Emergency public health event short-term OD passenger flow prediction method based on deep learning

Abstract

The invention discloses an emergent public health event short-term OD passenger flow prediction method based on deep learning. The method comprises the steps that OD demand matrixes of multiple modes and event related information observed at the t-th time step during an emergent public health event are acquired, and the OD demand matrixes of the multiple modes comprise a weekly OD demand matrix, a daily OD demand matrix and a real-time OD demand matrix; and based on the OD demand matrix of the multiple modes and the event related information, utilizing a trained deep learning model to predict an OD passenger flow matrix of a future time step, and obtaining a passenger flow prediction result of the target station. According to the invention, the passenger flow prediction accuracy and generalization ability of the rail transit in the major public health period are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of passenger flow prediction, and more specifically, to a short-term OD passenger flow prediction method for sudden public health events based on deep learning. Background Art

[0002] In recent years, with the rapid development of urban rail transit, short-term passenger flow prediction, which is closely related to the operation and management of urban rail transit systems, has received wide attention. During major sudden public health events, the passenger flow of urban rail transit has complex spatio-temporal characteristics, and the passenger flow pattern is highly random and sudden, posing huge challenges to the operation and management of rail transit. To improve the operation efficiency of rail transit and reduce the operation cost, operators need to timely adjust the train timetable and operation plan according to accurate passenger flow prediction results. Therefore, the short-term passenger flow prediction of rail transit during sudden public health events is particularly important.

[0003] After analysis, in the prior art, the short-term OD passenger flow prediction of urban rail transit during the impact of major sudden public health events mainly has the following defects:

[0004] 1) Passengers need a certain amount of time to travel from the starting station to the destination station, which results in the inability to obtain real-time OD flow immediately. Therefore, how to ensure the real-time and timeliness of data is an urgent problem to be solved;

[0005] 2) Due to the time delay and incompleteness of the OD demand matrix, the OD matrix of adjacent time periods cannot be directly used as the model input;

[0006] 3) Existing research rarely involves the sparsity problem of OD demand matrix data, that is, the number of OD pairs with zero or small OD flow is relatively large. The data sparsity problem will lead to a decrease in the prediction accuracy and stability of existing models;

[0007] 4) Existing research rarely considers short-term OD passenger flow prediction under abnormal conditions, especially under the influence of uncertain factors such as major sudden public health events. Most existing deep learning models only consider historical passenger flow data when predicting passenger flow. In fact, there are many factors affecting the change of passenger flow. Using relevant data to capture the influence of different factors on passenger flow is beneficial to improving the prediction accuracy and is worthy of further research. Summary of the Invention

[0008] The object of the present invention is to overcome the above defects of the prior art and provide a short-term OD passenger flow prediction method for sudden public health events based on deep learning. The method includes the following steps:

[0009] Obtain the OD demand matrices of multiple patterns and event-related information observed at the t-th time step during a public health emergency. The OD demand matrices of multiple patterns include the weekly OD demand matrix, the daily OD demand matrix, and the real-time OD demand matrix;

[0010] Based on the OD demand matrices of multiple patterns and the event-related information, use a trained deep learning model to predict the OD passenger flow matrix for future time steps and obtain the passenger flow prediction result of the target station.

[0011] Compared with the prior art, the advantages of the present invention are as follows: a graph neural network model based on multi-frequency domain spatio-temporal feature mining (MFST-GNN) is proposed for short-term OD passenger flow prediction of urban rail transit during major public health emergencies. This model integrates OD passenger flow data of long and short cycles, including real-time OD passenger flow, daily OD passenger flow, and weekly OD passenger flow, and extracts deep passenger flow features from three different modes. Moreover, a multi-frequency domain time feature extraction module is used to mine time series features, and a spatial feature extraction module is used to extract spatial features. In addition, the present invention collects event text information at different time periods, uses a heterogeneous information fusion module to integrate the information into the extracted deep passenger flow representation, and passes the fused features through a fully connected layer to output the predicted OD demand matrix. The present invention provides an effective tool for passenger flow data prediction of urban rail transit systems.

[0012] Other features and advantages of the present invention will become clear from the following detailed description of exemplary embodiments of the present invention with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The drawings incorporated in the specification and constituting a part of the specification illustrate embodiments of the present invention and, together with the description, are used to explain the principles of the present invention.

[0014] Figure 1 is a flowchart of a method for short-term OD passenger flow prediction of public health emergencies based on deep learning according to an embodiment of the present invention;

[0015] Figure 2 is a schematic diagram of a graph convolutional neural network according to an embodiment of the present invention;

[0016] Figure 3 is a schematic diagram of the MFST-GNN model framework according to an embodiment of the present invention;

[0017] Figure 4 is a schematic diagram of a multi-cycle time feature extraction module according to an embodiment of the present invention;

[0018] Figure 5 is a schematic diagram of an adaptive spatial feature extraction module according to an embodiment of the present invention;

[0019] Figure 6 It is a schematic diagram of a heterogeneous information fusion module according to an embodiment of the present invention;

[0020] Figure 7 It is a schematic diagram for comparing the performance of different models according to an embodiment of the present invention;

[0021] Figure 8 It is a schematic diagram of the prediction effect of the model at different time periods with a 15 - minute granularity according to an embodiment of the present invention;

[0022] Figure 9 It is a schematic diagram of the prediction effect of the model at different time periods with a 30 - minute granularity according to an embodiment of the present invention;

[0023] Figure 10 It is a schematic diagram of the prediction effect of the model at different time periods with a 60 - minute granularity according to an embodiment of the present invention. Detailed implementation manners

[0024] Now, various exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. It should be noted that: Unless otherwise specifically stated, the relative arrangements of components and steps, numerical expressions, and numerical values set forth in these embodiments do not limit the scope of the present invention.

[0025] The following description of at least one exemplary embodiment is merely illustrative in nature and in no way serves as a limitation to the present invention, its application, or its use.

[0026] Techniques, methods, and devices known to those of ordinary skill in the relevant art may not be discussed in detail, but where appropriate, such techniques, methods, and devices should be regarded as part of the specification.

[0027] In all the examples shown and discussed herein, any specific values should be construed as merely exemplary and not as a limitation. Thus, other examples of the exemplary embodiments may have different values.

[0028] It should be noted that: Similar reference numerals and letters denote similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further discussed in subsequent drawings.

[0029] The present invention constructs an effective deep - learning framework by means of a graph neural network based on multi - frequency - domain spatio - temporal feature mining, and at the same time organically fuses real - time passenger flow, daily passenger flow, weekly passenger flow, and relevant event text information, so as to fully study the impact of major public health emergencies on passenger flow changes, capture the complex spatio - temporal features of passenger flow dynamics, and improve the prediction accuracy of passenger flow during major public health emergencies while meeting the'real - time' requirement of short - term passenger flow prediction.

[0030] See Figure 1 As shown, generally speaking, the provided short-term OD passenger flow prediction method based on deep learning for public health emergencies includes: Step S110, obtaining the OD demand matrices of multiple patterns and event-related information observed at the t-th time step during a public health emergency, where the OD demand matrices of multiple patterns include the weekly OD demand matrix, the daily OD demand matrix, and the real-time OD demand matrix; Step S120, based on the OD demand matrices of multiple patterns and the event-related information, using a trained deep learning model to predict the OD passenger flow matrix at future time steps and obtain the passenger flow prediction results of the target stations.

[0031] The technical solution of the present invention is divided into the following parts: First, define the scientific problem to be solved; then, propose a deep learning model framework MFST-GNN, which includes a temporal feature extraction module and a spatial feature extraction module. Finally, verify the prediction effect of the model on a dataset of a certain year of the subway in Nanning, Guangxi. By comparing with the prediction performance of common short-term passenger flow prediction models, the accuracy of the prediction of the present invention and the rationality of the model structure are verified.

[0032] I. Problem Definition

[0033] The present invention aims to use historical AFC data and other reliable data sources, with the help of a deep learning model, to predict the short-term inbound passenger flow of all stations in the urban rail transit network during major public health emergencies.

[0034] Definition 1 (Urban Rail Transit Network): The purpose of the present invention is to predict the inbound passenger flow of all stations in the urban rail transit network at a certain time period simultaneously. The urban rail transit network is defined as a graph G=(V, E, A), where there are V subway stations in the graph, there are E edges between stations, and A∈R N×N represents the adjacency matrix of the network. For example, it has two elements, 0 and 1. 0 means that two stations are not adjacent, and 1 means that two stations are adjacent.

[0035] Definition 2 (OD Demand Matrix): The purpose of short-term OD demand prediction is to predict the short-term OD demand between stations in the rail transit network. Therefore, define the OD demand. From the above analysis, it can be seen that the real-time OD demand matrix is an incomplete matrix. Therefore, define the OD demand matrix X t ∈R N×N , which represents all OD demands of each station in the rail transit network during the time interval t. Taking all OD demands in the time interval t as an example, the corresponding real-time incomplete OD demand matrix can be expressed as:

[0036]

[0037] Among them, xt (i, j) represents the travel demand from station i to station j within the time interval t, where i, j = 1, 2, ..., X t ∈R N×N represents an incomplete OD demand matrix.

[0038] Definition 3 (Multiple OD Demand Patterns): Since passengers' travel behaviors have strong periodicity, the model considers multiple OD demand patterns that vary with time, namely the weekly OD pattern, the daily OD pattern, and the real-time OD pattern, to study the long-term and short-term periodic characteristics of the OD demand distribution respectively. Specifically, assuming the current time interval is t, the weekly OD pattern refers to the OD demand during the same time period of the previous week, denoted as X w , the daily OD pattern refers to the OD demand during the same time interval of the previous day, denoted as X d , and the real-time OD pattern refers to the historical OD demand at the current time t, which includes the OD demand of a total of T time steps, denoted as X = [X t-T+1 , …, X t-1 , X t .

[0039] Definition 4 (Preprocessing of OD Demand Matrix Compression): Since the passenger flow between some stations is usually small or even zero. In one embodiment, OD pairs with high passenger flow are considered. Specifically, for each origin station, analyze the destination distribution of all incoming passengers, then select the top k - 1 stations with the highest probability of passengers going to the destination, and aggregate the OD demands of the remaining all stations as the kth station. Therefore, an OD demand matrix X t ∈R N×k can be generated, where N is the number of stations.

[0040] Problem Definition: The goal of the deep learning model designed in the present invention is to predict the complete OD demand of the rail transit system at the next time interval X t+1 . Therefore, the prediction problem can be written as:

[0041] X t+1 = f([X w , X d , X t-T+1 , …, X t-1 , X t , G) (2)

[0042] In the formula: X t+1 represents the predicted OD demand matrix at the next time step, G represents the rail transit network diagram, X w represents the weekly OD demand matrix, X d represents the daily OD demand matrix, [X t-T+1 , …, X t-1, X t represents the real-time OD demand matrix at historical T time steps.

[0043] II. Regarding Graph Convolutional Network

[0044] The urban rail transit station network is a disordered and undirected network, and each station forms a complex network. Compared with the traditional two-dimensional convolution based on images to extract the regular characteristics of station passenger flow, using the Graph Convolutional Network (GCN) to process station passenger flow data can more effectively extract the patterns and features that are significant in terms of passenger flow space. See Figure 2 the graph convolutional network diagram shown, which generally includes an input layer, multiple hidden layers, an activation layer, an output layer, etc.

[0045] Graph structure representation: In the graph structure, nodes represent entities and edges represent the relationships between nodes. Assume graph G = (V, E, A), where V is the set of nodes, E is the set of edges, and the adjacency matrix A represents the connection relationships between nodes. In addition, each node v i has a feature vector x i , which represents the feature of the node.

[0046] Graph convolution operation: The core of the graph convolution operation is to update the feature representation of a node by aggregating the information of its neighboring nodes. Assume an input graph G and a node feature matrix X, where each row of X i corresponds to the feature of node v i . The graph convolution operation of GCN can be expressed as:

[0047]

[0048] In the formula: W is the weight matrix; D is the degree matrix, D ii = ∑ j A ij ; A is the adjacency matrix; σ is the activation function. Formula (3) can be interpreted as linearly combining the features of a node with the features of its neighboring nodes, then transforming through the weight matrix W, and finally applying the activation function.

[0049] Stacked layers: GCN is usually composed of multiple convolutional layers stacked together. Each layer uses the output of the previous layer as input and further extracts the features in the graph. By stacking multiple layers, GCN can capture graph structure information at different levels, similar to the stacking of convolutional layers in a convolutional neural network. Therefore, multi-layer graph convolution can be expressed as:

[0050]

[0051] In the formula: H (l) is the input of the previous layer, H(0) = X; H (l+1) is the output of the current layer; W (l) is the weight matrix of the l-th layer;

[0052] By learning the relationships and representations between nodes, GCN can effectively process graph data, is suitable for various graph data analysis tasks, and has a powerful ability to capture spatio-temporal and topological information.

[0053] In the embodiments of the present invention, variants and improvements of GCN can also be adopted to further improve performance and the diversity of application fields. For example, due to the development files in spectral graph f, Chebyshev polynomials, and first-order filters, the GCN layer has undergone significant improvements. The stacking of GCN layers with first-order filters can achieve similar effects to k-Chebyshev polynomial filters, while achieving significantly higher training speeds and prediction accuracies in most cases. Therefore, in an embodiment of the present invention, the GCN proposed by Kipf et al. is used, expressed as:

[0054]

[0055] In the formula: A is the adjacency matrix; I is the identity matrix; is the degree matrix of; X l , X l+1 are the feature matrices of the l-th layer and the (l + 1)-th layer respectively; W l is the weight matrix of the l-th layer; σ is the non-linear activation function.

[0056] III. Deep learning model structure

[0057] The present invention proposes a short-term OD passenger flow prediction method during major public health emergencies based on a multi-frequency spatio-temporal feature mining graph neural network (MFST-GNN). Refer to Figure 3 As shown, the overall framework of the MFST-GNN network includes a multi-period time feature extraction module (or multi-frequency time feature extraction module, Multi-frequency temporal feature extraction module), an adaptive spatial feature extraction module (Adaptive spatial feature extraction module), and a heterogeneous information fusion module (Heterogeneous data fusion module), etc. Its main idea is to learn the complex dynamic spatio-temporal dependencies of the inbound flow of urban rail transit during major public health emergencies to accurately predict the inbound flow of the subway during major public health emergencies.

[0058] First, the input of the model includes the weekly OD matrix, the daily OD matrix, and the real-time OD matrix. The overall model extracts deep passenger flow characteristics from multiple perspectives in different channel modes. Since during this event, the travel of residents shows randomness, and the laws of passenger flow movement in space change accordingly. The passenger flow prediction of the historical weekly and daily patterns can provide the spatial information that changes during the public health emergency. Therefore, an adaptive spatial feature extraction module is used to extract spatial features from the weekly and daily pattern OD matrices. Considering that the correlation between different time slices may be broken, and the passenger flow distribution in space also has uncertain dynamics. Therefore, a multi-frequency domain time feature extraction module is used for the real-time OD matrix to extract time features, and an adaptive spatial feature extraction module is used to extract spatial features. In addition, in order to better integrate the information of public health emergencies into the model, a heterogeneous information fusion module is designed to fuse event-related information into the OD matrix features. Finally, the predicted OD passenger flow matrix is output using the fully connected layer, and considering the relationship between the OD demand matrix and the inbound passenger flow, the PINN (Physics-Informed Neural Networks) loss function can be used to supervise the training of the model.

[0059] 1. Multi-period time feature extraction module

[0060] Due to the many uncertainties in the passenger flow time series correlation during major public health emergencies in abnormal situations, it is difficult for commonly used time series information mining networks to extract the regular information hidden in them. Therefore, in one embodiment, a multi-period time feature extraction module is proposed, which uses the weekly OD matrix, the daily OD matrix, and the real-time OD matrix as the model input for passenger flow feature extraction, and extracts the frequency domain information and time features of abnormal traffic time series data during major public health emergencies. Figure 4 It is the overall framework of the multi-period time feature extraction module. Since it is observed that time series can usually be decomposed into subsequences in multiple frequency domains. Among them, such subsequences overlap and interact with each other, making modeling difficult to handle. And for each subsequence, its change is still affected by the time pattern of the frequency domain of its adjacent subsequences.

[0061] Specifically, first, the input real-time OD matrix X ∈ R T×N×k×1 After passing through the feature encoding layer, the deep OD passenger flow features are output, and the formula is as follows:

[0062] X e = Embedding Layer(X) (6)

[0063] In the formula: X ∈ R T×N×k×1 is the real-time OD demand matrix, X e ∈ R T×N×k×FFor the OD passenger flow characteristics of the output, F is the feature dimension, and the Embedding Layer is the feature encoding layer.

[0064] Then, for X e perform a fast Fourier transform on the time dimension to extract the K main frequencies of the features, that is:

[0065] A = Avg(Amp(FFT(X e ))) (7)

[0066] f1,…,f k = argTopk(A) (8)

[0067] In the formula: A is the intensity of each frequency component in X e FFT is the Fourier transform, Amp is the calculation of the amplitude value, {f1,...,f k} are the k frequencies with the maximum intensity.

[0068] Next, based on the selected frequencies and the corresponding frequency lengths, fold the original one-dimensional time series X e After folding, a set of two-dimensional vectors is obtained Then use 2D convolution for time series feature extraction. The formula is as follows:

[0069]

[0070] In the formula: are the features of different frequencies after convolution.

[0071] Different amplitudes can reflect the relative importance of the selected frequencies. Therefore, use the Softmax function to calculate the weights for the amplitudes. The formula is as follows:

[0072]

[0073] In the formula: are the weights of the features at different frequencies.

[0074] Finally, multiply the final weights by the features and use the fully connected layer to obtain the final feature representation. The formula is as follows:

[0075]

[0076] In the formula: X T ∈R N×k×F is the feature output by the multi-frequency domain time feature extraction module.

[0077] 2. Adaptive Spatial Feature Extraction Module

[0078] During the occurrence of major public health emergencies, the stable pattern of the transportation system itself is disrupted, and the static spatial graph structure may lead to the aggregation of incorrect information. To adapt to the extraction of spatial information in such emergencies, the present invention designs an adaptive spatial feature extraction module that can extract spatial information from both dynamic adaptive and static perspectives. See Figure 5 as shown.

[0079] From Figure 5 it can be seen that the OD passenger flow characteristics are adaptively aggregated with the learnable node embedding vectors, and at the same time, the graph convolutional network is used to capture the spatial dependencies of the neighborhood relying on the traffic topological adjacency matrix. Specifically, the input of the adaptive spatial feature extraction module is divided into three parts: the OD passenger flow feature representation X T ∈R N×k×F output by the feature encoding layer, the learnable node embedding vector E a ∈R N×N , and the adjacency matrix A ∈ R N×N .

[0080] First, for the adaptive spatial feature aggregation part, a learnable node embedding vector is predefined, and then the relationship between nodes is measured by the similarity of the embedding vectors, expressed as:

[0081] Ω a =sofmax(ReLU(E a *E a T )) (12)

[0082] In the formula: Ω a ∈R N×N is the convolution kernel of the adaptive spatial aggregation part, E a ∈R N×N is the learnable node embedding vector, and E a T ∈R N×N is the transpose of the learnable node embedding vector.

[0083] The role of ReLU is to sparsify the matrix, and then softmax is used to normalize the matrix to obtain an aggregation convolution kernel similar to the attention matrix. The overall adaptive aggregation formula can be expressed as:

[0084] X a =(I N +Φ a )X T Φ a (13)

[0085] In the formula, Φ a is the parameter matrix used to transform the dimension of the feature, and I Nis the identity matrix, X a ∈R N×k×F is the output feature.

[0086] Secondly, the formula for the graph convolution part based on the neighborhood adjacency matrix can be expressed as:

[0087]

[0088] where: D is the degree matrix, Φ g is the parameter matrix, X g ∈R N×k×F is the output feature.

[0089] Then, in order to fuse the two spatial representations, first concatenate the features:

[0090] X ag = Concat(X a , X g ) (15)

[0091] Finally, the fully connected layer aggregates the two different spatial information:

[0092] X rs = Linear(X ag ) (16)

[0093] where, X rs is the feature representation output by the spatial feature extraction module.

[0094] Based on this, use X ws , X ds , X rs to represent the outputs after spatial feature extraction of the weekly pattern, daily pattern, and real-time passenger flow channels respectively.

[0095] Since both the weekly pattern and the daily pattern are passenger flow data corresponding to the predicted time period, such data provides more accurate spatial passenger flow analysis information. The real-time passenger flow data is passenger flow data similar to that in the predicted time period and has more accurate time-related scale information. Based on this, the three passenger flow characteristics are concatenated and fused through a linear layer to output the fused passenger flow characteristics.

[0096] Specifically, after the three-mode data are extracted by features, they will all output a feature matrix of N×k×F. Add and fuse the three OD feature matrices with different rules, and use a linear layer to eliminate the aliasing effect between features. The formula is as follows:

[0097] X rdw = X ws + X ds + X rs (17)

[0098]

[0099] In the formula, is the fused OD feature representation.

[0100] 3. Heterogeneous data fusion module

[0101] Since the confirmed case information can reflect the spread situation and risk level of the event, and help the rail transit operator to determine the key areas affected by the event and the stations in the areas of high-risk locations. In order to integrate the real-time event information into the model, a heterogeneous data fusion module is designed, as shown in Figure 6 shown.

[0102] Specifically, since the data of public health emergencies is only an infection quantity C ∈ R 1×1 , each OD pair will be affected by the public health emergency. Therefore, this data is extended, and a fully connected layer is used to extract deep features. The formula is as follows:

[0103] C′ = Linear(repeat(C)) (19)

[0104] In the formula, C ∈ R 1×1 is the number of confirmed cases, repeat is to extend the daily confirmed number to each moment, Linear is the fully connected layer, and C′ ∈ R N×k×F is the deep representation of the public health emergency data.

[0105] However, the severity of the impact of each OD pair by the public health emergency varies greatly. Intuitively, the larger the passenger flow, the greater the impact. Therefore, the impact weight of the public health emergency data on different stations is calculated, and this weight is weighted to the heterogeneous public health emergency data. The formula is as follows:

[0106]

[0107] C″ = C′ * At wdr (21)

[0108] In the formula, At wdr ∈ R N×k is the weight matrix, and C″ ∈ R N×k×F is the deep representation of the weighted public health emergency data.

[0109] Finally, the confirmed case data and the passenger flow data are concatenated in the feature dimension, and a fully connected layer is used to fuse the two heterogeneous data. The formula is as follows:

[0110]

[0111] In the formula, X′ rdw ∈ RN×k×F For the final passenger flow feature representation.

[0112] To further verify the effectiveness of the present invention, the prediction performance of MFST-GNN is verified on a real dataset. In the following, the dataset used in the research is first introduced, then the parameter settings and evaluation metrics of the model are introduced. Next, several conventional passenger flow prediction benchmark models selected in the research are introduced, and finally, the experimental results are analyzed from multiple perspectives.

[0113] 1. Dataset

[0114] OD demand data: A real-world large-scale dataset is used for short-term OD passenger flow prediction of urban rail transit during this event. This dataset consists of billions of AFC records from the rail transit system in Nanning, Guangxi, China. The dataset contains AFC data of all 62 stations in the rail transit network, and its time period is about three to four months of a certain year, including the main stages of the development of a sudden disease discovered in the previous year, namely the outbreak stage, the containment stage, and the stable stage. Affected by the sudden public health event, the OD demand of the rail transit system fluctuates greatly. Specifically, at the initial stage of the outbreak of the sudden public health event, due to the risk of infection among residents, the number of passengers taking public transportation decreases, and the OD demand drops. As the sudden public health event is brought under control, the OD demand gradually returns to normal. Therefore, the dataset in this paper is a typical rail transit passenger flow dataset with obvious characteristics of sudden public health events, which can fully verify the ability of the model in OD demand prediction during sudden public health events. According to the AFC records, the historical OD demand matrix and the real-time inflow vector for each time interval t can be generated. In this dataset, the time intervals are set to 15 minutes, 30 minutes, and 60 minutes respectively to effectively verify the prediction performance of the model in the short term. Finally, the dataset is divided into a training set, a validation set, and a test set with a ratio of 7:2:1.

[0115] Event-related data: The event-related dataset mainly consists of the confirmed data during the event. The data is sourced from the official website of the National Health Commission of China and includes the daily confirmed case data within a three- to four-month period in Nanning, which can directly reflect the evolution trend of the health event. The correlation between the event-related data and the OD demand data was studied, and the Pearson coefficient between the daily OD demand data and the daily event-related sequence data (daily confirmed case data) was calculated. Specifically, first, several pairs of OD pairs were randomly selected, and their time series data was extracted. Then, the Pearson coefficients between these OD demand sequence data and the event-related sequence data were calculated using SPSS respectively. It can be found that the absolute values of the Pearson coefficients of the OD demand sequence data and the event-related sequence data are all greater than 0.6, indicating a significant correlation between the OD demand sequence data and the event-related sequence data. Therefore, using the event information data to better predict the subway OD demand during the health event is a reasonable method.

[0116] 2. Model Configuration

[0117] The experimental verification used the PyTorch deep learning framework to build the model. The k in the multi-frequency domain time feature extraction module was set to 3. The feature dimension in the model was uniformly set to 64. The Xavier initializer was used to initialize the parameters related to the convolutional neural network. For example, the input historical time step was selected as 12, and the batch size was 32. The learning rate was set to 0.01. The "Adam" optimizer was used. The model code was implemented in the PyTorch framework, and all experiments were compiled and executed on the Windows system. And on a desktop computer with Core TM an i9-10900X processor, 32GB of running memory, and an NVIDIA GeForce RTX3080 GPU for computing.

[0118] 3. Evaluation Metrics

[0119] In the experimental verification, the Mean Square Error (MSE) was used as the loss function, and the Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and weighted Mean Absolute Percentage Error (WMAPE) were used as the evaluation metrics for the model prediction effect.

[0120] Considering the quantitative relationship between the OD demand matrix and the inbound flow, the present invention uses the PINN loss function as the loss function of the model, as shown in Formula (23). The root mean square error RMSE, mean absolute error MAE, and weighted mean absolute percentage error WMAPE are used as the model evaluation indicators, and the three indicators are calculated according to Formulas (24)-(26).

[0121]

[0122] In the formula, x t (i,j) is the true value, is the predicted value, is the real-time inbound flow sequence, W od is the influence weight of OD passenger flow, W in is the influence weight of inbound passenger flow, t represents different time periods, and N represents the number of stations.

[0123] 4. Benchmark Model

[0124] To comprehensively evaluate the prediction performance of MFSF-GNN, the prediction effects of MFSF-GNN and the benchmark model are compared on the passenger flow dataset during the sudden public health event in Nanning Metro. The details of the benchmark model are described as follows.

[0125] MLP (fully connected neural network). In this paper, a neural network composed of two layers of fully connected neural networks is constructed. It can be applied to OD demand prediction to extract the deep representation and time relationship of passenger flow between irregular regions in the city. The learning rate is 0.0001, the batch size is 32, and the feature dimension of the hidden layer is 64.

[0126] GRU (gated recurrent unit). This model is the most basic recurrent neural network, with a gating mechanism to model the time dependence of time series data. The learning rate is 0.0001, the batch size is 32, and the feature dimension of the hidden layer is 64.

[0127] DCRNN (diffusion convolutional recurrent neural network), which is a deep learning model for predicting traffic flow. It uses bidirectional random walks on the graph to capture spatial correlations and adopts a pre-determined sampling encoder-decoder architecture to capture time correlations. The input graph structure is a neighborhood graph, the learning rate is 0.0001, the batch size is 32, and the feature dimension of the hidden layer is 64.

[0128] ST-ResNet (spatiotemporal residual network), based on a multi-branch framework, uses a residual neural network to extract spatial features and dynamically integrates the outputs of the residual neural network branches. The learning rate is 0.0001, the batch size is 32, and the feature dimension of the hidden layer is 64.

[0129] STGCN (Spatio-Temporal Graph Convolutional Network) integrates graph convolution and gated temporal convolution in multiple convolutional blocks. In each block, two temporal gated convolutional layers are used to capture temporal dynamics, and one graph convolutional layer is used to capture spatial correlations. The input graph structure is a neighborhood graph. The learning rate is 0.0001, the batch size is 32, and the feature dimension of the hidden layer is 64.

[0130] AGCRN (Adaptive Graph Convolutional Recurrent Network). This method designs a node-adaptive parameter learning module to capture different node-specific patterns and adaptively generate a graph structure to infer the interdependencies between different traffic time series. The overall spatio-temporal modeling prediction is completed based on a recurrent network. The learning rate is 0.0001, the batch size is 32, and the feature dimension of the hidden layer is 64.

[0131] ODCRN (OD Convolutional Recurrent Network) integrates recurrence and two-dimensional graph convolutional neural networks to handle the highly complex spatio-temporal dependencies in the sequence OD matrix. And information on public health emergencies is incorporated into the network to extract the impact of public health emergency information on the OD passenger flow distribution. The learning rate during training is 0.0001, the batch size is 32, and the feature dimension of the hidden layer is 64.

[0132] 5. Result Analysis

[0133] 1) Comparison of Prediction Performance of Different Models

[0134] First, the prediction performance of the present invention and the mainstream OD passenger flow methods was compared on datasets with three granularities of 15 min, 30 min, and 60 min. The results are shown in Tables 1, 2, 3 and Figure 7 , where Figure 7 (a) corresponds to the MAE evaluation index, Figure 7 (b) corresponds to the RMSE evaluation index, Figure 7(c) Corresponding to the WMAPE evaluation index. It can be seen that the deep learning model MFST-GNN proposed by the present invention for short-term OD passenger flow prediction during major public health emergencies has higher accuracy compared with other baseline models. The MAE, RMSE, and WMAPE under the three time granularities are 0.7331, 1.2375, and 2.5354, 1.5399, 2.7119, and 6.2174, 0.6531, 0.5502, and 0.5635 respectively. Compared with other baseline models, the MAE, RMSE, and WMAPE have all improved. At the 15-minute granularity, compared with other baseline models, the MAE, RMSE, and WMAPE have increased by 7.99%, 15.25%, and 8.01% on average. At the 30-minute granularity, compared with other baseline models, the MAE, RMSE, and WMAPE have increased by 14.32%, 22.38%, and 13.57% on average. At the 60-minute granularity, compared with other baseline models, the MAE, RMSE, and WMAPE have increased by 7.11%, 12.91%, and 7.47% on average.

[0135] Table 1: Effect of the model under the 15-minute time granularity

[0136]

[0137] Table 2: Effect of the model under the 30-minute time granularity

[0138]

[0139] Table 3: Effect of the model under the 60-minute time granularity

[0140]

[0141] The deep learning models MLP and GRU learn the high-dimensional information of the passenger flow itself based on the neural network mechanism and explicitly model the spatial or temporal dependence of OD demand. However, since these models fail to learn the passenger flow change law during public health emergencies well, the results are slightly inferior. By simultaneously modeling the spatio-temporal correlation relationship of OD passenger flow, some mainstream spatio-temporal composite models (such as DCRNN, GWN, and STGCN) perform better than some basic deep learning models, which illustrates the importance of fully exploring complex spatio-temporal dependence for improving prediction performance. The mainstream OD demand prediction method AGCRN did not achieve satisfactory results on this dataset. The poor prediction performance may be due to the fact that these methods mainly focus on the passenger flow demand prediction between regions or within regions, and their spatio-temporal dependence is significantly different from the rail transit OD demand. In addition, ODCRN takes into account the relationship between the passenger flow at the origin and destination stations and the OD demand, so as to perform spatio-temporal modeling. However, due to the severe sparsity of the OD demand matrix during the event, this model did not achieve good prediction results.

[0142] Most existing models only focus on historical OD demand information, while ignoring external factors that affect the OD demand distribution during events and the changes in OD demand. Compared with the baseline model, the present invention not only deeply integrates multi-period OD passenger flow information to capture periodic spatio-temporal dependencies for modeling the OD demand distribution, but also introduces heterogeneous data sources, namely event-related data, which can mine more effective information and reflect the importance of confirmed data during events. Therefore, the model of the present invention outperforms other existing methods at all time granularities and achieves the lowest RMSE, MAE, and WMAPE in the OD demand prediction task of the urban rail transit network during the health event.

[0143] 2) Comparison of prediction effects in different time periods

[0144] During the occurrence of a major health event, the passenger flow prediction performance of the model is different in different time periods. Due to the social effects generated by the health event, residents may be more inclined to choose to travel during off-peak hours, which is different from normal travel. Therefore, to further study the prediction effect of MFST-GNN in different time periods, the MAE index of each time period of the model was calculated at granularities of 15 min, 30 min, and 60 min, and its prediction effect is as Figure 8 、 Figure 9 and Figure 10 shown.

[0145] As can be seen from the figure, the model of the present invention has relatively stable prediction accuracy in different time periods, indicating that the model has learned relatively fully about the irregular passenger flow distribution during the occurrence of the major health event, and can accurately capture the change trends and laws of OD passenger flow at different time intervals. The error between the prediction result and the true value is small, indicating that the accuracy of the model is high. At the same time, the prediction performance of the prediction model at different time intervals is better than that of the baseline model, indicating that the model has strong generalization ability and can achieve good prediction effects in different scenarios. Moreover, the performance of the prediction model at different time intervals is relatively stable, without obvious performance fluctuations or abnormal situations, which indicates that the modeling and training process of the model is reliable. Finally, compared with the baseline model, the prediction model of the present invention adopts more effective feature extraction, model structure design, and training optimization methods, so as to achieve better prediction results at different time intervals.

[0146] In summary, among the prediction results at different time intervals, the MAE index of the prediction model provided by the present invention is the smallest compared with different baseline models, indicating that the prediction model has high accuracy, strong generalization ability and good stability, and can obtain relatively accurate results for OD passenger flow prediction at different time intervals. Overall, it has good prediction effects both in individual time periods and as a whole. Whether it is the peak period or the stable period of passenger flow, MFST-GNN has good robustness and does not show large fluctuations, that is, it can adapt to the passenger flow characteristics of different time periods.

[0147] 3) Ablation experiment research

[0148] Since OD passenger flow has obvious long-period and short-period characteristics, and at the same time for the multi-period channel architecture proposed by the present invention, the change of the model's OD passenger flow prediction performance by considering periodic learning is considered. Specifically, the influence of multi-period OD demand on subway OD demand prediction is explored, that is, the long-term and short-term OD distribution information of the weekly OD pattern and the daily OD pattern is used to predict future OD demand. The following introduces several variants:

[0149] Real-time OD demand matrix: Only contains real-time passenger flow channels, that is, only the OD demand matrix is input. Then there is no operation of fusing the OD demand matrices of the three cycle types in the network.

[0150] Real-time + weekly OD demand matrix: Input the real-time OD demand matrix and the weekly OD demand matrix at the same time. Only the features of the real-time OD demand matrix and the weekly OD demand matrix are fused.

[0151] Real-time + daily OD demand matrix: Input the real-time OD demand matrix and the daily OD demand matrix at the same time. Only the features of the real-time OD demand matrix and the daily OD demand matrix are fused.

[0152] The prediction performance of all variables is summarized in Table 4. Compared with only the real-time OD demand matrix, for the 15-minute granularity dataset, the average increases in MAE, RMSE and WMAPE are 2.89%, 7.13% and 3.73% respectively. For the 30-minute granularity dataset, the average increases in MAE, RMSE and WMAPE are 4.82%, 9.44% and 7.76% respectively. For the 60-minute granularity dataset, the average increases in MAE, RMSE and WMAPE are 6.37%, 11.36% and 5.92% respectively.

[0153] Table 4: Ablation experiment of multi-period OD passenger flow

[0154]

[0155] It can be observed that the pure real-time OD has poor prediction performance in all experiments because this variable only uses limited real-time information for OD demand prediction and ignores the long-term and short-term periodicity of OD distribution. It can also be observed that since short-term real-time spatio-temporal information can more accurately describe the evolution pattern of periodic OD, the variable Network Real-time + Daily Passenger Flow is better than Network Real-time + Weekly Passenger Flow. Therefore, when considering the weekly OD pattern or the daily OD pattern, the prediction performance of the model has a certain improvement, indicating that periodic learning enables the model to better adapt to the periodic changes in OD passenger flow data. For example, in urban traffic, the passenger flow may show obvious morning and evening peaks within a day and different characteristics on weekdays and weekends within a week. Through periodic learning, the model can better capture the laws of such periodic changes, thereby improving the prediction performance. Moreover, periodic learning can improve the robustness of the model to periodic changes in the data. Since the fusion of multi-period data has learned the periodic laws in the data, even when the data changes to a certain extent or there are anomalies, the model can still maintain good prediction performance.

[0156] In summary, through periodic learning, the model can better capture the laws of such periodic changes. At the same time, periodic learning can have a positive impact on the OD passenger flow prediction performance of the model, helping the model better adapt to the periodic changes in the data, and improving the prediction accuracy, generalization ability and robustness.

[0157] In addition, in the present invention, a new heterogeneous information fusion module is introduced to fully mine the useful information hidden in the heterogeneous event-related data and fully study the impact of this event information on the passenger flow during major public health emergencies. Here, the impact of heterogeneous information obtained from multi-source data on the model prediction performance is further studied, and ablation experiments are respectively carried out with and without confirmed case information. The results are shown in Table 5.

[0158] As can be seen from Table 5, when considering heterogeneous information, the model of the present invention achieves the best prediction performance at all time intervals. Specifically, since the data related to major public health emergencies is significantly correlated with the rail transit passenger flow during major public health emergencies, it can affect the passenger flow distribution of OD from a global perspective.

[0159] In summary, adding heterogeneous data such as the number of confirmed cases to the deep learning model for OD passenger flow prediction can improve the prediction accuracy, generalization ability and decision-making support ability of the model, so as to better cope with the passenger flow prediction problem during major public health emergencies. Considering this heterogeneous information in the model of the present invention can effectively utilize the impact of the confirmed data of public health events on the OD passenger flow distribution.

[0160] Table 5: Ablation Experiment of Heterogeneous Information

[0161]

[0162] In summary, the present invention combines the multi-source data structure of passenger flow data with long and short periods and event text information, designs a deep learning framework based on multi-frequency domain spatio-temporal feature mining, and simultaneously integrates the temporal features and information in major public health emergencies for passenger flow prediction of rail transit during major public health events. Compared with the prior art, it has the following advantages:

[0163] 1) In terms of the model framework, the MFST-GNN model proposed by the present invention effectively captures the periodic characteristics of passenger flow by integrating real-time, daily, and weekly OD demand patterns. In particular, the multi-frequency domain time feature extraction module and the adaptive space feature extraction module can more accurately capture the complex features in passenger flow data. The heterogeneous information fusion module better adapts to the impact of events on passenger flow by deeply integrating event-related information into the model, while the prior art has not fully explored the potential value of event information for prediction.

[0164] 2) The present invention uses the multi-frequency domain time feature extraction module to mine the time features of OD demand from a multi-frequency domain perspective through Fourier transform and convolutional network, which can better cope with the randomness and uncertainty of time features during public health emergencies. At the same time, the adaptive space feature extraction module shows strong adaptability in abnormal scenarios. It combines dynamic adaptive convolutional kernels and static graph convolutional operations, fully considering the highly dynamic OD distribution during events.

[0165] 3) It has been verified that the prediction effect of the model proposed by the present invention is the best in both dataset experimental studies and ablation experimental studies. For example, for RMSE, the prediction index is improved by 15.25% compared with the currently optimal prediction model at a 15-minute granularity; the prediction index is improved by 22.38% compared with the currently optimal prediction model at a 30-minute granularity; the prediction index is improved by 12.91% compared with the currently optimal prediction model at a 60-minute granularity. For MAE, the prediction index is improved by 7.99% compared with the currently optimal prediction model at a 15-minute granularity; the prediction index is improved by 14.32% compared with the currently optimal prediction model at a 30-minute granularity; the prediction index is improved by 7.11% compared with the currently optimal prediction model at a 60-minute granularity. For WMAPE, the prediction index is improved by 8.01% compared with the currently optimal prediction model at a 15-minute granularity; the prediction index is improved by 13.57% compared with the currently optimal prediction model at a 30-minute granularity; the prediction index is improved by 7.47% compared with the currently optimal prediction model at a 60-minute granularity. From the comprehensive prediction results, it can be seen that MFST-GNN effectively improves the accuracy of passenger flow prediction during major emergencies and can be used to guide engineering practice.

[0166] The present invention may be a system, a method, and / or a computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions thereon for causing a processor to implement aspects of the present invention.

[0167] The flowcharts and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in the flowchart or block diagram may represent a module, a segment of a program, or a part of an instruction, which contains one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions noted in the blocks may occur in an order different from that noted in the figures. For example, two consecutive blocks may in fact be executed substantially in parallel, or they may sometimes be executed in the reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented by a dedicated hardware-based system that performs the specified functions or actions, or by a combination of dedicated hardware and computer instructions. As is well known to those skilled in the art, implementation by hardware, implementation by software, and implementation by a combination of software and hardware are all equivalent.

[0168] The embodiments of the present invention have been described above. The above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and variations will be apparent to those of ordinary skill in the art in the technical field without departing from the scope and spirit of the described embodiments. The selection of the terms used herein is intended to best explain the principles of the embodiments, the practical application, or the improvement of technologies in the market, or to enable other ordinary skilled persons in the technical field to understand the embodiments disclosed herein. The scope of the present invention is defined by the appended claims.

Claims

1. A method for predicting short-term OD passenger flow in public health emergencies based on deep learning, comprising the following steps: Obtaining multiple modes of OD demand matrices and event-related information observed at the t-th time step during a public health emergency, wherein the multiple modes of OD demand matrices include a weekly OD demand matrix, a daily OD demand matrix, and a real-time OD demand matrix; Based on the OD demand matrices of the multiple modes and the event-related information, the trained deep learning model is used to predict the OD passenger flow matrix of the future time step to obtain the passenger flow prediction result of the target site.

2. The method according to claim 1, characterized in that The deep learning model includes a feature coding layer, a multi-cycle time feature extraction module, an adaptive spatial feature extraction module, a heterogeneous information fusion module and a fully connected layer, wherein the feature coding layer is used to extract the deep passenger flow features corresponding to the weekly OD demand matrix, the daily OD demand matrix and the real-time OD demand matrix; for the deep passenger flow features corresponding to the weekly OD demand matrix and the daily OD demand matrix, the adaptive spatial feature extraction module is used to extract spatial features; for the deep passenger flow features corresponding to the real-time OD demand matrix, the multi-cycle time feature extraction module is used to extract time features, and the adaptive spatial feature extraction module is used to extract spatial features; the heterogeneous information fusion module is used to fuse event-related information into the fused OD features output by the adaptive spatial feature extraction module to obtain fused heterogeneous passenger flow features; the fully connected layer outputs a predicted OD passenger flow matrix based on the fused heterogeneous passenger flow features.

3. The method according to claim 2, characterized in that The multi-cycle time feature extraction module performs the following steps: Get the deep passenger flow features X corresponding to the daily OD demand matrix e , expressed as: X e =Embedding Layer(X) Among them, X is the real-time OD demand matrix, and Embedding Layer is the feature encoding layer; X e The time dimension is fast Fourier transformed to extract the K main frequencies of the features, which are expressed as: A=Avg(Amp(FFT(X e ))) f1,…,f k =argTopk(A) Where A is X e The intensity of each frequency component in, FFT is Fourier transform, Amp is the calculation of amplitude value, {f1,...,f k } are the k frequencies with the largest intensity; Based on the selected frequency and corresponding frequency length, the original one-dimensional time series X e Fold and get a set of two-dimensional vectors Then, 2D convolution is used to extract the features of the time series, which is expressed as: in, are the features of different frequencies after convolution; For different amplitudes The Softmax function is used to calculate the weight, which is expressed as: in, is the weight of the feature at different frequencies; Using the fully connected layer, we can obtain the temporal features, which can be expressed as: Among them, X T It is the time feature corresponding to the real-time OD demand matrix output by the multi-cycle time feature extraction module.

4. The method according to claim 2, characterized in that: The adaptive spatial feature extraction module performs the following steps: Get the convolution kernel Ω for adaptive spatial aggregation a , expressed as: Ω a =softmax(ReLU(E a *AND a T )) Among them, E a is a learnable node embedding vector, E a T is the learnable transpose of the node embedding vector; The output feature X of adaptive spatial aggregation is calculated using the following formula: a : X a =(I N +F a )X T F a Among them, Φ a is the parameter matrix used to transform the dimension of the feature, I N is the identity matrix, X T It is the OD passenger flow feature representation output by the feature encoding layer; The output feature X of the graph convolution is calculated according to the following formula g : Where D is the degree matrix, Φ g is the parameter matrix, A is the adjacency matrix of the traffic topology structure; According to the following formula, X a and X g Perform splicing to obtain splicing feature X ag : X ag =Concat(X a ,X g ) Using the fully connected layer to ag Aggregation is performed to obtain the output of the weekly mode, daily mode and real-time passenger flow channel extracted by the adaptive spatial feature extraction module, which are represented by X ws , X ds and X rs .

5. The method according to claim 4, characterized in that The fused OD features output by the adaptive spatial feature extraction module Expressed as in: X rdw =X ws +X ds +X rs Among them, X ws , X ds , X rs They respectively represent the outputs of the weekly mode, daily mode and real-time passenger flow channel after passing through the adaptive spatial feature extraction module.

6. The method according to claim 2, characterized in that The heterogeneous information fusion module performs the following steps to obtain the fused heterogeneous passenger flow features: Use the fully connected layer to extract the deep features C of the event related information ′ , expressed as: C ′ =Linear(repeat(C)) The deep feature C is calculated using the following formula ′ After weighting, the weighted deep feature C″ is obtained, which is expressed as: C″=C ′ *At wdr Among them, At wdr is the weight matrix, expressed as: The fused heterogeneous passenger flow feature X is obtained according to the following formula: ′ rdw , expressed as: Among them, C is the number of confirmed cases per day, repeat means expanding the number of confirmed cases per day to each moment, and Linear is the fully connected layer.

7. The method according to claim 1, characterized in that The OD demand matrices of the multiple modes are pre-processed by compression, and the compression method is: for each starting station, the destination distribution of all incoming passengers is analyzed, and then the first k-1 stations with the highest probability of passengers going to the destination are selected, and then the OD demands of all remaining stations are aggregated as the kth station.

8. The method according to claim 1, characterized in that The physical information neural network PINN loss function is used to supervise the training of the deep learning model.

9. A computer-readable storage medium having a computer program stored thereon, wherein: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

10. A computer device comprising a memory and a processor, wherein a computer program capable of running on the processor is stored in the memory, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

Images