Rail transit short-term OD passenger flow prediction method based on node-edge space-time fusion

Abstract

This invention discloses a short-term origin-destination (OD) passenger flow prediction method for rail transit based on node-edge spatiotemporal fusion. First, the OD prediction task is decoupled into two sub-tasks: deep spatiotemporal feature learning of nodes and temporal feature extraction of edges. These are processed separately by physically separated dual-path heterogeneous encoders. Next, an improved adaptive graph convolutional recurrent network with learnable embedded dynamically generated graph structures is used to fuse node spatiotemporal dependencies. Finally, at the aggregation layer, OD passenger flow predictions are generated based on origin node representations, destination node representations, edge temporal features, and OD relationship type features. This invention fundamentally reduces the number of core model parameters to the O(N) order of magnitude by decoupling heterogeneous data, asynchronously fusing dual-path information, and deeply fusing domain knowledge. It achieves high-precision, highly interpretable, and strongly generalizable short-term OD passenger flow prediction, making it particularly suitable for large-scale networks and for predicting sparse OD pairs and newly opened stations.

Description

Technical Field

[0001] This invention relates to the field of short-term passenger flow forecasting technology for rail transit, and in particular to a short-term OD passenger flow forecasting method for rail transit based on node-edge spatiotemporal fusion. Background Technology

[0002] With the rapid development of urban rail transit systems, passenger flow forecasting has become a key issue in rail transit management, serving as an important means to improve operational efficiency, alleviate traffic congestion, optimize resource allocation, and enhance passenger experience. Short-term origin-destination (OD) passenger flow forecasting is of significant practical importance for anticipating passenger demand at different lines and stations within a future timeframe, thereby guiding operational scheduling and resource allocation. However, traditional forecasting methods based on statistics and simple machine learning often overlook the complex spatiotemporal dependencies within rail transit networks and the dynamic relationships between stations, resulting in inadequate performance when facing complex passenger flow patterns and external influencing factors.

[0003] In recent years, deep learning, especially graph neural networks (GNNs), has made significant progress in processing spatiotemporal data with graph structures. Graph convolutional networks (GCNs) and their extended models in spatiotemporal prediction, such as spatiotemporal graph convolutional networks (STGCNs) and attention-based spatiotemporal graph convolutional networks (ASTGCNs), have become the mainstream technical approach for rail transit passenger flow prediction by combining graph convolution with time series models (such as CNNs and RNNs).

[0004] However, these existing methods still have some technical limitations when applied to short-term OD prediction:

[0005] 1. Large number of parameters and poor scalability: Models like STGCN typically predict or reconstruct the entire N×N dimensional OD matrix directly. The size of its core parameters is proportional to the square of the number of stations N, i.e., O(N²). As the network size increases, the number of model parameters increases dramatically, leading to low training efficiency and difficulty in scaling to large-scale urban rail networks.

[0006] 2. Rigid spatial dependency modeling: Classic methods such as GCN and STGCN rely heavily on predefined static adjacency matrices (usually based on physical distance or line connection), which makes it impossible to capture the functional relationships between stations that change dynamically with time and passenger flow patterns. For example, there may be passenger flow interaction between two non-adjacent commuter hub stations.

[0007] 3. Heterogeneous Feature Hybrid Processing: Existing methods generally concatenate or mix node features describing the static attributes of a site with edge features describing the dynamic changes between OD pairs at the input. This approach ignores the fundamental differences between node features (static / quasi-static) and edge features (strong temporal sequence) in terms of data nature and optimal processing paradigms, making it difficult for the model to fully explore and utilize these two types of heterogeneous information, thus limiting the improvement of prediction accuracy.

[0008] While recent research has explored dynamic graph construction and multi-task learning to address these issues, it has yet to propose a systematic solution based on the physical nature of the prediction task. Specifically, there is a lack of a complete framework that can thoroughly decouple the heterogeneous information flows of nodes and edges, design a matching asynchronous processing and fusion mechanism, and ultimately achieve efficient and interpretable predictions. Summary of the Invention

[0009] The purpose of this invention is to address the aforementioned problems by providing a short-term OD (Original Departure) passenger flow prediction method for rail transit based on node-edge spatiotemporal fusion. This method aims to solve the issues of existing methods' difficulty in perceiving high-order correlation features between heterogeneous nodes and their low accuracy in diagnosing minority and marginal samples. The goal is to improve the accuracy, efficiency, and interpretability of short-term OD passenger flow prediction by decoupling the OD prediction task into deep spatial modeling of node (station) features and temporal modeling of edge (OD pair) features. Finally, an innovative mechanism based on node representation is used to achieve the final OD flow prediction. A novel "decoupling-aggregation" prediction paradigm is proposed: First, node and edge features are thoroughly decoupled physically and logically at the input and processing paths. Then, an improved adaptive spatiotemporal fusion module is used to deeply mine the dynamic correlations between nodes. Finally, at the output layer, the starting node representation, ending node representation, edge temporal patterns, and prior relation type knowledge are aggregated to generate interpretable prediction results.

[0010] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows:

[0011] According to one aspect of the present invention, a method for short-term OD passenger flow prediction of rail transit based on node-edge spatiotemporal fusion is provided, comprising the following steps:

[0012] S1. Collect target rail transit network data, and extract node features and edge features from it. The node features include station static features and quasi-static features, and the edge features include the historical passenger flow sequence of each OD pair. Divide the feature dataset into training set, validation set and test set.

[0013] S2. Construct a node-edge dual-path spatiotemporal fusion graph learning model. The processing flow of the node-edge dual-path spatiotemporal fusion graph learning model includes the following steps:

[0014] S201. Input and feature decoupling: Physically separate the node features and edge features, and input them to the node processing path and edge processing path respectively;

[0015] S202, Dual-path asynchronous coding:

[0016] a. Node Path: Node features are mapped to initial node features via a multilayer perceptron encoder and input into the spatiotemporal fusion module. This module is an improved adaptive graph convolutional recurrent network. It dynamically generates an adaptive adjacency matrix using a learnable node embedding matrix and incorporates a gated recurrent mechanism to jointly model the dynamic spatial dependencies and temporal evolution between nodes, outputting the node's deep spatiotemporal features. The adaptive adjacency matrix is ​​represented by the following formula:

[0017]

[0018] in, An embedding matrix for learnable nodes; To modify the linear unit activation function to filter out negative or very weak similarity connections; The function is used to perform row-wise normalization;

[0019] b. Edge path: The historical passenger flow sequence of each OD pair is encoded by a long short-term memory network combined with an attention mechanism to extract the edge temporal features of each OD pair;

[0020] S203. OD Aggregation Prediction Based on Node Representation: The depth features of the starting node, the depth features of the ending node, the corresponding edge temporal features, and the predefined OD relationship type features of the OD pair to be predicted are concatenated. The predicted passenger flow of the OD pair is obtained through regression calculation. The predicted passenger flow of the OD pair is expressed by the following formula:

[0021]

[0022] in, Indicates the predicted location from the site Arrive at the station OD passenger traffic; and Representing the starting point and the end point Node depth features; Indicates OD pair The temporal characteristics; Indicates OD pair Pattern clustering embedding features; and Learnable weights and biases for the regression layer;

[0023] S3. Use the training set to train the model end-to-end, use the validation set to monitor the model performance and adjust the parameters, and use the test set to test and evaluate the model after training.

[0024] S4. Use the trained model to predict the OD passenger flow of the target rail transit network in the near future.

[0025] Preferably, in step S1, the node features include basic features, temporal features, clustering features, and POI features.

[0026] Preferably, the POI features are constructed using a dedicated classification system based on the mechanism of passenger flow generation and attraction in rail transit. The dedicated classification system includes categories such as transportation services, catering services, shopping services, healthcare services, accommodation services, scenic spots, commercial residences, government agencies and social organizations, public transportation facilities, financial and insurance services, companies and enterprises, living and public services, and cultural, leisure and scientific education functions.

[0027] Preferably, in step S202, the initial features of the node are represented by the following formula:

[0028]

[0029] in, For all The initial feature matrix of each site.

[0030] Preferably, in step S202, the long short-term memory network includes an input gating unit, a forgetting gating unit, and an output gating unit;

[0031] The input gating unit is represented by the following formula:

[0032]

[0033] The forget gate unit is represented by the following formula:

[0034]

[0035] The output gate is represented by the following formula:

[0036]

[0037]

[0038] in, , , , , , , , These are learnable parameters.

[0039] Preferably, in step S203, the edge timing feature is represented by the following formula:

[0040]

[0041] in, Indicates OD pair Temporal pattern characteristics; Indicates OD pair Historical passenger flow sequence; This represents the parameters of the LSTM network.

[0042] Preferably, in step S203, the OD relationship type feature is an embedding vector obtained by pattern clustering and mapping of historical OD data.

[0043] Preferably, in step S3, the model training is optimized using a robust objective function that combines Huber loss and relative error loss.

[0044] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0045] 1. To address the issues of large parameter count and poor scalability in existing methods, this invention innovatively proposes an OD aggregation prediction mechanism based on node representation. This mechanism indirectly predicts OD traffic by learning and fusing deep state representations of the starting and ending points, as well as the temporal pattern features of OD pairs. This reduces the number of core parameters and storage overhead of the model from the O(N²) level to the O(N) level, fundamentally solving the scalability problem in large-scale networks, and adhering to the physical nature of passenger flow originating from station interactions.

[0046] 2. To address the rigidity of existing spatial dependency modeling methods, this invention constructs an improved AGCRN as the core module for spatiotemporal fusion. It dynamically learns the spatial dependencies between sites through adaptive graph convolution and utilizes a gated recurrent mechanism to model the temporal evolution of node states, achieving deep fusion and accurate modeling of complex spatiotemporal features. This enables it to capture dynamic functional associations that transcend physical connections. In particular, the learnable adaptive adjacency matrix can discover and quantify geographically disproportionate but functionally strongly related site pairs (such as cross-regional commuter corridors) from the data—a capability lacking in static graph methods.

[0047] 3. To address the shortcomings of existing methods in handling heterogeneous features, this invention designs a dual-path feature extraction mechanism that separates node and edge features. It employs the optimally matched Multilayer Perceptron (MLP) and Long Short-Term Memory (LSTM) networks to specifically process static node attributes and dynamic edge temporal sequences, respectively, achieving full mining of heterogeneous information and avoiding feature confusion. This "decoupling" process lays the foundation for subsequent "asynchronous fusion," allowing both types of features to learn in their respective optimal representation spaces, with targeted aggregation only occurring in the final prediction stage.

[0048] 4. This invention deeply integrates knowledge from the rail transit field, constructs a dedicated POI classification system to accurately characterize station functions, and introduces two-level clustering analysis (node ​​function clustering and OD pattern clustering) to provide prior guidance for the model, specifically enhancing the model's generalization ability and robustness in dealing with practical challenges such as sparse OD pairs and newly opened stations. Among these features, the OD pattern clustering embedding feature, as a "relationship type prior," directly participates in the final prediction, which is one of the key design features of this invention to improve the model's semantic understanding and generalization ability. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0050] Figure 2 This is a schematic diagram of the model framework of the present invention;

[0051] Figure 3 This is a schematic diagram of the LSTM module of the present invention;

[0052] Figure 4 This is a schematic diagram of the AGCRN module of the present invention. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and preferred embodiments. However, it should be noted that many details listed in the specification are merely to provide the reader with a thorough understanding of one or more aspects of the invention, and these aspects of the invention can be achieved even without these specific details.

[0054] Please see Figures 1 to 4 This invention provides a method for predicting short-term origin-destination (OD) passenger flow in rail transit based on node-edge spatiotemporal fusion, and the technical solution is as follows:

[0055] A short-term OD passenger flow prediction method for rail transit based on node-edge spatiotemporal fusion includes the following steps:

[0056] S1. Data Acquisition and Preprocessing: Collect historical card swiping data, station information, and surrounding points of interest (POI) data of the target rail transit network, construct a historical OD matrix sequence, extract node features and edge features respectively, and divide the complete dataset into training set, validation set, and test set according to time.

[0057] Specifically, rail transit prediction data naturally includes two types of features: nodes (stations) and edges (OD pairs). Two heterogeneous data sources are collected: node features describing the station's own attributes and edge features describing passenger flow interactions between stations. Node features describing the station's own attributes include basic features, temporal features, clustering features, and POI features. Edge features describing passenger flow interactions between stations include historical OD passenger flow sequences. Among these, POI (Point of Interest) data is key information for characterizing land use and passenger attraction potential around stations. The POI classification system includes 13 functional categories: Transportation Services (sales, repair, maintenance, etc.), Catering Services, Shopping Services, Healthcare Services (disease prevention, health care, etc.), Accommodation Services, Scenic Spots, Commercial and Residential (business offices, residences, etc.), Government Agencies and Social Organizations (administration, social affairs, etc.), Public Transportation Facilities, Financial and Insurance Services, Corporate and Enterprise Services, Daily Life and Public Services (daily services, facility support, etc.), and Culture, Leisure, and Education (leisure and entertainment, education and research, etc.). The POI feature vector for each station is constructed by standardizing the number of POIs from these 13 categories, forming a 328×13 POI feature matrix. This feature system can finely characterize the functional composition around the station, providing rich semantic information for the model and helping to identify the passenger flow generation and attraction mechanisms of different types of stations.

[0058] S2. Model Construction: Construct a node-edge dual-path spatiotemporal fusion graph learning model. Encode node features and edge features using the node-edge dual-path spatiotemporal fusion graph learning model. Input the initial feature representation of the node after encoding the node features into the spatiotemporal fusion module of the model, and output the depth spatiotemporal features of the node. For any OD pair, the node depth features of its start and end points, the corresponding edge temporal features, and the predefined OD pattern clustering embedding are concatenated. The predicted passenger flow is calculated through a fully connected regression layer.

[0059] Specifically, a node-edge dual-path spatiotemporal fusion graph learning model is constructed. The data processing method for the node-edge dual-path spatiotemporal fusion graph learning model includes the following steps:

[0060] Data input and feature separation: Receive rail transit prediction data and separate and structure the raw heterogeneous data according to its physical nature. By physically separating these two types of features, they are input into two independent processing paths.

[0061] Dual-path feature encoding: Node features and edge features are input into two independent processing paths for feature encoding.

[0062] Node feature encoding path: The concatenated multi-source node features are input into a multilayer perceptron (MLP) encoder. This encoder maps high-dimensional, heterogeneous site attributes into a unified, low-dimensional, dense initial feature representation of the nodes. .

[0063] Specifically, the multi-source static and quasi-static attributes of the stations are transformed into dense numerical representations. For N stations in the network, four types of features are extracted and constructed:

[0064] 1) Basic traffic characteristics, such as historical statistics on inbound and outbound traffic;

[0065] 2) Temporal context features, such as the periodic encoding of the current moment, and the identification of holidays and peak periods;

[0066] 3) Site function clustering features, obtained through pre-clustering based on historical passenger flow pattern similarity;

[0067] 4) Rail transit-specific POI features are feature vectors formed by reclassifying and statistically analyzing the influence mechanism of points of interest around the station on passenger flow generation and attraction.

[0068] The above features, after being concatenated, are input into a multilayer perceptron (MLP) consisting of multiple fully connected layers and ReLU activation functions. Its mathematical expression is as follows:

[0069]

[0070] in, For all The initial feature matrix of each site, this path maps high-dimensional heterogeneous node features into a low-dimensional, dense and learnable semantic representation.

[0071] Edge feature encoding path: The historical passenger flow sequence of each OD pair is input into a Long Short-Term Memory (LSTM) network and subsequent processing layers. This process mines the dynamic evolution pattern of OD passenger flow over time and outputs edge feature representations characterizing the temporal patterns of each OD pair. .

[0072] Specifically, edge feature encoding paths are used to mine the dynamic temporal dependencies between OD pairs. For any OD pair in the network... , take it to the past The historical passenger flow at consecutive time steps is constructed into a sequence:

[0073]

[0074] All OD sequences are input into a Long Short-Term Memory (LSTM) network with shared parameters to capture long-term trends, cycles, and short-term fluctuations in the sequences, such as... Figure 3 As shown.

[0075] Long Short-Term Memory (LSTM) networks are a special type of recurrent neural network that controls the flow of information through gating mechanisms (input gate, forget gate, output gate), effectively solving the gradient vanishing problem of traditional RNNs and enabling the modeling of long-term temporal dependencies. In OD (Original Location) passenger flow forecasting, LSTM can capture the trend, periodicity, and seasonal patterns in OD traffic sequences. By using LSTM to process historical passenger flow sequences of OD pairs, dynamic temporal features of each OD pair can be extracted, providing rich temporal contextual information for subsequent forecasts.

[0076] Specifically, LSTM updates the cell state at each time step through its gating mechanism. With hidden state Specifically, in the time step LSTM receives the current input (i.e., the transformed passenger flow value) and the hidden state of the previous time step. With cell state The information flow is controlled by three gating units to generate a complete hidden state sequence. .

[0077] Forget Gate: Indicates the proportion of each dimension of the cell's state retained from the previous time step.

[0078]

[0079] Input Gate: Controls the proportion of new information flowing into the cell state, and simultaneously calculates candidate cell states.

[0080]

[0081] Cell State Update: This update combines signals from the forget gate and the input gate to update the cell state.

[0082]

[0083] Output Gate: Controls the output ratio based on the current cell state and outputs it to the hidden state sequence.

[0084]

[0085]

[0086] in, , , , , , , , These are learnable parameters.

[0087] The self-attention mechanism further processes the hidden state sequence output by the LSTM. By calculating the attention weights of query, key, and value, it assigns differentiated attention to different time steps. Q, K, and V are obtained by linear transformation of the hidden state sequence.

[0088] First, the hidden state sequence The query matrix is ​​obtained through linear transformation. Key matrix Sum matrix ,in , , These are learnable parameters.

[0089] Attention score is calculated as follows:

[0090]

[0091] in Let be the dimension of the key vector. Use a scaling factor to prevent gradient vanishing.

[0092] Context vector generation:

[0093]

[0094] This mechanism can adaptively capture the importance of different time steps, enhancing the ability to identify passenger flow patterns during key periods.

[0095] Finally, the context vector output by the attention mechanism is input into the projection layer to obtain the temporal features of the OD pair:

[0096]

[0097] in, Indicates OD pair Temporal pattern characteristics, Indicates OD pair Historical passenger flow sequence These are the parameters for the LSTM network. This path outputs a three-dimensional tensor. It encodes the historical evolution patterns of all OD pairs in the entire network. The dimension for edge temporal encoding.

[0098] Spatiotemporal fusion: representing the initial features of nodes output by the node path. The input is fed into the core spatiotemporal fusion module (an improved AGCRN). This module, through an adaptive graph convolutional recurrent network, simultaneously learns the dynamic spatial dependencies between sites and the evolution of node states over time steps, achieving deep fusion of complex spatiotemporal features and ultimately outputting node deep features rich in spatiotemporal semantics. .

[0099] Specifically, the spatiotemporal fusion module is responsible for performing in-depth spatiotemporal modeling of node features, and its structure is as follows: Figure 4 As shown, the improved adaptive graph convolutional recurrent network integrates adaptive graph convolution and gated recurrent mechanisms, enabling it to simultaneously learn dynamic spatial dependencies and the temporal evolution of node states.

[0100] Adaptive graph convolutional layer

[0101] Traditional graph convolution relies on a fixed physical adjacency matrix, making it difficult to capture the potential, dynamic functional relationships between sites. This invention introduces a learnable node embedding matrix. An adaptive adjacency matrix is ​​dynamically generated based on the similarity of node embeddings. :

[0102]

[0103] in, An embedding matrix for learnable nodes. To modify the linear unit activation function for filtering out negative or very weak similarity connections, The function is used for row-wise normalization, ensuring that the sum of the attention weights of each node to all its neighbors is 1. This mechanism enables the model to autonomously discover functional associations beyond physical connections, particularly identifying and quantifying counterintuitive pairings of geographically non-proximity but functionally strongly related sites (such as commuter corridors between large residential areas and core business districts). The resulting adaptive adjacency matrix... Developed entirely from data learning, this technology transcends physical connections, capturing potential functional relationships and hidden passenger flow interaction patterns between stations. This design enables the model to autonomously discover strong commuting connections, such as between "large residential area stations" and "core business district stations," even if they are not geographically or linearly adjacent, thus more accurately simulating the spatial distribution of passenger flow and responding to sudden surges in passenger volume.

[0104] Based on this adaptive adjacency matrix and node input features Graph convolution operations are performed in each layer. Defined as:

[0105]

[0106] in, Indicates the first The node feature matrix of the layer, Indicates the first A layer is a trainable parameter matrix. yes The degree matrix (diagonal elements are) ), used for symmetric normalization, It is a non-linear activation function. It is the first Output features of layer graph convolution.

[0107] This operation enables nodes to aggregate spatial information from their adaptive "neighbors," ensuring numerical stability through symmetric normalization and avoiding gradient vanishing or exploding problems.

[0108] Gated loop mechanism

[0109] The gated recurrent mechanism is a gated recurrent spatiotemporal state update unit. To model the evolution of node features over time steps, this invention deeply integrates graph convolutional layers with gated recurrent units (GRUs). At time steps... , This is a processing step within the model, corresponding to the prediction time period, for each node. The state update process is as follows:

[0110] 1. Obtain the spatial characteristics at the current moment: Let This represents the hidden state of the node in the previous time step. This represents the spatial features obtained after the node at the current time undergoes adaptive graph convolution (i.e., (The row vector of the corresponding node in the middle).

[0111] 2. Gating loop update:

[0112] Update Gate: Controls the proportion of the previous state that is retained in the current state.

[0113]

[0114] Reset Gate: Controls the contribution of the previous state to the candidate state.

[0115]

[0116] Candidate State: A new candidate state value calculated based on the reset gate and the current input.

[0117]

[0118] Current State: Controlled by the update gate, it merges the previous state and candidate states to obtain the final state at the current moment.

[0119]

[0120] in, and These represent updating the door and resetting the door's activation value, respectively. Indicates the candidate hidden state. Indicates the current time node The hidden state (i.e., the node depth features after spatiotemporal fusion). , , and , , For learnable weight matrices and bias vectors, For activation function, It is the hyperbolic tangent activation function.

[0121] 3. Output and Normalization: The updated hidden states are processed using LayerNorm and Dropout operations to obtain the final spatiotemporal feature representation of the nodes.

[0122]

[0123] This gated cyclic spatiotemporal state update unit achieves deep fusion of node spatiotemporal features, and can adaptively balance the influence of spatial dependence and temporal evolution, thereby more accurately capturing the dynamic changes in passenger flow patterns.

[0124] In the AGCRN module, graph convolution and GRU are deeply integrated to construct a gated cyclic spatiotemporal state update unit. This enables joint modeling and dynamic balance of spatial dependence and temporal evolution at the node state level, allowing the final feature representation of the node to simultaneously contain rich spatial correlation information and coherent temporal evolution information. This lays a solid feature foundation for subsequent OD prediction based on node representation.

[0125] Node-based OD prediction: In the output layer, the passenger flow prediction module executes a node-based prediction mechanism. For any OD pair... This mechanism utilizes the node depth features at its starting and ending points. Starting from this point, the temporal features of the OD pair extracted from the edge path are combined. and the characteristics of relational types guided by domain knowledge. By combining feature splicing and regression calculations, the future passenger flow of the OD pair can be indirectly predicted. By traversing all OD pairs, a complete short-term OD passenger flow forecast matrix is ​​generated. This step achieves the transformation from decoupled learning to interpretable prediction.

[0126] Specifically, the OD (Original Demand) passenger flow prediction module implements a mapping from node depth representation to OD traffic prediction, indirectly predicting edges by utilizing the roles and relationships of nodes. This module is efficient and physically interpretable. The specific steps include:

[0127] For the OD pairs to be predicted First, multi-source feature combination is performed:

[0128] 1. Node Role Characteristics: From the spatiotemporal feature matrix of AGCRN nodes Extracting the starting point and the end point Depth representation and .

[0129] 2. Relationship temporal characteristics: Output from edge-encoded paths Extract the temporal pattern features of the OD pair. .

[0130] 3. Relationship Type Features: Query the category of the OD pair in the pre-trained OD pattern clustering and introduce a predefined OD pattern clustering embedding. This serves as prior knowledge characterizing the interaction patterns of passenger flow in such OD scenarios.

[0131] After concatenating the above features, an OD flow is predicted through a regression layer:

[0132]

[0133] in, Indicates the predicted location from the site Arrive at the station OD passenger traffic and Representing the starting point and the end point The node depth features, Indicates OD pair The temporal characteristics, Indicates OD pair Pattern clustering embedding features, , These are the learnable weights and biases for the regression layer.

[0134] This mechanism reduces the number of core parameters and storage overhead of the model compared to traditional edge prediction. The magnitude was reduced to Magnitude, and prediction results This can be clearly attributed to the combined effects of starting point characteristics, ending point characteristics, historical patterns, and relationship models, providing a clear basis for decision-making in operational analysis.

[0135] S3. Model Training and Validation: A robust objective function combining Huber loss and relative error loss is used to train the model end-to-end with optimizers such as Adam. During training, the model performance is monitored using a validation set, and an early stopping strategy is implemented to prevent overfitting. After training, the model is evaluated on an independent, confidential test set, and its prediction accuracy is quantified using metrics such as mean absolute error (MAE) and root mean square error (RMSE).

[0136] Specifically, the entire process involves minimizing the predicted OD matrix. With the real OD matrix End-to-end training is performed using losses (such as mean squared error). The loss function employs a robust flow loss function, combined with Huber loss and relative error loss.

[0137]

[0138] in, Indicates actual OD traffic. Indicates the quantity of OD pairs. This represents the Huber loss function. and To balance hyperparameters, , .

[0139] The loss function is robust to outliers and takes into account the relative errors of different flow levels, so that the model can maintain good performance when predicting OD pairs of various flow levels.

[0140] S4. Predictive Application: The trained model is deployed to the actual operation system. The system receives real-time data streams and performs feature extraction and preprocessing following the same process as the training phase. This data is then input into the model, which outputs a short-term network-wide OD passenger flow prediction matrix. The prediction results can be presented as heatmaps or trend curves through a visual interface and directly integrated with the dispatching system. This serves dynamic capacity allocation during peak hours, emergency management under sudden surges in passenger flow, and real-time passenger flow guidance, thereby effectively translating predictive capabilities into improved operational efficiency and safety levels, achieving intelligent and refined operation management of the rail transit system.

[0141] To address the problems of existing short-term OD passenger flow prediction methods, such as large number of parameters leading to poor scalability, rigid spatial dependency modeling, mixed processing of heterogeneous features, and insufficient integration of domain knowledge, this invention introduces a model based on a node-edge dual-path spatiotemporal fusion graph learning framework.

[0142] The model comprises an input layer, a processing layer, and an output layer, designed to improve the accuracy, efficiency, and interpretability of short-term OD (Original Demand) passenger flow forecasting. Its structure is as follows: Figure 2 As shown, the core of the model is to decouple the OD prediction task into deep spatial modeling of node (site) features and temporal modeling of edge (OD pair) features, and achieve the final OD traffic prediction through an innovative node representation-based mechanism. By fully exploiting the essence of heterogeneous information through dual-path separation coding, accurately modeling complex dependencies through improved AGCRN spatiotemporal fusion, and finally completing the OD passenger flow prediction task in an efficient and interpretable manner through a node representation-based prediction mechanism, the model systematically improves prediction accuracy, efficiency, and practicality.

[0143] This invention reduces the number of core parameters and storage overhead of the model by decoupling the encoding of node and edge features and using a node representation-based prediction mechanism. Through deep node features, the prediction results can be traced back to the functional roles and specific spatiotemporal behavior patterns of the origin and destination stations, enhancing interpretability and providing clear and reliable evidence for operation managers to conduct passenger flow source analysis and composition assessment. Furthermore, by integrating prior domain knowledge such as OD clustering patterns and station POI (Point of Interest) semantics, the model improves overall prediction accuracy while significantly enhancing its robustness to sparsity and uncertainty in actual operational scenarios such as newly opened routes and sudden surges in passenger flow.

[0144] This invention first encodes node and edge features independently and professionally; then, it utilizes an improved Adaptive Graph Convolutional Recurrent Network (AGCRN) module for deep spatiotemporal dynamic dependency fusion; finally, it achieves OD traffic prediction based on the fused high-level node representation. This architecture achieves accurate matching from heterogeneous data characteristics to prediction task objectives, not only improving computational efficiency by separating redundant parameters but also effectively mitigating overfitting through a hierarchical fusion mechanism, thereby enhancing the model's generalization ability and stability.

[0145] To verify the effectiveness of this invention, comparative experiments were conducted on a real dataset. Card swipe data from Beijing Metro from February 28th to April 2nd, 2023, including 366 stations, was used. Odd origin (OD) traffic was aggregated at a daily granularity and divided into training, validation, and test sets according to time sequence. Linear regression, random forest, MLP neural network, support vector regression (SVR), decision tree, and XGBoost were used as baselines for comparison. Mean absolute error (MAE), root mean square error (RMSE), and symmetric mean absolute percentage error (SMAPE%) were used as evaluation metrics. Specific comparative experimental results are shown in Table 1.

[0146]

[0147] Note: The optimal hyperparameter configuration is: hidden_dim=128, embed_dim=32, num_layers=2, dropout=0.3, learning_rate=0.0001, cheb_k=2.

[0148] As shown in the table, the NEDP-GNN model, as the model of this invention, achieves the best performance in all three key metrics. In the MAE metric, which measures the mean error, NEDP-GNN (30.38) is significantly lower than all comparison baselines, demonstrating that the model possesses the most accurate average prediction capability. In the RMSE metric, which penalizes large errors more strictly, NEDP-GNN (40.54) shows a more significant advantage, indicating that it is more robust in predicting large-volume OD pairs and outliers with smaller error fluctuations. In the SMAPE metric, which has strong business interpretability, NEDP-GNN (31.80%) also maintains a leading position, reflecting that the model maintains good percentage error control across different traffic levels (especially high and low traffic), making it more practical.

[0149] Taking the Beijing subway network as an example, the complete application process of the short-term OD passenger flow forecasting method is as follows:

[0150] Data preprocessing stage: First, historical card swipe data, station information, and surrounding Points of Interest (POI) data of the target rail transit network are collected. Data is aggregated at fixed time intervals (e.g., 15 minutes) to construct a historical OD matrix sequence. Subsequently, node features and edge features are extracted: node features include basic traffic statistics for each station, time context encoding, station function clustering labels based on historical passenger flow patterns, and station surrounding function feature vectors generated according to a dedicated POI classification system; edge features are the historical passenger flow sequences for each OD pair over the past T consecutive time periods. Finally, the complete dataset, sorted by time, is divided into training, validation, and test sets.

[0151] Model Construction Phase: A node-edge dual-path spatiotemporal fusion graph learning model is built according to the framework of this invention. Node features are encoded into a unified initial representation using a multilayer perceptron (MLP); the historical passenger flow sequences of each OD pair are encoded by a shared-parameter Long Short-Term Memory (LSTM) network combined with a self-attention mechanism to extract its temporal pattern features. The core spatiotemporal fusion module employs an improved Adaptive Graph Convolutional Recurrent Network (AGCRN), which dynamically generates the graph structure through learnable node embeddings and utilizes gated recurrent units (GRUs) to deeply fuse spatial dependencies and temporal evolution, outputting the deep spatiotemporal features of nodes. During prediction, for any OD pair, the node depth features of its start and end points, the corresponding edge temporal features, and predefined OD pattern clustering embeddings are concatenated and the predicted passenger flow is calculated through a fully connected regression layer.

[0152] Model training and validation phases: A robust objective function combining Huber loss and relative error loss is used to train the model end-to-end with optimizers such as Adam. During training, the model performance is monitored using a validation set, and an early stopping strategy is implemented to prevent overfitting. After training, the model is evaluated on an independent, confidential test set, and its prediction accuracy is quantified using metrics such as mean absolute error (MAE) and root mean square error (RMSE).

[0153] Predictive Application Phase: The trained model is deployed to the actual operation system. The system receives real-time data streams and performs feature extraction and preprocessing following the same process as in the training phase. This data is then input into the model, which outputs a short-term network-wide OD passenger flow prediction matrix. The prediction results can be presented as heatmaps or trend curves through a visual interface and directly integrated with the dispatching system. This serves dynamic capacity allocation during peak hours, emergency management under sudden surges in passenger flow, and real-time passenger flow guidance, thereby effectively translating predictive capabilities into improved operational efficiency and safety levels, achieving intelligent and refined operation management of the rail transit system.

[0154] In summary, this invention systematically addresses the limitations of existing methods in terms of prediction accuracy, computational efficiency, model interpretability, and scenario robustness by decoupling heterogeneous features, innovating a node-edge dual-path prediction mechanism, and deeply integrating domain knowledge. Experimental results fully validate the advancement and effectiveness of this method. It not only provides a new theoretical research perspective and technical implementation path for short-term OD passenger flow prediction in urban rail transit, but also offers operators a more accurate, efficient, and reliable intelligent decision-making tool for refined scheduling, dynamic passenger flow guidance, and emergency resource allocation. This invention possesses significant theoretical innovation value and broad practical application prospects.

[0155] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A rail transit short-term OD passenger flow prediction method based on node-edge space-time fusion, characterized in that, Includes the following steps: S1. Collect target rail transit network data, and extract node features and edge features from it. The node features include station static features and quasi-static features, and the edge features include the historical passenger flow sequence of each OD pair. Divide the feature dataset into training set, validation set and test set. S2. Construct a node-edge dual-path spatiotemporal fusion graph learning model. The processing flow of the node-edge dual-path spatiotemporal fusion graph learning model includes the following steps: S201. Input and feature decoupling: Physically separate the node features and edge features, and input them to the node processing path and edge processing path respectively; S202, Dual-path asynchronous coding: a. Node Path: Node features are mapped to initial node features via a multilayer perceptron encoder and input into the spatiotemporal fusion module. This module is an improved adaptive graph convolutional recurrent network. It dynamically generates an adaptive adjacency matrix using a learnable node embedding matrix and incorporates a gated recurrent mechanism to jointly model the dynamic spatial dependencies and temporal evolution between nodes, outputting the node's deep spatiotemporal features. The adaptive adjacency matrix is ​​represented by the following formula: ， wherein, is a learnable node embedding matrix; is a rectified linear unit activation function used to filter out negative or very weak similarity connections; is a function used to perform row-wise normalization; b. Edge path: The historical passenger flow sequence of each OD pair is encoded by a long short-term memory network combined with an attention mechanism to extract the edge temporal features of each OD pair; S203. OD Aggregation Prediction Based on Node Representation: The depth features of the starting node, the depth features of the ending node, the corresponding edge temporal features, and the predefined OD relationship type features of the OD pair to be predicted are concatenated. The predicted passenger flow of the OD pair is obtained through regression calculation. The predicted passenger flow of the OD pair is expressed by the following formula: ， wherein, denotes the predicted OD passenger flow from site to site ; and denote the node depth features of the origin and destination sites, respectively; denotes the temporal feature of the OD pair ; denotes the mode cluster embedding feature of the OD pair ; and are the learnable weights and bias of the regression layer. S3. Use the training set to train the model end-to-end, use the validation set to monitor the model performance and adjust the parameters, and use the test set to test and evaluate the model after training. S4. Use the trained model to predict the OD passenger flow of the target rail transit network in the near future.

2. The method for predicting short-term OD passenger flow in rail transit based on node-edge spatiotemporal fusion according to claim 1, characterized in that: In step S1, the node features include basic features, temporal features, clustering features, and POI features.

3. The method for predicting short-term OD passenger flow in rail transit based on node-edge spatiotemporal fusion according to claim 2, characterized in that: The POI features are constructed using a dedicated classification system based on the mechanism of passenger flow generation and attraction in rail transit. This dedicated classification system includes categories such as transportation services, catering services, shopping services, healthcare services, accommodation services, scenic spots, commercial residences, government agencies and social organizations, public transportation facilities, financial and insurance services, companies and enterprises, living and public services, and cultural, leisure and scientific education functions.

4. The method for short-term OD passenger flow prediction of rail transit based on node-edge spatiotemporal fusion according to claim 1, characterized in that: In step S202, the initial characteristics of the node are represented by the following formula: ， in, For all The initial feature matrix of each site.

5. The method for predicting short-term OD passenger flow in rail transit based on node-edge spatiotemporal fusion according to claim 1, characterized in that: In step S202, the long short-term memory network includes an input gating unit, a forgetting gating unit, and an output gating unit; The input gating unit is represented by the following formula: ， The forget gate unit is represented by the following formula: ， The output gate is represented by the following formula: ， in, , , , , , , , These are learnable parameters.

6. The method for short-term OD passenger flow prediction of rail transit based on node-edge spatiotemporal fusion according to claim 1, characterized in that: In step S203, the edge timing characteristics are represented by the following formula: ， in, Indicates OD pair Temporal pattern characteristics; Indicates OD pair Historical passenger flow sequence; This represents the parameters of the LSTM network.

7. The method for short-term OD passenger flow prediction of rail transit based on node-edge spatiotemporal fusion according to claim 1, characterized in that: In step S203, the OD relationship type feature is an embedding vector obtained by pattern clustering and mapping of historical OD data.

8. The method for short-term OD passenger flow prediction of rail transit based on node-edge spatiotemporal fusion according to claim 1, characterized in that: In step S3, the model is trained using a robust objective function that combines Huber loss and relative error loss.

Images