High-speed rail second-level wind speed prediction and early warning method based on wind process dynamic division and physical guidance

Abstract

The application discloses a high-speed rail second-level wind speed prediction and early warning method based on wind process dynamic division and physical guidance, comprising the following steps: in stage one, wind process dynamic division and offline construction of a knowledge base, historical wind field data are divided into wind process types with clear physical meaning through a graph attention network and adaptive spectral clustering; in stage two, hierarchical online wind speed prediction based on physical guidance, macro-, meso- and micro-scale decoupling deduction is carried out based on a hierarchical state space model with physical guidance, and a multi-physical constraint loss function is used to ensure that the prediction conforms to the fluid mechanics law; in stage three, dynamic adaptive envelope line online early warning, a dynamic safety envelope line is generated by using the extreme value theory and real-time scene information, and scene-adaptive accurate hierarchical early warning is realized. The method can significantly improve the accuracy of high-speed rail wind speed prediction and the reliability of early warning.

Description

Technical Field

[0001] This invention relates to the field of high-speed railway safety early warning and wind speed prediction technology, and in particular to a high-speed railway second-level wind speed prediction and early warning method based on dynamic wind process division and physical guidance. Background Technology

[0002] The safe operation of high-speed railways is highly dependent on real-time perception and early warning of wind conditions along the line. Currently, most high-speed railway wind protection systems rely on fixed thresholds for alarms, lacking in-depth analysis of the dynamic evolution of wind speed over time and the ability to predict under physical constraints. This results in a high false alarm rate and delayed response. Especially in complex terrain and abrupt wind field changes, traditional methods struggle to achieve accurate, second-level wind speed predictions and scenario-adaptive risk warnings.

[0003] Existing wind speed forecasting methods mostly rely on statistical learning or deep learning models. While these methods have improved average performance, they generally lack physical interpretability and fail to effectively integrate fluid dynamics principles with terrain effects, leading to significant prediction errors in extreme weather or complex scenarios. Furthermore, existing early warning methods often use static thresholds, neglecting the dynamic coupling effects of train operation status, terrain changes, and meteorological background, making it difficult to achieve accurate tiered early warnings.

[0004] Therefore, there is an urgent need for a wind speed prediction and early warning method that can integrate physical laws, adapt to dynamic scenarios, and possess high interpretability and high accuracy, in order to improve the safety of high-speed rail operation and the level of intelligent early warning. Summary of the Invention

[0005] To address the problems of low prediction accuracy, poor physical consistency, and insufficient adaptability in existing high-speed rail wind speed warning systems, this invention proposes a second-level wind speed prediction and warning method for high-speed rail based on dynamic wind process segmentation and physical guidance. Through a three-stage framework of wind identification, wind calculation, and wind warning, it achieves intelligent segmentation of wind processes, hierarchical prediction guided by physical guidance, and dynamic warning with scenario adaptation.

[0006] This invention adopts the following technical solution: a method for predicting and warning of high-speed rail wind speed at the second level based on dynamic wind process division and physical guidance, comprising the following steps:

[0007] Step 1: Segment the historical long-series wind speed data, extract the morphological features, complexity features, and environmental features of each time segment, and construct a multi-dimensional feature vector;

[0008] Step 2: Construct a physical association graph based on time sequence segments, where nodes are time sequence segments and edge weights are defined by generalized physical distance;

[0009] Step 3: Use graph attention network to aggregate features, determine the adaptive number of clusters based on graph structure density and physical separability, divide the wind process into multiple categories through spectral clustering, and build a wind process knowledge base offline.

[0010] Step 4: Calculate the feature vector of the real-time wind speed sequence online, match it with the cluster center in the wind process knowledge base, and select the parameters of the physical-guided hierarchical state-space model.

[0011] Step 5: Use a physically-guided hierarchical state-space model to perform decoupled modeling at the macro, meso, and micro scales to predict wind speed;

[0012] Step 6: Train a physics-guided hierarchical state-space model using a total loss function that includes trend curvature constraints, energy conservation constraints, non-equilibrium physics constraints, and terrain constraints to ensure that the wind speed prediction results conform to physical laws.

[0013] Step 7: Based on the GRU network, predict the wind speed baseline, and dynamically calculate the envelope width by combining terrain roughness, wind speed variation coefficient, train speed and atmospheric stability to generate the upper and lower envelopes.

[0014] Step 8: Correct the wind speed forecast value based on whether it exceeds the envelope, calculate the continuous risk score, realize graded early warning, and output risk attribution analysis.

[0015] As a preferred embodiment, the multidimensional feature vector in step 1 is:

[0016] ;

[0017] in, For fragments Multidimensional feature vectors, superscript Indicates transpose. For morphological characteristics, As a feature of complexity, Environmental characteristics.

[0018] As a preferred option, the generalized physical distance described in step 2 is formulated as follows:

[0019] ;

[0020] in, For dynamic time-normalized distance, For physical environment feature vectors, The weighting coefficients and , For Gaussian kernel bandwidth parameters, For nodes and The generalized physical distance.

[0021] As a preferred embodiment, in the graph attention network described in step 3, the attention coefficient formula is:

[0022] ;

[0023] in, For a linear rectified activation function with leakage, For the attention mechanism, a learnable weight vector. The weight matrix is ​​a learnable matrix. For nodes and The initial feature vector, For nodes Its neighboring nodes Attention coefficient;

[0024] Get Node Aggregation representation The formula is:

[0025] ;

[0026] in, For nodes The set of first-order neighbor nodes, For nodes Attention coefficients of its first-order neighbor node n; This is the Sigmoid activation function, used to achieve non-linear mapping and normalization of features;

[0027] The adaptive clustering number The formula is determined as follows:

[0028] ;

[0029] in, For class index parameters, For inter-class physical feature similarity, For intra-class morphological similarity, The Jensen-Shannon divergence is the physical distribution. For balance coefficient, To determine the value of the independent variable that maximizes the objective function;

[0030] Calculate the similarity matrix And construct the Laplace matrix , For the degree matrix, select the first... K-means clustering is performed on the feature vectors.

[0031] As a preferred option, in step 4, the feature vector of the real-time wind speed sequence is calculated. With cluster center Mahalanobis distance Select the hierarchical state-space model parameters for physical guidance, corresponding to the category with the minimum distance; where... It is represented as Mahalanobis distance.

[0032] As a preferred option, in step 5, the wind speed prediction value is the sum of contributions from the macroscopic, mesoscopic, and microscopic scales:

[0033] ;

[0034] in, , , These represent velocities at the macroscopic, mesoscopic, and microscopic scales, respectively. This is the predicted wind speed value. For the current moment, To predict the time step;

[0035] As a preferred embodiment, the total loss function in step 6 is:

[0036] ;

[0037] in, For the total loss, For mean square error loss, As a trend curvature constraint, Due to the constraint of energy conservation, For non-equilibrium physical constraints, Due to terrain physical constraints, These are the weighting coefficients.

[0038] As a preferred option, in step 7, the wind speed baseline is predicted based on the GRU network. Calculate the dynamic envelope width ,generate Upper envelope of time and lower envelope The formula is:

[0039] ;

[0040] ;

[0041] in, This is the predicted value for the wind speed baseline. For standard normal distribution Quantiles Uncertainty in predictions for a physics-guided hierarchical state-space model. The vehicle-wind coupling coefficient is given.

[0042] As a preferred embodiment, in step 8, the continuous risk score is calculated using the following formula:

[0043] ;

[0044] in, This is the final warning wind speed value. and These are the upper and lower envelopes, respectively. The time sensitivity coefficient, For terrain risk weighting, Assuming a risk weight for vehicle speed, This is a continuous risk score.

[0045] If the predicted value exceeds the envelope range, it should be corrected to... ,according to Determine the warning level and output auxiliary decision-making information, including: predicted wind speed, uncertainty range, risk score and risk contribution.

[0046] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects:

[0047] 1. The high-speed rail second-level wind speed prediction and early warning method proposed in this invention uses both morphological and physical driving forces to adaptively construct historical wind field maps and establish a wind process knowledge base with clear physical meaning, laying a reliable foundation for online accurate matching and prediction. By adopting a physical-guided hierarchical state-space model, it innovatively decouples and couples macroscopic environmental forcing, mesoscopic topographic modulation and microscopic turbulent fluctuations through mechanism decoupling and coupling deduction, and designs a loss function that includes multiple hard constraints such as trend curvature, energy conservation and non-equilibrium physics, fundamentally ensuring the high accuracy of the prediction results under the premise of strictly following the laws of fluid mechanics.

[0048] 2. The method of this invention constructs a scenario-adaptive dynamic safety envelope early warning mechanism. By integrating multi-source information such as real-time terrain, train speed, and atmospheric stability, it dynamically adjusts the risk boundary and combines it with a continuous risk scoring formula to achieve intelligent hierarchical early warning from static threshold judgment to dynamic trend perception, providing interpretable decision support with risk attribution. The entire system adopts an efficient architecture of offline construction and online matching, with a clear modular design, combining real-time performance, reliability, scalability, and strong engineering practicality. It provides an integrated solution for wind speed safety early warning of high-speed rail and other complex transportation infrastructures through deep fusion of physical information. Attached Figure Description

[0049] Figure 1 This is the overall flowchart of the high-speed rail second-level wind speed prediction and early warning method of the present invention.

[0050] Figure 2This is a schematic diagram of the dynamic division of the wind process and the construction of the knowledge base.

[0051] Figure 3 It is a diagram of a physically-guided hierarchical state-space model structure.

[0052] Figure 4 It is a flowchart of dynamic adaptive envelope generation and early warning decision-making.

[0053] Figure 5 This is a comparison chart of wind speed predictions from the PH-SSM model. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the application will be further described in detail below with reference to the accompanying drawings. The described embodiments are only a part of the embodiments involved in this invention. All non-innovative embodiments based on these embodiments by other researchers in the art are within the protection scope of this invention. Furthermore, the step numbers in the embodiments of this invention are only set for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0055] This invention proposes a method for predicting and warning of high-speed rail wind speeds at the second level based on dynamic wind process division and physical guidance. The process is as follows: Figure 1 As shown, its core implementation framework is divided into three phases:

[0056] Wind process segmentation and knowledge base construction stage (offline): The input is historical long-sequence wind speed data, which is sequentially processed through a multi-scale feature extraction module, a time-series physical correlation graph construction module, and a graph attention aggregation and adaptive spectral clustering module, and outputs a wind process type knowledge base.

[0057] Physics-guided hierarchical wind speed prediction stage (online): The input is a real-time wind speed sequence. Combined with a wind process type knowledge base, the system goes through a scene matching module and macro-meso-micro three-scale decoupled deduction. The system is then optimized using a PH-SSM model with a multi-physical constraint loss function, and outputs second-level wind speed prediction values.

[0058] Dynamic adaptive envelope early warning stage (online): The input is a real-time wind speed sequence, combined with the second-level wind speed prediction value, and then performs GRU baseline prediction, dynamic safety envelope generation and risk score calculation in sequence, and outputs a multi-level (four-level) early warning level.

[0059] In one embodiment of the present invention, each stage is described in detail in conjunction with a typical application scenario of wind speed monitoring along a high-speed railway line.

[0060] Phase 1: Dynamic division of wind processes and construction of the knowledge base, the process is as follows: Figure 2As shown, the process includes: historical long-series wind speed data acquisition, multi-scale feature extraction, construction of temporal physical correlation graphs, graph attention aggregation and adaptive spectral clustering, and finally, the construction of a wind process type knowledge base.

[0061] Step 1: Collect historical second-level wind speed observation data for at least one full year (covering all four seasons) along the target high-speed railway line, along with synchronized meteorological data (such as air pressure and temperature), geographic information data (such as elevation and surface roughness), and train operation logs (speed and location). Unify the timestamps of all data to UTC time and spatially align them according to the monitoring station locations. Slide the preprocessed historical second-level wind speed long series into a series of time-series segments with a window length of 3 minutes (180 seconds) and a 50% overlap rate.

[0062] In this embodiment, feature extraction is performed for each segment, and the specific steps are as follows:

[0063] Step 1-1: Extract morphological features and calculate the average wind speed within the 180-second window. Wind speed standard deviation The wind speed trend within the window was fitted using linear regression, and its slope was used as the trend coefficient. Calculate the standard deviation of the residual sequence after wind speed detrending, and use it as the pulsation standard deviation. This constitutes the morphological feature vector. .

[0064] Steps 1-2: Extract complexity features and calculate the approximate entropy of the wind speed segment. As a feature of complexity .

[0065] Set mode length Similarity tolerance , The standard deviation of this segment is the sequence length. According to the formula Calculations are performed to quantify the randomness and unpredictability of the sequence.

[0066] Steps 1-3: Extract environmental features and obtain the environmental data corresponding to the midpoint of the segment, including: pressure gradient, roughness, wind-road angle, and temperature gradient.

[0067] Calculate the pressure gradient near this moment. (Obtained by the ratio of the pressure difference between adjacent weather stations to their distance); Query the terrain roughness coefficient of the location of this station. (From a geographic information database); Calculate the angle between the average wind direction at that moment and the orientation of the high-speed rail line. ; Calculate the vertical temperature gradient This constitutes the environmental feature vector. .

[0068] Finally, the three types of features are concatenated to obtain the multidimensional feature vector of this time series segment. :

[0069] .

[0070] Step 2: Treat all historical time series segments as nodes and construct an undirected weighted graph G.

[0071] In this embodiment, for any two nodes (fragments) and Calculate the weights of the edges connecting them, which are determined by the generalized physical distance. The reciprocal definition of (the smaller the distance, the higher the similarity, and the greater the weight).

[0072] The calculation is as follows:

[0073] ;

[0074] in, The morphological distance between two original wind speed sequence segments was calculated using the standard dynamic time warping algorithm. They are nodes Environmental feature vector The parameters are set as follows: , (Emphasizing that morphological similarity is slightly higher than environmental similarity) By calculating the distance between all pairs of nodes, the adjacency relationship of the graph is formed, and edges are established between nodes whose distance is less than a set threshold.

[0075] Step 3: Dynamic partitioning of wind processes and construction of a knowledge base, based on the constructed time-series physical correlation graph.

[0076] First, a graph attention network (GAT) is used to aggregate the neighborhood information of nodes to generate a high-dimensional representation that integrates the local graph structure and the node's own features. Then, an adaptive evaluation function that integrates intra-class morphological density and inter-class physical distribution separability is proposed to dynamically determine the optimal number of wind process categories in a data-driven manner. Finally, based on this number, spectral clustering is performed to divide all historical segments into several wind process categories with clear physical meaning, and the feature centers of each category are calculated and stored together with their corresponding prediction model parameters to form a structured wind process knowledge base.

[0077] In this embodiment, the specific steps are as follows:

[0078] Step 3-1, Graph Attention Feature Aggregation: Utilize a graph attention network to aggregate the multi-dimensional feature vectors of each node. Enhance the nodes. Calculate its relationship with neighboring nodes Attention coefficient :

[0079] ;

[0080] Then, weighted aggregation is performed to obtain the weighted aggregated node representation. :

[0081] ;

[0082] in, For nodes The set of first-order neighbor nodes, The Sigmoid activation function is used to achieve non-linear mapping and normalization of features, which allows the representation of each node to incorporate the similarity information of its local graph structure.

[0083] Step 3-2: Adaptively determine the number of clusters. To avoid bias from manually setting the number of clusters, an evaluation function is proposed:

[0084] ;

[0085] make In the interval Inner traversal, for each Spectral clustering partitioning is performed: one spectral clustering operation is performed (see next step) to obtain a temporary partition; the mean cosine similarity of physical features between classes is calculated. ; Calculate the average DTW distance for all samples within the class. ; Calculate the mean Jensen-Shannon divergence between the distributions of various environmental features and the overall distribution. .

[0086] Specifically, take Choose the option that maximizes the evaluation function value. As the final cluster number .

[0087] Step 3-3: Spectral clustering is performed and a knowledge base is built, based on the aggregated representation of all nodes. Calculate the similarity matrix , of which elements .

[0088] Construct the normalized Laplace matrix: , This is a degree matrix. For Perform eigenvalue decomposition and take its first eigenvalue. The eigenvectors corresponding to the smallest eigenvalues ​​form a new... The feature matrix. Finally, K-means clustering is performed on the row vectors of this matrix to obtain... There are several wind process categories. The cluster centers (mean values ​​of the feature vectors) for each category are calculated and stored in the wind process knowledge base.

[0089] Phase Two: Physically Guided Hierarchical Wind Speed ​​Prediction Phase. The structure of the Physically Guided Hierarchical State-Space Model (PH-SSM) is as follows: Figure 3 As shown, the model parameter transfer and update process includes the following steps:

[0090] Step 4: Online scene matching and model adaptation.

[0091] Based on real-time acquired second-level wind speed sequences, firstly, a three-minute real-time segment is extracted with the current moment as the endpoint, and its multi-dimensional feature vector is extracted according to the method in step 1, and input into the PH-SSM model; then, it is matched with the wind process type knowledge base, and the Mahalanobis distance between the feature vector and each cluster center in the wind process knowledge base is calculated. The optimal wind process category to which the current scene belongs is determined according to the minimum distance criterion; finally, the system automatically activates the pre-trained PH-SSM parameters corresponding to the category, so as to achieve accurate adaptation between the prediction model and the real-time wind field characteristics.

[0092] In this embodiment, the specific steps are as follows:

[0093] Step 4-1: Real-time feature extraction. Taking the current time as the endpoint, extract the wind speed sequence of the most recent 3 minutes (180 seconds) as a real-time segment. Following the method in Step 1, extract the multi-dimensional feature vector of this real-time segment. .

[0094] Step 4-2, Distance Calculation and Matching: Calculation With the knowledge base Cluster centers of each category The Mahalanobis distance between features takes into account the correlation between different dimensions of the feature. The formula is:

[0095] .

[0096] Step 4-3, Model Parameter Activation: Select and The corresponding wind process category. The system was pre-trained. Each set of PH-SSM parameters is optimized for a specific type of wind process. The system automatically loads and activates the PH-SSM model parameters corresponding to the matched category, completing the model adaptation.

[0097] Step 5: Use the activated PH-SSM model to predict wind speeds in the next second.

[0098] In this embodiment, the model decomposes the predicted value into the sum of three physical scales: macroscopic, mesoscopic, and microscopic. Each scale is modeled using different state-space equations, and they interact through coupling terms, as shown in the following formula:

[0099] ;

[0100] in, , , These represent velocities at the macroscopic, mesoscopic, and microscopic scales, respectively. This is the predicted wind speed value.

[0101] Macroscale: Simulating large-scale weather system forcing; macroscale state The evolution is governed by pressure gradient force, Coriolis force, and viscosity effect, with the background wind speed in the equation being... Feedback is provided by the mesoscale model, and the formula is:

[0102] ;

[0103] in, This refers to the macroscopic state. air density, For gradient operators, For air pressure, For Coriolis force parameters, Background wind speed, The turbulent viscosity coefficient, This refers to macroscopic-scale system noise.

[0104] Mesoscale: Simulating the modulation effect of terrain on airflow; mesoscale state In addition to the evolutionary advection term, the topographic gradient was added. Forced terms related to the square of wind speed The formula is:

[0105] ;

[0106] in, This is a mesoscopic state. Spatial coordinates along the wind direction, This is the terrain modulation coefficient. For terrain gradient, This represents noise in mesoscale systems.

[0107] Microscale: Simulating near-surface turbulent fluctuations; described using the Langevin equation, exhibiting a mean-oriented trend. The properties of regression are expressed in the formula:

[0108] ;

[0109] in, It is a directed mean. For the regression rate, For fluctuation intensity, For the standard Wiener process, For differential operators, This represents random turbulent impact.

[0110] During online forecasting, the system inputs current and historical wind speed and environmental observation data into the PH-SSM, solves the coupled equations through numerical integration (such as the Runge-Kutta method), and outputs second-level wind speed point predictions for the next 30 to 180 seconds. and its prediction uncertainty .

[0111] Step 6: Training the physical constraint model (completed offline).

[0112] Inputting a historical wind process category dataset, a physically guided hierarchical state-space model (PH-SSM) is constructed and optimized for each wind process category. A composite loss function is designed, fusing data fitting terms (such as mean squared error) with multiple physical constraint terms (including trend curvature constraints, energy conservation constraints, non-equilibrium Navier-Stokes equation constraints, and terrain similarity theory constraints). Gradient descent is used to iteratively optimize the model parameters. The training objective is to ensure that the model accurately fits historical data while its output strictly adheres to fundamental fluid dynamics laws, thereby obtaining a set of physically consistent predictive model parameters applicable to various wind processes, which is then stored for future use.

[0113] In this embodiment, the specific steps are as follows:

[0114] Step 6-1, regarding the result obtained in step 3 We categorized wind processes into distinct physical categories and extracted historical second-level wind speed sequences and environmental feature vectors for each category. The datasets for each category were randomly divided into training and validation sets in a 7:3 ratio to ensure consistency in wind process morphology and physical feature distribution between the training and validation sets.

[0115] Step 6-2, initialize the parameters of the PH-SSM macro-micro three-scale model:

[0116] For each wind process category, initialize the macro-, meso-, and micro-scale state equation parameters of the PH-SSM model. Macro-scale parameters include Coriolis force parameters. Background wind speed Turbulent viscosity coefficient etc.; mesoscale parameters include topographic modulation coefficients. Topographic gradient etc.; microscale parameters include regression rate. mean Fluctuation intensity Meanwhile, initialize system noise. , The prior distribution of .

[0117] Step 6-3: Construct a composite loss function that integrates data fitting and physical constraints, and define the total loss function. Based on the mean square error term as the data fitting basis, and superimposed with four constraint terms: trend curvature, energy conservation, non-equilibrium physics, and topographic physics, the formula is:

[0118] ;

[0119] in, For the total loss, For mean square error loss, As a trend curvature constraint, Due to the constraint of energy conservation, For non-equilibrium physical constraints, For terrain physical constraints.

[0120] Specifically, the calculation formulas for each constraint term are as follows:

[0121] Trend curvature constraint:

[0122] ;

[0123] Energy conservation constraint:

[0124] ;

[0125] Non-equilibrium physical constraints:

[0126] ;

[0127] Terrain physical constraints:

[0128] ;

[0129] in, The optimal values ​​for the constraint weights are determined using a grid search method.

[0130] Step 6-4: Iteratively optimize the model parameters using the gradient descent algorithm. With minimization as the optimization objective, the Adam gradient descent algorithm is used to iteratively update the parameters of the PH-SSM model, and the optimized PH-SSM model parameter set is output.

[0131] In each iteration, the wind speed prediction is first derived using the macro-, meso-, and micro-scale state equations, then the total loss function is calculated, the gradient is backpropagated, and the parameters are updated. The maximum number of iterations is set to 1000. If the change in the loss function value is less than a certain value over 20 consecutive iterations, the iteration is considered complete. Stop iterating when the time is right.

[0132] Step 6-5, Model Validation and Parameter Set Storage: The trained PH-SSM model is evaluated using the validation set to verify whether the mean square error of the predicted values ​​meets the preset accuracy requirements, and whether the output results meet all physical constraints. Model parameter combinations that do not meet physical consistency are removed, and the validated parameter sets of each category of PH-SSM model are stored in the wind process knowledge base for use in the online prediction phase.

[0133] Phase 3: Dynamic Adaptive Envelope Warning Phase, the process is as follows: Figure 4 As shown, the input is a real-time, second-level wind speed sequence environmental feature vector, and the output includes the warning level, corrected wind speed, and risk attribution. The specific steps are as follows:

[0134] Step 7: Dynamic safety envelope generation, to quantify risk boundaries for scenario-adaptive learning.

[0135] First, a gated cyclic unit (GRU) network is used to predict the baseline trend of future wind speed changes. Then, multi-source information such as real-time terrain roughness, wind speed variation coefficient, train speed, and atmospheric stability is integrated to calculate the dynamic envelope width using learnable coefficients. Next, dynamic upper and lower envelopes that evolve over time are generated by combining prediction uncertainty, vehicle-wind coupling effect, and extreme value theory. Finally, the predicted values ​​from the PH-SSM model are combined and corrected according to rules to obtain the final warning wind speed value. This comprehensive multi-dimensional calculation of continuous risk scoring and warning decision-making provides an adaptive and physically interpretable safety boundary for subsequent risk quantification and graded warning.

[0136] In this embodiment, the specific steps are as follows:

[0137] Step 7-1: Based on the GRU network, predict the wind speed baseline and input the real-time collected data. A continuous second-level wind speed sequence and the environmental feature vectors acquired synchronously The two types of data are input into the pre-trained GRU network.

[0138] By capturing the temporal dependencies of wind speed sequences through network gating mechanisms, the future wind speed sequence can be output. Wind speed baseline at any time :

[0139] .

[0140] Step 7-2: Calculate the dynamic envelope width by fusing multi-source information. First, calculate the real-time wind speed variation coefficient. The formula is:

[0141] ;

[0142] in, The standard deviation of real-time wind speed. This represents the real-time average wind speed.

[0143] Subsequently, the real-time terrain roughness index was obtained. Real-time train speed Atmospheric stability parameters Substituting into the envelope width calculation formula, we obtain the solution. Dynamic envelope width at time step :

[0144] ;

[0145] in, Learnable coefficients determined for offline training.

[0146] Step 7-3: Using extremum theory, generate dynamic upper and lower envelopes to obtain the standard normal distribution. quantiles Uncertainty in PH-SSM model predictions Vehicle-wind coupling coefficient .

[0147] Vehicle-wind coupling coefficient The calculation formula is:

[0148] ;

[0149] in, The calibration coefficient for vehicle-wind coupling is set to 0.8. The empirical coefficient is obtained by fitting measured vehicle-wind coupling data along the high-speed railway line, which characterizes the intensity of the interaction between the train and the wind field. The train's real-time operating speed, in km / h, is collected in real-time by the high-speed rail dispatching system. The train's reference operating speed is 300 km / h, which is a commonly used benchmark operating speed in high-speed rail design and serves as a reference threshold for the train-wind coupling effect.

[0150] Compare the above parameters with the wind speed baseline. Dynamic envelope width Substituting into the formula, we can calculate the results respectively. Upper envelope of time With lower envelope :

[0151] ;

[0152] ;

[0153] in, For GRU network output The baseline wind speed forecast at any given time represents the basic evolution trend of wind speed under the absence of sudden disturbances. middle A value of 0.05 corresponds to the 95th quantile of the standard normal distribution, while a value of 1.96 is the quantile value representing the 95% confidence level in extreme value theory. Output of the PH-SSM model The standard deviation of the wind speed prediction at any given time represents the degree of uncertainty in the model's prediction results; for The dynamic envelope width at any given time represents the range of safety boundary fluctuations determined by real-time scene factors.

[0154] Step 8: Intelligent early warning decision-making and risk quantification.

[0155] First, the point prediction output of the Physically Guided Model (PH-SSM) is boundary-corrected based on the dynamic envelope to obtain the final warning wind speed value. Then, a continuous risk score is calculated by comprehensively considering the degree of wind speed deviation from the envelope, the rate of deterioration of the trend, and the scenario risk weights of terrain and vehicle speed. Furthermore, a four-level (attention, caution, warning, alarm) graded warning is implemented based on the risk score threshold, and multi-condition triggering logic such as approach warning, over-limit warning, and emergency stop is set. Finally, the system outputs a complete decision information package containing corrected wind speed, risk score, warning level, and risk contribution attribution, providing interpretable and operable warning support for dispatchers.

[0156] In this embodiment, the specific steps are as follows:

[0157] Step 8-1: Correct the PH-SSM point prediction output based on the dynamic envelope to obtain the output of the PH-SSM model. Predicted wind speed at any time Compare it with the upper and lower envelopes generated in step 7-3. , Compare the values. Adjust according to the following rules to obtain the final warning wind speed value. :

[0158] ;

[0159] in, Physics-guided hierarchical state-space models for the future The raw, second-level wind speed point prediction values ​​are output at all times;

[0160] Step 8-2: Calculate the continuous risk score by integrating multiple dimensions, specifically calculating the values ​​for wind speed deviation, trend term, and scene weight. Multiply these three values ​​together to obtain the result. Continuous risk score at any time The formula is:

[0161] ;

[0162] Among them, wind speed deviation Used to quantify the degree of deviation of wind speed from the safety boundary; trend term Used to reflect the rate of deterioration of wind speed changes, The time sensitivity coefficient is used; the scene weight is determined by the terrain risk weight. Risk weighting based on vehicle speed Multiplying them together, we get the bridge and tunnel section. Values ​​greater than those in plains sections, high-speed conditions The value is greater than that of the low-speed operating condition.

[0163] Step 8-3: Determine the Level 4 warning level based on the risk scoring threshold, and then determine the level based on the continuous risk score. The numerical range is matched with the corresponding warning level, and the specific determination rule is as follows:

[0164] .

[0165] Step 8-4: Execute the multi-condition composite early warning triggering logic. Based on the early warning level and wind speed change trend, perform three types of early warning triggering judgments:

[0166] Approaching warning: When and At that time, a proximity warning was triggered, indicating that the wind speed risk was rising.

[0167] Over-limit warning: When Or the measured wind speed value is continuous When the second is located in the high-risk zone above the envelope, an over-limit warning is triggered.

[0168] Emergency Stop Warning: When Furthermore, if the predicted wind speed trend continues to worsen, an emergency shutdown warning will be triggered, and a shutdown command will be sent to the dispatch system.

[0169] Step 8-5: Output a complete decision support information package, integrating model calculations and early warning judgment results to generate and output a decision support information package containing four core components: Final predicted wind speed value at time and uncertainty range Continuous risk score at any time And the corresponding warning levels.

[0170] Perform risk contribution attribution analysis and output: terrain effect. Train speed Meteorological background All three conditions are met. ; issue targeted dispatching suggestions, such as speed reduction and temporary parking.

[0171] In this embodiment, the data comes from measured data along the Jiangsu section of the Beijing-Shanghai High-Speed ​​Railway. To verify the robustness of the model proposed in this invention and the effectiveness of each module, the following experiments are conducted, with specific examples provided:

[0172] I. To verify the effectiveness of the dynamic wind process segmentation module, annual wind speed data from the Jiangsu section of the Beijing-Shanghai High-Speed ​​Railway were processed. After graph attention network feature aggregation and adaptive spectral clustering, the historical wind field data were divided into four physically meaningful wind process categories:

[0173] Category A (Stable): The average wind speed is low, the fluctuations are gentle, the wind direction is stable, and it is mostly found in plains areas under clear weather conditions.

[0174] Category B (Gust type): The average wind speed is moderate, but the high-frequency pulsation is significant, and the wind direction changes rapidly. It is mostly related to local thermal turbulence or topographic flow.

[0175] Category C (Strong Gradient): Wind speeds show a clear upward or downward trend, usually accompanied by the passage of a significant weather system or frontal process.

[0176] Category D (Extreme Turbulence): High average wind speed, violent fluctuations, and the highest complexity characteristic (approximate entropy) value, often corresponding to strong wind events in special terrains such as bridges and tunnel entrances.

[0177] The silhouette coefficient of the clustering results reached 0.72, and the Jensen-Shannon divergence of physical features (such as pressure gradient and wind direction angle) between each category was high, indicating that this method can effectively decouple wind field data into different modes with high physical consistency based on the dual features of "morphology and physics", laying the foundation for subsequent accurate matching.

[0178] 2. To verify the prediction accuracy of the Physically Guided Hierarchical State-Space Model (PH-SSM), measured data from a typical site, K1066 (bridge-cut transition section), were selected for testing. The wind field at this site exhibits combined characteristics of background wind forcing and topographic disturbance.

[0179] The PH-SSM model was trained using pre-defined "Category C (Strong Gradient Type)" data, and then applied to a test dataset that was not used in the training and included a period of rapid wind speed increase. The prediction results are as follows: Figure 5As shown in the figure, the predicted curve of the PH-SSM model is in high agreement with the actual wind speed curve, especially at the beginning of the steep increase in wind speed and at its peak, the model can accurately capture the magnitude and timing of the change.

[0180] III. To verify the effectiveness and scenario adaptability of the dynamic early warning module, a simulated thunderstorm and strong wind event was conducted at station K1066. The online operation process of the system is as follows:

[0181] Real-time matching: The system extracts the current wind speed sequence features in real time, calculates the Mahalanobis distance between the system and each cluster center in the wind process knowledge base, quickly matches "Category D (Extreme Turbulence Type)", and loads the corresponding PH-SSM model parameters and warning parameter set.

[0182] Dynamic envelope generation: The GRU network provides a baseline wind speed prediction. Simultaneously, the system monitors in real time an increase in terrain roughness, a sharp increase in the wind speed variability coefficient (CV), a train operating at 300 km / h, and atmospheric stability parameters indicating a strongly unstable stratification. Based on this multi-source information, the dynamic envelope width calculation module automatically increases the envelope width. This generates a safety boundary that is significantly wider than the normal situation.

[0183] Risk Quantification and Tiered Early Warning: When the predicted value of PH-SSM briefly touches the dynamic upper envelope, the system triggers an early warning. This is based on the continuous risk scoring formula. The score rapidly increased from 0.4 to 0.75 within 60 seconds. Based on this, the system subsequently issued the following:

[0184] Level III (Caution) Warning: When When the wind speed exceeds 0.5 for the first time, the dispatcher will be notified that "the wind speed is showing a clear upward trend, please pay attention to monitoring."

[0185] Level II (Alert) Warning: When If the wind speed exceeds 0.7 and the predicted value is outside the envelope, the system will display the message "Wind speed is expected to exceed the limit, it is recommended to prepare for speed limits".

[0186] The warning information also included a risk attribution analysis: the contribution of topographic effects. Train speed contribution Meteorological background contribution This proves that early warning decisions are not based solely on absolute wind speed values, but rather deeply integrate multi-dimensional risk factors from real-time scenarios, achieving scenario-adaptive and accurate early warnings.

[0187] This invention also proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. It should be noted that each module in the above system corresponds to a specific step of the method provided in this invention, possessing the corresponding functional modules and beneficial effects for executing the method. Technical details not described in detail in this embodiment can be found in the method provided in this invention.

[0188] This invention also proposes a computer-readable storage medium storing a computer program. It should be noted that each module in the above system corresponds to a specific step of the method provided in this invention, possessing the corresponding functional modules and beneficial effects for executing the method. Technical details not described in detail in this embodiment can be found in the method provided in this invention.

[0189] This invention provides a method for predicting and warning of high-speed rail wind speeds at the second level based on dynamic wind process division and physical guidance. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A method for predicting and warning high-speed rail wind speeds at the second level based on dynamic wind process division and physical guidance, characterized in that, Includes the following steps: Step 1: Segment the historical long-series wind speed data, extract the morphological features, complexity features, and environmental features of each time segment, and construct a multi-dimensional feature vector; Step 2: Construct a physical association graph based on time sequence segments, where nodes are time sequence segments and edge weights are defined by generalized physical distance; Step 3: Use graph attention network to aggregate features, determine the adaptive number of clusters based on graph structure density and physical separability, divide the wind process into multiple categories through spectral clustering, and build a wind process knowledge base offline. Step 4: Calculate the feature vector of the real-time wind speed sequence online, match it with the cluster center in the wind process knowledge base, and select the parameters of the physical-guided hierarchical state-space model. Step 5: Use a physically-guided hierarchical state-space model to perform decoupled modeling at the macro, meso, and micro scales to predict wind speed; Wind speed forecasts are the sum of contributions from macroscopic, mesoscopic, and microscopic scales: ； in, , , These represent velocities at the macroscopic, mesoscopic, and microscopic scales, respectively. This is the predicted wind speed value. For the current moment, To predict the time step; The macroscopic evolution equation is: ； in, This refers to the macroscopic state. air density, For gradient operators, For air pressure, For Coriolis force parameters, Background wind speed, The turbulent viscosity coefficient, For macroscopic scale system noise; The mesoscale evolution equation is: ； in, This is a mesoscopic state. Spatial coordinates along the wind direction, This is the terrain modulation coefficient. For terrain gradient, This represents noise in mesoscale systems. The microscale evolution equation is: ； in, It is a directed mean. For the regression rate, For fluctuation intensity, For the standard Wiener process, For differential operators, Represents random turbulent impact; Step 6: Train a physics-guided hierarchical state-space model using a total loss function that includes trend curvature constraints, energy conservation constraints, non-equilibrium physics constraints, and terrain constraints to ensure that the wind speed prediction results conform to physical laws. Step 7: Based on the GRU network, predict the wind speed baseline, and dynamically calculate the envelope width by combining terrain roughness, wind speed variation coefficient, train speed and atmospheric stability to generate the upper and lower envelopes. Step 8: Correct the wind speed forecast value based on whether it exceeds the envelope, calculate the continuous risk score, realize graded early warning, and output risk attribution analysis; The formula for the continuous risk score is: ； in, This is the final warning wind speed value. and These are the upper and lower envelopes, respectively. The time sensitivity coefficient, For terrain risk weighting, Assuming a risk weight for vehicle speed, This is a continuous risk score. If the predicted value exceeds the envelope range, it should be corrected to... ,according to Determine the warning level and output auxiliary decision-making information, including: predicted wind speed, uncertainty range, risk score and risk contribution.

2. The high-speed rail second-level wind speed prediction and early warning method according to claim 1, characterized in that, The multidimensional feature vector mentioned in step 1 is: ； in, For fragments Multidimensional feature vectors, superscript Indicates transpose; Morphological characteristics ; The average wind speed, Standard deviation This is the trend coefficient. The standard deviation of the pulsation; Complex features By approximate entropy express, ; For pattern length, For similarity tolerance, For sequence length, For all wind speed time series segments of length , The distance between the vector template and other vector templates of the same length is less than The average probability; Environmental characteristics ; For pressure gradient, For terrain roughness coefficient, The angle between the wind direction and the line. For temperature gradient.

3. The high-speed rail second-level wind speed prediction and early warning method according to claim 1, characterized in that, The generalized physical distance mentioned in step 2 is calculated using the following formula: ； in, For dynamic time-normalized distance, For physical environment feature vectors, The weighting coefficients and , For Gaussian kernel bandwidth parameters, For nodes and The generalized physical distance.

4. The high-speed rail second-level wind speed prediction and early warning method according to claim 3, characterized in that, In the graph attention network described in step 3, the formula for the attention coefficient is: ； in, For a linear rectified activation function with leakage, For the attention mechanism, a learnable weight vector. The weight matrix is ​​a learnable matrix. For nodes Its neighboring nodes The initial feature vector, For nodes Its neighboring nodes Attention coefficient; Get Node Aggregation representation The formula is: ； in, For nodes The set of first-order neighbor nodes, For nodes Attention coefficients of its first-order neighbor node n; This is the Sigmoid activation function, used to achieve non-linear mapping and normalization of features; The adaptive clustering number The formula is determined as follows: ； in, For class index parameters, For inter-class physical feature similarity, For intra-class morphological similarity, The Jensen-Shannon divergence is the physical distribution. For balance coefficient, To determine the value of the independent variable that maximizes the objective function; Calculate the similarity matrix And construct the Laplace matrix , For the degree matrix, select the first... K-means clustering is performed on the feature vectors.

5. The high-speed rail second-level wind speed prediction and early warning method according to claim 1, characterized in that, In step 4, the feature vector of the real-time wind speed sequence is calculated. With cluster center Mahalanobis distance Select the category corresponding to the minimum distance to activate the parameters of the physically guided hierarchical state-space model; where, It is represented as Mahalanobis distance.

6. The high-speed rail second-level wind speed prediction and early warning method according to claim 1, characterized in that, The total loss function described in step 6 is: ； in, For the total loss, For mean square error loss, As a trend curvature constraint, Due to the constraint of energy conservation, For non-equilibrium physical constraints, Due to terrain physical constraints, These are the weighting coefficients.

7. The high-speed rail second-level wind speed prediction and early warning method according to claim 6, characterized in that, The trend curvature constraint is defined by the following formula: ； The energy conservation constraint is defined by the following formula: ； The non-equilibrium physical constraint is defined by the following formula: ； The terrain physical constraints are defined by the following formula: ； in, To predict sequence length, The maximum kinetic energy increment allowed by physics. It is a linear rectified activation function. Vertical height For friction speed, For von Kármán's constant, For stability function, This is the length of Obukhov.

8. The high-speed rail second-level wind speed prediction and early warning method according to claim 6, characterized in that, In step 7, the wind speed baseline is predicted based on the GRU network. ,include: Calculate the dynamic envelope width The formula is: ； in, The terrain roughness index, The coefficient of variation of wind speed. For the train's real-time speed, For atmospheric stability parameters, The learnable coefficient; generate Upper envelope of time and lower envelope : ； ； in, This is the predicted value for the wind speed baseline. For standard normal distribution Quantiles Uncertainty in predictions for a physics-guided hierarchical state-space model. The vehicle-wind coupling coefficient is given.

Images