Photovoltaic station landslide risk probability evaluation method and system

Abstract

The invention discloses a photovoltaic station landslide risk probability evaluation method and system, and the method comprises the steps: generating a high-quality pseudo sample library through independent identically distributed resampling and integrated clustering, extracting multi-scale spatial-temporal characteristics through a deep learning hybrid model, obtaining a risk evaluation result graph, and combining variance-mean value joint analysis to obtain a risk evaluation result graph. And carrying out danger level division on the danger evaluation result map, and carrying out quantitative analysis on the danger levels to obtain a field landslide danger spatial probability distribution map. Compared with a traditional method, the method has the advantages that the sample balance degree, the characteristic interpretation, the model reliability and the prediction precision are obviously improved; and a new technical normal form is provided for photovoltaic station full-life-cycle landslide danger management.

Description

Technical Field

[0001] This invention belongs to the field of geological hazard risk assessment technology, and relates to a method and system for assessing the probability of landslide risk at photovoltaic power plants. Background Technology

[0002] As the global energy structure accelerates its transformation towards a low-carbon model, photovoltaic (PV) power generation, as a core carrier of clean energy, has become a key path to achieving the "dual carbon" goal through its large-scale development. By the end of 2023, a certain country's cumulative installed capacity of PV power plants exceeded 520 million kilowatts, with new projects concentrated in the steep, geologically active southwestern mountainous regions and the Loess Plateau. These areas face a severe threat of landslides during construction and operation. Statistics show that in the past five years, the average annual increase in PV power plant equipment damage caused by landslides has reached 17%, resulting in direct economic losses exceeding 1 billion yuan. Traditional landslide hazard assessment methods are mostly based on static geological surveys and expert experience models, which are insufficient to meet the needs of multi-source data fusion and dynamic hazard prediction in complex geological environments.

[0003] Existing research indicates that the core challenge in landslide hazard assessment lies in the coordinated handling of sample imbalance and feature space heterogeneity. Traditional methods employ a single model (such as logistic regression or support vector machine) trained with a limited number of landslide samples. However, labeled landslide samples are scarce, while non-landslide samples constitute an excessively large proportion, leading to model overfitting and weak generalization ability. Furthermore, traditional indicator systems do not consider the spatial coupling effect between factors, and redundant features interfere with model weight allocation. In addition, existing models rely on static historical data for training, failing to capture the spatiotemporal evolution patterns of dynamic triggers such as extreme weather and engineering activities.

[0004] In recent years, some studies have attempted to optimize evaluation models through machine learning, but they still face a series of key bottlenecks, such as: traditional oversampling methods ignore the spatial distribution characteristics of geological data when generating pseudo-samples, leading to noise accumulation and model bias; conventional clustering algorithms have weak ability to model the nonlinear relationships of high-dimensional geological data, making it difficult to discover the local spatial patterns of key disaster-causing factors; and existing models lack probabilistic interpretation and stability assessment of prediction results, making it difficult to support engineering hazard decision-making. Summary of the Invention

[0005] The purpose of this invention is to solve the problem in the prior art that the uneven distribution of labeled landslide samples and non-landslide samples leads to model overfitting, weak generalization ability, and low prediction accuracy, and to provide a method and system for evaluating the probability of landslide hazard at photovoltaic power plants.

[0006] To achieve the above objectives, the present invention employs the following technical solution:

[0007] A method for assessing the probability of landslide hazard at photovoltaic power plants includes the following steps:

[0008] Obtain a sample set E of non-landslide hazards and a sample set A of landslide hazards for preliminary classification of hazard levels;

[0009] Sampling is performed on landslide disaster sample set A to obtain pseudo landslide disaster sample data A'. Sample set B is generated based on non-landslide disaster sample set E, landslide disaster sample set A, and pseudo landslide disaster sample data A'.

[0010] Cluster analysis is performed on sample set B to obtain pseudo landslide disaster data sample set X. Sample set X' is obtained by sampling from pseudo landslide disaster data sample set X. Based on sample set X', landslide disaster sample set A and non-landslide disaster sample set E, a mixed sample set D is generated.

[0011] A deep learning hybrid model is obtained, and the hybrid sample set D is used as the input of the deep learning hybrid model to obtain the hazard assessment result map. The hazard assessment result map is used to classify the hazard level to obtain the spatial probability distribution map of landslide hazard in the site and obtain the assessment result.

[0012] A further improvement of the present invention is that:

[0013] The acquisition of the landslide hazard sample set A for preliminary classification of hazard levels includes:

[0014] Obtain landslide disaster sample point data;

[0015] Construct a hazard assessment index system for sample disasters;

[0016] Based on the hazard assessment index system for sample disasters, each key factor of each landslide disaster sample point is rated to generate a landslide disaster sample set A.

[0017] The generation of the sample set B includes:

[0018] The landslide disaster sample set A was sampled using the independent and identically distributed sampling method, generating 10T sets of sample data, which constituted the pseudo-landslide disaster sample data A', where T represents the number of pseudo-positive samples taken in each set;

[0019] The non-landslide disaster sample set E, the landslide disaster sample set A, and the pseudo-landslide disaster sample data A' are combined to generate sample set B.

[0020] The acquisition of the pseudo-landslide disaster data sample set X includes:

[0021] Using multiple clustering methods, the pseudo-landslide disaster sample data A' is clustered sequentially. The landslide disaster sample set A is set as positive samples, and the non-landslide disaster sample set E is set as negative samples. The clustering results of the pseudo-landslide disaster sample data A' are then redistributed by combining the positive and negative samples to obtain the pseudo-landslide disaster data sample set X.

[0022] The generation of the mixed sample set D includes:

[0023] From the pseudo-landslide disaster data sample set X, T sets of data are randomly selected to form sample set X':

[0024]

[0025] In the formula, x' (i,j) This represents the level of the j-th key factor in the i-th group of sample data after extraction;

[0026] Sample set D is obtained by combining sample set X', sample set A, and non-landslide sample set E:

[0027]

[0028] The obtained spatial probability distribution map of landslide hazard in the site includes:

[0029] Calculate the average value of the landslide hazard assessment results based on the hazard assessment result map;

[0030] Based on the average value of the landslide hazard risk assessment results, and combined with the risk assessment grading threshold, the average value of the risk is classified into levels to obtain the risk assessment map of the average value of the risk assessment results;

[0031] Calculate the variance of the landslide hazard risk assessment results, classify the variance calculation results into levels, and obtain the variance level map of the hazard risk assessment results;

[0032] Based on the average value result hazard assessment map and the variance level map of the disaster hazard assessment result, the hazard assessment result map is divided into hazard areas to obtain the spatial probability distribution map of landslide hazard in the site and obtain the assessment results.

[0033] The process of dividing the hazard assessment result map into hazard zones based on the average value hazard assessment map and the variance level map of the hazard hazard assessment results includes:

[0034] The average result of the risk assessment, the risk assessment map, and the variance level map of the disaster risk assessment results are both level four;

[0035] If the variance value corresponding to the hazard assessment map is level three or four in the hazard level, the hazard level of the landslide hazard assessment result corresponding to the average value of the current hazard assessment map will be upgraded by one level;

[0036] If the variance value corresponding to the hazard assessment map is level one or level two in the hazard level, the hazard level of the landslide hazard assessment result corresponding to the average value of the current hazard assessment map will remain unchanged.

[0037] A landslide hazard probability assessment system for photovoltaic power plants includes:

[0038] The first dataset acquisition module is used to acquire a non-landslide hazard sample set E and a landslide hazard sample set A for preliminary classification of hazard levels.

[0039] The second dataset acquisition module is used to sample the landslide disaster sample set A to obtain pseudo landslide disaster sample data A', and generate sample set B based on the non-landslide disaster sample set E, the landslide disaster sample set A, and the pseudo landslide disaster sample data A';

[0040] The third dataset acquisition module is used to perform cluster analysis on sample set B to obtain pseudo landslide disaster data sample set X, sample the pseudo landslide disaster data sample set X to obtain sample set X', and generate a mixed sample set D based on sample set X', landslide disaster sample set A and non-landslide disaster sample set E;

[0041] The evaluation acquisition module is used to acquire a deep learning hybrid model. The hybrid sample set D is used as the input of the deep learning hybrid model to obtain a hazard assessment result map. The hazard assessment result map is used to classify the hazard level and obtain a spatial probability distribution map of landslide hazard in the site, thus obtaining the evaluation result.

[0042] A terminal device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of any of the methods described in this invention.

[0043] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of any of the methods described in this invention.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] This invention discloses a method for probabilistic assessment of landslide hazard at photovoltaic power plants. It generates a high-quality pseudo-sample library through independent and identically distributed resampling and ensemble clustering, extracts multi-scale spatiotemporal features using a deep learning hybrid model, obtains a hazard assessment result map, and classifies the hazard levels based on variance-mean joint analysis, achieving quantitative analysis of the hazard levels and obtaining a spatial probability distribution map of landslide hazard in the site. Compared to traditional methods, this method can increase the sample size when the sample number is small, employing multiple clustering methods to verify the reliability of new samples. The increased number of reliable samples significantly reduces errors, resulting in significant improvements in sample balance, interpretability of landslide hazard characteristics, model reliability, and prediction accuracy. This provides a new technical paradigm for landslide hazard management throughout the entire lifecycle of photovoltaic power plants. Attached Figure Description

[0046] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0049] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0050] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0051] In the description of the embodiments of the present invention, it should be noted that if terms such as "upper," "lower," "horizontal," or "inner" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of the invention is in use, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. Furthermore, terms such as "first" and "second" are only used to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0052] Furthermore, the use of the term "horizontal" does not imply that the component must be absolutely horizontal, but rather that it can be slightly tilted. For example, "horizontal" simply means that its direction is more horizontal than "vertical," and does not mean that the structure must be completely horizontal, but can be slightly tilted.

[0053] In the description of the embodiments of the present invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in the present invention according to the specific circumstances.

[0054] The present invention will now be described in further detail with reference to the accompanying drawings:

[0055] See Figure 1 This invention discloses a method for evaluating the probability of landslide hazard at photovoltaic power plants. By using landslide disaster sample enhancement technology, integrated clustering algorithm and CNN-LSTM deep learning hybrid model, a high-precision and adaptive landslide hazard probability evaluation framework is constructed, providing a scientific basis for the prevention and control of landslide disasters at photovoltaic power plants.

[0056] Specifically, the following steps are included:

[0057] Step 1: Determine the key factors for the acquired landslide hazard sample data (areas where disasters have already occurred) in photovoltaic power station areas, classify each key factor, and assign a level to all key factors for each sample to construct a landslide hazard risk assessment index system. The combination of all key factor levels for all samples constitutes sample set A.

[0058] Specifically, it includes:

[0059] 1.1 Constructing a landslide hazard risk assessment index system

[0060] Based on the existing data from n landslide hazard sample points, each sample was rated based on 15 key factors, including lithology, overburden thickness, distance from fault, distance from fold, slope, aspect, slope height, planar curvature, profile curvature, rainfall, distance from road, solar panel density, vegetation cover, soil erosion resistance, and surface deformation rate. The specific rating rules are shown in Table 1 below.

[0061] Table 1 Key Factor Rating Criteria

[0062]

[0063]

[0064]

[0065]

[0066] 1.2 Constructing Sample Set A

[0067] Based on the rating criteria of the above 15 key factors, each key factor of the n landslide disaster sample data is rated to generate sample set A:

[0068]

[0069] In the formula, a (i,j) This represents the j-th key factor level of the i-th sample data.

[0070] Step 2: It is necessary to add T sets of pseudo-landslide hazard sample data from the study area. Based on sample set A, use independent and identically distributed sampling to generate 10 sets of data to form sample set A'. Use a binary classification method to establish an ensemble clustering sample set B for the acquired landslide hazard data from the photovoltaic power station area.

[0071] Specifically, it includes the following:

[0072] 2.1 Independent and Identically Distributed Sampling

[0073] (1) Hierarchical experience probability

[0074] We calculate the frequency of occurrence of each level k for each key factor in the statistical sample set A. Since there are only n data points, the distribution estimation may be inaccurate; therefore, Laplace smoothing is introduced. The formula for calculating the empirical probability of each level is:

[0075]

[0076] In the formula, p jk N represents the empirical probability of level k in the j-th key factor; jk K represents the number of k levels in the j-th key factor; j This represents the total number of levels for key factor j. When performing empirical probability statistics on levels, it is assumed that any two key factors are independent of each other, and no correlation is considered.

[0077] (2) Generate new samples by independent and identically distributed sampling.

[0078] For key factor j, based on the empirical probability distribution {p} of the corresponding level j1 ,p j2 ,…,p jk Perform multinomial sampling:

[0079] X ij =Catcgorical(p j1 ,p j2 ,…,p jk (0.2)

[0080] (3) Sample data generation

[0081] Based on steps (1) and (2), 10T (T being the number of pseudo-positive samples in each group) sets of sample data are generated, forming sample set A':

[0082]

[0083] 2.2 Constructing an integrated clustering sample set

[0084] Obtain m (m≥3n) groups of non-landslide disaster sample point data, and rate each key factor of the sample point data according to the rating criteria of the 15 key factors in step one to generate a non-landslide disaster sample set E.

[0085] Sample set B is formed by combining sample set A, non-landslide disaster sample set E, and sample set A':

[0086]

[0087] Step 3: Perform cluster analysis on sample set B using six algorithms: K-means clustering, hierarchical clustering, LCA clustering, K-medoids clustering, spectral clustering, and self-organizing map neural network clustering. Output the pseudo-landslide disaster data sample set X obtained after the clustering algorithms are judged. The number of samples in sample set X is greater than T.

[0088] In step three, six clustering algorithms are used to perform cluster analysis on sample set B. Based on the comprehensive judgment results of the six clustering methods on the data of each group in sample set A', the pseudo-landslide disaster sample set X is obtained. It includes the following:

[0089] 3.1 K-means clustering analysis

[0090] (1) Data processing

[0091] Min-Man standardization is adopted:

[0092]

[0093] In the formula, b ik This represents the original rating value of the k-th replacement factor for the i-th sample; max k ,min k This represents the maximum and minimum original values ​​of the k-th key factor.

[0094] (2) Determine the number of clusters

[0095] Since there are two classes of samples, positive and negative, we set the number of clusters K = 2.

[0096] (3) Initialize cluster centers (K-means++ optimization)

[0097] Randomly select the first center point μ1, then perform cluster optimization, and select the center point farthest from the existing center points as the second center point μ2. At this point, the number of center points is equal to the number of clusters K, and the cluster optimization ends, outputting the two center point values.

[0098] (4) Iterative optimization

[0099] Assign each region i to the nearest cluster center μ. k :

[0100] Cluster(i) = argmin|S i -μ k | (0.6)

[0101] Then recalculate the center of each cluster:

[0102]

[0103] The newly generated cluster centers μ k Substitute into Formula 1.7 and repeat the above steps iteratively until the center point no longer changes, at which point the calculation stops. Output the final two cluster centers, which are the positive sample cluster center and the negative sample cluster center.

[0104] Based on the results of K-means clustering analysis in classifying each group of data in sample set A' as positive or negative samples, matrix C1 is obtained.

[0105]

[0106] In the formula, if the i-th group of data in the sample set A' is identified as positive sample data by K-means clustering analysis, then c (i,1) It equals 1, otherwise it equals 0.

[0107] 3.2 Hierarchical Cluster Analysis

[0108] (1) Data processing

[0109] Z-Score standardization is adopted:

[0110]

[0111] In the formula, b' ik b represents the standardized value of the edge; ik μ represents the original value of the k-th key factor for the i-th sample; k σ represents the mean of the k-th key factor; k This represents the standard deviation of the k-th critical factor.

[0112] (2) Calculate the distance matrix between samples

[0113] The distance between samples is calculated using the Euclidean distance formula, which is as follows:

[0114]

[0115] In the formula, d(i,j) represents the distance between the i-th sample and the j-th sample; x' ik This represents the standardized value of the k-th key factor for the i-th sample.

[0116] (3) Hierarchical clustering process

[0117] The Ward method is chosen to minimize the intra-cluster variance. Each sample is treated as a separate cluster:

[0118] N = {{x1}, {x2}, ..., {x} n+m+10N}}(0.11)

[0119] Calculate all inter-cluster distances and select the two clusters C that minimize the merge cost. p and C q :

[0120]

[0121] In the formula, |C p | represents cluster C p Number of samples; μC p Indicates cluster C p The centroid (mean vector of each key factor).

[0122] Merge C p and C q Update the distance matrix. The iteration ends when all samples cluster into one cluster.

[0123] (4) Determine the number of clusters (K=2)

[0124] The cut height is determined using a dendrogram, dividing the samples into two clusters. The cut height is selected from the position in the dendrogram with the largest distance jump (corresponding to significant differences between the two classes). Two clusters are output: C1 (positive sample candidate set) and C2 (negative sample candidate set).

[0125] (5) Cluster label allocation

[0126] Determine the positive and negative sample labels based on the known information, calculate the centroids of the two clusters, and compare the typical characteristics of landslides:

[0127]

[0128] If the centroid is significantly higher in the high-risk rating of the key factor, then N1 is a positive sample set.

[0129] Based on hierarchical cluster analysis, the data in sample set A' are classified as positive or negative samples, resulting in matrix C2.

[0130]

[0131] In the formula, if the i-th data in the sample set A' is identified as a positive sample data by hierarchical cluster analysis, then c (i,2) It equals 1, otherwise it equals 0.

[0132] 3.3 LCA Cluster Analysis

[0133] (1) Data processing

[0134] Each sample represents a 15-dimensional classification vector:

[0135] B i =(b i1 ,b i2 ,…,b i15 (0.15)

[0136] In the formula, b ik This represents the classification level of the k-th key factor for the i-th sample.

[0137] (2) LCA model assumptions

[0138] Assume there exists a latent categorical variable C, and all observed variables (15 key factors) are conditionally independent given the latent category (local independence assumption). The model formula is:

[0139]

[0140] In the formula, π c This represents the prior probability that a sample belongs to the latent class c (satisfying...). ); P(b ik |C=c) means that under category c, the value of the k-th key factor is b. ik The conditional probability.

[0141] (3) Parameter estimation (EM algorithm)

[0142] E-step (expectation step): Calculate the posterior probability that sample i belongs to class c:

[0143]

[0144] In the formula, γ ic This represents the probability that sample i belongs to category c.

[0145] M-step (maximization step): Prior probabilities of the categories:

[0146]

[0147] Conditional probability (for each key factor k and class C:)

[0148]

[0149] In the formula, I(b) ik =l) represents the knowledge function (when b) ik =1 when =l, otherwise 0); l is the specific key factor rating.

[0150] When the parameter change Δ < 10 -5 Stop iterating when the time comes.

[0151] (4) Determining the number of categories

[0152] The rationale for selecting two potential categories is verified using information criteria.

[0153] ①AIC (Akaike Information Criteria):

[0154] AIC = -2ln L + 2M (0.20)

[0155] ②BIC (Bayesian Information Criterion):

[0156] BIC = -2ln L + M ln N (0.21)

[0157] In the formula, L represents the model likelihood value; M represents the total number of model parameters. Choose a minimum number of categories of 2.

[0158] (5) Sample classification and label classification

[0159] Based on the posterior probability γ ic Assign samples to categories with higher probabilities:

[0160]

[0161] According to the allocation rules: ① If the typical key factor rating model of landslide disaster is known, compare it with the conditional probability distribution P(b) of each category. ik |C=c): The conditional probability of high-risk factors in the positive sample class should be significantly higher. ② If some samples have known labels, the classification results can be verified through the confusion matrix.

[0162] Wherein, the posterior probability γ i1 A sample set with a value greater than 0.5 is considered a positive sample set.

[0163] Based on the results of LCA cluster analysis in classifying each group of data in sample set A' as positive or negative samples, matrix C3 is obtained.

[0164]

[0165] In the formula, if the i-th group of data in the sample set A' is identified as positive sample data by LCA cluster analysis, then c (i,3) It equals 1, otherwise it equals 0.

[0166] 3.4K-medoids clustering analysis

[0167] (1) Data processing

[0168] Min-Man standardization is adopted:

[0169]

[0170] In the formula, b ik This represents the original rating value of the k-th replacement factor for the i-th sample; max k ,min k This represents the maximum and minimum original values ​​of the k-th key factor.

[0171] (2) Initialize cluster centers

[0172] Two samples are randomly selected as the initial centers, namely M1 and M2:

[0173] M1 = B p M2 = B q (0.25)

[0174] In the formula, p,q∈{1,2,3,……,n+m+10T}, and p≠q.

[0175] (3) Sample allocation and cluster partitioning

[0176] Calculate the distance of all samples to the two centroids using Euclidean distance, and assign the samples to the clusters with the closest distance:

[0177]

[0178] In the formula, X i M represents the standardized vector of the i-th sample; c m' represents the center point of the Cth cluster; ck Indicates center point M c The standardized value of the kth key factor.

[0179] Allocation rule: Sample B i Assigned to cluster C j If and only if:

[0180]

[0181] (4) Update cluster centers

[0182] For each cluster, Cj Reselect its center point M j This minimizes the total intra-cluster distance:

[0183]

[0184] In the formula, X m Indicates cluster C j Candidate samples in; d(B i B m ) represents sample X i With candidate center X m The distance.

[0185] Repeat steps (3) and (4) until the center points M1 and M2 no longer change, or the sample allocation within the cluster no longer changes, then stop the iteration and output two sample sets.

[0186] (5) Cluster label allocation

[0187] Comparing the two sample sets generated in (4), if M1 has a significantly higher rating than M2 on the high-risk key factor, then C1 is a positive sample set.

[0188] Based on the results of K-medoids clustering analysis to classify each group of data in sample set A' as positive or negative samples, matrix C4 is obtained.

[0189]

[0190] In the formula, if the i-th group of data in the sample set A' is identified as positive sample data by K-medoids clustering analysis, then c (i,4) It equals 1, otherwise it equals 0.

[0191] 3.5 Spectral Cluster Analysis

[0192] (1) Data processing

[0193] Z-Score standardization is adopted:

[0194]

[0195] In the formula, b' ik b represents the standardized value of the edge; ik μ represents the original value of the k-th key factor for the i-th sample; k σ represents the mean of the k-th key factor; k This represents the standard deviation of the k-th critical factor.

[0196] (2) Construct a similarity matrix

[0197] To calculate the similarity between samples, the Gaussian function is used to define the similarity score.

[0198]

[0199] In the formula, W ij B' represents the similarity weight between sample i and sample j; i Describes the vector of standardized sample i; ||B' i -B' j || represents the Euclidean distance between sample i and sample j; σ represents the kernel function bandwidth parameter, which controls the similarity decay rate.

[0200] (3) Calculate the Laplace matrix

[0201] Construct a normalized Laplacian matrix to enhance clustering results:

[0202] ① Degree matrix M: A diagonal matrix, where the diagonal elements are the sum of each row of the similarity matrix.

[0203]

[0204] ② Normalized Laplace matrix L:

[0205]

[0206] In the formula, Represents the inverse square root of the degree matrix (diagonal elements are 1). ).

[0207] (4) Feature decomposition and dimensionality reduction

[0208] Perform eigenvalue decomposition on the Laplacian matrix L, and extract the eigenvectors corresponding to the k smallest eigenvalues:

[0209] ① Solve the characteristic equation:

[0210] Lv m =λ m v m m = 1, 2, ..., n + m + 10T (0.34)

[0211] ② Select the eigenvectors v1 and v2 corresponding to the first K = 2 smallest non-zero eigenvalues ​​λ1 and λ2.

[0212] ③ Construct the feature matrix V∈R (n+m+10T)×2

[0213] V = [v1, v2] (0.35)

[0214] (5) Low-dimensional clustering

[0215] For each row of the feature vector matrix V, perform binary classification using the K-means algorithm:

[0216] ① Randomly initialize two cluster centers μ1 and μ2;

[0217] ② Assign samples to the nearest center:

[0218]

[0219] ③ Update the center point

[0220]

[0221] ④ Repeat the above steps until convergence. Output two sample sets.

[0222] (6) Cluster label allocation

[0223] Compare the two sample sets generated in (5) iteratively, calculate the original factor mean vector of each cluster, and if the mean of a cluster is significantly higher on the high-risk key factor, then mark it as a positive sample set.

[0224] Based on the results of spectral clustering analysis in classifying each group of data in sample set A' as positive or negative samples, matrix C5 is obtained.

[0225]

[0226] In the formula, if the i-th group of data in the sample set A' is identified as positive sample data A' by spectral clustering analysis... i Then c (i,5) It equals 1, otherwise it equals 0.

[0227] 3.6 Self-organizing map neural network clustering analysis

[0228] (1) Data processing

[0229] Z-Score standardization is adopted:

[0230]

[0231] In the formula, b' ik b represents the standardized value of the edge; ik μ represents the original value of the k-th key factor for the i-th sample; k σ represents the mean of the k-th key factor; k This represents the standard deviation of the k-th critical factor.

[0232] (2) Constructing the SOM network structure

[0233] Define a two-dimensional competition layer network; the number of nodes should be selected based on the data scale.

[0234] Network size: 10×10;

[0235] Each node contains a weight vector W.j =(w j1 ,w j2 ,……,w j15 The number of dimensions and key factors is the same.

[0236] Initialize weights: Randomly select samples as initial weights.

[0237] (3) Finding the best unit

[0238] For each input sample B i 'Calculate the Euclidean distance between the node and the weight vectors of all nodes, and select the node with the smallest distance as the BMU:'

[0239]

[0240] In the formula, ||B' i -W j || represents B' i Distance to the weight vector of node j.

[0241] (4) Neighborhood function update

[0242] Based on the location of the BMU, define the neighborhood radius σ(t) and the neighborhood function, and update the weights of the nodes within the neighborhood:

[0243]

[0244] In the formula, h jk (t) represents the neighborhood weight influence coefficient between node j and BMU; r j ,r c Let σ(t) represent the coordinates of node j in the network; σ(t) represents the decaying neighborhood radius over time.

[0245] (5) Weight Adjustment

[0246] Update the BMU and the weight vectors of nodes in the neighborhood based on the neighborhood function and the learning rate α(t):

[0247] W j (t+1)=W j (t)+α(t)·h jc (t)·(B' i -W j (t)) (0.42)

[0248] In the formula, α(t) represents the learning rate that decays over time; B' i This represents the standardized sample of the current input.

[0249] Repeat steps (3) to (5) until the number of iterations reaches 100, then stop iterating.

[0250] (6) Generation of clustering results

[0251] ① Mapping samples to BMU: Assign all samples to the corresponding BMU nodes

[0252] ② Node clustering: The K-means algorithm is used to perform binary classification (K=2) on the weight vectors of the nodes.

[0253] ③ Sample classification: Based on the node category to which the BMU belongs, the samples are divided into positive and negative sample sets.

[0254] (7) Cluster label allocation

[0255] Compare the two sample sets generated in (6) and the weight vector features of the two types of nodes. If the weight of a certain type of node on the high-risk key factor is significantly higher, it is marked as a positive sample set.

[0256] Based on the results of self-organizing map neural network clustering analysis to classify each group of data in sample set A' as positive or negative samples, matrix C6 is obtained.

[0257]

[0258] In the formula, if the i-th group of data in the sample set A' is identified as positive sample data by the self-organizing map neural network clustering analysis, then c (i,6) It equals 1, otherwise it equals 0.

[0259] 3.7 Sample Data Set of Pseudo-Landslide Disasters X

[0260] The above six clustering algorithms are used to analyze and comprehensively judge the results of each group of data in sample set A', resulting in matrix K:

[0261]

[0262] In the formula, if in sample set A', the i-th group of data is judged as positive sample data by all six clustering algorithms, i.e., c (i,1) ,…,c (i,6) If all six numbers are 1, then k i If k is 1, otherwise k i It is 0.

[0263] In matrix K, if k i If the number of 1s is less than T, then repeat steps two and three until k. i The loop stops when the number of k equal to 1 is greater than T (which is Q). i The data from each group in the sample set A' corresponding to 1 are combined to obtain the final pseudo landslide disaster data sample set X.

[0264]

[0265] Step 4: From the pseudo-landslide disaster data sample set X, randomly select T groups of corresponding sample data to form sample set X'. A hybrid model training sample set D is formed by sample set X', sample set A, and non-landslide sample set E. Based on the different sample set X', generate 100 different hybrid model training sample sets D.

[0266] In step four, T sets of data are randomly selected from the pseudo-landslide disaster sample set X and combined with sample set A and non-landslide sample set E to finally form sample set D. This sample set includes the following:

[0267] 4.1 Constructing the sample set X'

[0268] From the final pseudo-landslide disaster data sample set X obtained in the above steps, T sets of sample data are randomly selected to form sample set X':

[0269]

[0270] In the formula, x' (i,j) This represents the level of the j-th key factor in the i-th group of sample data after extraction.

[0271] 4.2 Constructing the sample set D

[0272] Sample set X' is combined with sample set A and non-slope sample set E to obtain sample set D.

[0273]

[0274] Step 5: Deep Learning Hybrid Model Construction. To overcome the limitations of the distance function combination weighting method, data preprocessing is performed based on the data in sample set D. A CNN model is constructed to extract spatial features (selecting the number of convolutional layers, kernel size, stride, and adding pooling layers for dimensionality reduction). Then, LSTM layer parameters are set, the CNN feature maps are converted into one-dimensional vectors and input into the LSTM, and then fully connected layers are added for fusion. 100 different hybrid models are trained on sample set D to obtain 100 CNN-LSTM deep learning hybrid models for the comprehensive landslide hazard assessment of photovoltaic power plants.

[0275] In step five, CNN-LSTM hybrid models are constructed from the 100 different training sample sets D generated in the previous step, resulting in 100 CNN-LSTM hybrid models. These models include the following:

[0276] 5.1 Data Preprocessing

[0277] Standardization: Perform Z-Score standardization on each factor in the sample set D:

[0278]

[0279] In the formula, x ik μ represents the original value of the k-th factor of the i-th sample; k ,σ k This represents the mean and labeled difference of the k-th factor.

[0280] Input data structure: Divide the data in dataset D into training set D. train and test set D test The ratio is 8:2.

[0281] 5.2 CNN Spatial Feature Extraction

[0282] This involves a dual-channel convolution-pooling module to extract local spatial features.

[0283] (1) First convolutional layer (Conv2D)

[0284] Output c1 =ReLU(W c1 *X i +b c1 (0.49)

[0285] In the formula, This indicates a convolution kernel (2×2 size, 32 filters); * indicates a convolution operation, stride = 1, padding = "same"; b c1 This indicates the offset top; the output dimension is 3×5×32.

[0286] (2) First pooling layer (MaxPool2D)

[0287] Output p1 =MaxpPool(Output) c1 (0.50)

[0288] In the formula, the output dimension is 3×2×32;

[0289] (3) Second convolutional layer (Conv2D)

[0290] Output c2 =ReLU(W c2 *Output p1 +b c2 (0.51)

[0291] In the formula, This represents the convolution kernel (1×2 size, 64 filters); the output dimension is 3×1×64.

[0292] (4) Flatten

[0293] Convert the feature map into a one-dimensional sequence:

[0294]

[0295] 5.3 LSTM Temporal Feature Fusion

[0296] LSTM layer parameters:

[0297] h t ,c t =LSTM(Output) flat ,h t-1 ,c t-1 (0.53)

[0298] In the formula, the input sequence consists of 3 time steps, each with 64 dimensions; the number of hidden units is 128; and the output is the hidden state of the last time step.

[0299] 5.4 Hazard Score for Fully Connected Layers

[0300] Fully connected layer (Dense):

[0301] y=σ(W f ·h3+b f (0.54)

[0302] In the formula, σ represents the weight matrix; σ denotes the Sigmoid activation function, which outputs the probability of the genomic subtype (y∈[0,1]).

[0303] 5.5 Loss Function and Optimization

[0304] Loss function (binary classification cross-entropy):

[0305]

[0306] In the formula, y i This represents the true label (1 for a positive sample, 0 for a negative sample); This represents the probability predicted by the model.

[0307] Optimize it: Adam algorithm parameter update:

[0308]

[0309] In the formula, η represents the learning rate.

[0310] 5.6 Model Evaluation

[0311] Evaluation indicators

[0312] (1) Accuracy

[0313] Measuring the proportion of correct predictions made by the model overall:

[0314]

[0315] (2) Precision

[0316] Measuring the probability of a positive sample prediction (how many of the positive predictions are actually true disasters):

[0317]

[0318] (3) Recall (sensitivity)

[0319] Measuring the model's ability to cover real positive samples (the proportion of samples correctly identified in real disasters):

[0320]

[0321] (4) F1 score

[0322] The harmonic mean of precision and recall balances the contradiction between the two:

[0323]

[0324] (5) ROC-AUC: measures the ability of a classifier to distinguish between positive and negative samples.

[0325] ROC curve: A curve plotted with the true positive rate (TPR) and false positive rate (FPR) at different classification thresholds as axes.

[0326]

[0327] AUC: Area under the ROC curve. The closer the value is to 1, the stronger the model's ability to distinguish between positive and negative samples.

[0328] AUC∈[0,1], ideal value=1(0.62)

[0329] Step 6: Import the full-area data of the photovoltaic power station into 100 CNN-LSTM deep learning hybrid models to train and obtain 100 landslide hazard risk assessment result maps of the entire photovoltaic power station area. In ArcGIS, classify the risk assessment result maps into four levels (low risk, medium risk, high risk and extremely high risk).

[0330] In step six, based on the 100 CNN-LSTM deep learning hybrid models obtained above, the photovoltaic power station area data is input into each of the 100 hybrid models to generate 100 landslide hazard risk assessment result maps for the entire photovoltaic power station area. This includes the following:

[0331] Using the 100 CNN-LSTM deep learning hybrid models trained in step five above, a risk assessment of the entire photovoltaic power station data was conducted, resulting in 100 corresponding landslide hazard assessment result maps for the entire photovoltaic power station area.

[0332] The predicted probability for each sample represents the landslide hazard risk, y i ∈[0,1], where y i The closer it is to 1, the higher the risk.

[0333] The probability value matrix is ​​converted to an ArcGIS-compatible raster data format (GeoTIFF) and associated with a geographic coordinate system based on the spatial coordinate information.

[0334] The hazard classification threshold is determined using the equal-interval classification method:

[0335]

[0336] The resulting color mapping in the image corresponds to: low risk - green; medium risk - yellow; high risk - orange; and very high risk - red.

[0337] Step 7: Calculate the average and variance of the hazard assessment results from 100 photovoltaic power plant area landslide hazard assessment result maps. If the variance of the assessment result at a certain point is relatively large, consider raising the assessment result level of that point by one level or indicating that this point needs special attention. Finally, obtain the spatial probability distribution map of landslide hazard in the site area through probability density function analysis.

[0338] In step seven, based on the 100 landslide hazard risk assessment results maps of the entire photovoltaic power station area, the average and variance of the landslide hazard risk assessment results for each region in the 100 maps are calculated. A comprehensive judgment is then made on the landslide hazard risk in the maps to obtain a spatial probability distribution map of landslide hazard risk for the entire photovoltaic power station area. This includes the following:

[0339] 7.1 Calculation of Average Value Results

[0340] For each pixel location (i,j), calculate the average of the 100 landslide hazard risk assessment results:

[0341]

[0342] In the formula, y ijk This represents the pixel value (probability of danger) at the k-th layer position (i,j); This represents the average hazard value at location (i,j).

[0343] After analysis and calculation, the area was further divided into four levels based on the hazard assessment threshold corresponding to the average value results. This resulted in 100 hazard assessment maps with the average value results of the slope hazard assessment.

[0344] 7.2 Calculation of Variance Results Plot

[0345] For each pixel location (i,j), calculate the variance of its 100 slope hazard risk assessment results:

[0346]

[0347] In the formula, The hazard variance at location (i,j) is represented by an unbiased estimate with a denominator of 99.

[0348] After analysis and calculation, the variance results were divided into four levels. A variance level diagram of 100 slope hazard risk assessment results was obtained.

[0349] 7.3 Comprehensive assessment of landslide hazard risk

[0350] The above-obtained average values ​​are used to comprehensively assess the hazard rating chart and variance rank chart. The specific rules for this assessment are as follows (as shown in the hazard matrix chart):

[0351] (1) If the variance is large (level 3 or 4), the landslide hazard risk assessment result corresponding to the average value will be upgraded by one level (up to level 4);

[0352] (2) If the variance is relatively small (level 1 or level 2), the level corresponding to the landslide hazard risk assessment result corresponding to the average value will remain unchanged, see Table 2.

[0353] Table 2 Average Risk Level Table

[0356] Based on the comprehensive assessment results, the landslide hazard risk assessment map of the photovoltaic power station area is divided into zones and colored. Finally, a spatial probability distribution map of landslide hazard risk in the photovoltaic power station area is output.

[0357] This embodiment also discloses a landslide hazard probability assessment system for photovoltaic power plants, including:

[0358] The first dataset acquisition module is used to acquire a non-landslide hazard sample set E and a landslide hazard sample set A for preliminary classification of hazard levels.

[0359] The second dataset acquisition module is used to sample the landslide disaster sample set A to obtain pseudo landslide disaster sample data A', and generate sample set B based on the non-landslide disaster sample set E, the landslide disaster sample set A, and the pseudo landslide disaster sample data A';

[0360] The third dataset acquisition module is used to perform cluster analysis on sample set B to obtain pseudo landslide disaster data sample set X, sample the pseudo landslide disaster data sample set X to obtain sample set X', and generate a mixed sample set D based on sample set X', landslide disaster sample set A and non-landslide disaster sample set E;

[0361] The evaluation acquisition module is used to acquire a deep learning hybrid model. The hybrid sample set D is used as the input of the deep learning hybrid model to obtain a hazard assessment result map. The hazard assessment result map is used to classify the hazard level and obtain a spatial probability distribution map of landslide hazard in the site, thus obtaining the evaluation result.

[0362] A schematic diagram of a terminal device according to an embodiment of the present invention. The terminal device of this embodiment includes: a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps in the various method embodiments described above. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the various device embodiments described above.

[0363] The computer program can be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention.

[0364] The terminal device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0365] The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0366] The memory can be used to store the computer program and / or module. The processor implements various functions of the terminal device by running or executing the computer program and / or module stored in the memory and calling the data stored in the memory.

[0367] If the modules / units integrated into the terminal device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0368] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for assessing the probability of landslide hazard at photovoltaic power plants, characterized in that, Includes the following steps: Obtain a sample set E of non-landslide hazards and a sample set A of landslide hazards for preliminary classification of hazard levels; Sampling is performed on landslide disaster sample set A to obtain pseudo landslide disaster sample data A'. Sample set B is generated based on non-landslide disaster sample set E, landslide disaster sample set A, and pseudo landslide disaster sample data A'. Cluster analysis is performed on sample set B to obtain pseudo landslide disaster data sample set X. Sample set X' is obtained by sampling from pseudo landslide disaster data sample set X. Based on sample set X', landslide disaster sample set A and non-landslide disaster sample set E, a mixed sample set D is generated. A deep learning hybrid model is obtained, and the hybrid sample set D is used as the input of the deep learning hybrid model to obtain the hazard assessment result map. The hazard assessment result map is used to classify the hazard level to obtain the spatial probability distribution map of landslide hazard in the site and obtain the assessment result.

2. The method for assessing the probability of landslide hazard at photovoltaic power plants according to claim 1, characterized in that, The acquisition of the landslide hazard sample set A for preliminary classification of hazard levels includes: Obtain landslide disaster sample point data; Construct a hazard assessment index system for sample disasters; Based on the hazard assessment index system for sample disasters, each key factor of each landslide disaster sample point is rated to generate a landslide disaster sample set A.

3. The method for assessing the probability of landslide hazard at photovoltaic power plants according to claim 1, characterized in that, The generation of the sample set B includes: The landslide disaster sample set A was sampled using the independent and identically distributed sampling method, generating 10T sets of sample data, which constituted the pseudo-landslide disaster sample data A', where T represents the number of pseudo-positive samples taken in each set; The non-landslide disaster sample set E, the landslide disaster sample set A, and the pseudo-landslide disaster sample data A' are combined to generate sample set B.

4. The method for assessing the probability of landslide hazard at photovoltaic power plants according to claim 1, characterized in that, The acquisition of the pseudo-landslide disaster data sample set X includes: Using multiple clustering methods, the pseudo-landslide disaster sample data A' is clustered sequentially. The landslide disaster sample set A is set as positive samples, and the non-landslide disaster sample set E is set as negative samples. The clustering results of the pseudo-landslide disaster sample data A' are then redistributed by combining the positive and negative samples to obtain the pseudo-landslide disaster data sample set X.

5. The method for assessing the probability of landslide hazard at a photovoltaic power station according to claim 1, characterized in that, The generation of the mixed sample set D includes: From the pseudo-landslide disaster data sample set X, T sets of data are randomly selected to form sample set X': In the formula, x (i,j) This represents the level of the j-th key factor in the i-th group of sample data after extraction; Sample set D is obtained by combining sample set X', sample set A, and non-landslide sample set E:

6. The method for assessing the probability of landslide hazard at a photovoltaic power station according to claim 1, characterized in that, The obtained spatial probability distribution map of landslide hazard in the site includes: Calculate the average value of the landslide hazard assessment results based on the hazard assessment result map; Based on the average value of the landslide hazard risk assessment results, and combined with the risk assessment grading threshold, the average value of the risk is classified into levels to obtain the risk assessment map of the average value of the risk assessment results; Calculate the variance of the landslide hazard risk assessment results, classify the variance calculation results into levels, and obtain the variance level map of the hazard risk assessment results; Based on the average value result hazard assessment map and the variance level map of the disaster hazard assessment result, the hazard assessment result map is divided into hazard areas to obtain the spatial probability distribution map of landslide hazard in the site and obtain the assessment results.

7. The method for evaluating the probability of landslide hazard at a photovoltaic power station according to claim 6, characterized in that, The process of dividing the hazard assessment result map into hazard zones based on the average value hazard assessment map and the variance level map of the hazard hazard assessment results includes: The average result of the risk assessment, the risk assessment map, and the variance level map of the disaster risk assessment results are both level four; If the variance value corresponding to the hazard assessment map is level three or four in the hazard level, the hazard level of the landslide hazard assessment result corresponding to the average value of the current hazard assessment map will be upgraded by one level; If the variance value corresponding to the hazard assessment map is level one or level two in the hazard level, the hazard level of the landslide hazard assessment result corresponding to the average value of the current hazard assessment map will remain unchanged.

8. A landslide hazard probability assessment system for photovoltaic power plants, characterized in that, include: The first dataset acquisition module is used to acquire a non-landslide hazard sample set E and a landslide hazard sample set A for preliminary classification of hazard levels. The second dataset acquisition module is used to sample the landslide disaster sample set A to obtain pseudo landslide disaster sample data A', and generate sample set B based on the non-landslide disaster sample set E, the landslide disaster sample set A, and the pseudo landslide disaster sample data A'; The third dataset acquisition module is used to perform cluster analysis on sample set B to obtain pseudo landslide disaster data sample set X, sample the pseudo landslide disaster data sample set X to obtain sample set X', and generate a mixed sample set D based on sample set X', landslide disaster sample set A and non-landslide disaster sample set E; The evaluation acquisition module is used to acquire a deep learning hybrid model. The hybrid sample set D is used as the input of the deep learning hybrid model to obtain a hazard assessment result map. The hazard assessment result map is used to classify the hazard level and obtain a spatial probability distribution map of landslide hazard in the site, thus obtaining the evaluation result.

9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-7.

Images