Tracking operation safety protection method and system for virtual marshalling train

Abstract

The invention discloses a tracking operation safety protection method and system for a virtual marshalling train, and relates to the technical field of urban rail transit management. Comprising the following steps: acquiring a historical data feature set of a leading train and an actual acceleration of the leading train in a historical operation process of a virtual marshalling train; and inputting the historical data feature set into the acceleration prediction model to obtain the predicted acceleration of the leader train. Training the acceleration prediction model with the purpose of minimizing the error between the predicted acceleration of the leader train and the actual acceleration of the leader train to obtain a trained acceleration prediction model. And inputting the real-time data feature set into the trained acceleration prediction model to obtain a predicted acceleration and a predicted speed of the leader train at the next time step, and further obtaining a safety protection distance between the following train at the next time step and the leader train. According to the method, the accuracy of the safety protection distance can be remarkably improved, and safe and efficient operation of the virtual marshalling train is guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of urban rail transit management, and particularly to a method and system for tracking and running safety protection of virtual formation trains. Background Art

[0002] With the development of urban rail transit, the demand for residents' travel continues to climb, which poses a severe challenge to the operation organization of rail transit. Especially during the morning and evening rush hours, a large number of passengers gather at stations, resulting in continuous overload of trains. In recent years, the virtual formation operation organization mode is considered an effective means to solve or alleviate the tight transportation capacity in some line sections during peak periods. Based on technologies such as information interaction and automatic control via wireless communication, virtual formation trains form large formation trains by coupling multiple trains and can be decoupled into multiple independent running trains when needed, which can meet the complex and time-varying urban rail operation requirements. As a new technology, virtual formation trains require appropriate tracking and running safety protection methods to ensure their safe and efficient operation, which is a necessary condition for their wide application.

[0003] In the urban rail transit moving block system, the existing train running safety protection mainly adopts two methods: the absolute braking distance method and the relative braking distance method. When the train is running, the distance between trains must be greater than the calculated safety protection distance. In the absolute braking distance method, the safety protection distance is calculated assuming that the braking distance of the leading train is 0 and the following train is in the most unfavorable braking condition. In the relative braking distance method, the braking distance of the leading train is subtracted when calculating the safety protection distance, rather than assuming it to be 0.

[0004] The absolute braking distance method assumes that the braking distance of the preceding train is 0, that is, the following train calculates the safety protection distance only based on its own emergency braking guarantee rate. However, in virtual formation, the leading train and the following train achieve dynamic coupling through wireless communication and coordinated control, and the actual braking distance of the leading train cannot be 0, resulting in a smaller calculated safety protection distance, thus increasing the risk of rear-end collision. In complex working conditions such as slopes and curves, the train needs to decelerate, which will affect the braking distance and lead to inaccurate calculation of the safety protection distance in the relative braking distance method. It can be seen that the existing train running safety protection cannot accurately determine the safety protection distance between the leading train and the following train, resulting in an increased train running interval and reduced line capacity and operation efficiency of rail transit. Summary of the Invention

[0005] Based on this, it is necessary to provide a method and system for tracking and running safety protection of virtual formation trains in view of the above technical problems.

[0006] An embodiment of the present invention provides a method for tracking and running safety protection of virtual formation trains, including: Obtain the historical data feature set of the leading train and the actual acceleration of the leading train during the historical operation of the virtual formation train; Input the historical data feature set into the acceleration prediction model, which includes a time convolutional network, a long short-term memory network, and a multi-head self-attention mechanism connected in sequence; capture the long-distance time dependence relationship of the historical data feature set through the time convolutional network to obtain the temporal context features; model the temporal dynamic changes of the temporal context features through the long short-term memory network to generate temporal dynamic features; weight the temporal dynamic features through the multi-head self-attention mechanism to obtain the attention-weighted features; obtain the predicted acceleration of the leading train according to the attention-weighted features; train the acceleration prediction model with the goal of minimizing the error between the predicted acceleration of the leading train and the actual acceleration of the leading train to obtain the trained acceleration prediction model; Obtain the real-time data feature set of the leading train at the current time step during the tracking operation of the virtual formation train, input the real-time data feature set into the trained acceleration prediction model to obtain the predicted acceleration of the leading train at the next time step, and calculate the predicted speed of the leading train at the next time step according to the predicted acceleration; obtain the safety protection distance between the following train and the leading train at the next time step according to the predicted acceleration and predicted speed of the leading train at the next time step.

[0007] Optionally, the data feature set includes acceleration state data and line condition data; The acceleration state data includes: the train speed at the previous time step, the train acceleration at the previous time step, and the train traction and braking level at the previous time step; The line condition data includes: the line gradient value and the vertical curve radius of the line.

[0008] Optionally, obtaining the safety protection distance between the following train and the leading train at the next time step according to the predicted acceleration and predicted speed of the leading train at the next time step specifically includes: Determine the safety protection distance calculation formula through the following formula: ; where, is the safety protection distance; is the operating space requirement of the following train at the next time step; is the speed of the following train at time ; is the time step length; is the desired interval distance when adjacent trains stop at the platform; is the dynamic distance margin term of the following train at time ; Input the predicted acceleration and predicted speed of the leading train at the next time step into the safety protection distance calculation formula to obtain the safety protection distance between the following train and the leading train at the next time step.

[0009] Optionally, determine the dynamic distance margin term, specifically including: Determine the potential approaching distance margin term based on the following formula: ; ; ; Among them, is the running distance for the following train to adjust its speed with the maximum acceleration at the next time step; is the running distance for the leading train to adjust its speed with the predicted acceleration of the leading train at the next time step at the next time step; is the predicted acceleration of the leading train at the next time step; Determine the communication delay margin term based on the following formula: ; Among them, is the communication delay time of the following train at time , and its time-varying range is ; is the maximum communication delay time of the train; is the predicted speed of the leading train at the next time step; Determine the control delay margin term based on the following formula: ; Among them, is the control delay time of the following train at time , and its time-varying range is ; is the control delay time of the leading train at time , and its time-varying range is ; is the maximum control delay time of the train; Determine the dynamic distance margin term through the potential approaching distance margin term, communication delay margin term, and control delay margin term based on the following formula: ; Among them, is the potential approaching distance margin term, is the communication delay margin term, is the control delay margin term.

[0010] Optionally, the acceleration prediction model further includes an input layer, which normalizes the data feature set through the input layer to obtain preprocessed features, specifically including: Eliminate the data outliers in the data feature set to obtain a preprocessed data feature set; Perform global normalization on the preprocessed data feature set to scale the data interval of the preprocessed data feature set to [-1, 1], eliminate the dimensional differences between different features, and obtain preprocessed features.

[0011] Optionally, the variables affecting the train acceleration in the time-series dynamic features are weighted and fused through a multi-head self-attention mechanism to obtain attention-weighted features, specifically including: Reconstruct the time-series dynamic features into a vector form through a multi-head self-attention mechanism; Generate query, key, and value matrices in vector form through linear transformation, and respectively determine the attention weights of the query, key, and value matrices; Perform weighted fusion on the attention weights of the query, key, and value matrices to obtain attention-weighted features.

[0012] Optionally, the acceleration prediction model further includes an output layer, and obtaining the predicted acceleration of the leading train according to the attention-weighted features specifically includes: Perform denormalization on the attention-weighted features through the output layer to restore the attention-weighted features to acceleration values, and obtain the predicted acceleration of the leading train.

[0013] An embodiment of the present invention further provides a tracking operation safety protection system for a virtual formation train, including: A data acquisition module for acquiring the historical data feature set of the leading train and the actual acceleration of the leading train during the historical operation of the virtual formation train; A data processing module for inputting the historical data feature set into an acceleration prediction model, which includes a time convolutional network, a long short-term memory network, and a multi-head self-attention mechanism connected in sequence; capturing the long-distance time dependence relationship of the historical data feature set through the time convolutional network to obtain time-series context features; modeling the time-series dynamic changes of the time-series context features through the long short-term memory network to generate time-series dynamic features; weighting the time-series dynamic features through the multi-head self-attention mechanism to obtain attention-weighted features; obtaining the predicted acceleration of the leading train according to the attention-weighted features; training the acceleration prediction model with the goal of minimizing the error between the predicted acceleration of the leading train and the actual acceleration of the leading train to obtain a trained acceleration prediction model; The safety protection distance output module is used to obtain the real-time data feature set of the leading train at the current time step during the tracking operation of the virtual formation train, input the real-time data feature set into the trained acceleration prediction model to obtain the predicted acceleration of the leading train at the next time step, and calculate the predicted speed of the leading train at the next time step according to the predicted acceleration; according to the predicted acceleration and predicted speed of the leading train at the next time step, obtain the safety protection distance between the following train and the leading train at the next time step.

[0014] The above-mentioned method and system for tracking operation safety protection of virtual formation trains provided by the embodiments of the present invention have the following beneficial effects compared with the prior art: The acceleration prediction model proposed by the present invention captures long-distance time dependencies through a temporal convolutional network, models temporal dynamic changes using a long short-term memory network, and also strengthens the weight fusion of key variables through a multi-head self-attention mechanism, comprehensively capturing long-term dependencies, dynamic changes, and the influence of key variables in the time series, realizing high-precision prediction of the acceleration of the leading train; determining the safety protection distance based on the acceleration prediction result of the leading train can dynamically adapt to complex working conditions and changes in operating states, overcome the calculation deviations caused by the traditional method assuming the braking distance of the leading vehicle is 0 or not fully considering complex working conditions, significantly improve the accuracy of the safety protection distance, and ensure the safe and efficient operation of virtual formation trains. Description of the Drawings

[0015] Figure 1 It is a schematic flowchart of a method for tracking operation safety protection of virtual formation trains provided in an embodiment; Figure 2 It is a composition diagram of a hybrid TLMA model of a method for tracking operation safety protection of virtual formation trains provided in an embodiment; Figure 3 It is a curve graph of the acceleration of the leading train of a method for tracking operation safety protection of virtual formation trains provided in an embodiment; Figure 4 It is a curve graph of the speed of the leading train of a method for tracking operation safety protection of virtual formation trains provided in an embodiment; Figure 5 It is a schematic diagram of the train trajectory under the potential approaching distance margin term of a method for tracking operation safety protection of virtual formation trains provided in an embodiment; Figure 6 It is a comparison diagram of methods of a method for tracking operation safety protection of virtual formation trains provided in an embodiment. Detailed Embodiments

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

[0017] First Embodiment An embodiment of the present invention provides a method for tracking and running safety protection of a virtual formation train. Figure 1 The following shows a schematic flowchart of the method for tracking and running safety protection of the virtual formation train. As Figure 1 shown, it includes the following steps:

[0018] Obtain the historical data feature set of the leading train and the actual acceleration of the leading train during the historical operation of the virtual formation train; Input the historical data feature set into an acceleration prediction model. The acceleration prediction model includes a temporal convolutional network, a long short-term memory network, and a multi-head self-attention mechanism connected in sequence. Capture the long-distance temporal dependence relationship of the historical data feature set through the temporal convolutional network to obtain temporal context features. Model the temporal dynamic changes of the temporal context features through the long short-term memory network to generate temporal dynamic features. Weight the temporal dynamic features through the multi-head self-attention mechanism to obtain attention-weighted features. Obtain the predicted acceleration of the leading train according to the attention-weighted features. Train the acceleration prediction model with the goal of minimizing the error between the predicted acceleration of the leading train and the actual acceleration of the leading train to obtain the trained acceleration prediction model; Obtain the real-time data feature set of the leading train at the current time step during the tracking operation of the virtual formation train. Input the real-time data feature set into the trained acceleration prediction model to obtain the predicted acceleration of the leading train at the next time step, and calculate the predicted speed of the leading train at the next time step according to the predicted acceleration. Obtain the safety protection distance between the following train and the leading train at the next time step according to the predicted acceleration and predicted speed of the leading train at the next time step.

[0019] Among them, the acceleration prediction model further includes an input layer. Normalize the data feature set through the input layer to obtain preprocessed features, specifically including: removing the data outliers in the data feature set to obtain the preprocessed data feature set. Perform global normalization on the preprocessed data feature set to scale the data interval of the preprocessed data feature set to [-1, 1] to eliminate the dimensional difference between different features and obtain the preprocessed features.

[0020] Among them, the acceleration prediction model further includes an output layer. Denormalize the attention-weighted features through the output layer to restore the attention-weighted features to actual values and obtain the predicted acceleration of the leading train.

[0021] The implementation process includes: Step S1, obtaining a real-time data feature set for real-time prediction of train acceleration.

[0022] Step S2, construct an acceleration prediction model (TLMA) based on the temporal convolutional network (TCN), long short-term memory network (LSTM) and multi-head self-attention mechanism (MATT).

[0023] Before using TLMA for prediction, model parameter training is required. The collected historical data feature set is standardized and input into TLMA. During training, 30 seconds of historical data are used to predict the acceleration of the next 1 second. The optimizer uses the Adam optimizer, and the training cycle is 300 rounds. The trained TLMA model is deployed on the following train, and the real-time data feature set collected in real time by input sensors and wireless communication technology is used to obtain the predicted acceleration and predicted speed of the leading train in the next time step.

[0024] Step S3, a safety protection distance calculation formula is proposed, which takes into account the running space requirement of the following train in the next time step, the safety requirement when the train stops at the station, and the safety requirement of the following train under adverse conditions.

[0025] Step S4, substituting the predicted acceleration and predicted speed of the leading train into the safety protection distance calculation formula, and calculating the safety protection distance between the following train and the leading train in the next time step.

[0026] Furthermore, in step S1, the obtained ATO time of train operation is subjected to basic analysis and screening, outliers are eliminated, and the acceleration state data variable of the data feature set in the prediction model is determined to be the train speed at the previous time step, in combination with the characteristics of the data set required for train acceleration prediction. , the train acceleration at the previous time step , the train traction brake level at the previous time step .

[0027] During the operation of the virtual marshaling train, in addition to the train's own state, it is also affected by the line conditions. According to the train dynamics model, the line condition data variable of the data feature set in the prediction model is selected as the line slope value. , Line vertical curve radius .

[0028] Some variable values ​​input into the prediction model are shown in Table 1.

[0029] Table 1 Prediction model input data set Furthermore, in step S2, an acceleration prediction model TLMA combining a temporal convolutional network, a long short-term memory network, and a multi-head self-attention mechanism is proposed. The main advantage of this model is that it can overcome the limitations of traditional prediction models in dealing with long-term dependencies and capturing non-linear relationships, while enhancing the model's attention to key information.

[0030] Figure 2 The composition diagram of TLMA is shown as follows. In the input layer, data outliers in the data feature set are removed to obtain a preprocessed data feature set. The preprocessed data feature set is globally normalized to scale the data interval of the preprocessed data feature set to [-1, 1], eliminating the dimensionality differences between different features and obtaining preprocessed features. In the TCN, the dilation factor in the dilated convolution is adjusted to expand the time window to increase the receptive field of the convolutional kernel, thereby capturing long-distance temporal dependencies. Causal convolution is also used to avoid the leakage of future information, and residual connections are combined to prevent gradient vanishing. Then, the high-order features output by the TCN are input into the LSTM. The LSTM uses a gating mechanism and cell state design to model the temporal dynamic changes of temporal context features, obtaining temporal dynamic features, which can more effectively process temporal dynamic data, making the model more effective in processing long sequence data. In addition, the MATT reconstructs the temporal dynamic features into a vector form. Query, key, and value matrices are generated through linear transformation, the attention weights of the query, key, and value matrices are calculated respectively, and the calculation results of the attention weights are weighted and fused to obtain attention-weighted features. Thereby reflecting the relative importance of relevant variables and optimizing the prediction results of the model. Finally, in the output layer, the attention-weighted features are anti-normalized to restore the attention-weighted features to actual values, obtaining the predicted acceleration of the leading train.

[0031] To measure the effectiveness of the proposed TLMA, four evaluation indicators, namely mean squared error, root mean squared error, mean absolute error, and coefficient of determination, are used to compare the prediction results. The comparison values of the prediction results with some existing prediction models are shown in Table 2.

[0032] Table 2 Comparison results of prediction models The mean squared error value of the proposed TLMA is 0.0404, the root mean squared error value is 0.0053, the mean absolute error value is 0.0730, and the coefficient of determination value is 0.9653. The results show that the predicted acceleration value of TLMA has good consistency with the actual acceleration value. Therefore, TLMA is an effective and accurate short-term prediction model for the acceleration of the leading train.

[0033] Figure 3 and Figure 4Shown are the acceleration curve of the leading train predicted by TLMA and the further calculated speed curve of the leading train. It can be seen from the figure that the predicted acceleration is close to the actual acceleration, and the predicted speed calculated based on the predicted acceleration fits well with the actual speed. The predicted acceleration and speed results will be used for the safety protection distance calculation in step S4.

[0034] Further, in step S3, a train tracking operation safety protection method based on acceleration prediction is proposed. Existing safety protection methods require the distance interval between trains to be greater than the braking distance difference between adjacent trains. However, when the real-time operating state of the train can be predicted, ensuring that the distance interval between trains is greater than the safety protection distance in the next time step is sufficient to ensure the safe operation of the train, and there is no need to calculate the braking distance difference.

[0035] The safety protection method proposed by the present invention mainly includes three parts. The first part is the interval distance for the following train to run at the current speed in the next time step, which can basically meet the running space requirement of the following train in the next time step. The second part is the desired stop interval distance between adjacent trains in the platform to ensure the safety requirement when the train stops. The third part is the dynamic distance margin term, which can ensure the safety requirement of the following train under adverse conditions. The calculation formula for the train safety protection method is:

[0036] (1) Wherein, is the safety protection distance calculated by the proposed operation safety protection method; is the running space requirement of the following train in the next time step; is the speed of the following train at time ; is the time step length; is the desired interval distance for adjacent trains to stay in the platform to ensure the safety requirement when the train stops; is the dynamic distance margin term of the following train at time to ensure the safety requirement of the following train under adverse conditions.

[0037] The dynamic distance margin term consists of three parts: the potential approaching distance margin term, the communication delay margin term, and the control delay margin term. The potential approaching distance margin refers to the reduced spacing between the trailing train and the leading train when the trailing train accelerates at the maximum acceleration in the next time step. This margin term can avoid the impact of extreme acceleration behavior of the trailing train. In addition, wireless communication and autonomous driving are key technologies for virtual formation trains. Due to different train operating environments, communication and control delays may occur, which can cause fluctuations in the spacing between trains. Therefore, when calculating the safety protection distance, the impact of communication and control delays is considered, and the corresponding margin terms are added. The calculation method of the dynamic distance margin term is as follows:

[0038] (2) Among them, is the potential approaching distance margin term; is the communication delay margin term; is the control delay margin term.

[0039] Figure 5 The figure shows the train trajectory schematic diagram of the potential approaching distance margin term. In the next time step, if the running distance of the trailing train accelerating at the maximum acceleration is still less than the predicted running distance of the leading train, the potential approaching distance margin is 0. Otherwise, the extreme acceleration behavior of the trailing train will shorten the spacing between trains. Therefore, the potential approaching distance needs to be compensated. The calculation method of the potential approaching distance margin term is as follows:

[0040] (3) (4) (5) Among them, is the running distance of the trailing train for speed adjustment at the maximum acceleration in the next time step; is the running distance of the leading train for speed adjustment with the predicted acceleration of the leading train in the next time step; is the predicted acceleration of the leading train in the next time step.

[0041] If the speed of the leading train is greater than the speed of the trailing train, the communication delay will not affect the safety protection distance. However, if the speed of the leading train is less than the speed of the trailing train, the communication delay will shorten the spacing between trains. In this case, an additional distance margin is required to compensate for the additional running distance of the trailing train during the communication delay time. The calculation method of the communication delay margin term is as follows:

[0042] (6) Among them, To follow the communication delay time of the train at time whose time-varying range is ; is the maximum communication delay time of the train; is the predicted speed of the leading train at the next time step.

[0043] Control delay refers to the time for the train to calculate and execute the appropriate instructions required for its operation. If the control delay time of the leading train is longer than that of the following train, the control delay does not affect the safe protection distance of the train. However, if the control delay time of the following train is longer than that of the leading train, the interval distance between the trains will be shortened during the control delay time, and the distance margin needs to be increased. The calculation method of the control delay margin term is:

[0044] (7) where is the control delay time of the following train at time whose time-varying range is ; is the control delay time of the leading train at time whose time-varying range is ; is the maximum control delay time of the train.

[0045] Furthermore, in step S4, the acceleration and speed results of the leading train predicted by TLMA are substituted into the safe protection distance calculation formula to obtain the safe protection distance between the following train and the leading train at the next time step.

[0046] Figure 6 The figure shows the comparison chart of the safe protection distance calculation results between the safe protection method proposed by the present invention and the existing absolute braking distance method and relative braking distance method. As can be seen from the figure, the average value of the safe protection distance between the following train and the leading train under the absolute braking distance method is 110.68 m, and the maximum safe protection distance is 325.22 m. The average value of the safe protection distance between the following train and the leading train under the relative braking distance method is 26.87 m, and the maximum safe protection distance is 78.63 m. The average value of the safe protection distance between the following train and the leading train under the safe protection method based on acceleration prediction proposed by the present invention is 17.02 m, and the maximum safe protection distance is 36.53 m. Under the condition of ensuring safe operation, the average value of the safe protection distance calculated by the proposed safe protection method is reduced by 84.6% and 36.7% respectively compared with the absolute braking distance method and the relative braking distance method.

[0047] Second Embodiment This embodiment provides a tracking operation safety protection system for virtual formation trains. The system includes: a data acquisition module, a data processing module, and a safety protection distance output module.

[0048] Among them, the data acquisition module is used to acquire the historical data feature set of the leading train and the actual acceleration of the leading train during the historical operation of the virtual formation train.

[0049] The data processing module is used to input the historical data feature set into an acceleration prediction model. The acceleration prediction model includes a temporal convolutional network, a long short-term memory network, and a multi-head self-attention mechanism connected in sequence; capture the long-distance temporal dependence relationship of the historical data feature set through the temporal convolutional network to obtain temporal context features; model the temporal dynamic changes of the temporal context features through the long short-term memory network to generate temporal dynamic features; weight the temporal dynamic features through the multi-head self-attention mechanism to obtain attention-weighted features; obtain the predicted acceleration of the leading train according to the attention-weighted features; and train the acceleration prediction model with the goal of minimizing the error between the predicted acceleration of the leading train and the actual acceleration of the leading train to obtain the trained acceleration prediction model.

[0050] The safety protection distance output module is used to acquire the real-time data feature set of the leading train at the current time step during the tracking operation of the virtual formation train, input the real-time data feature set into the trained acceleration prediction model to obtain the predicted acceleration of the leading train at the next time step, and calculate the predicted speed of the leading train at the next time step according to the predicted acceleration; obtain the safety protection distance between the following train and the leading train at the next time step according to the predicted acceleration and predicted speed of the leading train at the next time step.

[0051] The above-mentioned tracking operation safety protection system for virtual formation trains is used to implement the tracking operation safety protection method described in the first embodiment, and there is a one-to-one correspondence with the method part. The above descriptions and limitations of the method also apply to the virtual formation train tracking operation safety protection system of this embodiment, and will not be elaborated here.

[0052] The above embodiments only represent several implementation manners of the present invention. The descriptions are relatively specific and detailed, but should not be construed as limiting the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention.

Claims

1. A method for tracking and running safety protection of a virtual formation train, characterized in that Including: Obtain the historical data feature set of the leading train and the actual acceleration of the leading train during the historical operation of the virtual formation train; Input the historical data feature set into the acceleration prediction model, which includes a time convolutional network, a long short-term memory network, and a multi-head self-attention mechanism connected in sequence; capture the long-distance time dependence of the historical data feature set through the time convolutional network to obtain temporal context features; model the temporal dynamic changes of the temporal context features through the long short-term memory network to generate temporal dynamic features; Weight the temporal dynamic features through the multi-head self-attention mechanism to obtain attention-weighted features; obtain the predicted acceleration of the leading train according to the attention-weighted features; Train the acceleration prediction model with the goal of minimizing the error between the predicted acceleration of the leading train and the actual acceleration of the leading train to obtain the trained acceleration prediction model; Obtain the real-time data feature set of the leading train at the current time step during the tracking operation of the virtual formation train, input the real-time data feature set into the trained acceleration prediction model to obtain the predicted acceleration of the leading train at the next time step, and calculate the predicted speed of the leading train at the next time step according to the predicted acceleration; obtain the safety protection distance between the following train and the leading train at the next time step according to the predicted acceleration and predicted speed of the leading train at the next time step.

2. The tracking operation safety protection method for a virtual formation train according to claim 1, characterized in that, The data feature set includes acceleration state data and line condition data; The acceleration state data includes: the train speed at the previous time step, the train acceleration at the previous time step, and the train traction and braking level at the previous time step; The line condition data includes: the line gradient value and the vertical curve radius of the line.

3. The tracking operation safety protection method for a virtual formation train according to claim 1, characterized in that, The obtaining the safety protection distance between the following train and the leading train at the next time step according to the predicted acceleration and predicted speed of the leading train at the next time step specifically includes: Determine the safety protection distance calculation formula through the following formula: ； Among them, is the safety protection distance; is the operating space requirement of the following train at the next time step; is the following train at time speed; is the time step length; is the desired interval distance when adjacent trains stop at the platform; is the following train at time dynamic distance margin term; Input the predicted acceleration and predicted speed of the leading train at the next time step into the safety protection distance calculation formula to obtain the safety protection distance between the following train and the leading train at the next time step.

4. The tracking operation safety protection method for a virtual formation train according to claim 3, characterized in that, Determine the dynamic distance margin term, specifically including: Determine the potential approaching distance margin term based on the following formula: ； ； ； Among them, is the running distance for the following train to adjust its speed with the maximum acceleration in the next time step; is the running distance for the leading train to adjust its speed with the predicted acceleration of the leading train in the next time step; is the predicted acceleration of the leading train in the next time step; Determine the communication delay margin term based on the following formula: ； Among them, is the communication delay time following the train at time whose time-varying range is ; is the maximum communication delay time of the train; is the predicted speed of the leading train at the next time step; Determine the control delay margin term based on the following formula: ； Among them, is the control delay time for the following train at time , and its time-varying range is ; is the control delay time for the leading train at time , and its time-varying range is ; is the maximum control delay time of the train; Determine the dynamic distance margin term through the potential approaching distance margin term, the communication delay margin term, and the control delay margin term based on the following formula: ； Among them, is the potential proximity margin term, is the communication delay margin term, is the control delay margin term.

5. The tracking operation safety protection method for a virtual formation train according to claim 1, characterized in that, The acceleration prediction model further includes an input layer, which normalizes the data feature set through the input layer to obtain preprocessed features, specifically including: Eliminate the data outliers in the data feature set to obtain the preprocessed data feature set; Perform global normalization on the preprocessed data feature set to scale the data interval of the preprocessed data feature set to [-1, 1] to eliminate the dimensional differences between different features and obtain the preprocessed features.

6. The tracking operation safety protection method for a virtual formation train according to claim 1, characterized in that, The weight fusion of the variables affecting the train acceleration in the temporal dynamic features through the multi-head self-attention mechanism to obtain the attention-weighted features specifically includes: Reconstruct the temporal dynamic features into a vector form through the multi-head self-attention mechanism; Generate query, key, and value matrices in vector form through linear transformation, and determine the attention weights of the query, key, and value matrices respectively; Perform weighted fusion on the attention weights of the query, key, and value matrices to obtain attention-weighted features.

7. A safety protection method for the tracking operation of a virtual formation train according to claim 1, characterized in that, The acceleration prediction model further includes an output layer. The specific process of obtaining the predicted acceleration of the leading train based on the attention-weighted features includes: Perform inverse normalization on the attention-weighted features through the output layer to restore the attention-weighted features to acceleration values, and obtain the predicted acceleration of the leading train.

8. A safety protection system for the tracking operation of a virtual formation train, characterized in that, It includes: A data acquisition module for acquiring the historical data feature set of the leading train and the actual acceleration of the leading train during the historical operation of the virtual formation train; A data processing module for inputting the historical data feature set into the acceleration prediction model. The acceleration prediction model includes a temporal convolutional network, a long short-term memory network, and a multi-head self-attention mechanism connected in sequence; capture the long-distance temporal dependence relationship of the historical data feature set through the temporal convolutional network to obtain temporal context features; model the temporal dynamic changes of the temporal context features through the long short-term memory network to generate temporal dynamic features; Perform weighting on the temporal dynamic features through the multi-head self-attention mechanism to obtain attention-weighted features; obtain the predicted acceleration of the leading train based on the attention-weighted features; Train the acceleration prediction model with the goal of minimizing the error between the predicted acceleration of the leading train and the actual acceleration of the leading train to obtain the trained acceleration prediction model; A safety protection distance output module for acquiring the real-time data feature set of the leading train at the current time step during the tracking operation of the virtual formation train, inputting the real-time data feature set into the trained acceleration prediction model to obtain the predicted acceleration of the leading train at the next time step, and calculating the predicted speed of the leading train at the next time step based on the predicted acceleration; obtain the safety protection distance between the following train and the leading train at the next time step according to the predicted acceleration and predicted speed of the leading train at the next time step.

Images