A high-speed rail network delay prediction method based on multi-source heterogeneous data processing

Abstract

The present invention discloses a high-speed railway network delay deduction method based on multi-source heterogeneous data processing. Aiming at the problem of railway network delay evolution, the present invention comprehensively considers the interaction between different types of entities in the high-speed railway network, and designs a new neural network model SAGE-HetGNN model by combining SAGE homogeneous graph neural network and HetGNN heterogeneous graph neural network structure. The SAGE-HetGNN model can process different types of nodes and capture the interaction between nodes. Based on the model, the railway network delay evolution can be studied by predicting the delay of running trains at future timestamps, and the prediction accuracy is relatively high.

Description

Technical Field

[0001] The present invention relates to the field of rail train technology, and in particular to a high-speed rail network delay deduction method based on multi-source heterogeneous data processing. Background Art

[0002] The railway network is a complex system composed of many fixed facilities (such as stations, signal systems) and mobile objects (such as trains). The operating status of mobile objects is the result of the interaction between facilities and mobile objects in the system and the influence of external environment such as bad weather. As an indicator to measure the difference between the actual operation plan and the planned operation plan, delays are usually used to evaluate the status of the railway network. As the most common and most important mobile object in the railway network, the delay of running trains (RT) in the network is an excellent indicator for evaluating the status of the railway network. Predicting the delay of RT in the railway network can reveal the law of the evolution of the delay status of the railway network, thereby helping dispatchers to formulate more reasonable and scientific dispatching strategy adjustment plans; it can also support dispatchers to conduct real-time network status assessment to make appropriate short-term and long-term decisions.

[0003] The train delay prediction problem can be divided into train-oriented and network-oriented research according to the target to be predicted. Train-oriented research mainly targets each train and predicts the delay of each train at the next station. Network-oriented research takes the high-speed railway network as the research object and mainly studies the operating status of the railway network at different timestamps. Most of the existing delay prediction research is train-oriented. However, in the actual transportation organization process, dispatchers usually pay more attention to the network status from a systematic perspective, that is, the delay status of the railway network after a certain period of time. Therefore, compared with the train-oriented research method, the network-oriented method is more systematic. The network-oriented method regards all trains at each moment as a whole. At a given moment, it explores the evolution of railway network delays by predicting the delays of all trains in the network after a certain time interval. Therefore, the network-oriented method is more suitable for dispatchers to grasp the status of the entire railway network and formulate a more comprehensive and systematic adjustment strategy. Therefore, the main problem studied in this scheme is: under the interaction of different entities in the transportation network (for example, stations, trains, etc.), what kind of delay status does the network present after the train delay has evolved for a certain period of time.

[0004] Although delays are the result of interactions between different entities in the railway network, not all trains are affected by the same entity at the same time; for example, at a given timestamp, some trains in the railway network are affected by equipment failure, but other trains may not be affected. Therefore, in delay prediction, the input features of the trains may be different, that is, the input features are inconsistent. Existing delay prediction research is mainly based on traditional machine learning algorithms (such as RF, SVR, etc.) and graph neural networks (GNN). Machine learning models usually use data in rectangular or grid format as model input. In other words, machine learning models cannot overcome the inconsistency of input features. They require all samples to have the same input features, and those with some missing input features must be filled manually. For example, trains to be predicted without equipment failure-related features need to be artificially filled with some specified values ​​(e.g., 0) to ensure that they have the same input features as trains with facility failure-related features. This artificial filling method requires some prior knowledge and domain knowledge as support, and these filling values ​​will also affect the final prediction results to a certain extent.

[0005] GNN updates node information based on edges and adjacent nodes, and allows nodes to have different numbers of adjacent nodes. Therefore, treating influencing factors and trains to be predicted as nodes and connecting them to capture the impact of influencing factors on trains to be predicted is a new idea to address the shortcomings of traditional machine learning research. In this way, trains with specific influencing factors only need to be connected to specific influencing factors, and all samples can have different input features. However, existing GNN-based studies only apply homogeneous graph neural networks (Homogeneous Graph Neural Network, HomoGNN), which requires all nodes in GNN to have the same attributes; therefore, these studies cannot treat influencing factors and trains to be predicted as nodes at the same time, they can only treat trains as nodes, or assign the impact of other entities on trains as attributes of train nodes. Therefore, if there is a network-oriented method that can handle the inconsistency of input features, it can make up for the shortcomings of existing railway network delay evolution methods. Summary of the invention

[0006] In order to solve the problems existing in the prior art, the purpose of the present invention is to provide a high-speed rail network delay deduction method based on multi-source heterogeneous data processing. The present invention solves the limitation that the existing HomoGNN model cannot process different types of nodes, and the existing machine learning model requires input features in rectangular or grid format, that is, each sample input feature must be consistent.

[0007] To achieve the above object, the technical solution adopted by the present invention is: a high-speed rail network delay deduction method based on multi-source heterogeneous data processing, comprising the following steps:

[0008] Step 1: Based on the HetGNN heterogeneous graph neural network and the GraphSAGE homogeneous graph neural network, the SAGE-HetGNN model is constructed by combining them;

[0009] Step 2: Obtain sample data of the high-speed railway network and construct training sets, validation sets, and test sets. Abstract entities in the high-speed railway network into nodes with different attributes, use meta-paths to describe the relationships between nodes, and construct a heterogeneous graph containing multiple node types and multiple meta-paths at each timestamp.

[0010] Step 3: Based on the training set, validation set and test set, the constructed SAGE-HetGNN model is trained, validated, and the model hyperparameters are adjusted and tested to obtain the final SAGE-HetGNN model;

[0011] Step 4: Based on the delay status of the high-speed railway network at the current timestamp, the final SAGE-HetGNN model is used to obtain the delay prediction value of all running trains contained in the high-speed railway network at the next timestamp after a preset time interval.

[0012] As a further improvement of the present invention, in step 2, the meta-path includes RT-rr-RT, TT-tr-RT, Station-ss-Station and Station-sr-RT; RT-rr-RT represents the connection relationship between two adjacent RT nodes, reflecting the interaction between RT nodes; TT-tr-RT represents the relationship between the preceding TT node and the succeeding RT node; the Station-ss-Station meta-path reflects the physical structure of the railway network, that is, the station in front affects the station behind; the Station-sr-RT meta-path connects the PS / TS node with the RT node, indicating the relationship between the train and the next station; wherein: the RT node represents the running train, the TT node represents the terminal train, the PS node represents the passing station, and the TS node represents the terminal station.

[0013] As a further improvement of the present invention, in step 2, entities in the high-speed railway network are abstracted into nodes with different attributes including running train nodes, terminal train nodes, passing station nodes and terminal station nodes; wherein:

[0014] The properties of the running train node include: C: the delay of the train to be predicted at the current timestamp; I: the interval time between the train to be predicted and the first preceding train; S: when the train is at the station, it is the planned running time of the train to be predicted in the next section; when the train is in the section, it is the planned running time of the train to be predicted in the current section; M: when the train is at the station, it is the minimum running time of the train to be predicted in the next section; when the train is in the section, it is the minimum running time of the train to be predicted in the current section; R: when the train is at the station, it is the remaining running time of the train to be predicted in the next section; when the train is in the section, it is the remaining running time of the train to be predicted in the current section;

[0015] The attributes of the terminal train node include: C: the delay of the train to be predicted at the current timestamp;

[0016] The attributes of the passing station node include: N: the number of arrival and departure lines of the station;

[0017] The attributes of the terminal node include: N: the number of arrival and departure lines of the corresponding arrival and departure yard.

[0018] As a further improvement of the present invention, the step 4 is specifically as follows:

[0019] Four different types of initial node information are mapped to the same dimensional space, and all nodes are mapped to 256-dimensional vectors; a four-layer SAGE neural network is used to update the node information in each meta-path, each layer of the neural network has 256 neurons, and the ReLU function is used for activation between two adjacent layers; the updated node information is aggregated by using the sum aggregation function; finally, the delay prediction value of each running train node in the current high-speed railway network at the next timestamp is obtained.

[0020] As a further improvement of the present invention, in step 4, if the train arrives ahead of schedule, the delay value of the train is a negative value; if the train arrives on time, the delay value of the train is 0; if the train arrives late, the delay value of the train is a positive value.

[0021] As a further improvement of the present invention, the step 3 is specifically as follows:

[0022] Use mean absolute error (MAE) as the loss function to train the SAGE-HetGNN model;

[0023] The Adam optimizer with an initial learning rate of 0.001 was used as the model optimizer. When the loss function of the validation dataset dropped less than 0.01 within 10 iterations, the global learning rate of the optimizer was reduced to 10% of the original value. To prevent overfitting of the model, the model training was stopped when the loss function of the validation dataset dropped less than 0.01 in 30 iterations. To save training time and computing resources, the maximum number of iterations was set to 300, and the dataloader technology with a batchsize of 64 was used.

[0024] As a further improvement of the present invention, it also includes:

[0025] By comparing the prediction accuracy of the SAGE-HetGNN model when different meta-paths are eliminated, the effect of the interaction between different nodes on the evolution of the delay status of the high-speed railway network is revealed from the spatial dimension.

[0026] As a further improvement of the present invention, it also includes:

[0027] By comparing the prediction accuracy of the SAGE-HetGNN model at different prediction time intervals, the prediction effect of the SAGE-HetGNN model on the delay status of the high-speed railway network at different prediction time intervals is verified from the time dimension.

[0028] The beneficial effects of the present invention are:

[0029] In response to the problem of railway network delay evolution, the present invention comprehensively considers the interaction between different types of entities in the high-speed railway network, and designs a new neural network model SAGE-HetGNN model by combining the SAGE homogeneous graph neural network and the HetGNN structure. The SAGE-HetGNN model can process different types of nodes and capture the interaction between nodes. Based on the SAGE-HetGNN model, the evolution of railway network delays can be studied by predicting the delays of running trains at future timestamps, and it has high prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 is a flow chart of an embodiment of the present invention;

[0031] Figure 2 A modeling flow chart of a heterogeneous graph neural network in an embodiment of the present invention;

[0032] Figure 3 Schematic diagram of the evolution of the delay state of the railway network in an embodiment of the present invention;

[0033] Figure 4 Schematic diagram of a heterogeneous graph structure of a railway network at each timestamp in an embodiment of the present invention;

[0034] Figure 5 Schematic diagram of Guangzhou South Network (GZS-Net) and Changsha South Network (CSS-Net) in an embodiment of the present invention;

[0035] Figure 6 Schematic diagram of the SAGE-Het model and its hyperparameters in an embodiment of the present invention;

[0036] Figure 7 A comparison chart of the prediction accuracy of SAGE-Het and other benchmark models in an embodiment of the present invention;

[0037] Figure 8 It is a comparison chart of the actual delay and the predicted value based on SAGE-Het in the embodiment of the present invention. DETAILED DESCRIPTION

[0038] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0039] Example

[0040] The research objective of the present invention is to comprehensively consider the interactions among different types of entities in the transportation network, establish a model to predict the delay status of a running train at a given time after a certain period of time, and reveal the evolution of the network delay status.

[0041] The present invention uses a HetGNN structure, which can capture the information of different types of nodes and the interactions between nodes. Then, a SAGE-HetGNN model is proposed by combining the HetGNN structure with a commonly used and efficient homogeneous graph neural network GraphSAGE. The present invention considers four railway network entities, namely, running trains (RT), terminal trains (TT), passing stations (PS) and terminal stations (TS), and describes the relationship between them with meta-paths. The evolution of the railway network delay state is predicted by processing these four nodes using the SAGE-HetGNN model and extracting their relationships.

[0042] like Figure 1 As shown, a high-speed rail network delay prediction method based on multi-source heterogeneous data processing includes:

[0043] In step S101, the SAGE-HetGNN model is constructed based on the combination of the HetGNN heterogeneous graph neural network and the GraphSAGE homogeneous graph neural network.

[0044] 1. HetGNN Principle

[0045] In order to predict the delay of each train in the railway network, an effective network-oriented method is needed, which should be able to handle different types of nodes and different sample inputs in the railway network. As an important branch of GNN, HetGNN can meet these two requirements. Figure 2 Describes the entire process of the HetGNN framework when performing prediction tasks.

[0046] Assume a HetGNN is G = (V, E), where V and E represent the set of nodes and edges respectively. In HetGNN, edges can also be represented by metapaths; metapaths contain the relationships between nodes (the types of nodes in a metapath can be the same or different).

[0047] In HetGNN, since it contains attributes of different types of nodes, these nodes may have different attributes and different data dimensions, so in the first step ( Figure 2 In step (a), we first need to map these input nodes to the same dimensional space. In other words, given a node v∈V of type A, in step (a),

[0048]

[0049] Among them is is the original feature vector of a node of type A, with dimension d A . h′ v is the mapping result after the linear transformation in step (a), with a dimension of d. A is a d×d A The matrix is ​​used to linearly transform the nodes of node type A. Subsequently, in step (b), the node information in the corresponding meta-path will be updated by some HomoGNN models. The updating process of the node information in the meta-path can be expressed as follows:

[0050]

[0051] in, Represents the feature vector of node v in the hth hidden layer. θ (l,r) is an instance of each pair of meta-paths, θ refers to f θ (l,r) The set of all parameters, r represents a meta-path, R is the set of all meta-paths, r∈R. N r(v) represents the set of neighbor nodes of node v. Only the information of nodes adjacent to v will be used when the node is updated. The node adjacency relationship can be set to first-order adjacency (that is, points are directly adjacent to each other) or to multiple-order adjacency (that is, points are not directly adjacent to each other, but can be adjacent based on passing through other points). It means to aggregate the information of nodes v in different meta-paths, which can be sum, mean or maximum value.

[0052] Many excellent models can be used as the basic HomoGNN in step (b), such as graph attention network GAT (graph attention network), graph convolutional network GCN (graph convolutional network) and SAGE (GraphSAGE). In these HomoGNNs, the node features of node i in layer k are updated by the following operation:

[0053]

[0054] where e ji Represents the adjacency matrix of the edge from node j to node i, and when they are connected, the corresponding matrix position is 1, otherwise it is 0. N(i) represents the set of neighboring nodes of node i. Function □ is a differentiable permutation invariant function (for example, sum, mean, or max). In addition, φ and r represent differentiable functions such as MLP and GCN. Therefore, the output of step (a) will be updated by step (b). It is worth noting that the HomoGNN structures processing different meta-paths in step (b) can be the same or different, but the output dimensions of these HomoGNNs must be the same.

[0055] After all nodes go through step (b), various nodes update the node information through different meta-paths and benchmark HomoGNN. Then, in step (c), the output results of each node after step (b) are aggregated through an aggregation function (e.g., sum or mean). Finally, in step (d), the final prediction task is performed, which can be a prediction of a node, or a prediction of a graph and an edge.

[0056] 2. GraphSAGE Principle

[0057] The traditional GCN algorithm is a method of learning node information by using the graph structure and the information of neighboring nodes. When each node updates its information, the Laplacian matrix of the entire graph is required. When a new node is added to the graph and the information of the node needs to be updated, the original Laplacian matrix also needs to be updated. In other words, when GCN performs convolution operations, all nodes in the entire graph need to participate in a calculation. Once some nodes in the graph change, the graph needs to be retrained and learned. When the scale of the graph is large, this training method may require a lot of training resources.

[0058] GraphSAGE is a classic graph convolution algorithm in the spatial domain. Compared with the traditional GCN algorithm, it updates the node information by randomly sampling the neighbors of the node during training. This method can control the number of nodes involved in the calculation, greatly reduce the training resources, and make distributed training in large-scale graphs possible. In addition, this training method makes SAGE have good generalization. When a new node is added, the information can be updated based on its neighbor aggregation without the need to re-iterate the entire graph. On the other hand, SAGE studies various ways of aggregating neighbor nodes and verifies the advantages and disadvantages of different aggregation methods.

[0059] The node update process of SAGE can be shown in the following pseudo code table 1-1:

[0060] Table 1-1 Pseudo code for SAGE node update

[0061]

[0062]

[0063] In the above Table 1-1, the 1st to 7th lines of the pseudo code are for sampling nodes. Since SAGE samples from the outside to the inside, the traversal in the 2nd line is from k to 1. The 8th to 15th lines of the pseudo code are the aggregation process of the model. The node features are aggregated from the outside to the inside, and finally the feature vector of the central node is obtained. The core code of the aggregation process is from lines 11 to 13; in line 11, the neighbor nodes of the node are aggregated using the aggregation function, and then the aggregated output of these nodes is obtained; and in line 12, these aggregated neighbor features and the features of the previous layer of the central node are spliced, and then input into an MLP layer to obtain a new feature vector; the model normalizes the feature vector in line 13. The aggregation function in SAGE can be selected from average aggregation (mean), pooling aggregation, and LSTM. When the aggregation function is LSTM, since LSTM does not satisfy the permutation invariance and the calculation is complex, it is necessary to first randomly arrange the order of neighbors, and then input the neighbors of the random neighbor sequence into the LSTM.

[0064] The present invention proposes a SAGE-based HetGNN model (named SAGE-HetGNN), which combines HetGNN, delay propagation domain knowledge and homogeneous graph neural network SAGE to study the evolution of railway network delays. GraphSAGE (SAGE) is a classic HomoGNN algorithm that has achieved satisfactory performance in many prediction tasks. It is selected as the benchmark HomoGNN of the present invention and is applied to update all meta-paths in the present invention. SAGE is not a GNN with weighted edges, which means that when node information is updated, the node only captures information from adjacent nodes and cannot consider the attributes of the edges connected to the node. Therefore, the interaction relationship between two different nodes should be assigned to the corresponding nodes, while for the weighted HomoGNN, the interaction attributes between points can be assigned as edge attributes.

[0065] like Figure 1 As shown, in step S102, high-speed railway network sample data is obtained and a training set, a validation set and a test set are constructed, including abstracting entities in the high-speed railway network into nodes containing different attributes, and using meta-paths to describe the relationship between nodes, and constructing a heterogeneous graph containing multiple node types and multiple meta-paths at each timestamp, wherein the node types include running trains, terminating trains, passing stations and terminating stations.

[0066] The railway system is composed of various types of entities (such as stations, trains, tracks, etc.) and the external environment. Under the interaction between various entities, the key issue studied in this paper is how the delays of running trains change after the railway network evolves over a period of time.

[0067] The present invention abstracts entities in the high-speed rail network into nodes with different attributes. The transportation network at any given time is assumed to be a heterogeneous graph consisting of four nodes with different attributes: RT, TT, PS and TS. Metapaths (edges) are used to connect nodes to represent the interaction relationship between nodes. Predicting the delay state of RT nodes in the network at a given time after a certain time interval reveals the evolution of the network delay state.

[0068] 1. Node selection

[0069] The research target schematic diagram of the present invention is as follows Figure 3 As shown, Figure 3 (A) is the current (i.e., time t1) railway network status, and the railway network status to be predicted (i.e., time t2) is as follows: Figure 3 (B) At t1, trains 0-7 are RT, running from the passing station (PS) to the terminal station (TS) TS1; and train 8 has arrived at TS1, so it is the final train (TT). TS1 has three arrival and departure stations ( Figure 3TS1-1, TS1-2, TS1-3 in (B) provide arrival and departure sites for trains on four railway lines, and these arrival and departure sites are interconnected.

[0070] Train interactions are a key factor in the evolution of delays; for example, Figure 3 In (A), if train 1 is delayed and the interval between it and train 0 is very small, for safety reasons, train 0 may be delayed in order to maintain the minimum interval time. Therefore, trains, as entities that affect the evolution of delays, should be regarded as nodes. RT is a running train and is the most suitable entity to represent the delay status of the railway network. Therefore, in the present invention, they are regarded as nodes to be predicted to explore the evolution of the delay status of the railway network. TT is a train that has arrived at the terminal at the current moment, so there is no need to predict their delay status at the next timestamp. Train-train interaction can be used to reflect the impact of adjacent trains on the evolution of delays, including the interaction between RT and TT and between RT and RT. In the present invention, a train will only update its delay status at a station, so the remaining running time of a train in the interval at a given moment is also a key factor in the evolution of delays. In addition, the interval connects adjacent stations, so the minimum running time of the train interval can be used as an indicator to reflect the impact of the interaction between stations on the evolution of delays. For example, in Figure 3 In the example, trains 4 and 5 are running in the same section at t1; at t2, train 4 has already run to the next section, while train 5 is still running in the original section (i.e., train 4 has updated its delayed status, but train 5’s delayed status has not been updated). In this case, the minimum running time of the section is greater than the predicted time interval (Δt). If the remaining running time of the train in the section, the planned running time in the section, and the minimum running time of the section are not considered, the prediction model may not be able to identify whether the delay of train 5 has not changed because it has not arrived at the next station (i.e., it has not arrived at the delay update record point), or whether train 5 has arrived at the next station but the delay has not changed. The former case will not cause the train to change its delay at the next timestamp for all samples, but the latter case only applies to some trains. Therefore, stations must also be regarded as nodes.

[0071] The connection between a station and a train can reflect the interaction between a train and a station, while the interaction between stations can be represented by connecting adjacent stations. The present invention considers two types of stations, namely PS and TS. A PS is crossed by only one railway line in one direction of operation (e.g., Figure 3 There is only one railway line from PS1 to PS2, so PS is considered a node. TS is the convergence point of multiple lines, usually with several arrival and departure stations connected to different railway lines (for example, Figure 3TS1 has TS1-1, TS1-2, and TS1-3 in the example. However, the destination station is the actual entity that affects RT and PS, so the present invention regards the destination station as a TS node. Therefore, the present invention considers four types of nodes, namely, RT nodes (e.g. Figure 3 Train 0 in (A), TT nodes (e.g. Figure 3 (A) Train 8), PS nodes (e.g. Figure 3 (A) PS1) and TS nodes (e.g. Figure 3 (A) TS1-1).

[0072] 2. Node Analysis

[0073] At each timestamp, the present invention abstracts the entity corresponding to the timestamp as a node, and uses the corresponding edge (meta-path) to connect the nodes. Finally, the HetGNN obtained by the present invention is as follows Figure 4 As shown. The railway network heterogeneous graph of the present invention includes four types of nodes, namely RT, TT, PS and TS nodes. According to the delay propagation mechanism analysis and the existing delay prediction research, the present invention allocates the corresponding delay propagation influencing factors to the nodes and designates them as node attributes for studying the evolution of railway network delay status.

[0074] For RT nodes, the current delay (marked as C) is the basis for the delayed state at the next timestamp in the delay evolution. For example, if the current delayed train does not have enough redundant time for it to recover within the prediction time interval, the train will also be delayed at the next timestamp, so C is selected as the attribute of the RT node. The running interval (marked as I) between the train to be predicted and the preceding train reflects the interaction between trains, which is another key factor in the evolution of delays. According to the knowledge of delay propagation, if the train interval is too short, the delayed preceding train is very likely to affect the following train, causing the delay of the following train. On the contrary, when the train interval is large enough, it means that there is sufficient redundant time between trains, and the probability of the following train being affected by the delay of the preceding train is small. The redundant time available for the delayed train to recover can be calculated by combining the interval planned running time (marked as S) and the interval minimum running time (marked as M), that is, the redundant time is S minus M; therefore, S and M are also regarded as influencing factors. It is worth noting that the train considered by the present invention may be at the station at the current timestamp or may be running in the interval. When the train is at the station, the S and M considered by the present invention should be the running time of the next section; and when the train is in the section, it should be the scheduled running time and the minimum running time of the train in the current section. In the present invention, the running (delayed) status of the train will be updated only when the train arrives / leaves the station, that is, the present invention only has one delay record update point at the station. Therefore, some trains may not update their status at two adjacent timestamps, for example, Figure 3Train 5 in the figure does not update the delay at t2 because it runs in the same section at both timestamps. Therefore, the remaining running time of the section (marked as R) should also be considered as an influencing factor. R can be calculated by subtracting the current timestamp from the planned arrival time of the train at the next station. Like S and M, when the train is at the station, R should be the remaining running time of the next section; when the train is in the section, it should be the remaining running time of the train in the current section. All of the above influencing factors are related to the train, so they are all considered node attributes. The attributes of the RT node are summarized as follows:

[0075] C: The delay of the train to be predicted at the current timestamp;

[0076] I: The interval between the train to be predicted and the first preceding train;

[0077] S: When the train is at the station, it is the planned running time of the train to be predicted in the next section; when the train is in the section, it is the planned running time of the train to be predicted in the current section;

[0078] M: When the train is at the station, it is the minimum running time of the train to be predicted in the next section; when the train is in the section, it is the minimum running time of the train to be predicted in the current section;

[0079] R: When the train is at the station, it is the remaining running time of the train to be predicted in the next section; when the train is in the section, it is the remaining running time of the train to be predicted in the current section.

[0080] TT nodes represent the final trains that have arrived at the final station, which means that they do not have the attributes I, S, M and R. Therefore, only the influencing factor C is considered for the TT nodes. The other two types of nodes, namely PS and TS nodes, are station nodes. In the present invention, the number of arrival and departure lines (marked as N) is used as the attribute of the PS node. As mentioned above, the arrival and departure yard of the final station is regarded as a TS node, so the number of arrival and departure lines of the corresponding arrival and departure yard is regarded as an attribute of the TS node. M represents the minimum running time of the train in the interval, which can be used as an attribute of both the station and the train node. When M is designated as a station attribute, it reflects the relationship between the two stations. It should be the attribute of the arrival end station of the interval, because the arrival end station will affect the departure end station of the interval. As the minimum running time of the train in the interval, M can also be expressed as a factor reflecting the operation of the train, so it can also be regarded as a train node attribute. In the present invention, M is regarded as a train node attribute.

[0081] 3. Meta-path analysis

[0082] There are four meta-paths in the railway network heterogeneous graph considered by the present invention: RT-rr-RT, TT-tr-RT, Station-ss-Station and Station-sr-RT. RT-rr-RT is the connection relationship between two adjacent RT nodes, reflecting the interaction between RTs. Since the train-train interaction should be the influence of the preceding train on the following train, the meta-path is a directed edge (for example, Figure 4 RT2 in points to RT1). Another train-related meta-path is TT-tr-RT, which reveals the relationship between the preceding TT node and the succeeding RT node, so it is also directed (e.g., Figure 4 TT1 in the figure points to RT2); since RT and TT are different types of nodes, they have different properties, so the meta-paths RT-rr-RT and TT-tr-RT are different. The Station-ss-Station meta-path reflects the physical structure of the railway network, that is, the station in front should affect the station behind, so it is also directed (for example, Figure 4 The arrival and departure yard of the terminal station is connected to other yards in the form of undirected edges of Station-ss-Station (for example, Figure 4 TS1-1 to TS1-2 in the meta-path). The Station-sr-RT meta-path connects the PS / TS nodes with the RT nodes to reveal the relationship between the train and the next station. When the train is located in the previous section or the previous station, the Station-sr-RT should be a directed edge pointing from the station in front to the train. For example, RT1 runs in the section between PS1 and PS2, PS2 should point to RT1; and RT5 is located at PS7, then TS1-3 should point to RT5. Since SAGE is not a HomoGNN with weighted edges, the present invention only uses 1 and 0 as edge attributes, that is, when the nodes in the meta-path are connected, the attributes of rr, rt, sr, and ss are equal to 1.

[0083] The types and attributes of nodes and edges in the present invention are summarized in Table 1-2:

[0084] Table 1-2 Introduction to nodes and edges and their corresponding attributes

[0085]

[0086] 4. Dataset Construction

[0087] The data used in this paper comes from China Guangzhan Railway Group, and the data period is from March 24, 2015 to November 10, 2016. Since the research object of this paper is the railway network, this paper selects two local high-speed rail networks - Guangzhou South Network (GZS-Net) and Changsha South Network (CSS-Net) as cases to verify the effectiveness and accuracy of the SAGE-HetGNN model proposed in this paper.

[0088] The schematic diagram of the GZS-Net and CSS-Net networks used in the present invention is as follows Figure 5 shown.

[0089] Guangzhou South is the terminus of the Beijing-Guangzhou High-speed Railway, the Nanning-Guangzhou High-speed Railway, the Guiyang-Guangzhou High-speed Railway, the Guangzhou-Zhuhai High-speed Railway, and the Guangzhou-Shenzhen High-speed Railway, while Changsha South is a multi-line connection station, which can be regarded as a connecting station for the Beijing-Guangzhou High-speed Railway and the Shanghai-Kunming High-speed Railway. In the Guangzhou South network, the present invention has selected a total of 12 stations, including Qingyuan, Guangzhou North, Qingsheng, Humen, Bijiang, Beijiao, Foshan West, Sanshui South, Zhaoqing East, Yunfu East, Guangning, and Guangzhou South. Among them, Guangzhou South Station includes two arrival and departure yards (Beijing Square and Guangzhou-Zhuhai Yard). The present invention divides five TS nodes according to the lines connected to Guangzhou South Station. If trains from two lines may share an arrival and departure yard, the present invention connects the corresponding TS nodes; for example, trains from the Beijing-Guangzhou and Nanning-Guangzhou High-speed Railways may arrive at the same arrival and departure yard, so the present invention connects the two TS nodes of "Guangzhou South-Nanning-Guangzhou" and "Guangzhou South-Beijing-Guangzhou".

[0090] The Changsha South network includes 8 stations, namely Shaoshan South, Xiangtan North, Hengshan West, Zhuzhou West, Liling East, Miluo East, Yueyang East, and Changsha South. Changsha South Station also includes two arrival and departure yards (Wu Square and Shanghai-Kunming Station) to receive and send trains from the Beijing-Guangzhou and Shanghai-Kunming high-speed railways. Since the cross-line trains at Changsha South Station basically only stay at the arrival and departure yards corresponding to the incoming lines, the present invention provides the following connection methods for the TS nodes at Changsha South Station: Figure 5 The present invention only considers the train heading to the final destination.

[0091] The samples used in the present invention should contain constant station nodes (PS and TS) and different numbers of RT and TT nodes at each timestamp, such as Figure 4As shown. The present invention sets the prediction time interval to 20 minutes, that is, based on the current delay status of the railway network, the delay of all RTs in the network after 20 minutes is predicted. TT has arrived at the final destination at each timestamp, so there is no need to predict its delay in the next timestamp. The prediction period selected by the present invention only considers 8:00 to 23:00 every day (that is, the timestamps corresponding to all heterogeneous graphs should be 8:00, 8:20, ..., 23:00), because there are too few trains running in the railway network during other periods (or no trains running), which is not representative of the evolution of network delays. Samples with RT node delays of more than 90 minutes are deleted. On the one hand, excessive delays may cause the train to be cancelled when arriving at the station; on the other hand, the proportion of these delays is too small to be fully trained. This deletion method can ensure the statistical significance of the SAGE-HetGNN model and obtain sufficient training, and once there is enough delay data greater than 90 minutes, the range of delays considered for deletion can be extended. The delay is 0 minutes when arriving on time, and it is a negative value when arriving early. Therefore, a "delay" can be assigned to each train. The delays considered in the present invention include early arrival, on-time arrival and late arrival. It is necessary to consider these three situations at the same time because the delay of the train to be predicted is unknown before the next timestamp. Therefore, if only the late arrival is applied to the training model, in actual application, some early and on-time arrivals will also be predicted as late arrivals. Finally, the present invention shuffles the order of all the obtained data sets, uses 60% of the data for model training, 20% of the data as a validation data set to adjust the model hyperparameters, and the remaining 20% ​​for model testing.

[0092] like Figure 1 As shown, in step S103, the constructed SAGE-HetGNN model is trained, verified, and the model hyperparameters are adjusted and tested based on the training set, validation set, and test set to obtain the final SAGE-HetGNN model.

[0093] The SAGE-HetGNN model and its related hyperparameters used in the present invention are as follows: Figure 6 As shown. In the first step, the four different types of nodes are mapped to the same dimensional space, that is, all nodes are mapped to 256-dimensional vectors; then, a four-layer SAGE neural network is used to update the node information in each meta-path, each layer of the neural network has 256 neurons, and the ReLU function is used between two adjacent layers for activation; then, the updated node information is aggregated by using the "sum" aggregation function; finally, the predicted value of each RT node in the current network at the next timestamp (i.e., 20 minutes later) is obtained. The present invention uses the mean absolute error (MAE) as the loss function to train the SAGE-HetGNN model, and the expression of MAE is shown in the following formula.

[0094]

[0095] Among them, y i and i ' corresponds to the actual delay and predicted delay of the RT node respectively, and N represents the sample size of the data set. The present invention uses the Adam optimizer with an initial learning rate of 0.001 as the model optimizer. When the loss function of the verification data set decreases by less than 0.01 within 10 iterations (epochs), the global learning rate of the optimizer is reduced to 10% of the original. In order to prevent the model from overfitting, the present invention also uses the "Earlystopping" technology during model training. When the loss function of the verification data set decreases by less than 0.01 in 30 iterations, the model stops training. In addition, in order to save training time and computing resources, the present invention sets the maximum number of iterations to 300, and uses the dataloader technology with a batchsize of 64.

[0096] like Figure 1 As shown, in step S104, based on the delay status of the high-speed railway network at the current timestamp, the obtained SAGE-HetGNN model is used to obtain the delay prediction value of all running trains contained in the high-speed railway network at the next timestamp after a preset time interval.

[0097] 1. Late prediction

[0098] The present invention sets the prediction time interval to 20 minutes, that is, based on the delay status of the current railway network, predicts the delay of all RTs in the network 20 minutes later, and obtains the delay prediction value of all RTs in the network through the trained SAGE-HetGNN model.

[0099] 2. Comparative analysis of the prediction performance of SAGE-Het and existing research methods

[0100] Furthermore, in order to verify the high accuracy of the SAGE-Het model of the present invention, the prediction performance of the SAGE-HetGNN model was compared with existing late prediction models (Keep Constant, RF, SVR and ANN) and some HetGNN models (HGT and HAN) that have achieved great success in other prediction tasks. The prediction performance of SAGE-Het and other benchmark models is shown in Figure 2. Figure 7 shown.

[0101] The results show that SAGE-Het has the best prediction performance on GZS network and CSS network. Keep Constant has the worst prediction effect, because train operation is a dynamic process, and dispatchers adjust the train operation plan according to the real-time delay situation and timetable structure to reduce delays. Therefore, it is unreasonable to regard delays as constant. Among the commonly used machine learning algorithms (i.e., RF, SVR, and ANN), RF performs quite well in RMSE in GZS-Net, but SAGE-Het has a 15% decrease in MAE compared to RF. For CSS-Net, ANN has better prediction effect than RF and SVR. Compared with ANN, SAGE-Het decreased by 11.2% and 20.7% in RMSE and MAE, respectively. The two existing efficient HetGNNs (i.e., HAN and HGT) considered by the present invention do not have better prediction accuracy than SAGE-Het. This may be because the present invention does not consider the influence of too many entities when considering train influencing factors, so simpler models can achieve better performance.

[0102] 3. Comparison between predicted value and true value

[0103] In order to study the prediction performance of SAGE-Het at different delay lengths, Figure 8 The actual delays and the predicted delays based on SAGE-Het are compared in Figure 2. The results show that almost all points are distributed near the 45-degree oblique line, which indicates that the difference between the actual value and the predicted value is very small, that is, SAGE-Het has a very superior prediction ability. In addition, as the actual delay increases, the corresponding predicted delay does not deviate from the 45-degree oblique line, proving that SAGE-Het has stable prediction performance for both short and long delays.

[0104] 4. Analysis of the impact of meta-path on delay evolution

[0105] This section reveals the effect of the interaction between different nodes on the evolution of network delay status from a spatial dimension by comparing the prediction accuracy when different meta-paths are removed from the SAGE-Het model. First, the prediction accuracy of the model is explored when the interaction between trains is not considered (i.e., meta-paths RT-rr-RT and TT-tr-RT), and the effect of the interval time between trains on the interaction between trains is explored. Subsequently, the prediction accuracy of the SAGE-Het model is explored when the interaction between stations and trains (i.e., meta-path Station-sr-RT) and between stations and stations (i.e., meta-path Station-ss-Station) is not considered, revealing their effects on the evolution of network delays.

[0106] From the delay propagation mechanism and the results of the previous chapters, we can see that the interaction between adjacent trains is a key factor in the delay propagation process. However, the interaction between trains is greatly affected by the interval time between adjacent trains. For example, when the interval time is small, even if the leading train is slightly delayed, the following train is easily affected by the leading train and delayed; when the interval time between trains is large, even if the leading train is significantly delayed, there is enough redundant time between trains, which can ensure that the following train maintains a sufficient interval time with the leading train, thus maintaining normal operation.

[0107] Heterogeneous graph structures allow different numbers and types of nodes to be connected to reflect the interaction relationship between nodes. Therefore, the present invention will verify the effectiveness of train interaction in different situations by using different connection methods between trains. In the SAGE-Het model, two types of interaction relationships between trains are considered, namely the relationship between adjacent RTs and the relationship between the preceding TT and the following RT. First, the present invention studies whether the interaction relationship between trains has an impact on delay propagation. Therefore, the present invention cuts off all connections between trains (i.e., deletes the meta-paths RT-rr-RT and TT-tr-RT). Based on this method, GNN will not update the train-related meta-paths (RT-rr-RT and TT-tr-RT) when updating the RT node information. At this time, the RT information update will only use the information of its own current timestamp and the impact of the station on the RT from the meta-path Station-sr-RT. The present invention marks this method as SAGE-Het-selflink.

[0108] According to the above analysis of the delay propagation mechanism, if the interval between adjacent trains is large enough, the possibility of the following train being affected by the preceding train is small (i.e., in this case, the train interaction may be insignificant to the following train). Therefore, the present invention sets different thresholds, and when the interval between adjacent trains exceeds this threshold, the connection between the two trains is disconnected. The present invention defines this method as SAGE-Het-cut-J, where J represents this threshold. For example, SAGE-Het-cut-10 means that when the interval between adjacent RTs and between the preceding TT and the following RT exceeds 10 minutes, the connection between the edges of these trains will be deleted. The minimum tracking interval between adjacent trains can reach 3 minutes (for example, when two passing trains are running in the section), and most of the time, the tracking interval will still be maintained at least 5 minutes. The present invention regards the 10-minute interval as a medium interval between the two trains, and regards the 20-minute interval as a long interval between the two trains. Therefore, the present invention verifies the prediction effect when the thresholds are 3, 5, 10 and 20 minutes, and the prediction results corresponding to different thresholds are shown in Table 1-3.

[0109] Table 1-3 Prediction results of SAGE-het under different edge connection modes (GZS-net)

[0110]

[0111] Note: The increase in related indicators is the growth percentage of the corresponding model relative to SAGE-Het.

[0112] Table 1-4 Prediction results of SAGE-het under different edge connection modes (CSS-net)

[0113]

[0114] Note: The increase in related indicators is the growth percentage of the corresponding model relative to SAGE-Het.

[0115] Tables 1-3 and 1-4 show that the prediction performance of SAGE-Het_selflink is significantly lower than that of SAGE-Het, which indicates that the interaction between trains is crucial to the propagation of delays. When the intervals between different trains are disconnected, the prediction performance of SAGE-Het-cut-3 and SAGE-Het-cut-5 is greatly reduced compared with SAGE-Het. This is because when the interval between adjacent trains is very small, the following train can only slow down its own operation to maintain the minimum tracking interval time, at which time the leading train will have a significant impact on the following train. As the threshold of disconnection of the present invention increases, the prediction accuracy of the corresponding model gradually increases. The performance of SAGE-Het-cut-20 is almost equivalent to that of the SAGE-Het model, which indicates that if the interval time between trains is sufficient, the interaction between trains no longer has a significant impact on the trains.

[0116] In existing studies, no matter how long the interval between trains is, it is considered that there is interaction between trains. Because in order to ensure the consistency of all samples, the existing research methods, when considering the role of the preceding train, all samples must input the relevant information of the preceding train into the model regardless of how long the interval between trains is, the SAGE-Het model proposed in the present invention can solve this shortcoming by modifying the connection method of the trains. In the present invention, all adjacent trains (stations) are considered to be interconnected, but in future studies, the connection method between edge points can be learned by adding an encoder module to the model to obtain the optimal connection between points.

[0117] In addition, the present invention also analyzes the effect of different meta-paths on the evolution of network delay status. In the SAGE-Het model, there are meta-paths RT-rr-RT and TT-tr-RT that reflect the interaction between trains, the meta-path Station-sr-RT that reflects the interaction between trains and stations, and the meta-path Station-ss-Station that reflects the interaction between stations and stations. By comparing the prediction performance of SAGE-Het-selflink and SAGE-Het, the importance of the interaction between trains and trains to the evolution of network delay status is verified. Using the same research idea, the present invention is based on the SAGE-Het model, and does not consider the interaction between stations and trains, and obtains the model SAGE-Het-NoStaTra, that is, in the SAGE-Het-NoStaTra model, the meta-path Station-sr-RT is not considered. Similarly, the model that does not consider the interaction between stations can be represented by the SAGE-Het-NoStaSta model, in which the meta-path Station-ss-Station will not be considered. The present invention compares the prediction accuracy of models SAGE-Het, SAGE-Het-selflink, SAGE-Het-NoStaTra, and SAGE-Het-NoStaSta to reveal the impact of different meta-paths on the evolution of network delay states, and the results are shown in Tables 1-5.

[0118] Table 1-5 does not consider the prediction effect of different meta-paths

[0119]

[0120]

[0121] As can be seen from Tables 1-5, SAGE-Het has the highest prediction accuracy. This result indicates that the prediction accuracy of the model that does not consider any meta-path is lower than that of the model that considers all meta-paths, which verifies that the various meta-paths considered in the present invention all play an important role in the evolution of network delay status. By comparing the prediction accuracy of the SAGE-Het-selflink, SAGE-Het-NoStaTra, and SAGE-Het-NoStaSta models, it is found that the prediction effect of SAGE-Het-selflink is the worst, proving that the SAGE-Het model has the greatest impact when the interaction between trains is not considered; the prediction accuracy of the SAGE-Het-NoStaTra and SAGE-Het-NoStaSta models is similar, proving that the interaction between stations and trains, and between stations has a similar impact on the SAGE-Het model.

[0122] 5. Analysis of the impact of different prediction intervals on the prediction effect of delay evolution

[0123] The present invention predicts the delay in 20 minutes based on the constructed SAGE-Het model according to the current railway network delay status. In order to study the prediction effect of SAGE-Het after different prediction time intervals, the present invention also studies the prediction accuracy of the SAGE-Het model and other benchmark models at 10-minute and 30-minute prediction time intervals, as shown in Tables 1-6 and 1-7.

[0124] Table 1-6 Model prediction accuracy at different prediction time intervals (GZS-Net)

[0125]

[0126]

[0127] Table 1-7 Model prediction accuracy at different prediction time intervals (CSS-Net)

[0128]

[0129] As can be seen from Tables 1-6 and 1-7, on GZS-Net and CSS-Net, the prediction performance of all models decreases as the prediction time interval increases. This is because the longer the advance prediction time, the greater the uncertainty of the system, and the greater the prediction difficulty at this time. However, in the present invention, SAGE-Het is better than other benchmark models at three different prediction time intervals, proving the excellence and stability of SAGE-Het's prediction performance.

[0130] In order to better understand the absolute prediction performance of SAGE-Het at different prediction time intervals, the residual distribution at different prediction time intervals and the cumulative distribution function (CDF) of the absolute value of the residuals are also calculated. The results show that SAGE-Het achieves satisfactory performance at different prediction time intervals for both railway networks, and the residual distribution almost follows a bell-shaped distribution, which indicates that the residuals satisfy the zero mean and normal distribution assumptions.

[0131] The embodiment of the present invention provides a high-speed railway network delay deduction method based on multi-source heterogeneous data processing, which is based on combining the HetGNN heterogeneous graph neural network with the GraphSAGE homogeneous graph neural network to construct a SAGE-HetGNN model; obtaining high-speed railway network sample data and constructing a training set, a validation set and a test set, including abstracting entities in the high-speed railway network into nodes containing different attributes, and using meta-paths to describe the relationship between nodes, and constructing a heterogeneous graph containing multiple node types and multiple meta-paths at each timestamp, wherein the node types include running trains, terminal trains, passing stations and terminal stations; based on the training set, validation set and test set, the constructed SAGE-HetGNN model is trained, validated, and the model hyperparameters are adjusted and tested to obtain the final SAGE-HetGNN model; based on the delay status of the high-speed railway network at the current timestamp, the obtained SAGE-HetGNN model is used to obtain the delay prediction value of all running trains contained in the high-speed railway network at the next timestamp after a preset time interval. In response to the problem of railway network delay evolution, the present invention comprehensively considers the interaction between different types of entities in the high-speed railway network, and designs a new neural network model SAGE-HetGNN model by combining the SAGE homogeneous graph neural network and the HetGNN structure. The SAGE-HetGNN model can process different types of nodes and capture the interaction between nodes. Based on the SAGE-HetGNN model, the evolution of railway network delays can be studied by predicting the delays of running trains at future timestamps, and it has high prediction accuracy.

[0132] The above-mentioned embodiments only express the specific implementation of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A high-speed rail network delay prediction method based on multi-source heterogeneous data processing, characterized in that: The following steps are involved: Step 1: Based on the HetGNN heterogeneous graph neural network and the GraphSAGE homogeneous graph neural network, the SAGE-HetGNN model is constructed by combining them; Step 2: Obtain sample data of the high-speed railway network and construct training sets, validation sets, and test sets. Abstract entities in the high-speed railway network into nodes with different attributes, use meta-paths to describe the relationships between nodes, and construct a heterogeneous graph containing multiple node types and multiple meta-paths at each timestamp. In step 2, the meta-path includes RT-rr-RT, TT-tr-RT, Station-ss-Station and Station-sr-RT; RT-rr-RT represents the connection relationship between two adjacent RT nodes, reflecting the interaction between RT nodes; TT-tr-RT represents the relationship between the preceding TT node and the succeeding RT node; the Station-ss-Station meta-path reflects the physical structure of the railway network, that is, the station in front affects the station in the rear; the Station-sr-RT meta-path connects the PS / TS node with the RT node, indicating the relationship between the train and the next station; wherein: the RT node represents the running train, the TT node represents the terminal train, the PS node represents the passing station, and the TS node represents the terminal station; In step 2, the entities in the high-speed railway network are abstracted into nodes with different attributes, including running train nodes, terminal train nodes, passing station nodes and terminal station nodes; wherein: The properties of the running train node include: C: the delay of the train to be predicted at the current timestamp; I: the interval time between the train to be predicted and the first preceding train; S: when the train is at the station, it is the planned running time of the train to be predicted in the next section; when the train is in the section, it is the planned running time of the train to be predicted in the current section; M: when the train is at the station, it is the minimum running time of the train to be predicted in the next section; when the train is in the section, it is the minimum running time of the train to be predicted in the current section; R: when the train is at the station, it is the remaining running time of the train to be predicted in the next section; when the train is in the section, it is the remaining running time of the train to be predicted in the current section; The attributes of the terminal train node include: C: the delay of the train to be predicted at the current timestamp; The attributes of the passing station node include: N: the number of arrival and departure lines of the station; The attributes of the terminal node include: N: the number of arrival and departure lines of the corresponding arrival and departure yard; Step 3: Based on the training set, validation set and test set, the constructed SAGE-HetGNN model is trained, validated, and the model hyperparameters are adjusted and tested to obtain the final SAGE-HetGNN model; Step 4: Based on the delay status of the high-speed railway network at the current timestamp, the final SAGE-HetGNN model is used to obtain the delay prediction value of all running trains contained in the high-speed railway network at the next timestamp after a preset time interval.

2. The high-speed rail network delay deduction method based on multi-source heterogeneous data processing according to claim 1 is characterized in that: The step 4 is specifically as follows: Four different types of initial node information are mapped to the same dimensional space, and all nodes are mapped to 256-dimensional vectors; a four-layer SAGE neural network is used to update the node information in each meta-path, each layer of the neural network has 256 neurons, and the ReLU function is used for activation between two adjacent layers; the updated node information is aggregated by using the sum aggregation function; finally, the delay prediction value of each running train node in the current high-speed railway network at the next timestamp is obtained.

3. The high-speed rail network delay deduction method based on multi-source heterogeneous data processing according to claim 2 is characterized in that: In step 4, if the train arrives ahead of time, the delay value of the train is a negative value; if the train arrives on time, the delay value of the train is 0; if the train arrives late, the delay value of the train is a positive value.

4. The high-speed rail network delay deduction method based on multi-source heterogeneous data processing according to claim 1 is characterized in that: The step 3 is as follows: Use mean absolute error (MAE) as the loss function to train the SAGE-HetGNN model; The Adam optimizer with an initial learning rate of 0.001 was used as the model optimizer. When the loss function of the validation dataset dropped less than 0.01 within 10 iterations, the global learning rate of the optimizer was reduced to 10% of the original value. To prevent overfitting of the model, the model training was stopped when the loss function of the validation dataset dropped less than 0.01 in 30 iterations. To save training time and computing resources, the maximum number of iterations was set to 300, and the dataloader technology with a batchsize of 64 was used.

5. The high-speed rail network delay deduction method based on multi-source heterogeneous data processing according to claim 1 is characterized in that: Also includes: By comparing the prediction accuracy of the SAGE-HetGNN model when different meta-paths are eliminated, the effect of the interaction between different nodes on the evolution of the delay status of the high-speed railway network is revealed from the spatial dimension.

6. The high-speed rail network delay deduction method based on multi-source heterogeneous data processing according to claim 1 or 5 is characterized in that: Also includes: By comparing the prediction accuracy of the SAGE-HetGNN model at different prediction time intervals, the prediction effect of the SAGE-HetGNN model on the delay status of the high-speed railway network at different prediction time intervals is verified from the time dimension.

Images