CN-122021483-A - High-speed rail second-level wind speed prediction and early warning method based on wind process dynamic division and physical guidance

CN122021483ACN 122021483 ACN122021483 ACN 122021483ACN-122021483-A

Abstract

The invention discloses a high-speed rail second-level wind speed prediction and early warning method based on wind process dynamic division and physical guidance, which comprises the steps of off-line construction of a stage one wind process dynamic division and a knowledge base, dividing historical wind field data into wind process types with definite physical meaning through a graph attention network and self-adaptive spectral clustering, performing macro-, middle-and micro-scale decoupling deduction based on a physical guidance layering state space model in stage two, performing macro-, middle-and micro-scale decoupling deduction, ensuring that the prediction accords with a fluid mechanics rule through a multi-physical constraint loss function, and generating a dynamic safety envelope line by utilizing an extremum theory and real-time scene information in stage three, namely, performing on-line early warning of a dynamic self-adaptive envelope line, and realizing scene self-adaptive accurate grading early warning. The method can remarkably improve the accuracy of wind speed prediction along the high-speed rail and the reliability of early warning.

Inventors

HE YUEWEN
XIONG XIONG
WANG YIJIA
Dai Baolun
WANG SHIHAO
WANG YUYANG

Assignees

南京信息工程大学

Dates

Publication Date: 20260512
Application Date: 20260416

Claims (10)

1. A high-speed rail second-level wind speed prediction and early warning method based on wind process dynamic division and physical guidance is characterized by comprising the following steps: step 1, cutting historical long-sequence wind speed data, extracting morphological characteristics, complexity characteristics and environmental characteristics of each time sequence segment, and constructing a multidimensional characteristic vector; Step 2, constructing a physical association diagram based on the time sequence segment, wherein nodes are the time sequence segment, and the edge weight is defined by generalized physical distance; step 3, performing feature aggregation by using a graph attention network, determining an adaptive clustering number based on graph structure compactness and physical separability, dividing a wind process into a plurality of categories through spectral clustering, and constructing a wind process knowledge base offline; step 4, online calculating a real-time wind speed sequence feature vector, matching with a clustering center in a wind process knowledge base, and selecting parameters of a physical-guided hierarchical state space model; Step 5, performing macro, medium and micro three-scale decoupling modeling by adopting a physical guided layering state space model, and predicting wind speed; Step 6, training a layering state space model of physical guidance through a total loss function comprising trend curvature constraint, energy conservation constraint, unbalanced physical constraint and terrain constraint, and ensuring that a wind speed prediction result accords with a physical rule; Step 7, predicting a wind speed datum line based on the GRU network, and dynamically calculating the width of an envelope curve by combining the topographic roughness, the wind speed variation coefficient, the train speed and the atmospheric stability to generate an upper envelope curve and a lower envelope curve; and step 8, correcting according to whether the wind speed predicted value exceeds the envelope curve, calculating continuous risk scores, realizing grading early warning, and outputting risk attribution analysis.
2. The method for predicting and pre-warning the second-level wind speed of the high-speed rail according to claim 1, wherein the multidimensional feature vector in the step 1 is: ; Wherein, the Is a segment Is marked with a reference to the multidimensional feature vector of (2) Representing a transpose; Morphological features ; For the average wind speed, Is the standard deviation of the two-dimensional image, As a coefficient of the trend the number of the trend, Is the standard deviation of pulsation; Complexity features By approximating entropy The representation is made of a combination of a first and a second color, ; For the pattern length it is possible that, For a similar tolerance to be achieved, For the length of the sequence, For all lengths in the wind speed time sequence segment Is less than the distance between the vector templates of the (a) and other vector templates with the same length Probability average of (2); Environmental features ; For the purpose of a pressure gradient, In order to provide a topographical roughness coefficient, Is the included angle between the wind direction and the line, Is a temperature gradient.
3. The method for predicting and pre-warning the second-level wind speed of the high-speed rail according to claim 1, wherein the generalized physical distance in the step 2 is as follows: ; Wherein, the For a dynamic time-warping distance, As a feature vector of the physical environment, Is a weight coefficient and , As a gaussian kernel bandwidth parameter, Is a node And Is a broad physical distance of (c).
4. The method for predicting and pre-warning high-speed rail second-level wind speed according to claim 3, wherein in the graph attention network in step 3, an attention coefficient formula is as follows: ; Wherein, the For a linear rectifying activation function with leakage, As a learnable weight vector for the attention mechanism, As a matrix of weights that can be learned, Is a node With its neighbors Is used to determine the initial feature vector of (a), Is a node With its neighbors Is a concentration factor of (2); Obtaining a node Is an aggregate representation of (2) The formula is: ; Wherein, the Is a node Is a set of first-order neighbor nodes of (a), Is a node Attention coefficients with its first order neighbor node n; the method is a Sigmoid activation function and is used for realizing nonlinear mapping and normalization of the features; The adaptive clustering number The determination formula is: ; Wherein, the For the class index parameter to be a class index parameter, For the similarity of the physical characteristics between classes, Is the similarity of the morphology in the class, For the physical distribution Jensen-Shannon divergence, In order for the coefficient of balance to be present, Taking the value of the independent variable which enables the objective function to take the maximum value; calculating a similarity matrix And construct the Laplace matrix , For the degree matrix, before selecting K-means clustering is carried out on the feature vectors.
5. The method for predicting and pre-warning second-level wind speed of high-speed rail according to claim 1, wherein in step 4, a real-time wind speed sequence feature vector is calculated And cluster center Mahalanobis distance of (v) Selecting the hierarchical state space model parameters of the category activation physical guidance corresponding to the minimum distance, wherein, Expressed as mahalanobis distance.
6. The method for predicting and pre-warning the second-level wind speed of high-speed rail according to claim 1, wherein in the step 5, the predicted value of the wind speed is the sum of macroscopic, mesoscopic and microscopic contributions: ; Wherein, the 、、 Respectively represents macroscopic, mesoscopic and microscopic scale speeds, As a predicted value of the wind speed, For the current moment of time, For a predicted time step; The macroscopic scale evolution equation is: ; Wherein, the In the macroscopic state of the body, In order to achieve an air density of the air, For the purpose of the gradient operator, Is the air pressure of the air, and the air pressure is the air pressure, As a result of the coriolis force parameter, For the background wind speed, For the turbulent viscosity coefficient, Is macro-scale system noise; Mesoscale evolution equation is: ; Wherein, the In the mesoscopic state, the device is in a mesoscopic state, For the spatial coordinates along the wind direction, For the topographical modulation coefficients, In the form of a topographical gradient, Is mesoscale system noise; The microscale evolution equation is: ; Wherein, the As a mean value of the direction (direction), In order to return to the rate of speed, In order for the intensity of the fluctuations to be uniform, In order to be a standard wiener process, In the case of a differential operator, Representing random turbulence impact.
7. The method for predicting and pre-warning the second-level wind speed of high-speed rail according to claim 6, wherein the total loss function in step 6 is: ; Wherein, the In order to account for the total loss, In order to be a mean square error loss, For the trend curvature constraint to be a trend, As a constraint on the conservation of energy, In the event of an unbalanced physical constraint, In order to be a physical constraint of the terrain, Is a weight coefficient.
8. The method for predicting and pre-warning high-speed rail second-level wind speed according to claim 7, wherein the trend curvature constraint is as follows: ; the energy conservation constraint is as follows: ; The unbalanced physical constraint is expressed as: ; The terrain physical constraint is as follows: ; Wherein, the In order to predict the length of the sequence, For the maximum kinetic energy increment physically allowed, For a linear rectification activation function, Is the height in the vertical direction, and the height is equal to the height in the vertical direction, In order to achieve a friction speed, the friction speed, Is a von willebrand constant, As a function of the degree of stability, Is Obukhov length.
9. The method for predicting and pre-warning second-level wind speed of high-speed rail of claim 7, wherein in step 7, a wind speed datum line is predicted based on a GRU network Comprising: calculating dynamic envelope width The formula is: ; Wherein, the For the terrain roughness index, As the coefficient of variation of the wind speed, For the real-time speed of the train, As a parameter of the stability of the atmosphere, Is a learnable coefficient; Generating Upper envelope of time of day And lower envelope curve : ; ; Wherein, the For the wind speed baseline prediction value, Is normally distributed in standard The number of the sub-digits is calculated, Uncertainty of predictions for the physically guided hierarchical state space model, Is the car-wind coupling coefficient.
10. The method for predicting and pre-warning a second-level wind speed of a high-speed rail according to claim 1, wherein in the step 8, the continuous risk score is given by the formula: ; Wherein, the For the final pre-warning wind speed value, And The upper and lower envelope lines are respectively arranged, As a function of the time sensitivity coefficient, As a terrain risk weight, As a risk weight for the vehicle speed, Scoring a value for continuous risk; If the predicted value exceeds the envelope range, the predicted value is corrected as According to And determining an early warning level, and outputting auxiliary decision information, wherein the auxiliary decision information comprises a predicted wind speed value, an uncertainty interval, a risk score and a risk contribution degree.

Description

High-speed rail second-level wind speed prediction and early warning method based on wind process dynamic division and physical guidance Technical Field The invention relates to the technical field of high-speed railway safety early warning and wind speed prediction, in particular to a high-speed railway second-level wind speed prediction and early warning method based on wind process dynamic division and physical guidance. Background The operation safety of the high-speed railway is highly dependent on real-time sensing and early warning of the wind environment along the line. The existing high-speed rail windproof systems are mainly used for alarming based on fixed thresholds, and lack of prediction capability under deep excavation and physical constraint on wind speed time sequence dynamic evolution rules, so that early warning false alarm rate is high and response is lagged. Particularly, under complex terrains and abrupt wind fields, the traditional method is difficult to realize accurate second-level wind speed prediction and scene self-adaptive risk early warning. The existing wind speed prediction method is mainly focused on a statistical learning or deep learning model, although the average performance is improved, the physical interpretability is generally lacking, and the fluid mechanics rule and the topography effect cannot be effectively fused, so that the prediction deviation is larger in extreme weather or complex scenes. In addition, the existing early warning method mostly adopts a static threshold value, the dynamic coupling influence of the train running state, the terrain change and the meteorological background is not considered, and accurate hierarchical early warning is difficult to realize. Therefore, a method for predicting and early warning wind speed, which can integrate physical laws, adapt to dynamic scenes and has high interpretation and high precision, is needed to improve the running safety and the early warning intellectualization level of the high-speed rail. Disclosure of Invention Aiming at the problems of low prediction precision, poor physical consistency, insufficient early warning adaptability and the like in the existing early warning of the wind speed of the high-speed rail, the invention provides a high-speed rail second-level wind speed prediction and early warning method based on the dynamic division and physical guidance of the wind process, and the intelligent division of the wind process, the hierarchical prediction and the scene self-adaption dynamic early warning of the physical guidance are realized through a three-stage frame of wind identification, wind calculation and police conduct. The invention adopts the following technical scheme that the high-speed rail second-level wind speed prediction and early warning method based on wind process dynamic division and physical guidance comprises the following steps: step 1, cutting historical long-sequence wind speed data, extracting morphological characteristics, complexity characteristics and environmental characteristics of each time sequence segment, and constructing a multidimensional characteristic vector; Step 2, constructing a physical association diagram based on the time sequence segment, wherein nodes are the time sequence segment, and the edge weight is defined by generalized physical distance; step 3, performing feature aggregation by using a graph attention network, determining an adaptive clustering number based on graph structure compactness and physical separability, dividing a wind process into a plurality of categories through spectral clustering, and constructing a wind process knowledge base offline; step 4, online calculating a real-time wind speed sequence feature vector, matching with a clustering center in a wind process knowledge base, and selecting parameters of a physical-guided hierarchical state space model; Step 5, performing macro, medium and micro three-scale decoupling modeling by adopting a physical guided layering state space model, and predicting wind speed; Step 6, training a layering state space model of physical guidance through a total loss function comprising trend curvature constraint, energy conservation constraint, unbalanced physical constraint and terrain constraint, and ensuring that a wind speed prediction result accords with a physical rule; Step 7, predicting a wind speed datum line based on the GRU network, and dynamically calculating the width of an envelope curve by combining the topographic roughness, the wind speed variation coefficient, the train speed and the atmospheric stability to generate an upper envelope curve and a lower envelope curve; and step 8, correcting according to whether the wind speed predicted value exceeds the envelope curve, calculating continuous risk scores, realizing grading early warning, and outputting risk attribution analysis. Preferably, the multidimensional feature vector in step 1 is: ; Wherein, the Is a segmentIs marked wit