;
step 22: calculating the fitness of each particle in the initialized exploring sub-population in the step 21, and selecting the position EBest of the exploring sub-population particle with the minimum fitness;
step 23: randomly generating n in the Ebest neighborhood range obtained in the step 22 2 The particle is used as development sub-population, and the position x of the ith particle in the development sub-population is initialized i Historical optimal solution p of particles i Velocity v of particle i The formula of (2) is as follows:
v i，j ＝r(Vmax j -Vmin j )+Vmin j (4)
wherein r is a [0,1 ]]Random number, x on i，j Representing the position of particle i in the j-th dimension, v i，j Representing the velocity of particle i in the j-th dimension, EBest j A value representing the position of the least fitness particle in the search sub-population in the j-th dimension;
step 24: calculating the fitness of each particle of the initialized development sub-population in the step 23;
step 3: if the cluster concentration of the explored sub-species reaches the threshold value, executing a collision rebound operator, otherwise executing the step 4; the specific judging method for exploring whether the grouping degree of the sub-seeds is larger than a threshold value comprises the following steps:
randomly selecting a particle from the current exploring sub-population, calculating the particle number of the exploring sub-population contained in the particle neighborhood, and marking as neib; the particle i neighborhood is defined as [ x ] i，j -R j ，x i，j +R j ]J e {1,2,., j, D }; if neib > λn 1 Then executing collision rebound operator, specifically using formula (6) to update the speed of all particles in the exploring sub-population, using formula (5) to update the positions of the particles, and obtaining the exploring sub-population after updating;
x i，j ＝max(min(x i，j +v i，j ，ub j )，lb j ) (5)
v i，j ＝max(min(u′ i，j ，Vmax j )，Vmin j ) (6)
the collSet in formula (7) i Representing the collection of particles within the neighborhood of particle i, u' i ＝[u′ i，1 ，u′ i，2 ，...，u′ i，j ，...，u′ i，D ]Is a D-dimensional vector, where u' i，j Is u' i Scalar in the j-th dimension; v in formula (8) i And v k Representing the velocity of particles i and k, m i And m k The mass of particles i and k is two real numbers, which are calculated by the formula (9); fit in (9) i Representing the fitness of particle i, minfit representing the minimum fitness value in the search sub-population, δ being an amount preventing the denominator from being 0, δ=1;
step 4: updating the position and speed of the probe population, calculating the fitness and updating the historical optimal solution; the method specifically comprises the following substeps:
step 41, updating the speeds and positions of all particles in the updated exploring sub-population obtained in the step 3 by using the formula (11) and the formula (13) respectively to obtain the updated exploring sub-population;
wherein C is E1 ，C E2 Is two [0,2 ]]Constant C of the upper part E1 ＝C E2 ＝2；
Is a [0.4,1 ]]The random number represents the inertia weight of the exploring sub-population at the t-th iteration; r is (r) 1 ，r 2 Is two [0,1 ]]A random number on the table; velocity for arbitrary particle i
Scalar +.>
Share the same r 1 ，r 2 ；/>
Is->
Is a scalar quantity of the j-th dimension of (c),
the method is characterized in that one of the positions of EB particles with the minimum fitness in the sub-population is explored at the t-th iteration, k is randomly selected from the EB particles, EB is preset in advance, EB epsilon {1, 2., n 1 }，EB＝3；/>
Is->
J-th dimensional scalar,/->
Representing the optimal position found by the particles i through the t iteration; />
Is- >
Scalar on the j-th dimension, +.>
Is->
Scalar in the j-th dimension;
step 42: calculating the fitness of each particle in the updated exploring sub-population obtained in the step 41;
step 43: updating the historical optimal solutions of all particles of the probe population;
the method specifically comprises the following steps: traversing the updated search sub-population particles i obtained in step 41, if
Make->
Otherwise->
Performing elite retention policies on the first EB best particles, i.e. if the particles adapt after updating than when not updatedBigger, then->
Is the position corresponding to one of the best RB particles;
step 5: calculating and developing the Ramahalanobis factor of each particle in the sub-population; the method specifically comprises the following steps:
step 51: judging the set corresponding to all particles of the current development sub-population i If the two are empty sets, executing the step, then executing the step 6, otherwise executing the step 52;
the specific operation is as follows: traversing each particle in the development sub-population; randomly selecting two particles k and particles 1 from the development sub-population; if it is
Make->
No->
set l ＝set l ∪i；
Step 52: calculating and developing a Ramahalanobis factor eta of each particle in the sub-population;
the method specifically comprises the following steps: (1) Calculate all satisfaction
The lamac factor of the particles of (a);
wherein the method comprises the steps of
Indicating the fitness of the particle k at the t-th iteration; / >
Represents the lamac factor of particle i at the t-th iteration;
(2) To satisfy the following
Particles i, eta thereof i Equal to the mean of all η's greater than 0 at the t-th iteration, the formula:
wherein avg represents an averaging function;
step 53: judging
If the number of the particles is less than 2, executing the step, then executing the step 6, otherwise executing the step 54;
the method specifically comprises the following steps: the set corresponding to all particles of the current development sub-population is first set i Setting the sample as an empty set; then traversing each particle in the development sub-population, and randomly selecting two particles k and particle 1 from the particles; if it is
Order of principle
set k ＝set k U-shaped U i; otherwise->
set l ＝set l U-shaped U i; executing the step 6;
step 54: selection of each particle of a current development sub-population by a tournament mechanism
The specific operation is as follows: firstly, the set corresponding to all particles of the current development sub-population is set i Setting an empty set, and then traversing each particle in the current development sub-population; from the slave
Randomly selecting two particles k and 1 from the particles of (a); if it is
Make->
Otherwise->
set l ＝set l U-shaped U i; executing the step 6;
step 6: updating the speed and the position of the particles of the development sub-population, calculating the adaptability of each particle in the development sub-population, and updating the historical optimal solution; the method specifically comprises the following substeps:
Step 61: updating the speed and the position of the particle i by using the formulas (17) and (19) for each particle in the development sub-population;
wherein C is M1 、C M2 Is two [0,2 ]]Constant of C M1 ＝C M2 ＝2；
Is one of [0.2,0.8 ]]The random number represents the inertia weight of the development sub-population at the t-th iteration; r is (r) 1 、r 2 Is two [0,1 ]]A random number on the table; speed for arbitrary particle i>
Scalar +.>
Share the same r 1 ，r 2 ；/>
Is MBest t Jth dimensional scalar, MBest t Representing the position of the particle with the smallest adaptability in the development sub-population at the t-th iteration; />
Is->
Scalar in the j-th dimension; />
Is->
Scalar in the j-th dimension;
step 62: calculating the fitness of each particle in the updated development sub-population obtained in the step 61;
step 63: updating and developing historical optimal solutions of all particles of the sub-population;
the method specifically comprises the following steps: traversing the updated particles i of the development sub-population obtained in step 62, if fitness
Make->
Otherwise->
Step 7: judging whether the development sub-population converges or not, if so, selecting an optimal target area from the current exploration sub-population through an intuitive fuzzy decision, otherwise, executing the step 8;
step 8: judging whether the iteration ending condition is met, if not, adding 1 to the iteration times, returning to the step 3, and if so, entering the step 9;
Step 9: interpreting the optimal solution obtained by searching as LSTM; training the LSTM model by using a training sample to obtain a final prediction model, wherein the method specifically comprises the following substeps:
step 91: selecting particles with the smallest fitness from the exploring sub-population updated in the step 4 and the developing sub-population updated in the step 6 as optimal particles Best, and taking the position x of Best best The value in each dimension is converted into the corresponding super parameter in the LSTM model;
step 92: training the model by using the training sample obtained in the step 1 to obtain a trained LSTM model;
step 93: substituting the characteristic data of the landslide displacement to be tested into a trained LSTM model to obtain the predicted landslide displacement distance.
2. The landslide displacement prediction method based on the intuitionistic fuzzy-mother PSO-LSTM model as claimed in claim 1, wherein in the step 1:
the establishing the initial LSTM model comprises the following steps: the first layer is an LSTM model, the last seven layers are full-connection layers FC, wherein the last layer has only one unit, the number of hidden layer units serving as output units is 1, the activation function is fixed to be tanh, the number of super parameters is 14, namely, the search dimension D=14, the first 7 represents the number of hidden units of the corresponding layer, and the last 7 indicates the activation function used by the corresponding layer;
The setting parameters for the IFMHDPSO algorithm comprises the following steps: setting population size N, i.e. exploring the number of particles N of sub-population 1 Developing the particle number n of the sub-population 2 ＝N-n 1 And n is 1 ＞n 2 The method comprises the steps of carrying out a first treatment on the surface of the Maximum iteration number maxIter; developing a sub-population convergence decision number s e {1, 2.,. MaxIter }; judgment error epsilon [0,1 ]]The method comprises the steps of carrying out a first treatment on the surface of the Intuitive fuzzy decision of membership degree of each attribute weight
And hesitation->
And->
Two attributes a 1 And a 2 The method comprises the steps of carrying out a first treatment on the surface of the Collision triggering ratio lambda epsilon 0,1]The method comprises the steps of carrying out a first treatment on the surface of the A neighborhood range R; the speed of the probe sub-population is at the upper and lower limit Vmax of the j dimension j And Vmin j The method comprises the steps of carrying out a first treatment on the surface of the Search space upper bound ub, lower bound lb; searching dimension D; initially, the set of all particles i +.>
Iteration number t=1, forbidden set +.>
3. The landslide displacement prediction method based on the intuitive fuzzy-based PSO-LSTM model as claimed in claim 1, wherein the step 7 specifically comprises the following steps:
judging whether the development sub-population converges or not, specifically: if fit (MBest) t-n )-fit(MBest t ) And E, considering that the development sub-population is converged and using the position MBest of the optimal particles of the development sub-population obtained in the step 6 t Replacing the position of the particle with the greatest adaptability of the probe population; otherwise, executing the step 8;
the method comprises the following steps of:
Step 81: normalizing the fitness of all particles obtained in the step 4, and specifically adopting logarithmic normalization shown in a formula 20:
in the formula (20), nrzfit i The fitness of the particle i after logarithmic normalization is represented; the EWorstFit represents the maximum fitness of the particles of the exploring sub-population, and is obtained by executing a max function on the fitness of all the particles of the exploring sub-population; EBestFit represents the minimum fitness of the particles in the exploring sub-population, and is obtained by executing a min function on the fitness of all the particles in the exploring sub-population; fit i Representing the fitness of the particles i in the exploring sub-population; e is a natural constant;
step 82: converting the fitness normalized in the step 81 into an intuitive fuzzy set of fitness attributes of the corresponding particles, namely calculating and exploring the properties a of the sub-population particles i 1 Degree of membership on
Non-membership->
And hesitation->
Exp in the above formula represents an exponential function based on e;
step 83: calculating the number of particles of other exploring sub-populations in the neighborhood of each particle in the exploring sub-population, and marking the number of particles in the neighborhood of the particle i as agg i The method comprises the steps of carrying out a first treatment on the surface of the The neighborhood R is obtained by the step 4, and the range of the neighborhood of any particle i in the j-th dimension is expressed as [ agg ] i，j -R j ，agg i，j +R j ]；
Step 84: converting the particle number in the neighborhood of each particle i obtained in step 83 into an intuitive fuzzy set of the aggregation degree attribute, namely calculating the attribute a of the particle i 2 Membership degree of (C)
Non-membership->
And hesitation->
Max agg in the above i Indicating that the largest agg is selected from the particles of the exploring sub-population, the agg being calculated in step 83;
step 85: the two properties of particle i are combined into a comprehensive intuitional fuzzy set, denoted as A i The method comprises the steps of carrying out a first treatment on the surface of the The calculation formula is shown as follows;
in the formula (28)
And->
Respectively representing the membership and the hesitation of the attribute J; weight +.A weight of property J is synthesized in equation (28) using symmetric weight coefficients>
K in the formulas (29) and (30) represents the number of attributes participating in decision making, and K is 2; in formula (30), particle i is in attribute a J Membership degree->
And non-membership->
Calculated from steps 82 and 84;
step 86: calculating approximation degree xi of particles i and B according to ideal solution B and negative ideal solution G i The method comprises the steps of carrying out a first treatment on the surface of the The calculation method comprises the following steps:
in the formula, ideal solution B= { mu B ＝1，π B ＝0，γ B =0 }, non-ideal solution g= { μ } G ＝0，π G ＝0，γ G =1 } is two fixed constant parameters; in the above formula D (X, Y) is used to calculate the distance between the intuitionistic ambiguity set X and Y, here using hamming distance calculation, the formula is as follows:
step 87: and (3) performing zeta corresponding to the particle i in the exploring sub-population obtained in the step (86) i Sequencing from small to large; checking the corresponding x in order i Whether in the forbidden set, if so, check the next x i Otherwise let target=x i Step 88 is entered; if the examination is over and none of the particles satisfies the condition, then let target=x 1 Where x is 1 Representing particles of the exploring sub-population ranked first;
step 88: storing the target obtained in the step 87 into a forbidden set; i.e. forbidset=forbidset;
step 89: by position x of target target Velocity v target History optimal solution p target And respectively replacing the position x, the speed v and the historical optimal solution p of the optimal particles of the development sub-population.
4. The landslide displacement prediction method based on the intuitive fuzzy mother PSO-LSTM model of claim 1, wherein in the 2 nd, 4 th and 6 th steps, the fitness of the particles is calculated by:
position x of particle i i The value in each dimension is converted into the corresponding super parameter in the LSTM model; training the model by using the training sample obtained in the step 1; the test samples are then substituted and finally the sum of the errors of all test samples is multiplied by 100 as the fitness of the particle i.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110244201.6A CN113033879B (en) | 2021-03-05 | 2021-03-05 | Landslide displacement prediction method based on intuitive fuzzy denominator PSO-LSTM |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110244201.6A CN113033879B (en) | 2021-03-05 | 2021-03-05 | Landslide displacement prediction method based on intuitive fuzzy denominator PSO-LSTM |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113033879A CN113033879A (en) | 2021-06-25 |
| CN113033879B true CN113033879B (en) | 2023-07-07 |
Family
ID=76468003
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110244201.6A Active CN113033879B (en) | 2021-03-05 | 2021-03-05 | Landslide displacement prediction method based on intuitive fuzzy denominator PSO-LSTM |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113033879B (en) |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106503788A (en) * | 2016-10-28 | 2017-03-15 | 中国矿业大学 | Least square method supporting vector machine Forecasting Methodology based on self adaptation population |
| CN110427654A (en) * | 2019-07-09 | 2019-11-08 | 中国地质大学（武汉） | A kind of predictive model of landslide construction method and system based on sensitiveness |
| CN111310883A (en) * | 2020-01-16 | 2020-06-19 | 东北石油大学 | Improved particle swarm algorithm for solving high-dimensional optimization problem |
| CN112131794A (en) * | 2020-09-25 | 2020-12-25 | 天津大学 | Hydraulic structure multi-effect optimization prediction and visualization method based on LSTM network |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107590603B (en) * | 2017-09-11 | 2018-11-13 | 合肥工业大学 | Based on the dispatching method and system for improving change neighborhood search and differential evolution algorithm |
-
2021
- 2021-03-05 CN CN202110244201.6A patent/CN113033879B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106503788A (en) * | 2016-10-28 | 2017-03-15 | 中国矿业大学 | Least square method supporting vector machine Forecasting Methodology based on self adaptation population |
| CN110427654A (en) * | 2019-07-09 | 2019-11-08 | 中国地质大学（武汉） | A kind of predictive model of landslide construction method and system based on sensitiveness |
| CN111310883A (en) * | 2020-01-16 | 2020-06-19 | 东北石油大学 | Improved particle swarm algorithm for solving high-dimensional optimization problem |
| CN112131794A (en) * | 2020-09-25 | 2020-12-25 | 天津大学 | Hydraulic structure multi-effect optimization prediction and visualization method based on LSTM network |
Non-Patent Citations (2)
| Title |
|---|
| Prediction on the Highest Price of the Stock Based on PSO-LSTM Neural Network;YUSHAN ZHANG，and etc;《2019 3rd International Conference on Electronic Information Technology and Computer Engineering (EITCE)》;第1565-1569页 * |
| 基于直觉模糊熵的粒子群多目标优化;苏丁为;周创明;王毅;;计算机科学(第08期);第267-271页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113033879A (en) | 2021-06-25 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Haji et al. | Comparison of optimization techniques based on gradient descent algorithm: A review | |
| Kashef et al. | An advanced ACO algorithm for feature subset selection | |
| WO2022121289A1 (en) | Methods and systems for mining minority-class data samples for training neural network | |
| Neal | Connectionist learning of belief networks | |
| EP1598751B1 (en) | Estimation of distribution algorithm (EDA) | |
| Wang et al. | Sequential inference for deep Gaussian process | |
| CN113138555A (en) | GRNN electric spindle thermal error modeling method based on genetic algorithm optimization | |
| CN113469426A (en) | Photovoltaic output power prediction method and system based on improved BP neural network | |
| Eskandari et al. | Evolutionary multiobjective optimization in noisy problem environments | |
| CN110515845B (en) | Combined test case optimization generation method based on improved IPO strategy | |
| CN111723909A (en) | Optimization method and system of fuzzy neural network model | |
| CN116170328A (en) | Method and device for predicting bandwidth used for graphic coding | |
| Zhang et al. | A learning automata-based particle swarm optimization algorithm for noisy environment | |
| CN113033879B (en) | Landslide displacement prediction method based on intuitive fuzzy denominator PSO-LSTM | |
| CN113821025A (en) | Mobile robot path planning method for optimizing heuristic function through neural network | |
| Gao et al. | Optimization methods for large-scale machine learning | |
| CN116415177A (en) | Classifier parameter identification method based on extreme learning machine | |
| CN115829097A (en) | Air conditioner ultra-short term load prediction method based on VMD and KELM | |
| Maryam et al. | Artificial neural network weights optimization based on imperialist competitive algorithm | |
| CN109919389B (en) | Sea surface temperature time sequence prediction method based on PSO (particle swarm optimization) double-objective optimization | |
| CN113836818A (en) | Ocean current motion prediction algorithm based on BP neural network prediction model | |
| CN113177675A (en) | Air conditioner cold load prediction method based on optimization neural network of longicorn group algorithm | |
| Du et al. | Learning deep generative models with doubly stochastic mcmc | |
| Murphy | Hidden states, hidden structures: Bayesian learning in time series models | |
| YONAR et al. | Evaluation and comparison of metaheuristic methods to estimate the parameters of gamma distribution |
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 |