Method, device and equipment for predicting short-term wind speed along railway and readable storage medium

Abstract

The invention relates to a short-term wind speed prediction method, device and equipment along a railway and a readable storage medium. The method comprises the following steps: decomposing a wind speed data sequence in a preset time period into a plurality of initial subsequences, extracting potential incidence matrixes among the subsequences, and constructing a first graph structure; generating a backtracking value of each initial subsequence wind speed value and a first prediction value of each initial subsequence wind speed value at the next moment based on the first graph structure; calculating a difference value between the wind speed value of each initial sub-sequence and each backtracking value to obtain a backtracking error of the wind speed value of each initial sub-sequence; constructing a second graph structure based on the first graph structure and the backtracking error; generating a second predicted value of each initial sub-sequence backtracking error at the next moment based on the second graph structure; and generating a wind speed prediction value of the wind speed data sequence at the next moment based on the first prediction value and the second prediction value of each initial sub-sequence. By adopting the method, high-precision short-term wind speed prediction along the railway can be realized.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and particularly to a method, device, equipment and readable storage medium for short-term wind speed prediction along railway lines. Background Art

[0002] High wind meteorological disasters are one of the adverse factors affecting the operation safety of high-speed railways. Under the action of strong crosswinds, the aerodynamic performance of trains deteriorates. Not only do the air resistance, lift and lateral force of the train increase rapidly, but it also affects the lateral movement stability of the train. In severe cases, it will cause the train to overturn. Compared with traditional trains, high-speed trains are more sensitive to wind. Therefore, establishing accurate and reliable wind speed prediction and timely warning the dispatcher to slow down or stop the train before the occurrence of high wind disasters is of great significance for ensuring the operation safety of high-speed railways. Summary of the Invention

[0003] Based on this, in view of the above technical problems, it is necessary to provide a method, device, equipment and readable storage medium for short-term wind speed prediction along railway lines that can fully consider the temporal correlation between wind speed data signals, explore the relationship between the characteristics of input wind speed data and prediction performance, and improve the accuracy of short-term wind speed prediction.

[0004] In a first aspect, the present application provides a method for short-term wind speed prediction along railway lines, the method comprising: Decompose the wind speed data sequence within a preset time period collected into a plurality of initial subsequences, and generate the wind speed values of each of the initial subsequences; Use a two-layer gated recurrent unit to extract the temporal characteristics of the wind speed values of each of the initial subsequences, and perform regularization processing on the temporal characteristics of the wind speed values of each of the initial subsequences to generate a potential correlation matrix between each of the initial subsequences; Based on the wind speed values of each of the initial subsequences and the potential correlation matrix, construct a first graph structure of the wind speed data sequence; Based on the first graph structure, extract the spectral characteristics and hidden dependency relationships of the wind speed values of each of the initial subsequences, and based on the spectral characteristics and hidden dependency relationships of the wind speed values of each of the initial subsequences, generate a backtracking value corresponding to the wind speed value of each of the initial subsequences and a first predicted value for the next moment corresponding to the wind speed value of each of the initial subsequences; Calculate the difference between the wind speed value of each of the initial subsequences and the corresponding backtracking value to obtain the backtracking error of the wind speed value of each of the initial subsequences, and based on each of the backtracking errors, the potential correlation matrix, and the spectral characteristics and hidden dependency relationships of the wind speed values of each of the initial subsequences, construct a second graph structure of the wind speed data sequence; Based on the second graph structure, extract the spectral features and hidden dependencies of the backtracking errors of each of the initial subsequences, and generate the second predicted value at the next moment corresponding to the backtracking error of each of the initial subsequences according to the spectral features and hidden dependencies of the backtracking errors of each of the initial subsequences; Based on the first predicted value and the corresponding second predicted value of each of the initial subsequences, generate the wind speed predicted value at the next moment of the wind speed data sequence.

[0005] In one embodiment, the decomposing the wind speed data sequence within a preset time period collected into a plurality of initial subsequences and generating the wind speed value of each of the initial subsequences includes: Decompose the wind speed data sequence into a plurality of initial mode functions, perform Hilbert transform on each of the initial mode functions to obtain the one-sided spectrum of each initial mode function; use the exponential mixing modulation method to move the one-sided spectrum of each of the initial mode functions to their respective estimated center frequencies and perform demodulation to obtain the estimated bandwidth of each initial mode function; perform unconstrained variational on the estimated bandwidths of each of the initial mode functions, and use the multiplicative operator alternating direction method for iterative update to obtain a plurality of the initial subsequences and the wind speed value of each of the initial subsequences.

[0006] In one embodiment, the using two-layer gated recurrent units to extract the temporal features of the wind speed values of each of the initial subsequences and perform regularization processing on the temporal features of the wind speed values of each of the initial subsequences to generate the potential correlation matrix between each of the initial subsequences includes: Obtain the time stamp order corresponding to each of the initial subsequences; use two-layer gated recurrent units to calculate the temporal features of the wind speed values of each of the initial subsequences according to the time stamp order; use a regularization layer to perform regularization processing on the temporal features of the wind speed values of each of the initial subsequences, and use a self-attention mechanism to perform linear projection calculation on the temporal features of the wind speed values of each of the initial subsequences after regularization to generate the potential correlation matrix.

[0007] In one embodiment, extracting the spectral features and hidden dependencies of the wind speed values of each of the initial subsequences based on the first graph structure includes: performing a graph Fourier transform on the first graph structure to obtain a first spectral matrix of the wind speed data sequence; performing a discrete Fourier transform on the first spectral matrix of the wind speed data sequence to obtain a first frequency domain representation of the wind speed values of each of the initial subsequences; passing the first frequency domain representations of the wind speed values through a one-dimensional convolution and a gated linear unit to obtain the spectral features of the wind speed values of each of the initial subsequences; performing an inverse discrete Fourier transform on the first frequency domain representations of the wind speed values to obtain a first time domain representation of each of the wind speed values; performing a graph convolution on the first time domain representations of the wind speed values to obtain a second frequency domain representation of each of the wind speed values and obtain the hidden dependencies of the wind speed values of each of the initial subsequences; and performing an inverse graph Fourier transform on the second frequency domain representations of the wind speed values to obtain a second time domain representation of each of the wind speed values.

[0008] In one embodiment, extracting the spectral features and hidden dependencies of the backtracking errors of each of the initial subsequences based on the second graph structure includes: performing a graph Fourier transform on the second graph structure to obtain a second spectral matrix of the wind speed data sequence; performing a discrete Fourier transform on the second spectral matrix of the wind speed data sequence to obtain a third frequency domain representation of the backtracking errors of each of the initial subsequences; passing the third frequency domain representations of the backtracking errors through a one-dimensional convolution and a gated linear unit to obtain the spectral features of the backtracking errors; performing an inverse discrete Fourier transform on the third frequency domain representations of the backtracking errors to obtain a third time domain representation of each of the backtracking errors; performing a graph convolution on the third time domain representations of the backtracking errors to obtain a fourth frequency domain representation of each of the backtracking errors and obtain the hidden dependencies of the backtracking errors of each of the initial subsequences; and performing an inverse graph Fourier transform on the fourth frequency domain representations of the backtracking errors to obtain a fourth time domain representation of each of the backtracking errors.

[0009] In one embodiment, generating a wind speed prediction value for the next moment of the wind speed data sequence based on the first prediction value and the corresponding second prediction value of each of the initial subsequences includes: Adding the first prediction value and the corresponding second prediction value of each of the initial subsequences to obtain a wind speed prediction value for each initial subsequence at the next moment; and adding the wind speed prediction values for each initial subsequence at the next moment to generate a wind speed prediction value for the next moment of the wind speed data sequence.

[0010] In a second aspect, the present application further provides a short-term wind speed prediction device for a railway line. The device includes: A sequence decomposition module, configured to decompose the wind speed data sequence collected within a preset time period into a plurality of initial subsequences, and generate the wind speed values of each of the initial subsequences; A potential association module, configured to use a two-layer gated recurrent unit to extract the temporal features of the wind speed values of each of the initial subsequences, and perform regularization processing on the temporal features of the wind speed values of each of the initial subsequences to generate a potential association matrix between each of the initial subsequences; based on the wind speed values of each of the initial subsequences and the potential association matrix, construct a first graph structure of the wind speed data sequence; A first prediction module, configured to extract the spectral features and hidden dependencies of the wind speed values of each of the initial subsequences based on the first graph structure, and generate a backtracking value corresponding to the wind speed value of each of the initial subsequences and a first prediction value for the next moment corresponding to the wind speed value of each of the initial subsequences based on the spectral features and hidden dependencies of the wind speed values of each of the initial subsequences; The potential association module is further configured to calculate the difference between the wind speed value of each of the initial subsequences and the corresponding backtracking value to obtain the backtracking error of the wind speed value of each of the initial subsequences, and construct a second graph structure of the wind speed data sequence based on the backtracking errors and the potential association matrix; A second prediction module, configured to extract the spectral features and hidden dependencies of the backtracking errors of each of the initial subsequences based on the second graph structure, and generate a second prediction value for the next moment corresponding to the backtracking error of each of the initial subsequences according to the spectral features and hidden dependencies of the backtracking errors of each of the initial subsequences; A prediction output module, configured to generate a wind speed prediction value for the next moment of the wind speed data sequence based on the first prediction value and the corresponding second prediction value of each of the initial subsequences.

[0011] In a third aspect, the present application further provides a computer device, including a memory and a processor, where the memory stores a computer program, and when the processor executes the computer program, the steps corresponding to the short-term wind speed prediction method along the railway line in the first aspect are implemented.

[0012] In a fourth aspect, the present application further provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the content in the first aspect is implemented.

[0013] The above-mentioned short-term wind speed prediction method, device, equipment and readable storage medium for the railway line collect a wind speed data sequence within a preset time period, and decompose the wind speed data sequence into a plurality of initial subsequences; based on the plurality of initial subsequences, extract the potential correlation matrix between the initial subsequences, and construct a first graph structure; based on the first graph structure, extract the spectral features and hidden dependencies of the wind speed values of the initial subsequences, generate a backtracking value corresponding to the wind speed value of each initial subsequence and a first prediction value for the next moment corresponding to the wind speed value of each initial subsequence; calculate the difference between the wind speed value of each initial subsequence and the corresponding backtracking value to obtain the backtracking error of the wind speed value of each initial subsequence, and based on the backtracking errors, the potential correlation matrix, and the spectral features and hidden dependencies of the wind speed values of the initial subsequences, construct a second graph structure of the wind speed data sequence; based on the second graph structure, extract the spectral features and hidden dependencies of the backtracking errors of the initial subsequences, and generate a second prediction value for the next moment corresponding to the backtracking error of each initial subsequence according to the spectral features and hidden dependencies of the backtracking errors of the initial subsequences; based on the first prediction value and the corresponding second prediction value of each initial subsequence, generate a wind speed prediction value for the next moment of the wind speed data sequence, realizing high-precision short-term wind speed prediction for the railway line. Description of the Drawings

[0014] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following will briefly introduce the drawings required to be used in the description of the embodiments of the present application or related technologies.

[0015] Figure 1 It is an application environment diagram of the short-term wind speed prediction method for the railway line in an embodiment; Figure 2 It is a flowchart of the short-term wind speed prediction method for the railway line in an embodiment; Figure 3 It is a flowchart of the step of generating the wind speed prediction value for the next moment of the wind speed data sequence in an embodiment; Figure 4 It is a flowchart of the step of decomposing the initial wind speed data sequence into a plurality of subsequences in an embodiment; Figure 5 It is a flowchart of the step of constructing the potential correlation matrix between the subsequences in an embodiment; Figure 6 It is a flowchart of the step of extracting the spectral features and hidden dependencies of the wind speed values of the initial subsequences in an embodiment; Figure 7 It is a structural block diagram of the short-term wind speed prediction device for the railway line in an embodiment. Detailed implementation manners

[0016] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0017] Unless otherwise defined, the technical terms or scientific terms involved in the present application shall have the ordinary meanings understood by those with ordinary skills in the technical field to which the present application belongs. The terms "a", "one", "kind", "the" and similar words involved in the present application do not indicate a quantity limitation and may represent a single or plural number. The terms "include", "comprise", "have" and any variations thereof involved in the present application are intended to cover non-exclusive inclusion. The "plurality" involved in the present application refers to two or more. "And / or" describes the association relationship of associated objects and indicates that three relationships may exist. For example, "A and / or B" may represent: A exists alone, A and B exist simultaneously, and B exists alone. The terms "first", "second", "third", etc. involved in the present application are only used to distinguish similar objects and do not represent a specific order for the objects.

[0018] The short-term wind speed prediction method for railway lines provided by the embodiments of the present application can be applied to an application environment as Figure 1 shown. Among them, the terminal 102 communicates with the server 104 through a network. The data storage system can store the data that the server 104 needs to process. The data storage system can be integrated on the server 104, or can be placed in the cloud or other network servers.

[0019] On server 104, the wind speed data sequence collected within a preset time period is decomposed into a number of initial subsequences, and the wind speed value of each of the initial subsequences is generated; using a two-layer gated recurrent unit, the temporal features of the wind speed values of each of the initial subsequences are extracted, and the temporal features of the wind speed values of each of the initial subsequences are regularized to generate a potential correlation matrix between each of the initial subsequences; based on the wind speed values of each of the initial subsequences and the potential correlation matrix, a first graph structure of the wind speed data sequence is constructed; based on the first graph structure, the spectral features and hidden dependencies of the wind speed values of each of the initial subsequences are extracted, and based on the spectral features and hidden dependencies of the wind speed values of each of the initial subsequences, a backtracking value corresponding to the wind speed value of each of the initial subsequences and a first predicted value for the next moment corresponding to the wind speed value of each of the initial subsequences are generated; the difference between the wind speed value of each of the initial subsequences and the corresponding backtracking value is calculated to obtain the backtracking error of the wind speed value of each of the initial subsequences, and based on the backtracking errors and the potential correlation matrix, a second graph structure of the wind speed data sequence is constructed; based on the second graph structure, the spectral features and hidden dependencies of the backtracking errors of each of the initial subsequences are extracted, and according to the spectral features and hidden dependencies of the backtracking errors of each of the initial subsequences, a second predicted value for the next moment corresponding to the backtracking error of each of the initial subsequences is generated; based on the first predicted value and the corresponding second predicted value of each of the initial subsequences, a wind speed predicted value for the next moment of the wind speed data sequence is generated.

[0020] Among them, the terminal 102 can be, but is not limited to, various personal computers, laptop computers, smart phones, tablet computers, projection devices, smart watches, smart bracelets, head-mounted devices, etc. The server 104 can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0021] In an exemplary embodiment, as Figure 2 shown, a short-term wind speed prediction method along a railway is provided. Taking the application of this method to the Figure 1 server side as an example, it includes the following steps 201 to step 207. Among them: Step 201, decompose the wind speed data sequence collected within a preset time period into a number of initial subsequences, and generate the wind speed value of each of the initial subsequences.

[0022] Since the wind speed data has high volatility and randomness, by collecting a single wind speed data sequence with original frequency aliasing within a preset time period, the wind speed data sequence is decomposed into a group of initial subsequences with several frequency separations , where \(t\) represents the moment, and the wind speed value \(X\) of each of the initial subsequences is generated i , which refines the frequency characteristics of the original wind speed data sequence, where \(X\) i ∈ , and \(n\) represents the number of initial subsequences.

[0023] Step 202: Use a two-layer gated recurrent unit to extract the temporal characteristics of the wind speed values of each of the initial subsequences, and perform regularization processing on the temporal characteristics of the wind speed values of each of the initial subsequences to generate a potential correlation matrix between each of the initial subsequences.

[0024] Specifically, the temporal characteristics of the wind speed value of each initial subsequence are extracted through a two-layer gated recurrent unit, and the temporal characteristics are regularized to prevent overfitting of the data. Then, a self-attention mechanism is used to construct a potential correlation matrix \(W\) between each initial subsequence.

[0025] Step 203: Based on the wind speed values of each of the initial subsequences and the potential correlation matrix, construct a first graph structure of the wind speed data sequence.

[0026] Among them, the first graph structure \(G1=(X, W)\), where \(X = [X1, X2,..., X\) n represents the set of wind speed values of each of the initial subsequences, and \(W\) represents the potential correlation matrix between each initial subsequence.

[0027] Step 204: Based on the first graph structure, extract the spectral characteristics and hidden dependencies of the wind speed values of each of the initial subsequences, and based on the spectral characteristics and hidden dependencies of the wind speed values of each of the initial subsequences, generate a retrospective value corresponding to the wind speed value of each of the initial subsequences and a first predicted value for the next moment corresponding to the wind speed value of each of the initial subsequences.

[0028] Specifically, perform a graph Fourier transform on the first graph structure \(G1=(X, W)\) to convert the wind speed values of each of the initial subsequences into a first spectral matrix. Based on the first spectral matrix, extract the spectral characteristics and hidden dependencies of the wind speed values of each of the initial subsequences. According to the spectral characteristics and hidden dependencies of the wind speed values of each of the initial subsequences, generate a retrospective value corresponding to the wind speed value of each of the initial subsequences and a first predicted value for the next moment corresponding to the wind speed value of each of the initial subsequences. Among them, the meaning of the retrospective value is the calculated magnitude of the wind speed value of each initial subsequence inversely deduced based on the spectral characteristics and hidden dependencies of the wind speed value of the initial subsequence. Since the wind speed value \(X\) of each initial subsequence i corresponds to a sequence timestamp, the meaning of the first predicted value for the next moment corresponding to the wind speed value of each initial subsequence is the wind speed value predicted at time \(t\) given the wind speed value at time \(t - 1\) of the known initial subsequence.

[0029] Step 205: Calculate the difference between the wind speed value of each of the initial subsequences and the corresponding backtracking value, obtain the backtracking error of the wind speed value of each initial subsequence, and construct a second graph structure of the wind speed data sequence based on each of the backtracking errors, the potential correlation matrix, and the spectral characteristics and hidden dependencies of the wind speed values of each initial subsequence.

[0030] Specifically, calculate the difference between the wind speed value of each of the initial subsequences and the corresponding backtracking value to obtain the backtracking error of the wind speed value of each initial subsequence. At this time, the second graph structure G2=(X - , W), where represents the backtracking value corresponding to the wind speed value of each initial subsequence. X - contains the spectral characteristics and hidden dependencies of the wind speed values of each initial subsequence extracted from the first graph structure G1.

[0031] Step 206: Based on the second graph structure, extract the spectral characteristics and hidden dependencies of the backtracking error of each initial subsequence, and generate a second predicted value for the next moment corresponding to the backtracking error of each initial subsequence according to the spectral characteristics and hidden dependencies of the prediction error of each initial subsequence.

[0032] Specifically, perform a graph Fourier transform on the second graph structure G2=(X - , W) to convert the wind speed values of each initial subsequence into a second spectral matrix. Based on the second spectral matrix, extract the spectral characteristics and hidden dependencies of the backtracking error of each initial subsequence. According to the spectral characteristics and hidden dependencies of the backtracking error of each initial subsequence, generate a second predicted value for the next moment corresponding to the backtracking error of each initial subsequence. The meaning of the second predicted value for the next moment corresponding to the backtracking error of each initial subsequence is the error value of the first predicted value at the t time stamp predicted after knowing the wind speed value at the t - 1 time stamp of the known initial subsequence and the first predicted value predicted at the t time stamp.

[0033] Step 207: Generate a wind speed predicted value for the next moment of the wind speed data sequence based on the first predicted value and the corresponding second predicted value of each initial subsequence.

[0034] In the above short-term wind speed prediction method along the railway, the wind speed data sequence is decomposed into several initial subsequences with separated frequencies; based on the initial subsequences, a potential correlation matrix between the initial subsequences is generated, thereby constructing a first graph structure for prediction at the next moment. Based on the first graph structure, the spectral features and hidden dependencies of the wind speed values of the initial subsequences are extracted, and a backtracking value corresponding to the wind speed value of each initial subsequence and a first prediction value at the next moment corresponding to the wind speed value of each initial subsequence are generated; the difference between the wind speed value of each initial subsequence and the corresponding backtracking value is calculated to obtain the backtracking error of the wind speed value of each initial subsequence, and based on the backtracking errors, the potential correlation matrix, and the spectral features and hidden dependencies of the wind speed values of the initial subsequences, a second graph structure of the wind speed data sequence is constructed; based on the second graph structure, the spectral features and hidden dependencies of the backtracking errors of the initial subsequences are extracted, and according to the spectral features and hidden dependencies of the backtracking errors of the initial subsequences, a second prediction value at the next moment corresponding to the backtracking error of each initial subsequence is generated; based on the first prediction value and the corresponding second prediction value of each initial subsequence, a wind speed prediction value at the next moment of the wind speed data sequence is generated, realizing high-precision short-term wind speed prediction along the railway.

[0035] In one embodiment, as Figure 3 shown, step 207, based on the first prediction value and the corresponding second prediction value of each initial subsequence, generating a wind speed prediction value at the next moment of the wind speed data sequence specifically includes the following steps 301 to step 302: Step 301, sum the first prediction value and the corresponding second prediction value of each initial subsequence to obtain the wind speed prediction value of each initial subsequence at the next moment.

[0036] Step 302, sum the wind speed prediction values of the initial subsequences at the next moment to generate the wind speed prediction value at the next moment of the wind speed data sequence.

[0037] Exemplarily, the first prediction value of each initial subsequence at the next moment , where n represents the number of initial subsequences. The second prediction value of each initial subsequence at the next moment . Sum the first prediction value and the corresponding second prediction value of each initial subsequence to obtain the wind speed prediction value of each initial subsequence at the next moment , where Y 1 = a 11 + b 11 ,......,Y n = a n1 + b n1 Sum the predicted wind speed values of each of the initial subsequences at the next moment to obtain the predicted wind speed value at the next moment t of the wind speed data sequence. : ; where n represents the number of initial subsequences.

[0038] It should be noted that in this embodiment, it is only described that the predicted wind speed value of each initial subsequence is composed of the sum of the first predicted value obtained by the first prediction module based on the first graph structure and the second predicted value obtained by the second prediction module based on the second graph structure. While generating the second predicted value of the backtracking error for each initial subsequence by the second prediction module, a backtracking value of the backtracking error is also generated. Therefore, a third prediction module can be constructed based on the backtracking value of the backtracking error according to actual needs, and the third prediction module predicts the third predicted value for the third graph structure including the backtracking value of the backtracking error. At this time, the predicted wind speed value of each initial subsequence is the sum of the first predicted value, the second predicted value, and the third predicted value. Preferably, new prediction modules can be continuously generated by recursion according to actual needs, and the Nth predicted value for the Nth graph structure can be continuously predicted.

[0039] Preferably, the method provided in this embodiment can finally use a sliding window for predicting the wind speed at the next moment. For example, to predict the wind speed at the next moment t, take as the input within the sliding window, M is the window size, is the magnitude of the wind speed value of the initial subsequence at time t - M, and the predicted wind speed value at time t is output as ; Change the input to for predicting the wind speed value at time t + 1. By continuously applying this strategy, the wind speed prediction values at any subsequent timestamp can be obtained.

[0040] In an exemplary embodiment, as Figure 4 shown, step 201 decomposes the wind speed data sequence within a preset time period collected into a number of initial subsequences and generates the wind speed value of each of the initial subsequences, specifically including the following steps 401 to step 403. Wherein: Step 401, decompose the initial wind speed data sequence into a number of initial mode functions, and perform Hilbert transform on each of the initial mode functions to obtain the single-sided spectrum of each initial mode function.

[0041] Step 402: Using the exponential hybrid modulation method, shift the single-sided spectrum of each of the initial mode functions to their respective estimated center frequencies and perform demodulation to obtain the estimated bandwidth of each initial mode function.

[0042] Step 403: Perform unconstrained variational on the estimated bandwidths of the initial mode functions, and use the alternating direction method of multipliers for iterative update to obtain a number of the initial subsequences and the wind speed values of each of the initial subsequences.

[0043] Specifically, decompose the wind speed data sequence x(t) into n initial mode functions v j , j = 1, 2,..., n. Perform Hilbert transform on each initial mode function v j to obtain the single-sided spectrum of each initial mode function v j .

[0044] Using the exponential hybrid modulation method, shift the single-sided spectrum of each of the initial mode functions to their respective estimated center frequencies.

[0045] According to the Gaussian smoothness and gradient square criterion, perform demodulation on each initial mode function v j to obtain the estimated bandwidth of each initial mode function v j .

[0046] Using the quadratic penalty term and Lagrange multiplier, perform unconstrained variational on the estimated bandwidths of the initial mode functions v j , and use the alternating direction method of multipliers to iteratively update the saddle points of the estimated bandwidths of the initial mode functions v j after unconstrained variational. When the convergence condition is satisfied, terminate the iteration to obtain n narrowband components v j , that is, the wind speed data sequence x(t) is decomposed into n initial subsequences .

[0047] In this embodiment, by decomposing the wind speed data sequence, the original wind speed sequence with frequency aliasing is decomposed into a group of stationary subsequences with frequency separation, which solves the complex relationship in time and frequency brought by the high volatility and randomness of the wind speed data, and improves the accuracy of subsequent wind speed prediction using the subsequences.

[0048] In one of the embodiments, as Figure 5 shown, step 202 uses a two-layer gated recurrent unit to extract the temporal features of the wind speed values of each of the initial subsequences, and perform regularization processing on the temporal features of the wind speed values of each of the initial subsequences to generate a potential correlation matrix between each of the initial subsequences, which specifically includes the following steps 501 to step 503.

[0049] Step 501: Obtain the timestamp order corresponding to each of the initial subsequences.

[0050] Step 502: Use a two-layer gated recurrent unit to calculate the temporal features of the wind speed values of each of the initial subsequences according to the timestamp order.

[0051] Step 503: Use a regularization layer to perform regularization processing on the temporal features of the wind speed values of each of the initial subsequences, and use a self-attention mechanism to perform a linear projection calculation on the temporal features of the wind speed values of each of the initial subsequences after regularization to generate the potential correlation matrix.

[0052] Specifically, take each of the initial subsequences as a prior input to construct the nodes of the first graph structure G1, and regard the wind speed value X of each initial subsequence as the signal of the corresponding node on the first graph structure G1. Sequentially use a two-layer gated recurrent unit to calculate the temporal features corresponding to the wind speed values at different timestamps according to the timestamp order of the wind speed value X of each initial subsequence, and at the same time add regularization processing to reduce the fitting of the data. Use a self-attention mechanism to perform a linear projection calculation on the temporal features of the wind speed values of each of the initial subsequences after regularization in the first graph structure G1 to generate the potential correlation matrix W: ; In the formula, . Q and K respectively represent the query vector and the key vector, and they can be linearly projected and calculated respectively using the learnable parameters W Q and W K ; d is the hidden dimension of Q and K.

[0053] In this embodiment, through a two-layer gated recurrent unit and regularization means, the temporal correlation between the initial subsequences is fully considered, the learning ability and generalization ability of the temporal features between the initial subsequences are enhanced, the accuracy of the potential correlation matrix is improved, and thus the accuracy of wind speed prediction is improved.

[0054] In one of the embodiments, as Figure 6 shown, step 204 extracts the spectral features and hidden dependencies of the wind speed values of each of the initial subsequences based on the first graph structure, specifically including the following steps 601 to step 606.

[0055] Step 601: Perform a graph Fourier transform on the first graph structure to obtain the first spectral matrix of the wind speed data sequence.

[0056] Step 602: Perform a discrete Fourier transform on the first spectral matrix of the wind speed data sequence to obtain the first frequency domain representation of the wind speed values of each of the initial subsequences.

[0057] Step 603: Obtain the spectral features of the wind speed values of each of the initial subsequences by performing one-dimensional convolution and gated linear units on the first frequency domain representations of the wind speed values.

[0058] Step 604: Perform inverse discrete Fourier transform on the first frequency domain representations of the wind speed values to obtain the first time domain representations of the wind speed values.

[0059] Step 605: Perform graph convolution on the first time domain representations of the wind speed values to obtain the second frequency domain representations of the wind speed values and obtain the hidden dependency relationships of the wind speed values of each of the initial subsequences.

[0060] Step 606: Perform inverse graph Fourier transform on the second frequency domain representations of the wind speed values to obtain the second time domain representations of the wind speed values.

[0061] Specifically, the potential association matrix W and the wind speed values X of each of the initial subsequences form a first graph structure G1 = (X, W). Perform graph Fourier transform on the first graph structure G1 to obtain the first spectral matrix of each of the initial subsequences . Perform discrete Fourier transform on the first spectral matrix of each of the initial subsequences to obtain the frequency domain results of the wind speed values of each of the initial subsequences, that is, the first frequency domain representations .

[0062] Pass the first frequency domain representations of each of the initial subsequences through one-dimensional convolution and gated linear units to learn the frequency domain features of each of the initial subsequences .

[0063] Use inverse discrete Fourier transform to convert the frequency domain features of the wind speed values of each of the initial subsequences into the first time domain representations of each of the initial subsequences . Perform graph convolution on the first time domain representations of the wind speed values of each of the initial subsequences to obtain the second frequency domain representations of the wind speed values and obtain the hidden dependency relationships of the wind speed values of each of the initial subsequences.

[0064] Perform convolution on the coupled association spectral feature matrix between nodes in the first graph structure. Perform inverse graph Fourier transform on the second frequency domain representations of the wind speed values to obtain the second time domain representations of the wind speed values, that is, obtain the updated subsequences Z corresponding to each of the initial subsequences j .

[0065] More preferably, in step 204, generating the backtracking value corresponding to the wind speed value of each of the initial subsequences and the first prediction value of the next moment corresponding to the wind speed value of each of the initial subsequences based on the spectral characteristics and hidden dependencies of the wind speed values of the initial subsequences includes the following: The updated subsequence Z output by each channel j is concatenated to form a wind speed subsequence set Z. The wind speed subsequence set Z is respectively passed through a prediction channel to generate the first prediction value of the next moment corresponding to the wind speed value of each of the initial subsequences and through a backtracking channel to generate the backtracking value corresponding to the wind speed value of each of the initial subsequences.

[0066] Among them, both the prediction channel and the backtracking channel are composed of a gated linear unit and a fully connected layer.

[0067] Specifically, the prediction value output after passing through the gated linear unit and the fully connected layer of the prediction channel ; the backtracking value output after passing through the gated linear unit and the fully connected layer of the backtracking channel . Among them, represents the basis vector; , is the expansion basis coefficient generated by the fully connected layer based on the wind speed subsequence set Z; k represents the kth prediction module, which is used to predict the kth graph structure.

[0068] In this embodiment, by performing discrete Fourier transform and inverse discrete Fourier transform on the first graph structure, the self-correlation characteristics of different modes and the same module in different time stamps in each initial subsequence are captured, and the initial subsequence in the original time domain is converted to the frequency domain, making the subsequence easier to be recognized by convolution. Finally, it is converted back to the time domain, which can effectively capture the local information and global topological structure of the graph structure data, realize the signal transmission between the initial subsequences, and learn the time series data from the perspective of frequency selection, and can denoise the wind speed value without loss of accuracy, thereby improving the accuracy of the first prediction value of the next moment corresponding to the wind speed value of each of the generated initial subsequences.

[0069] In one of the embodiments, in step 206, extracting the spectral characteristics and hidden dependencies of the backtracking errors of the initial subsequences based on the second graph structure specifically includes steps 701 to 706.

[0070] Step 701, perform graph Fourier transform on the second graph structure to obtain the second spectral matrix of the wind speed data sequence.

[0071] Step 702, perform discrete Fourier transform on the second spectral matrix of the wind speed data sequence to obtain the third frequency domain representation of the backtracking errors of the initial subsequences.

[0072] Step 703: Obtain the spectral features of each of the backtracking errors by performing one-dimensional convolution and gated linear units on the third frequency domain representations of the backtracking errors.

[0073] Step 704: Perform inverse discrete Fourier transform on the third frequency domain representations of the backtracking errors to obtain the third time domain representations of the backtracking errors.

[0074] Step 705: Perform graph convolution on the third time domain representations of the backtracking errors to obtain the fourth frequency domain representations of the backtracking errors and acquire the hidden dependencies of the backtracking errors of the initial subsequences.

[0075] Step 706: Perform inverse graph Fourier transform on the fourth frequency domain representations of the backtracking errors to obtain the fourth time domain representations of the backtracking errors.

[0076] In this embodiment, the processing method of step 206 is the same as that of step 204 as Figure 6 shown, except that the graph structure processed is different, which is the second graph structure. Therefore, the method defined in this embodiment can refer to the description of steps 601 to 606 in the above embodiment. However, in steps 701 to 706, the second graph structure G2 is used to predict the second predicted value, that is, to predict the backtracking error part of the wind speed value of each initial subsequence, so as to improve the accuracy of the wind speed prediction value of the final wind speed data sequence.

[0077] In one embodiment, a short-term wind speed prediction device along a railway is provided, that is, a short-term wind speed prediction model along a railway. As Figure 7 shown, the wind speed prediction model includes a sequence decomposition module 21, a potential association module 22, a first prediction module 23, a second prediction module 24, and a prediction output module 25.

[0078] Among them, the first prediction module 23 and the second prediction module 24 have the same function. The number of prediction modules in the short-term wind speed prediction model along the railway can be increased or decreased according to actual needs. Figure 7 The two prediction modules included are the optimal number of modules for short-term wind speed prediction along the railway obtained from the experimental results of this application.

[0079] The sequence decomposition module 21 is used to decompose the wind speed data sequence within a preset time period collected into a number of initial subsequences and generate the wind speed values of each of the initial subsequences.

[0080] The potential correlation module 22 is used to extract the temporal features of the wind speed values of each of the initial subsequences by using a two-layer gated recurrent unit, and perform regularization processing on the temporal features of the wind speed values of each of the initial subsequences to generate a potential correlation matrix between each of the initial subsequences; based on the wind speed values of each of the initial subsequences and the potential correlation matrix, a first graph structure of the wind speed data sequence is constructed.

[0081] The first prediction module 23 is used to extract the spectral features and hidden dependencies of the wind speed values of each of the initial subsequences based on the first graph structure, and generate a backtracking value corresponding to the wind speed value of each of the initial subsequences and a first prediction value for the next moment corresponding to the wind speed value of each of the initial subsequences based on the spectral features and hidden dependencies of the wind speed values of each of the initial subsequences.

[0082] The potential correlation module 22 is further used to calculate the difference between the wind speed value of each of the initial subsequences and the corresponding backtracking value to obtain the prediction error of the wind speed value of each of the initial subsequences, and construct a second graph structure of the wind speed data sequence based on each of the prediction errors and the potential correlation matrix.

[0083] The second prediction module 24 is used to extract the spectral features and hidden dependencies of the prediction errors of each of the initial subsequences based on the second graph structure, and generate a second prediction value for the next moment corresponding to the backtracking error of each of the initial subsequences according to the spectral features and hidden dependencies of the prediction errors of each of the initial subsequences.

[0084] The prediction output module 25 is used to generate a wind speed prediction value for the next moment of the wind speed data sequence based on the first prediction value and the corresponding second prediction value of each of the initial subsequences.

[0085] In one embodiment, the sequence decomposition module 21 is further used to: decompose the wind speed data sequence into a plurality of initial mode functions, perform Hilbert transform on each of the initial mode functions to obtain the single-sided spectrum of each initial mode function; use the exponential mixing modulation method to move the single-sided spectrum of each of the initial mode functions to their respective estimated center frequencies and perform demodulation to obtain the estimated bandwidth of each of the initial mode functions; perform unconstrained variational on the estimated bandwidths of each of the initial mode functions, and use the alternating direction method of multiplicative operators for iterative update to obtain a plurality of the initial subsequences and the wind speed value of each of the initial subsequences.

[0086] In one embodiment, the potential association module 22 is further configured to: obtain the timestamp order corresponding to each of the initial subsequences; use a two-layer gated recurrent unit to calculate the temporal features of the wind speed values of each of the initial subsequences according to the timestamp order; use a regularization layer to perform regularization processing on the temporal features of the wind speed values of each of the initial subsequences, and use a self-attention mechanism to perform a linear projection calculation on the temporal features of the wind speed values of each of the initial subsequences after regularization to generate the potential association matrix.

[0087] In one embodiment, the first prediction module 23 is further configured to: perform a graph Fourier transform on the first graph structure to obtain the first spectral matrix of the wind speed data sequence; perform a discrete Fourier transform on the first spectral matrix of the wind speed data sequence to obtain the first frequency domain representation of the wind speed values of each of the initial subsequences; pass the first frequency domain representation of each of the wind speed values through a one-dimensional convolution and a gated linear unit to obtain the spectral features of the wind speed values of each of the initial subsequences; perform an inverse discrete Fourier transform on the first frequency domain representation of each of the wind speed values to obtain the first time domain representation of each of the wind speed values; perform a graph convolution on the first time domain representation of each of the wind speed values to obtain the second frequency domain representation of each of the wind speed values and obtain the hidden dependencies of the wind speed values of each of the initial subsequences; perform an inverse graph Fourier transform on the second frequency domain representation of each of the wind speed values to obtain the second time domain representation of each of the wind speed values.

[0088] In one embodiment, the first prediction module 23 is further configured to: perform a graph Fourier transform on the second graph structure to obtain the second spectral matrix of the wind speed data sequence; perform a discrete Fourier transform on the second spectral matrix of the wind speed data sequence to obtain the third frequency domain representation of the backtracking errors of each of the initial subsequences; pass the third frequency domain representation of each of the backtracking errors through a one-dimensional convolution and a gated linear unit to obtain the spectral features of each of the backtracking errors; perform an inverse discrete Fourier transform on the third frequency domain representation of each of the backtracking errors to obtain the third time domain representation of each of the backtracking errors; perform a graph convolution on the third time domain representation of each of the backtracking errors to obtain the fourth frequency domain representation of each of the backtracking errors and obtain the hidden dependencies of the backtracking errors of each of the initial subsequences; perform an inverse graph Fourier transform on the fourth frequency domain representation of each of the backtracking errors to obtain the fourth time domain representation of each of the backtracking errors.

[0089] In one embodiment, the prediction output module 25 is configured to: sum the first prediction value of each of the initial subsequences and the corresponding second prediction value to obtain the wind speed prediction value of each initial subsequence at the next moment; sum the wind speed prediction values of each of the initial subsequences at the next moment to generate the wind speed prediction value of the wind speed data sequence at the next moment.

[0090] In an exemplary embodiment, a training method for a wind speed prediction model is provided, which is applied to a short-term wind speed prediction model along a railway as shown in Figure 7 . The specific content includes the following: Step 801: Use the sequence decomposition module 21 to collect the real wind speed data sequence within a preset time period, and decompose the real wind speed data sequence into several initial subsequences and the wind speed values of each initial subsequence.

[0091] Step 802: The two-layer gated recurrent unit in the potential association module 22 calculates the temporal characteristics of the wind speed values of each initial subsequence according to the time stamp order corresponding to each initial subsequence. Use the regularization layer in the potential association module 22 to perform regularization processing on the temporal characteristics of the wind speed values of each initial subsequence, and use the self-attention mechanism to perform linear projection calculation on the temporal characteristics of the wind speed values of each initial subsequence after regularization to generate a potential association matrix. Based on the wind speed values of each initial subsequence and the potential association matrix, construct the first graph structure of the wind speed data sequence.

[0092] Step 803: Use the first prediction module 23 to perform graph Fourier transform on the first graph structure to obtain the first spectral matrix of each initial subsequence; for each initial subsequence. Obtain the spectral characteristics of the wind speed values of each initial subsequence through one-dimensional convolution and gated linear unit for the first frequency domain representation of each wind speed value. Perform inverse discrete Fourier transform on the first frequency domain representation of each wind speed value to obtain the first time domain representation of each wind speed value. Perform graph convolution on the first time domain representation of each wind speed value to obtain the second frequency domain representation of each wind speed value and obtain the hidden dependency relationship of the wind speed values of each initial subsequence. Perform inverse graph Fourier transform on the second frequency domain representation of each wind speed value to obtain the second time domain representation of each wind speed value, generate the first backtracking value corresponding to the wind speed value of each initial subsequence and the first prediction value of the next moment corresponding to the wind speed value of each initial subsequence.

[0093] Step 804: Calculate the difference between the wind speed value of each initial subsequence and the corresponding backtracking value to obtain the backtracking error of the wind speed value of each initial subsequence. According to each backtracking error, the potential association matrix, the spectral characteristics of the wind speed values of each initial subsequence, and the hidden dependency relationship, construct the second graph structure of the wind speed data sequence.

[0094] Step 805: Perform graph Fourier transform on the second graph structure to obtain the second spectral matrix of the wind speed data sequence. Perform discrete Fourier transform on the second spectral matrix of the wind speed data sequence to obtain the third frequency domain representation of the backtracking error of each initial subsequence. Pass the third frequency domain representation of each backtracking error through a one-dimensional convolution and a gated linear unit to obtain the spectral features of each backtracking error. Perform inverse discrete Fourier transform on the third frequency domain representation of each backtracking error to obtain the third time domain representation of each backtracking error. Perform graph convolution on the third time domain representation of each backtracking error to obtain the fourth frequency domain representation of each backtracking error and obtain the hidden dependencies of the backtracking errors of each initial subsequence. Perform inverse graph Fourier transform on the fourth frequency domain representation of each backtracking error to obtain the fourth time domain representation of each backtracking error, generate the second predicted value corresponding to the backtracking error of each initial subsequence at the next moment and the second backtracking value corresponding to the backtracking error of each initial subsequence.

[0095] Step 806: Sum the first predicted value of each initial subsequence and the corresponding second predicted value to obtain the wind speed predicted value of each initial subsequence at the next moment. Sum the first backtracking value of each initial subsequence and the second backtracking value to obtain the wind speed backtracking value of each initial subsequence.

[0096] Since during the model training process, the wind speed value of each initial subsequence and the wind speed predicted value at the next moment are both known quantities, i.e., real data.

[0097] Step 807: Calculate the mean absolute error between the wind speed predicted value and the wind speed backtracking value of each initial subsequence and the real data respectively to obtain the prediction loss and the inversion loss. And construct a loss function based on the prediction loss and the inversion loss : ; where, is the prediction loss, is the inversion loss. For each time stamp t, X(t) is the actual known real data set, is the wind speed predicted value of each initial subsequence, is the wind speed backtracking value of each initial subsequence, represents all the parameters in the short-term prediction model along the high-speed rail line.

[0098] Step 808: Adjust the network model parameters in the short-term wind speed prediction model along the high-speed rail line according to the loss result solved by the loss function until the loss result is less than or equal to the preset loss value.

[0099] In this embodiment, the wind speed prediction model is optimized through the prediction data and inversion data of the wind speed prediction model, which is beneficial to realizing accurate estimation of the wind speed.

[0100] Preferably, asFigure 7 The short - term wind speed prediction model along the high - speed railway shown operates on the Anaconda (Python 3.9) platform and the Intel(R) Core(TM) i5 - 8250U hardware environment. The prediction accuracy of the short - term wind speed prediction model along the railway in this application is compared with other models, specifically including the following content: First, obtain the original data along a certain high - speed railway. The original data is collected once per second by an ultrasonic anemometer located 4 meters above the rail surface of the catenary pole. Randomly select the measured wind speed time - series data at point A to construct an initial wind speed data sequence DataSet1 with an interval of 1 min for simulation testing. All data uses 1440 consecutive wind speed data as the simulation object, where 70% of the data (1008) is used as the training set, 20% of the data (288) is used as the validation set, and the remaining 10% of the data (144) is used as the test set.

[0101] Input the wind speed data sequence DataSet1 into the short - term wind speed prediction model along the railway in this application and other existing prediction models. According to the mean squared error (MSE), root mean squared error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE) of the model, and the overall orthogonality index (IOO), where the overall orthogonality index (IOO) describes the degree of frequency aliasing of the subsequence, and the larger the IOO value, the more serious the frequency aliasing. The following comparison of prediction errors as shown in Table 1 is obtained: Table 1 It can be seen from this that in the short - term wind speed prediction model along the railway in this application, the modal aliasing degree IOO of the subsequences decomposed in the sequence decomposition module 21 is the smallest. The two - layer gated recurrent and regularization operations performed by the potential association module 22 on each subsequence improve the capture of the time - series correlation, effectively enhancing the overall performance of the model. Compared with other models, the accuracy of wind speed prediction is improved.

[0102] Randomly select the wind speed data sequence DataSet2 with an interval of 1 min collected at point B to verify the accuracy and generalization ability of the short - term wind speed prediction model along the railway in this application. Input the wind speed data sequence DataSet2 into the short - term wind speed prediction model along the railway in this application and other existing prediction models, and the following comparison of prediction errors as shown in Table 2 is obtained: Table 2 It can be concluded that the wind speed prediction model in this application performs well on different data sets and has good generalization ability.

[0103] It should be understood that although the steps in the flowcharts involved in the above-described embodiments do not necessarily execute in the order indicated by the arrows. Unless there is a clear description in this article, there is no strict order limit for the execution of these steps, and these steps can be executed in other orders. Moreover, at least some of the steps in the flowcharts involved in the above-described embodiments may include multiple steps or multiple stages. These steps or stages do not necessarily execute and complete at the same moment, but can execute at different moments. The execution order of these steps or stages is not necessarily sequential either, but can execute alternately or in turn with at least some of the steps or stages in other steps or other steps.

[0104] In an exemplary embodiment, a computer device is provided, including a memory and a processor. A computer program is stored in the memory, and when the processor executes the computer program, the steps corresponding to the short-term wind speed prediction method along the railway as described in the above embodiments are implemented.

[0105] In an embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps corresponding to the short-term wind speed prediction method along the railway as described in the above embodiments are implemented.

[0106] In an embodiment, a computer program product is provided, including a computer program. When the computer program is executed by a processor, the steps corresponding to the short-term wind speed prediction method along the railway as described in the above embodiments are implemented.

[0107] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this application.

[0108] The above-described embodiments merely represent several implementation manners of this application. Their descriptions are relatively specific and detailed, but they should not be construed as limiting the patent scope of this application. It should be noted that for those of ordinary skill in the art, without departing from the concept of this application, several modifications and improvements can still be made, and these all belong to the protection scope of this application. Therefore, the protection scope of this application should be subject to the appended claims.

Claims

1. A method for predicting short-term wind speed along a railway, characterized in that: The method comprises: Decomposing the wind speed data sequence collected within a preset time period into a plurality of initial subsequences, and generating a wind speed value of each of the initial subsequences; Obtaining the timestamp sequence corresponding to each of the initial subsequences; using a two-layer gated recurrent unit, calculating the time series features of the wind speed values ​​of each of the initial subsequences according to the timestamp sequence; using a regularization layer, regularizing the time series features of the wind speed values ​​of each of the initial subsequences, and using a self-attention mechanism, performing linear projection calculation on the time series features of the wind speed values ​​of each of the initial subsequences after regularization, and generating a potential correlation matrix between each of the initial subsequences; Based on the wind speed values ​​of each of the initial subsequences and the potential association matrix, constructing a first graph structure of the wind speed data sequence; Based on the first graph structure, extract the frequency spectrum characteristics and hidden dependency relationship of the wind speed values ​​of each of the initial subsequences, and based on the frequency spectrum characteristics and hidden dependency relationship of the wind speed values ​​of each of the initial subsequences, generate a backtracking value corresponding to the wind speed value of each of the initial subsequences and a first predicted value at the next moment corresponding to the wind speed value of each of the initial subsequences; Calculate the difference between the wind speed value of each of the initial subsequences and the corresponding backtracking values ​​to obtain the backtracking error of the wind speed value of each of the initial subsequences, and construct a second graph structure of the wind speed data sequence based on the backtracking errors, the potential association matrix, and the spectral characteristics and hidden dependencies of the wind speed values ​​of each of the initial subsequences; Based on the second graph structure, extract the frequency spectrum characteristics and hidden dependency relationship of the backtracking errors of each of the initial subsequences, and generate a second prediction value at the next moment corresponding to the backtracking error of each of the initial subsequences according to the frequency spectrum characteristics and hidden dependency relationship of the backtracking errors of each of the initial subsequences; Based on the first prediction value of each of the initial subsequences and the corresponding second prediction value, a wind speed prediction value at the next moment of the wind speed data sequence is generated.

2. The method for predicting short-term wind speed along a railway according to claim 1, characterized in that: Decomposing the collected wind speed data sequence within the preset time period into a plurality of initial subsequences and generating a wind speed value of each initial subsequence comprises: Decomposing the wind speed data sequence into a plurality of initial modal functions, and performing Hilbert transform on each of the initial modal functions to obtain a single-sided spectrum of each initial modal function; Using an exponential hybrid modulation method, the single-sided spectrum of each of the initial modal functions is moved to the respective estimated center frequency and demodulated to obtain an estimated bandwidth of each initial modal function; The estimated bandwidth of each of the initial modal functions is subjected to unconstrained variation and iteratively updated using a multiplication operator alternating direction method to obtain a plurality of the initial subsequences and a wind speed value of each of the initial subsequences.

3. The method for predicting short-term wind speed along a railway according to claim 1, characterized in that: The extracting the frequency spectrum features and hidden dependencies of the wind speed values ​​of each of the initial subsequences based on the first graph structure includes: Performing a graph Fourier transform on the first graph structure to obtain a first spectrum matrix of the wind speed data sequence; Performing a discrete Fourier transform on the first spectrum matrix of the wind speed data sequence to obtain a first frequency domain representation of the wind speed value of each of the initial subsequences; The first frequency domain representation of each wind speed value is subjected to one-dimensional convolution and a gated linear unit to obtain the frequency spectrum characteristics of the wind speed value of each initial subsequence; Performing an inverse discrete Fourier transform on the first frequency domain representation of each wind speed value to obtain a first time domain representation of each wind speed value; Performing graph convolution on the first time domain representation of each wind speed value to obtain a second frequency domain representation of each wind speed value and acquiring a hidden dependency relationship of the wind speed values ​​of each initial subsequence; An inverse Fourier transform is performed on the second frequency domain representation of each wind speed value to obtain a second time domain representation of each wind speed value.

4. The method for predicting short-term wind speed along a railway according to claim 1, characterized in that: The extracting, based on the second graph structure, the spectrum features and hidden dependencies of the backtracking errors of each of the initial subsequences comprises: Performing a graph Fourier transform on the second graph structure to obtain a second spectrum matrix of the wind speed data sequence; Performing discrete Fourier transform on the second spectrum matrix of the wind speed data sequence to obtain a third frequency domain representation of the backtracking error of each of the initial subsequences; The third frequency domain representation of each of the backtracking errors is subjected to one-dimensional convolution and a gated linear unit to obtain a frequency spectrum feature of each of the backtracking errors; Performing an inverse discrete Fourier transform on the third frequency domain representation of each of the backtracking errors to obtain a third time domain representation of each of the backtracking errors; Performing graph convolution on the third time domain representation of each of the backtracking errors to obtain a fourth frequency domain representation of each of the backtracking errors and acquiring a hidden dependency relationship of the backtracking errors of each of the initial subsequences; An inverse Fourier transform is performed on the fourth frequency domain representation of each of the back-tracing errors to obtain a fourth time domain representation of each of the back-tracing errors.

5. The method for predicting short-term wind speed along a railway according to claim 1, characterized in that: The step of generating a wind speed prediction value at the next moment of the wind speed data sequence based on the first prediction value of each of the initial subsequences and the corresponding second prediction value comprises: The first predicted value of each of the initial subsequences is summed with the corresponding second predicted value to obtain a wind speed predicted value of each initial subsequence at the next moment; The wind speed prediction values ​​of each of the initial subsequences at the next moment are summed to generate the wind speed prediction value of the wind speed data sequence at the next moment.

6. A short-term wind speed prediction device along a railway, characterized in that: The device comprises: A sequence decomposition module is used to decompose the wind speed data sequence collected within a preset time period into a plurality of initial subsequences, and generate a wind speed value for each of the initial subsequences; A potential association module is used to obtain the timestamp sequence corresponding to each of the initial subsequences; using two layers of gated recurrent units, according to the timestamp sequence, calculate the time series characteristics of the wind speed values ​​of each of the initial subsequences; using a regularization layer, regularize the time series characteristics of the wind speed values ​​of each of the initial subsequences, and use a self-attention mechanism to perform linear projection calculation on the time series characteristics of the wind speed values ​​of each of the initial subsequences after regularization, and generate a potential association matrix between each of the initial subsequences; based on the wind speed values ​​of each of the initial subsequences and the potential association matrix, construct a first graph structure of the wind speed data sequence; A first prediction module, configured to extract, based on the first graph structure, frequency spectrum features and hidden dependencies of the wind speed values ​​of each of the initial subsequences, and generate, based on the frequency spectrum features and hidden dependencies of the wind speed values ​​of each of the initial subsequences, a backtracking value corresponding to the wind speed value of each of the initial subsequences and a first prediction value at a next moment corresponding to the wind speed value of each of the initial subsequences; The potential association module is further used to calculate the difference between the wind speed value of each of the initial subsequences and the corresponding backtracking values, obtain the backtracking error of the wind speed value of each of the initial subsequences, and construct a second graph structure of the wind speed data sequence based on each of the backtracking errors and the potential association matrix; a second prediction module, configured to extract, based on the second graph structure, frequency spectrum features and hidden dependencies of the backtracking errors of each of the initial subsequences, and generate, according to the frequency spectrum features and hidden dependencies of the backtracking errors of each of the initial subsequences, a second prediction value at a next moment corresponding to the backtracking error of each of the initial subsequences; The prediction output module is used to generate the wind speed prediction value of the wind speed data sequence at the next moment based on the first prediction value of each initial subsequence and the corresponding second prediction value.

7. The short-term wind speed prediction device along the railway according to claim 6, characterized in that: The sequence decomposition module is also used to: decompose the wind speed data sequence into a number of initial modal functions, and perform Hilbert transform on each of the initial modal functions to obtain a single-sided spectrum of each initial modal function; use an exponential hybrid modulation method to move the single-sided spectrum of each initial modal function to its respective estimated center frequency and perform demodulation to obtain an estimated bandwidth of each initial modal function; perform unconstrained variation on the estimated bandwidth of each initial modal function, and perform iterative update using a multiplication operator alternating direction method to obtain a number of the initial subsequences and a wind speed value of each of the initial subsequences.

8. The short-term wind speed prediction device along the railway according to claim 6, characterized in that: The first prediction module is also used to: perform a graph Fourier transform on the first graph structure to obtain a first spectrum matrix of the wind speed data sequence; perform a discrete Fourier transform on the first spectrum matrix of the wind speed data sequence to obtain a first frequency domain representation of the wind speed values ​​of each of the initial subsequences; perform a one-dimensional convolution and a gated linear unit on the first frequency domain representation of each of the wind speed values ​​to obtain the frequency spectrum characteristics of the wind speed values ​​of each of the initial subsequences; perform an inverse discrete Fourier transform on the first frequency domain representation of each of the wind speed values ​​to obtain a first time domain representation of each of the wind speed values; perform a graph convolution on the first time domain representation of each of the wind speed values ​​to obtain a second frequency domain representation of each of the wind speed values ​​and obtain hidden dependencies of the wind speed values ​​of each of the initial subsequences; An inverse Fourier transform is performed on the second frequency domain representation of each wind speed value to obtain a second time domain representation of each wind speed value.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 5 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images