System and method for predicting gale along high-speed rail

Abstract

The invention discloses a high-speed rail line gale prediction system and method, and relates to the technical field of high-speed rail gale prediction. The system for predicting the gale along the high-speed rail comprises a multi-scale grouping representation module, a comprehensive prediction module and a data enhancement module, the multi-scale grouping representation module is used for generating a wind speed preliminary prediction value; the comprehensive prediction module is used for generating a final wind speed prediction value; the data enhancement module expands the diversity of the original wind speed data set. Based on the high-speed rail line gale prediction system, the high-speed rail line gale prediction method comprises the following steps: grouping multi-scale data; carrying out multi-dimensional attention feature fusion; calculating a preliminary predicted value; calculating a final predicted value; carrying out combined optimization training; expanding an original wind speed sample; and performing consensus enhancement training. The method has the advantages that through collaborative design of multi-scale feature fusion, dynamic adaptive weighting and unsupervised enhanced training, the three core problems of sudden wind speed report missing, noise sensitivity and marking dependence in high-speed rail gale prediction are solved, and the operation efficiency and safety of high-speed rail lines are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of high-speed railway gale prediction, and in particular to a gale prediction system and method along a high-speed railway line. Background Art

[0002] As an integral part of the modern transportation backbone network, the safety and operational efficiency of high-speed rail lines are directly linked to socioeconomic benefits. However, sudden strong winds (instantaneous wind speeds ≥ 15 m / s) in complex meteorological environments exhibit significant temporal and spatial randomness and nonlinear characteristics, posing a serious threat to train operations. Existing high-speed rail gale prevention and control mechanisms primarily rely on real-time wind speed data from fixed monitoring points, employing a threshold-triggered passive response model. This presents the following key challenges:

[0003] 1. Braking delay risk: Threshold-based passive response requires waiting for wind speeds to reach dangerous levels before initiating deceleration. This can result in trains being unable to brake in time in dangerous areas, posing a safety hazard.

[0004] 2. Inefficient speed limits: Conservative standards for lifting speed limits often result in excessively long speed limit periods in sections, reducing line capacity (statistics show that redundant speed limit periods can account for 15%-20% of total operating time).

[0005] 3. Insufficient prediction accuracy: Traditional statistical models (such as ARIMA and wind speed propagation models) have difficulty capturing the long-range dependence and non-stationary characteristics of wind speed series, especially their poor robustness to noisy data, and the prediction error rate generally exceeds 20%.

[0006] Although existing deep learning technologies have performed well in time series forecasting, they face the following unique challenges in high-speed rail scenarios:

[0007] 1. Multi-scale temporal dependence: The evolution of strong winds involves minute-level mutations and hour-level trends, requiring simultaneous modeling of local details and global patterns;

[0008] 2. Data noise interference: Sensor data is often contaminated by environmental noise (such as mechanical vibration and electromagnetic interference), which affects the model's generalization ability;

[0009] 3. Scarcity of labeled data: The noise samples generated by enhancement lack real labels, making it difficult for traditional supervised learning to effectively utilize them.

[0010] Based on the above reasons, it is necessary to study a high-speed rail gale prediction system and method that can integrate multi-scale time series features, resist noise interference, and does not rely on enhanced sample annotation. Summary of the Invention

[0011] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide a high-speed rail line gale prediction system and prediction method, which solves the problems in the existing technology of being unable to effectively capture the multi-scale time series dependence of wind speed, weak anti-noise ability, and relying on enhanced sample annotation.

[0012] The technical solution of the present invention is: a high-speed rail line gale prediction system, including a multi-scale grouping representation module, a comprehensive prediction module and a data enhancement module; the multi-scale grouping representation module is used to extract the multi-scale temporal dependency of historical wind speed sequences and generate preliminary prediction values ​​of future wind speeds; the comprehensive prediction module is used to aggregate the preliminary prediction values ​​to generate final wind speed prediction values; the data enhancement module expands the scale and diversity of the original wind speed data set by generating synthetic data, thereby improving the model's generalization ability and anti-interference ability in real complex environments.

[0013] A further technical solution of the present invention is: the multi-scale grouping representation module includes a multi-scale grouping module, a fusion attention module and a multi-scale prediction module; the multi-scale grouping module is used to divide the original wind speed sequence into multiple sub-blocks according to different time window lengths to obtain a sub-block set; the fusion attention module uses an intra-block attention mechanism and an inter-block attention mechanism to process the divided sub-blocks, and then fuses the processing results; the multi-scale prediction module is used to input the fusion representations of different scales into an independent multi-layer perception machine to calculate and obtain a preliminary prediction value of the future wind speed.

[0014] A further technical solution of the present invention is: the comprehensive prediction module includes a dynamic weighted fusion module, an aggregation result optimization module and a joint loss function module; the dynamic weighted fusion module dynamically assigns weights according to the importance of preliminary prediction values ​​at different scales to achieve the fusion of multi-scale information; the aggregation result optimization module optimizes the aggregation results through nonlinear transformation to generate the final prediction value of future wind speed; the joint loss function module combines full supervision loss and consistency loss constraints to improve the stability and generalization ability of the model.

[0015] A further technical solution of the present invention is: the data enhancement module includes a time warping module, an interpolation enhancement module, a noise injection module and a consensus enhancement training module; the time warping module simulates the acceleration or deceleration process of wind speed evolution by adjusting the rhythm of time samples; the interpolation enhancement module is used to insert virtual sampling points between original data points to enrich the detailed features of the samples; the noise injection module is used to add random noise to the original data to simulate sensor measurement errors or environmental interference; the consensus enhancement training module forces each enhanced prediction value to approach the mean of all enhanced prediction values ​​through consistency constraints to achieve unsupervised optimization.

[0016] The technical solution of the present invention is: a method for predicting high winds along high-speed railway lines, based on a high wind prediction system along high-speed railway lines, and the steps are as follows:

[0017] S01, multi-scale data grouping: Use the multi-scale grouping module to group the original wind speed sequence S = {s1, s2, ..., s T} is divided into multiple sub-blocks according to different time window lengths, and the sub-block set {B1, B2, ..., B n}, the number of time steps of each sub-block is L i Determined by the window length corresponding to the i-th scale;

[0018] S02, multi-dimensional attention feature fusion: the fusion attention module is used to process the divided sub-blocks using the intra-block attention mechanism and the inter-block attention mechanism, and then the processing results are fused; the intra-block attention mechanism is used for each sub-block B i Calculate the self-attention weight matrix , output block representation ; Among them, the inter-block attention mechanism is used to calculate the global correlation matrix across sub-blocks , output inter-block representation ; Intra-block representation and inter-block representation Generate fused representation by element-wise addition ;

[0019] S03, calculate the preliminary prediction value: use a multi-scale prediction module to represent the fusion of different scales [p1, p2, ..., p n ] Input the independent multi-layer perception machine to calculate the preliminary prediction value of future wind speed [x1, x2, ..., x n ];

[0020] S04, calculate the final prediction value: use the attention coefficient generator in the dynamic weighted fusion module , is the initial predicted value x of each scale i Generate attention weights ; Then the initial predicted value x of each scale i The corresponding weight Multiply and sum to get the aggregate result ; Aggregate the results The input aggregation result optimization module performs nonlinear transformation and outputs the final predicted value x of future wind speed;

[0021] S05, joint optimization training: Calculate the mean square error between the final prediction value x and the true label y based on the full supervision loss term in the joint loss function, and calculate the preliminary prediction value x at each scale based on the consistency loss term in the joint loss function. i The difference between the predicted value and the final predicted value x; finally, the two are weighted and fused to balance the constraint effects of the two;

[0022] S06, expand the original wind speed sample: perform the following three operations in no particular order:

[0023] A. Adjust the rhythm of time samples based on the time warping module to simulate the acceleration or deceleration process of wind speed evolution; the time warping module first randomly selects a scaling factor for the original wind speed sample , generate a new sample length ; Then, based on the data points of the original wind speed sample, interpolate according to the new time axis to generate a regularized enhanced sample ; Enhanced samples Input to the multi-scale grouping module and share the model with the original wind speed series;

[0024] B. Based on the interpolation enhancement module, virtual sampling points are inserted between the original data points to enrich the detailed features of the sample; the interpolation enhancement module first randomly selects a subinterval in the original wind speed sample, and then inserts n new data points in the subinterval to generate a smooth enhanced sample ; Enhanced samples Input to the multi-scale grouping module and share the model with the original wind speed series;

[0025] C. The noise injection module is used to add random noise to the original data to simulate sensor measurement errors or environmental interference; the noise injection module first generates random noise from a Gaussian distribution. Sampling noise, is the standard deviation of the original sample, and the enhanced sample is generated ; Enhanced samples Input to the multi-scale grouping module and share the model with the original wind speed series;

[0026] S07, consensus enhancement training: First, the enhancement training module is used to generate m enhanced samples for each original sample S , then input all enhanced samples into the model to obtain enhanced prediction values Collection , calculate the mean of the enhanced prediction values ,Then a consistency loss function is used to force all enhanced prediction values ​​to converge to the group mean, achieving unsupervised optimization.

[0027] Compared with the prior art, the present invention has the following advantages:

[0028] 1. Through the collaborative design of multi-scale feature fusion, dynamic adaptive weighting, and unsupervised reinforcement training, it solves the three core problems of sudden wind speed omission, noise sensitivity, and labeling dependence in high-speed rail gale forecasting. It can support the triggering of train speed limit instructions 10-15 minutes in advance, reduce the frequency of emergency braking by more than 50%, and directly improve the operational efficiency and safety of high-speed rail lines.

[0029] 2. Multi-scale Time Series Modeling: A multi-scale grouping strategy simultaneously captures both transient wind speed fluctuations (such as gusts) and long-term trends (such as sustained high winds), addressing the inadequacy of traditional methods in modeling nonlinear time series dependencies and reducing the prediction error rate by over 30%.

[0030] 3. Dynamic fusion and adaptive optimization: The dynamic weighted fusion module based on the attention mechanism automatically learns the importance weights of prediction results at different time scales, adapting to complex wind speed scenarios (such as strong winds in mountainous areas and gentle winds on plains), and improving the robustness of prediction results in scenarios with diverse spatial distributions.

[0031] 4. Noise resistance and strong generalization capabilities: The data augmentation module simulates real-world interference (such as sensor errors and sudden weather events) through time warping, interpolation, and noise injection. Combined with the unsupervised consistency constraints of consensus-enhanced training, this module improves the model's predictive stability by 40% under noisy data.

[0032] 5. End-to-end joint optimization: A joint loss function (full supervision loss + multi-scale consistency loss) balances prediction accuracy and model stability, preventing the impact of single-scale prediction bias on the overall results. The underreporting rate for extreme wind speed scenarios (≥20m / s) is reduced to less than 5%.

[0033] 6. Low dependency on labeled data: The data augmentation module generates synthetic samples without real labels. It achieves unsupervised optimization through consensus training, solving the pain points of high cost and incomplete coverage of meteorological data annotation along high-speed rail lines.

[0034] The advantages of the multi-scale grouping representation module are as follows:

[0035] 1. Refined feature extraction: Divide the wind speed series into different time windows (5 minutes, 30 minutes, and 1 hour), modeling local fluctuations and global trends separately, addressing the problem of delayed response to sudden strong winds in traditional single-scale models.

[0036] 2. Cross-scale attention fusion: Intra-block attention focuses on key time points in sub-blocks (such as when wind speed suddenly increases), while inter-block attention captures cross-scale correlations (such as the impact of minute-level mutations on hour-level trends), reducing feature representation error by 25%.

[0037] 3. Independent multi-scale prediction: An independent MLP prediction head is assigned to each time scale to avoid multi-task interference. The preliminary prediction results can be used as an explanatory indicator of local wind speed evolution.

[0038] The advantages of the comprehensive prediction module are as follows:

[0039] 1. Dynamic weight allocation: A dynamic weighting strategy based on the attention coefficient generator automatically strengthens the dominant scale (for example, focusing on minute-level predictions during sudden strong winds) and weakens the contribution of noise scales, reducing the adaptive error rate of weight allocation by 18%.

[0040] 2. Nonlinear result optimization: The aggregation result optimization module corrects the fusion deviation through MLP nonlinear transformation, solving the problem that linear weighting is unable to express complex patterns. The extreme wind speed prediction deviation is controlled within ±2m / s.

[0041] 3. Multi-scale consistency constraint: The joint loss function forces the initial prediction value to converge with the final prediction result, suppressing overfitting at a single scale and improving the model's generalization ability by 35% under sparse data.

[0042] The advantages of the data enhancement module are as follows:

[0043] 1. Physical logic-driven data expansion: Time regularization simulates changes in wind speed evolution rate, interpolation enhancement captures microscopic fluctuation patterns, noise injection restores sensor errors, and enhanced sample coverage exceeds 95% of real interference scenarios.

[0044] 2. Unlabeled consensus training: By forcing the predicted values ​​to converge toward the group mean (Formula 7), the model can be optimized for interference resistance without augmenting the true labels of the samples, reducing the amount of labeled data required by 60%.

[0045] 3. Multimodal Enhancement Collaboration: Three enhancement strategies, namely time warping, interpolation, and noise injection, generate complementary samples in parallel, overcoming the limitations of a single enhancement method (such as simply adding noise). This reduces the model's generalization error by 22% when deployed across regions.

[0046] The present invention is further described below with reference to the figures and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a connection relationship diagram of each module in the present invention. DETAILED DESCRIPTION

[0048] Example 1:

[0049] like Figure 1 As shown in the figure, the high-speed rail line strong wind prediction system includes a multi-scale grouping characterization module, a comprehensive prediction module and a data enhancement module.

[0050] The multi-scale grouping representation module is used to extract the multi-scale temporal dependencies of the historical wind speed series and generate preliminary predictions of future wind speeds. The multi-scale grouping representation module includes a multi-scale grouping module, a fusion attention module, and a multi-scale prediction module.

[0051] The multi-scale grouping module is used to transform the original wind speed sequence S = {s1, s2, ..., s T} is divided into multiple sub-blocks according to different time window lengths, and the sub-block set {B1, B2, ..., B n}, the number of time steps of each sub-block is L i It is determined by the window length corresponding to the i-th scale (for example, if the first scale is 5 minutes (data sampling interval is 1 minute), then L1=5; if the second scale is 30 minutes (data sampling interval is 1 minute), then L2=30).

[0052] The fusion attention module uses the intra-block attention mechanism and the inter-block attention mechanism to process the divided sub-blocks, and then fuses the processing results. Among them, the intra-block attention mechanism is used for each sub-block B. i Calculate the self-attention weight matrix (The superscript i indicates that the weight matrix belongs to the sub-block of the i-th scale. For example, if there are three scales of 5min, 30min, and 1h, then each scale sub-block will generate an independent intra-block attention matrix. 、 、 ), focusing on local key time points and outputting intra-block representations (The superscript i represents the sub-block of the i-th scale). The inter-block attention mechanism is used to calculate the global correlation matrix across sub-blocks. , identify pattern associations between different scales and output inter-block representation (Superscript i represents the sub-block of the i-th scale); where the intra-block representation and inter-block representation Generate fused representation by element-wise addition .

[0053] The multi-scale prediction module is used to represent the fusion of different scales [p1, p2, ..., p n ] Input the independent multi-layer perception machine (each layer of perception head contains 3 fully connected layers, the activation function is LeakyReLU), and calculate the preliminary prediction value of the future wind speed [x1, x2, ..., x n ].

[0054] The comprehensive prediction module is used to aggregate the preliminary prediction values ​​to generate the final wind speed prediction value. The comprehensive prediction module includes a dynamic weighted fusion module, an aggregation result optimization module, and a joint loss function module.

[0055] The dynamic weighted fusion module dynamically assigns weights based on the importance of the initial prediction values ​​at different scales to achieve optimal fusion of multi-scale information. The dynamic weighted fusion module designs the attention coefficient generator. (Using two fully connected layers, the activation function is LeakyReLU), which is the initial prediction value x for each scale i Generate attention weights ; Then the initial predicted value x of each scale i The corresponding weight Multiply and sum to get the aggregate result (For example, there are three scales: 5min, 30min and 1h, with weights of 0.4, 0.3 and 0.3 respectively, then the aggregation result ).

[0056] The aggregation result optimization module optimizes the aggregation results through nonlinear transformation , generating the final predicted value x for future wind speed. The aggregation result optimization module designs a global detection head. The global detection head uses a multi-layer perception machine, consisting of 2-3 fully connected layers, with the last layer being a linear layer that directly outputs the final predicted value. Each layer is followed by a LeakyReLU activation function and a Dropout layer (dropout rate 0.2) to prevent overfitting.

[0057] The joint loss function module combines full supervision loss and consistency loss constraints to improve the stability and generalization ability of the model. The joint loss function module contains a joint loss function, in which the full supervision loss is used to calculate the mean square error between the final predicted value x and the true label y, and the consistency loss is used to constrain the initial predicted value x at each scale. i The difference between the final predicted value x and the initial predicted value x is used to reduce the discreteness of the prediction result. Finally, the two are weighted and fused to balance the constraint effects of the two. The data flow transmission path in the joint loss function module includes two types: forward propagation and back propagation. In the forward propagation stage, the three input items are "final predicted value x, preliminary predicted value x i The sources of the real label y are respectively “aggregation result optimization module, multi-scale prediction module and external supervision signal”, and combined with x, x i and y, calculate the joint loss. In the back propagation stage, through the full supervision loss term The gradient update aggregation result optimization module (Multi-layer Perceptron (MLP) parameters) and the dynamic weighted fusion module (Attention coefficient generator parameters), through the consistency loss term The gradient updates the multi-scale prediction module (multi-layer perception machine (MLP) parameters at each scale) and the fusion attention module (intra-block and inter-block attention parameters).

[0058] The data augmentation module expands the size and diversity of the original wind speed dataset by generating synthetic data, improving the model's generalization and interference resistance in real-world complex environments. The data augmentation module includes a time warping module, an interpolation augmentation module, a noise injection module, and a consensus augmentation training module.

[0059] The time warping module simulates the acceleration or deceleration process of wind speed evolution by adjusting the rhythm of time samples. The time warping module first calculates the original wind speed samples (i.e., the original wind speed sequence) S = {s1, s2, ..., s T}, randomly select the scaling factor , The value range is 0.8-1.2, generating a new sample length ; Then, based on the data points of the original wind speed sample, interpolate according to the new time axis to generate a regularized enhanced sample Enhanced Sample Input to the multi-scale grouping module (sharing the model with the original wind speed series) for subsequent processing. For example, if The original 60-minute sample is compressed into 54 minutes to simulate the scene of sudden increase in wind speed. ,The original 60-min sample was expanded to 66 minutes to simulate the scenario of gentle changes in wind speed.

[0060] The interpolation enhancement module is used to insert virtual sampling points between the original data points (i.e. the data points in the original wind speed sample) to enrich the detailed features of the sample in order to capture the microscopic pattern of wind speed changes and alleviate the problem of data scarcity. The interpolation enhancement module first randomly selects a subinterval in the original sample. , then insert n new data points into the subinterval to generate smooth enhanced samples Enhanced Sample The data is fed into the multi-scale grouping module (which shares the same model as the original wind speed series) for subsequent processing. For example, inserting four new data points into the original data at 5-minute intervals generates a high-density sample at 1-minute intervals.

[0061] The noise injection module is used to add random noise to the original data to simulate sensor measurement errors or environmental interference. The noise injection module first Sampling noise, is the standard deviation of the original sample, and the enhanced sample is generated Enhanced Sample Input to the multi-scale grouping module (sharing the model with the original wind speed sequence) and then perform subsequent processing. For example, if the original wind speed mean is 15m / s and the standard deviation is , the noise intensity is 0.2m / s.

[0062] The consensus enhancement training module uses consistency constraints to force each enhanced prediction value to approach the mean of all enhanced prediction values, achieving unsupervised optimization. The enhancement training module first generates m enhanced samples for each original sample S , then input all enhanced samples into the model to obtain a set of enhanced prediction values , calculate the mean of the enhanced prediction values (i.e. the mean of all enhanced training values), and then use the consistency loss function to force the predicted value of each enhanced sample to approach the mean of the enhanced predicted value (That is, forcing the prediction results of all enhanced samples to converge to the group mean.) The consensus enhancement training module connects the time warping module, difference enhancement module, noise injection module, multi-scale grouping representation module and comprehensive prediction module through multi-segment data flow and gradient back propagation to achieve unsupervised consistency optimization. The data flow transmission path of the consensus enhancement training module includes forward propagation and back propagation. In the forward propagation stage, all enhanced prediction values ​​are collected from the aggregation result optimization module. , first calculate the mean of the enhanced prediction value , and then calculate the consistency loss (based on Formula 7). In the back propagation stage, the consistency loss gradient updates the dynamic weighted fusion module (attention coefficient generator parameters), aggregation result optimization module (multi-layer perceptron (MLP) parameters), multi-scale prediction module (multi-layer perceptron (MLP) parameters at each scale) and fusion attention module (intra-block and inter-block attention parameters).

[0063] A method for predicting high winds along a high-speed railway line, based on the high-speed railway line high wind prediction system, comprises the following steps:

[0064] S01, multi-scale data grouping: Use the multi-scale grouping module to group the original wind speed sequence S = {s1, s2, ..., s T} is divided into multiple sub-blocks according to different time window lengths, and the sub-block set {B1, B2, ..., B n}, the number of time steps of each sub-block is L i It is determined by the window length corresponding to the i-th scale.

[0065] S02, multi-dimensional attention feature fusion: the fusion attention module is used to process the divided sub-blocks using the intra-block attention mechanism and the inter-block attention mechanism, and then the processing results are fused; the intra-block attention mechanism is used for each sub-block B i Calculate the self-attention weight matrix , output block representation ; Among them, the inter-block attention mechanism is used to calculate the global correlation matrix across sub-blocks , output inter-block representation ; Intra-block representation and inter-block representation Generate fused representation by element-wise addition .

[0066] In this step, the fusion representation The calculation of is shown in formula 1; in step S03, the preliminary prediction value of the i-th scale is The calculation of is shown in formula 2;

[0067] Formula 1: ;

[0068] Formula 2: ;

[0069] in, Represents the multi-layer perceptron corresponding to the sub-block of the i-th scale, and the subscript i represents the i-th scale.

[0070] S03, calculate the preliminary prediction value: use a multi-scale prediction module to represent the fusion of different scales [p1, p2, ..., p n ] Input the independent multi-layer perception machine to calculate the preliminary prediction value of future wind speed [x1, x2, ..., x n ].

[0071] S04, calculate the final prediction value: use the attention coefficient generator in the dynamic weighted fusion module , is the initial predicted value x of each scale i Generate attention weights ; Then the initial predicted value x of each scale i The corresponding weight Multiply and sum to get the aggregate result ; Aggregate the results The input aggregation result optimization module performs nonlinear transformation and outputs the final predicted value x of future wind speed.

[0072] In this step, the aggregation results The calculation formulas for and the final predicted value x are respectively shown in Formula 3-4;

[0073] Formula 3: ;

[0074] Formula 4: ;

[0075] in, is the initial predicted value of the i-th scale The corresponding attention weight, is the preliminary predicted value of the i-th scale; is the multi-layer perception machine corresponding to the global detection head; n is the number of sub-blocks of multi-scale grouping (i.e., the number of types of time scales).

[0076] S05, joint optimization training: Calculate the mean square error between the final prediction value x and the true label y based on the full supervision loss term in the joint loss function, and calculate the preliminary prediction value x at each scale based on the consistency loss term in the joint loss function. iThe difference between the final predicted value x and the final predicted value x; finally, the two are weighted and fused to balance the constraint effects of the two.

[0077] In this step, the joint loss function is shown in Formula 5;

[0078] Formula 5: ;

[0079] Where y is the true future wind speed label, and n is the number of sub-blocks in the multi-scale grouping (i.e., the number of time scale types); is the fully supervised loss term, which is used to force the final predicted value x to approach the true label y; is the consistency loss term, which is used to make the initial prediction value of each scale Approximate the final prediction value x and suppress the differences between multi-scale predictions.

[0080] S06, expand the original wind speed sample: perform the following three operations in no particular order:

[0081] A. Adjust the rhythm of time samples based on the time warping module to simulate the acceleration or deceleration process of wind speed evolution; the time warping module first randomly selects a scaling factor for the original wind speed sample , generate a new sample length ; Then, based on the data points of the original wind speed sample, interpolate according to the new time axis to generate a regularized enhanced sample ;

[0082] B. Based on the interpolation enhancement module, virtual sampling points are inserted between the original data points to enrich the detailed features of the sample; the interpolation enhancement module first randomly selects a subinterval in the original wind speed sample, and then inserts n new data points in the subinterval to generate a smooth enhanced sample ;

[0083] C. The noise injection module is used to add random noise to the original data to simulate sensor measurement errors or environmental interference; the noise injection module first generates random noise from a Gaussian distribution. Sampling noise, is the standard deviation of the original sample, and the enhanced sample is generated ;

[0084] S07, consensus enhancement training: First, the enhancement training module is used to generate m enhanced samples for each original sample S , then input all enhanced samples into the model to obtain enhanced prediction values Collection , calculate the mean of the enhanced prediction values ,Then a consistency loss function is used to force all enhanced prediction values ​​to converge to the group mean, achieving unsupervised optimization.

[0085] In this step, the mean of the enhanced prediction value is The calculation formula of is shown in Formula 6;

[0086] Formula 6: .

[0087] In this step, the consistency loss function is shown in Formula 7;

[0088] Formula 7: ;

[0089] Among them, m is the total number of enhanced samples, is the enhanced prediction value of the j-th enhanced sample; is a consistency term, which is used to force the enhanced prediction value of each enhanced sample to approach the mean of the enhanced prediction value , improving the model’s robustness to noise.

Claims

1. The high-speed rail line high wind forecast system is characterized by: It includes a multi-scale grouping representation module, a comprehensive prediction module and a data enhancement module; the multi-scale grouping representation module is used to extract the multi-scale temporal dependency of the historical wind speed sequence and generate a preliminary prediction value of the future wind speed; the comprehensive prediction module is used to aggregate the preliminary prediction values ​​to generate the final wind speed prediction value; The data enhancement module expands the scale and diversity of the original wind speed dataset by generating synthetic data, thereby improving the generalization ability and anti-interference ability of the model in real complex environments.

2. The high-speed rail line gale prediction system according to claim 1, wherein: The multi-scale grouping representation module includes a multi-scale grouping module, a fusion attention module and a multi-scale prediction module; the multi-scale grouping module is used to divide the original wind speed sequence into multiple sub-blocks according to different time window lengths to obtain a sub-block set; the fusion attention module uses the intra-block attention mechanism and the inter-block attention mechanism to process the divided sub-blocks, and then fuses the processing results; the multi-scale prediction module is used to input the fusion representations of different scales into an independent multi-layer perception machine to calculate the preliminary prediction value of the future wind speed.

3. The high-speed rail line gale prediction system according to claim 2, wherein: The comprehensive prediction module includes a dynamic weighted fusion module, an aggregation result optimization module and a joint loss function module; the dynamic weighted fusion module dynamically assigns weights according to the importance of preliminary prediction values ​​at different scales to achieve the fusion of multi-scale information; the aggregation result optimization module optimizes the aggregation results through nonlinear transformation to generate the final prediction value of future wind speed; the joint loss function module combines full supervision loss and consistency loss constraints to improve the stability and generalization ability of the model.

4. The high-speed rail line gale prediction system according to claim 3, wherein: The data enhancement module includes a time warping module, an interpolation enhancement module, a noise injection module and a consensus enhancement training module; the time warping module simulates the acceleration or deceleration process of wind speed evolution by adjusting the rhythm of time samples; the interpolation enhancement module is used to insert virtual sampling points between the original data points to enrich the detailed features of the samples; the noise injection module is used to add random noise to the original data to simulate sensor measurement errors or environmental interference; the consensus enhancement training module uses consistency constraints to force each enhanced prediction value to approach the mean of all enhanced prediction values, thereby achieving unsupervised optimization.

5. A method for predicting high winds along a high-speed railway line, based on the high wind prediction system along a high-speed railway line according to claim 4, characterized in that the steps as follows: S01, multi-scale data grouping: Use the multi-scale grouping module to group the original wind speed sequence S = {s1, s2, ..., s T } is divided into multiple sub-blocks according to different time window lengths, and the sub-block set {B1, B2, ..., B n }, the number of time steps of each sub-block is L i Determined by the window length corresponding to the i-th scale; S02, multi-dimensional attention feature fusion: the fusion attention module is used to process the divided sub-blocks using the intra-block attention mechanism and the inter-block attention mechanism, and then the processing results are fused; the intra-block attention mechanism is used for each sub-block B i Calculate the self-attention weight matrix , output block representation ; Among them, the inter-block attention mechanism is used to calculate the global correlation matrix across sub-blocks , output inter-block representation ; Intra-block representation and inter-block representation Generate fused representation by element-wise addition ; S03, calculate the preliminary prediction value: use a multi-scale prediction module to represent the fusion of different scales [p1, p2, ..., p n ] Input the independent multi-layer perception machine to calculate the preliminary prediction value of future wind speed [x1, x2, ..., x n ]; S04, calculate the final prediction value: use the attention coefficient generator in the dynamic weighted fusion module , is the initial predicted value x of each scale i Generate attention weights ; Then the initial predicted value x of each scale i The corresponding weight Multiply and sum to get the aggregate result ; Aggregate the results The input aggregation result optimization module performs nonlinear transformation and outputs the final predicted value x of future wind speed; S05, joint optimization training: Calculate the mean square error between the final prediction value x and the true label y based on the full supervision loss term in the joint loss function, and calculate the preliminary prediction value x at each scale based on the consistency loss term in the joint loss function. i The difference between the predicted value and the final predicted value x; finally, the two are weighted and fused to balance the constraint effects of the two; S06, expand the original wind speed sample: perform the following three operations in no particular order: A. Adjust the rhythm of time samples based on the time warping module to simulate the acceleration or deceleration process of wind speed evolution; the time warping module first randomly selects a scaling factor for the original wind speed sample , generate a new sample length ; Then, based on the data points of the original wind speed sample, interpolate according to the new time axis to generate a regularized enhanced sample ; Enhanced samples Input to the multi-scale grouping module and share the model with the original wind speed series; B. Based on the interpolation enhancement module, virtual sampling points are inserted between the original data points to enrich the detailed features of the sample; the interpolation enhancement module first randomly selects a subinterval in the original wind speed sample, and then inserts n new data points in the subinterval to generate a smooth enhanced sample ; Enhanced samples Input to the multi-scale grouping module and share the model with the original wind speed series; C. The noise injection module is used to add random noise to the original data to simulate sensor measurement errors or environmental interference; the noise injection module first generates random noise from a Gaussian distribution. Sampling noise, is the standard deviation of the original sample, and the enhanced sample is generated ; Enhanced samples Input to the multi-scale grouping module and share the model with the original wind speed series; S07, consensus enhancement training: First, the enhancement training module is used to generate m enhanced samples for each original sample S , then input all enhanced samples into the model to obtain enhanced prediction values Collection , calculate the mean of the enhanced prediction values ,Then a consistency loss function is used to force all enhanced prediction values ​​to converge to the group mean, achieving unsupervised optimization.

6. The method for predicting high winds along a high-speed railway as claimed in claim 5, wherein: In step 02, fusion representation The calculation of is shown in formula 1; In step S03, the initial predicted value of the i-th scale The calculation of is shown in formula 2; Formula 1: ; Formula 2: ; in, Represents the multi-layer perceptron corresponding to the sub-block of the i-th scale, and the subscript i represents the i-th scale.

7. The method for predicting high winds along a high-speed railway as claimed in claim 6, wherein: In step 04, aggregate the results The calculation formulas for and the final predicted value x are respectively shown in Formula 3-4; Formula 3: ; Formula 4: ; in, is the initial predicted value of the i-th scale The corresponding attention weight, is the preliminary predicted value of the i-th scale; is the multi-layer perception machine corresponding to the global detection head; n is the number of sub-blocks in the multi-scale grouping.

8. The method for predicting high winds along a high-speed railway according to claim 7, wherein: In step S05, the joint loss function is shown in Formula 5; Formula 5: ; Where y is the true future wind speed label, and n is the number of sub-blocks in the multi-scale grouping (i.e., the number of time scale types); is the fully supervised loss term, which is used to force the final predicted value x to approach the true label y; is the consistency loss term, which is used to make the initial prediction value of each scale Approximate the final predicted value x.

9. The method for predicting high winds along a high-speed railway as claimed in claim 8, wherein: In step 07, the mean of the predicted value is enhanced The calculation formula of is shown in Formula 6; Formula 6: .

10. The method for predicting high winds along a high-speed railway as claimed in claim 9, wherein: The loss function is shown in Formula 7; Formula 7: ; Among them, m is the total number of enhanced samples, is the enhanced prediction value of the j-th enhanced sample; is a consistency term, which is used to force the enhanced prediction value of each enhanced sample to approach the mean of the enhanced prediction value .

Images