Train arrival delay prediction method based on knowledge data collaborative drive and its application

By combining knowledge-driven and data-driven methods in high-speed rail operations, we build a train operation situation deduction model and perform data enhancement. We use deep learning models to predict train delay time, solving the problem of low prediction accuracy under small sample data, and achieving more Accurate late prediction.

CN118536681BActive Publication Date: 2024-11-08SIGNAL & COMM RES INST OF CHINA ACAD OF RAILWAY SCI +2
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411002010.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-07-25
Publication Date
2024-11-08
Estimated Expiration
2044-07-25

AI Technical Summary

Technical Problem

The existing train delay prediction model has overfitting problems in high-speed rail operations, especially in small sample data, which makes it difficult to accurately predict the long delay time, resulting in low prediction accuracy.

Method used

Using a method based on knowledge data collaboratively drives, by constructing a train operating situation deduction model, combining dynamic knowledge and deep learning model, using convolutional neural networks and long and short-term memory networks for data augmentation and prediction, and integrating knowledge-driven and data-driven methods to improve prediction accuracy.

Benefits of technology

The accuracy of train delay prediction is significantly improved, especially in the case of large delays, the prediction errors of MAE and RMSE are reduced by 80%, which is suitable for situations where the punctuality rate is high but the late data is scarce in high-speed rail operations.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118536681B_ABST
    Figure CN118536681B_ABST
Patent Text Reader

Abstract

The train arrival delay prediction method based on knowledge data collaborative drive includes the following steps: constructing a train operation situation deduction model according to the train operation constraints, working condition conversion relationship and train dynamics model under the temporary speed limit of the section, and obtaining the train delay time under different speed limit conditions simulated by the model; combining the delay data obtained by the model with the original train data to obtain enhanced data; constructing a data-driven train delay prediction model, using the enhanced data as the input of the neural network, and using deep learning to predict the train arrival delay time at the target station. The present invention realizes data enhancement by integrating the train delay time obtained by model deduction into the historical operation data, which solves the problem of low prediction accuracy of sparse values ​​in the case of small samples, and has certain reference value for the prediction of high-speed rail delays in my country, where the punctuality rate is currently high, the large delay situation is less, and the line and vehicle parameters are known.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of high-speed rail arrival delay prediction, and relates to a train arrival delay prediction method based on knowledge data collaborative drive and its application. Background Art

[0002] By the end of 2023, China's high-speed rail operating mileage will reach 45,000 km, with a maximum operating speed of 350 km / h, and high-speed rail technology will lead the world. my country's high-speed rail has become a bright national business card with its high speed and high punctuality. In order to achieve high-level scientific and technological self-reliance in the railway and further meet the travel efficiency needs of passengers, the CR450 scientific and technological innovation project led by the China Railway Group is currently being fully promoted, and the era of high-speed rail speeds of 400 km / h is coming. With the continuous increase in train speeds, the emergency response time of trains is shortened, and the requirements for railway traffic dispatching and command are becoming higher and higher.

[0003] When dealing with some emergencies, such as bad weather and facility failures, trains need to run at limited speeds, which causes them to deviate from the operation diagram. At this time, the dispatcher needs to quickly formulate an operation adjustment strategy based on the train delay time and experience. Therefore, a higher standard intelligent train dispatching and command system is urgently needed to accurately and real-time predict the delay time under emergencies and assist dispatchers in formulating scientific and effective adjustment plans.

[0004] The current train delay prediction models can be divided into two main types: model-based deduction methods and data-based machine learning methods. Due to the complex train operation situation, model-based situation deduction can only predict delays by building a train operation model under ideal conditions, which is difficult to characterize the complex real-world operation environment. The data-based machine learning method needs to rely on a large amount of delay data for model training to obtain more accurate results. In the daily operation of high-speed rail, the train punctuality rate is high, and delay data only accounts for a small part. According to the actual operation data statistics of the Beijing-Tianjin line trains in a certain period of time in 2020, there are a total of 86,315 train operation records, of which 7,749 are delay records, accounting for only 8.98% of the total data, and there are even fewer delay records of more than 30 minutes, only 0.27%. There are few data samples for large delays, which makes the model more prone to overfitting. This situation makes it difficult to use data-driven delay prediction. Summary of the invention

[0005] In order to solve the defects in the prior art, the present invention discloses a method, and its technical solution is as follows:

[0006] The train arrival delay prediction method based on knowledge data collaborative drive is characterized by: Step S100: constructing a train operation situation deduction model according to the train operation constraints, working condition conversion relationship and train dynamics model under the temporary speed limit of the section, and obtaining the train delay time under different speed limit conditions simulated by the model.

[0007] Step S200: combining the delay data obtained by the model with the original train data to obtain enhanced data.

[0008] Step S300: construct a data-driven train delay prediction model. The enhanced data is used as the input of the neural network, and the train arrival delay time at the target station is predicted by deep learning.

[0009] Beneficial Effects

[0010] The present invention proposes a train arrival delay prediction model based on knowledge and data collaborative driving. In the knowledge-driven model, a train situation deduction model is established based on the knowledge of train dynamics to calculate the predicted delay time of the train under different speed limit conditions; in the data-driven model, a convolutional neural network and a long-short time memory network are introduced to calculate the predicted delay time of the train in real time based on the historical operation data of the train. The train delay time obtained by model deduction is integrated into the historical operation data to achieve data enhancement, which solves the problem of low prediction accuracy of sparse values ​​in the case of small samples. It has certain reference value for the prediction of high-speed rail delays in my country, where the punctuality rate is relatively high, the major delays are relatively few, and the line and vehicle parameters are known. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 This is a processing flow chart of a train arrival delay prediction method driven by knowledge data collaboration provided by the present invention;

[0012] Figure 2 It is a simulation diagram of a train operation curve (vt) under speed limit conditions in a preferred example of a train arrival delay prediction method driven by knowledge data collaboration provided by the present invention.

[0013] Figure 3 It is a simulation diagram of a train operation curve (vs) under speed limit conditions in a preferred example of a train arrival delay prediction method driven by knowledge data collaboration provided by the present invention.

[0014] Figure 4 It is a scatter plot of predicted train delays and actual train delays in a preferred example of a knowledge data collaboratively driven train arrival delay prediction method provided by the present invention. DETAILED DESCRIPTION

[0015] The present invention proposes a delay prediction method driven by knowledge data collaboration, such as Figure 1As shown in the figure, the simulation data obtained by model deduction is mainly fused with the original data for data enhancement, and then the neural network model is used to predict train delays. The specific steps include:

[0016] Step S100: construct a train operation status deduction model according to the train operation constraints, working condition conversion relationship and train dynamics model under the temporary speed limit of the section, and obtain the train delay time under different speed limit conditions simulated by the model.

[0017] Step S200: combining the delay data obtained by the model with the original train data to obtain enhanced data.

[0018] Step S300: construct a data-driven train delay prediction model. The enhanced data is used as the input of the neural network, and the train arrival delay time at the target station is predicted by deep learning.

[0019] Furthermore, the step S100 includes:

[0020] Step S110: Input the mass, length, line slope, total length and maximum speed limit of the train into the train dynamics model, and obtain the acceleration, speed and passing time of the train at various locations according to the distance step method to form a simulated operation curve of the train and obtain the simulated operation time of the train under normal operating conditions.

[0021] (1) Acceleration calculation:

[0022] (1)

[0023] In the formula, and The trains are located at and The acceleration at the point is in m / s 2 ; m is the mass of the train, in kg; k is the gyration coefficient, which indicates the effect of horizontal deviation on the acceleration of the train during operation. The maximum acceleration allowed by the train, in m / s 2 ; The maximum impact rate allowed by the train, in m / s 3 ; The distance between trains in the previous The running time, in seconds. For trains in position The total force at the location is in N and is calculated as follows:

[0024] (2)

[0025] In the formula, , , , They are respectively the traction, braking force, basic resistance and slope resistance experienced by the train at position i+1. and It is related to the working condition of the train. , The train is in maximum traction condition; , The train is in the maximum braking condition; , When the train is in coasting condition; when and Take a constant such that The train is in cruising condition.

[0026] (2) Speed ​​calculation:

[0027] (3)

[0028] In the formula, and The trains are located at and The speed at the point, in m / s; For trains in position The speed limit value at the position is in m / s. Is the train in position and The running distance between.

[0029] (3) Calculation by time point:

[0030] (4)

[0031] In the formula, The train arrival location moment, The train arrival location The moment, in s; and The trains are located at and The speed at the point, in m / s; For trains in position The acceleration at the point is in m / s 2 .

[0032] In the specific deduction process, the maximum traction curve of the train is first calculated from the starting point of acceleration to the right until the speed does not exceed the speed limit and enters cruising; secondly, the maximum braking curve of the train is calculated from the end point of deceleration to the left until the speed does not exceed the speed limit and enters cruising. If the two curves intersect, the minimum speed of the two is taken as the speed of that point.

[0033] Step S120: setting various possible speed limit interval starting points, end points and speed limits to obtain the train simulation running time under different speed limit conditions.

[0034] Step S130: According to the difference between the running time under normal conditions and the simulated running time of the train under speed limit conditions, the simulated delay time of the train under different speed limit conditions is obtained.

[0035] Furthermore, the step S200 includes:

[0036] Step S210: The train delay time obtained by the model simulation is screened according to the distribution of the actual delay data, and only the simulation data within the actual maximum delay range is retained. The simulation data with the same delay situation at each station is replaced by the mean to reduce redundancy.

[0037] Step S220: The delay data obtained by the model is combined with the original train data in different proportions to construct different enhanced data.

[0038] Furthermore, the step S300 includes:

[0039] Step S310: CNN, Bi-LSTM and attention mechanism are integrated to build a neural network model CBLA-net for train delay prediction. The number of LSTM layers is set to 2, with 64 neurons in each layer, and the optimizer is Adam, with a random dropout rate of 0.1.

[0040] CNN consists of multiple convolutional layers, pooling layers, and fully connected layers, and has powerful feature extraction capabilities. By using convolution kernels of different sizes, CNN can effectively extract local important information. Input data is compressed through the pooling layer to simplify feature information and reduce computational complexity. Therefore, CNN can effectively process and identify multi-dimensional data information generated during the actual operation of the train.

[0041] Bi-LSTM is composed of forward LSTM and backward LSTM. LSTM consists of memory cells and control gates, which enable the network to better capture and remember long-term dependencies in the train feature sequence. It is divided into three main parts according to its function: input gate, forget gate and output gate. Equations (5)-(10) describe the working principle of LSTM.

[0042] (5)

[0043] (6)

[0044] (7)

[0045] (8)

[0046] (9)

[0047] (10)

[0048] In formulas (5)-(10), the train feature information to be processed in the current time step is the retained historical information and updated input information The weighted combination of and .in, is the output of the forget gate, which controls the historical train information that should be retained , the value range is from 0 to 1, 0 means complete forgetting, and 1 means complete retention. is the output of the input gate, which controls the input information updated to the model .at last, is the output of the output gate, controlling the train feature information that will be passed to the next time step . , and All are past time step information and current input information The function of , , and They are the bias items of the forget gate, input gate, candidate memory unit and output gate respectively. , , and They are the weight matrices of the forget gate, input gate, candidate memory unit, and output gate respectively.

[0049] In traditional unidirectional LSTM, information propagation is unidirectional, and only the impact of information from past time steps on the current time step is considered. Bi-LSTM introduces an additional backward layer to propagate information in the opposite direction along the time series. Bi-LSTM considers information from both past and future time steps at the same time, which helps to more fully understand the deep relationships in the train operation time series data.

[0050] The attention mechanism allows the model to dynamically weight different parts of the input sequence when generating each output. This enables the model to selectively focus on different parts of the features related to the current train delay, reduce the impact of irrelevant information, and improve the accuracy of the prediction. In this model, the input train operation data is processed through a fully connected layer to calculate the attention weights. This approach allows the model to dynamically adjust the weights based on the content of the input sequence, allowing it to focus more on the key information related to the current train operation.

[0051] Step S320: Determine the input structure of the neural network. Select the station with more passenger operations as the target prediction station. Use the arrival delay time of a train at all stations before the target station as the feature input. In order to better extract the spatiotemporal characteristics of the data, the information of the previous stations is divided into 2 sequences in order, with a time step of 2. Each step inputs a sequence, and two steps correspond to a prediction output.

[0052] Step S330: Select enhanced data with different fusion ratios as training sets, and randomly select 15% of the original train data by month as test sets. Use the training sets constructed with different enhanced data to train the model, and compare the prediction results under the same test set.

[0053] Step S330: Calculate the prediction performance of the training sets with different fusion ratios under different delay conditions, expressed as mean absolute error (MAE) and root mean square error (RMSE). According to the influence of the fusion ratio of model data and real data on the prediction performance, select the fusion ratio with the best prediction performance.

[0054] The present invention also provides an embodiment for exemplarily showing the process and results of delay prediction using the method provided by the present invention. The specific steps are as follows:

[0055] S100: Generation of train delay simulation data.

[0056] S110: Obtain specific parameters of the train and line.

[0057] The delay data is simulated and generated based on the planned operation diagram of the down train of Beijing-Tianjin Intercity Railway (Beijing South-Tianjin) and the traction and braking characteristic curves of the train. The train model is CR400BF with 8 cars, and the train parameters are shown in Table 1.

[0058] Table 1 Simulation train parameters

[0059] parameter Value Train mass / kg 501000 Train length / m 209 Maximum speed / (km / h) 350 <![CDATA[Maximum acceleration / (m / s 2 )]]> 1 <![CDATA[Maximum impact rate / (m / s 3 )]]> 0.5

[0060] When a train encounters extreme weather or emergencies during operation, the train will run at a limited speed. The speed limit in the interval is one of the important reasons for train delays. Common temporary speed limit setting rules under extreme weather conditions are shown in Table 2:

[0061] Table 2 Speed ​​limit settings in strong wind scenarios

[0062] High wind warning level Wind speed / (m / s) Speed ​​limit / (km / h) Level 1 (Blue Warning) (15,20] 300 Level 2 (yellow warning) (20,25] 200 Level 3 (Orange Warning) (25,30] 120 Level 4 (Red Alert) (30,+∞) Detain train

[0063] The total length of the line is 117040m. There are six stations along the line. The central kilometer mark of each station is shown in Table 3. There are 56 block sections in total. The endpoint values ​​of a total of 57 block sections are recorded as the locations for setting speed limit sections.

[0064] Table 3 Center kilometer mark of each station along the Beijing-Tianjin line

[0065] Station Name Center kilometer mark Beijing South 1105 Yizhuang 22285 Yongle 46565 Wuqing 84454 South Warehouse 107969 Tianjin 118114

[0066] Step S120: Setting different speed limit conditions.

[0067] In order to simulate the speed limit of trains in different scenarios, the simulation sets the interval speed limit value to 8 levels, including 45km / h, 60km / h, 80km / h, 100km / h, 120km / h, 160km / h, 200km / h, and 250km / h, and the speed limit in the station is set to 80km / h. The minimum speed limit interval is set to 2, the starting point is any value among the first 54 endpoints of all block sections from Beijing to Tianjin, and the end point is set to any value that is more than two intervals away from the starting point and does not exceed the maximum interval endpoint, and all possible speed limit areas are obtained. And these 8 speed limit values ​​are set to these speed limit areas respectively, and all possible train delay times are obtained through model simulation.

[0068] Step S130: generating simulation delay data.

[0069] The train operation situation was simulated on a computer with Intel(R) Core(TM) i9-13900 2.00 GHz configuration. A total of 12,232 delay data were obtained from the simulation, some of which are shown in Table 4:

[0070] Table 4 Example of model simulation train delay data

[0071] Speed ​​limit starting point km / m Speed ​​limit end kilometer sign / m Speed ​​limit / (km / h) Train running time / s Speed ​​limit starting point Speed ​​limit terminal Yizhuang Delayed / min Yongle Delay / min Wuqing late / min Nancang Delay / min Tianjin Delay / min 6900 83551 100 3616.37 Beijing South Wuqing 8 18 35 35 35 29700 116403 80 1544.28 Yizhuang Tianjin 0 13 35 48 50 43700 113700 250 1831.71 Yizhuang Tianjin 0 0 4 5 5 85170 116403 160 1485.24 Wuqing Tianjin 0 0 0 6 6 18300 116403 160 1485.24 Beijing South Tianjin 3 8 15 19 20

[0072] Among them, the speed limit range is (43700, 113700), and the running situation of the train with a speed limit of 250km / h in the section is as follows Figure 2 , Figure 3 shown.

[0073] Step S200: The train simulation delay data is integrated with the actual operation delay data to perform data enhancement.

[0074] Step S210: Simulation data processing.

[0075] The model generated 12,674 data points, which were screened according to the distribution of real delay data. Only the model data within the maximum delay range of the real data were retained, and the data with the same delay situation at each station were replaced by the mean to reduce redundancy. After processing, a total of 11,395 model-generated delay data points were obtained, and the ratio of original data to model data was about 1:8.

[0076] Step S220: merging the simulation data with the original data.

[0077] Considering that the ratio of model-generated data to original data may affect the prediction performance, a stratified sampling method is used according to the distribution of model data to randomly extract model delay data according to different categories. The stratification standards are shown in Table 5:

[0078] Table 5 Model data distribution statistics

[0079] category Sample size / item The proportion of total samples All five stations are delayed 3188 27.98% Only four stations were delayed 3268 28.68% Only three stations were delayed 3715 32.60% Only two stations were delayed 1056 9.27% Only one station is delayed 168 1.47%

[0080] Five different augmented data sizes were randomly extracted, so that the ratio of model data to original data was 8:1, 6:1, 4:1, 2:1 and 1:1 respectively. The training set was constructed using data of different proportions.

[0081] Step S300: Perform delay prediction based on deep learning.

[0082] Step S310: construct a delay prediction model.

[0083] The delay prediction model uses CBLA-net, the number of LSTM layers is set to 2, each layer has 64 neurons, the time step is set to 2, the optimizer is Adam, the random dropout rate is 0.1, the training batch size is 64, and the iteration is 1500 rounds.

[0084] Step S320: Train delay prediction.

[0085] First, the original data was used for delay prediction. A total of 1,725 ​​records of delayed trains with Beijing as the starting station and Tianjin as the terminal from June 2019 to September 2020 were collected. Approximately 15% of them per month, totaling 252 data, were randomly sampled as the test set to verify the prediction performance of the model, and the remaining 1,473 data were used as the training set.

[0086] Secondly, the delay prediction is performed by combining the model with the original data. The delay data generated by the processed model is fused with the original data as the training set, and the model performance is verified using the same test set as the original data. The delay prediction results are shown in Figure 2. Figure 4 shown.

[0087] Step S320: Determine the optimal data fusion ratio.

[0088] The model data was randomly extracted in layers and combined with the original data to obtain five different sizes of enhanced data, so that the ratio of the model data to the original data was 8:1, 6:1, 4:1, 2:1 and 1:1. The training set was constructed using data fused in different proportions, and its prediction performance was verified on the same test set.

[0089] According to the different delay times, the delay data of the test set are divided into small delay (0-10min), medium delay (10-30min), and large delay (greater than 30min) and performance analysis is performed on these delay situations respectively.

[0090] The root mean square error (RMSE) and mean absolute error (MAE) are selected as the prediction performance evaluation indicators. Their calculation formulas are (11) and (12) respectively:

[0091] (11)

[0092] (12)

[0093] The specific results are shown in Table 6:

[0094] Table 6 Delay prediction performance of data augmentation methods with different data ratios

[0095]

[0096] In summary, the present invention proposes a high-speed rail delay prediction method based on knowledge and data collaborative driving. The method first constructs a train operation situation deduction model based on knowledge driving, and the model constructs a train dynamics model according to the relevant parameters of the train and the line, and combines the operation-related constraints to deduce and simulate to obtain the train operation curve, and obtains different train delay times by setting different speed limit scenarios. Aiming at the problem that the accuracy of delay prediction by the data-driven model under small sample data is not high, the train delay prediction method driven by knowledge model collaborative driving is used to first obtain the delay time generated based on the train operation model, and then combine the model simulation data with the actual delay data for data enhancement, and determine the optimal ratio of model data and original data through experiments, which significantly improves the performance of train delay prediction, especially in the case of large delays greater than 30 minutes, the prediction error MAE and RMSE are reduced by 80% after using model data enhancement, which has certain reference value for the high-speed rail delay prediction in my country with high punctuality rate, few large delays and known line and vehicle parameters.

[0097] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions only describe the principles of the present invention. The present invention may be subject to various changes and improvements without departing from the spirit and scope of the present invention. These changes and improvements fall within the scope of the present invention. The scope of protection claimed by the present invention is defined by the attached claims and their equivalents.

Claims

1. A train arrival delay prediction method based on knowledge data collaborative drive, which is characterized by: step S100: constructing a train operation situation deduction model according to the train operation constraints under the temporary speed limit of the section, the working condition conversion relationship and the train dynamics model, and obtaining the train delay time under different speed limit conditions simulated by the model; Step S110: The mass, length, line slope, total length and maximum speed limit of the train are input into the train dynamics model, and the acceleration, speed and passing time of the train at various locations are obtained according to the distance step method to form a simulation operation curve of the train, and obtain the simulation operation time of the train under normal operating conditions; Step S120: setting various possible speed limit interval starting points, end points and speed limits to obtain the train simulation running time under different speed limit conditions; Step S130: obtaining the simulated delay time of the train under different speed limit conditions according to the difference between the running time under normal state and the simulated running time of the train under speed limit conditions; Calculation of the acceleration: (1) In the formula, and The trains are located at and The acceleration at the point is in m / s 2 ; m is the mass of the train, in kg; k is the gyration coefficient, which indicates the effect of horizontal deviation on the acceleration of the train during operation. The maximum acceleration allowed by the train, in m / s 2 ; The maximum impact rate allowed by the train, in m / s 3 ; The distance between trains in the previous The running time, in seconds; For trains in position The total force at the location is in N and is calculated as follows: (2) In the formula, , , , are the traction, braking force, basic resistance and slope resistance of the train at position i+1 respectively; and It is related to the working condition of the train. , The train is in maximum traction condition; when , The train is in the maximum braking condition; , When the train is in coasting condition; when and Take a constant such that When , the train is in cruising condition; Speed ​​calculation: (3) In the formula, and The trains are located at and The speed at the point, in m / s; For trains in position The speed limit value at the position is in m / s; Is the train in position and The distance between them; Calculate by time point: (4) In the formula, The train arrival location moment, The train arrival location The moment, in s; and The trains are located at and The speed at the point, in m / s; For trains in position The acceleration at the point is in m / s 2 ; Step S200: combining the delay data obtained by the model with the original train data to obtain enhanced data; Step S210: The train delay time obtained by the model simulation is screened according to the distribution of the real delay data, and only the simulation data within the real maximum delay range is retained, and the simulation data with the same delay situation at each station is replaced by the mean value to reduce redundancy; Step S220: combining the delay data obtained by the model with the original train data in different proportions to construct different enhanced data; Step S300: constructing a data-driven train delay prediction model, using the enhanced data as the input of the neural network, and using deep learning to predict the arrival delay time of the train at the target station; Step S310: integrating CNN, Bi-LSTM and attention mechanism to build a neural network model CBLA-net for train delay prediction; Bi-LSTM is composed of forward LSTM and backward LSTM. LSTM consists of memory cells and control gates, which enable the network to better capture and remember long-term dependencies in the train feature sequence. It is divided into three main parts according to its function: input gate, forget gate and output gate. Equations (5)-(10) describe the working principle of LSTM: (5) (6) (7) (8) (9) (10) In formulas (5)-(10), the train feature information to be processed in the current time step is the retained historical information and updated input information The weighted combination of and ;in, is the output of the forget gate, which controls the historical train information that should be retained , the value range is from 0 to 1, 0 means completely forgotten, 1 means completely retained; is the output of the input gate, which controls the input information updated to the model ;at last, is the output of the output gate, controlling the train feature information that will be passed to the next time step ; , , and All are past time step information and current input information The function of , , and They are the bias items of the forget gate, input gate, candidate memory unit and output gate respectively. , , and They are the weight matrices of the forget gate, input gate, candidate memory unit, and output gate respectively; Step S320: Determine the input structure of the neural network: select the station with more passenger operations as the target prediction station; use the arrival delay time of a train at all stations before the target station as the feature input; Step S330: Select enhanced data with different fusion ratios as training sets, and randomly select 15% of the original train data by month as test sets, use the training sets constructed with different enhanced data to train the model, and compare the prediction effects under the same test set; Step S340: Calculate the prediction performance of the training sets with different fusion ratios under different delay conditions, and express it in terms of mean absolute error (MAE) and root mean square error (RMSE).

2. The train arrival delay prediction method based on knowledge data collaborative drive according to claim 1 is characterized by: The number of LSTM layers is set to 2, with 64 neurons in each layer, Adam is selected as the optimizer, and the random dropout rate is 0.

1.

3. A non-volatile storage medium, characterized in that: The non-volatile storage medium includes a stored program, wherein when the program is executed, the device where the non-volatile storage medium is located is controlled to execute the method according to any one of claims 1 to 2.

4. An electronic device, characterized in that: It comprises a processor and a memory; the memory stores computer-readable instructions, and the processor is used to execute the computer-readable instructions, wherein the computer-readable instructions execute the method described in any one of claims 1 to 2 when executed.