Rail transit passenger flow prediction method and device

Abstract

The invention relates to a rail transit passenger flow prediction method and device, and the method comprises the steps: constructing a plurality of historical passenger flow data sequences for each rail transit station in a rail transit system based on the number of people entering the station in each time interval in a real-time window, the tth historical passenger flow data sequence comprises the number of passengers entering the station before the tth time interval; segmenting each historical passenger flow data sequence of each rail transit station according to a preset time period to obtain a plurality of historical passenger flow data subsequences, and constructing the plurality of historical passenger flow data subsequences into a two-dimensional matrix to obtain a historical passenger flow data matrix; processing each historical passenger flow data matrix of each rail transit station through a convolution residual block to obtain each time feature of each rail transit station; and obtaining a passenger flow prediction value of the rail transit system based on each time characteristic of each rail transit station through the multi-graph convolutional neural network.

Description

Technical Field

[0001] The present application relates to the field of traffic forecasting, and in particular to a method and device for predicting rail transit passenger flow. Background Art

[0002] The operating efficiency and service quality of rail transit systems are highly dependent on the accuracy of passenger flow prediction. By accurately predicting the passenger flow at time intervals such as 15 minutes, 30 minutes, 45 minutes or one hour in the future, appropriate operating strategies can be implemented to prevent overcrowding and mitigate safety hazards. The passenger flow of urban subway stations in the future can be regarded as a spatiotemporal prediction problem. Solving such problems mainly involves analyzing two aspects. The first aspect is the time dimension. The passenger flow of a specific station in the future time period is affected by the passenger flow in a certain time span in the past. However, the existing rail transit passenger flow prediction methods fail to make reasonable use of historical passenger flow data, and it is difficult to fully extract temporal features from historical passenger flow data, resulting in low prediction accuracy. The other aspect is the spatial dimension. Graph convolutional networks have a strong ability to mine the spatial correlation and topological information of graph structures, and have received increasing attention in recent years. However, the existing rail transit passenger flow prediction methods usually directly use only the physical topological relationship between stations for prediction, ignoring the non-Euclidean relationship between stations, resulting in insufficient extraction of spatial features and low prediction accuracy.

[0003] In summary, the existing rail transit passenger flow prediction methods have the problem of low prediction accuracy. Summary of the invention

[0004] The present invention provides a rail transit passenger flow prediction method and device to solve the problem of low prediction accuracy in existing rail transit passenger flow prediction methods.

[0005] In a first aspect, the present invention provides a rail transit passenger flow prediction method, comprising:

[0006] For each rail transit station in the rail transit system, multiple historical passenger flow data sequences are constructed based on the number of people entering the station in each time interval in the real-time window, and the t-th historical passenger flow data sequence includes the number of people entering the station before the t-th time interval;

[0007] For each historical passenger flow data sequence of each rail transit station, segment it according to a preset time period to obtain multiple historical passenger flow data subsequences, and construct the multiple historical passenger flow data subsequences into a two-dimensional matrix to obtain a historical passenger flow data matrix;

[0008] The historical passenger flow data matrix of each rail transit station is processed through the convolution residual block to obtain the time characteristics of each rail transit station;

[0009] The passenger flow prediction value of the rail transit system is obtained based on the time characteristics of each rail transit station through a multi-graph convolutional neural network.

[0010] In a second aspect, the present invention provides a rail transit passenger flow prediction device, comprising:

[0011] A sequence construction module, for each rail transit station in the rail transit system, constructs multiple historical passenger flow data sequences based on the number of people entering the station in each time interval in the real-time window, and the t-th historical passenger flow data sequence includes the number of people entering the station before the t-th time interval;

[0012] A matrix construction module is used to divide each historical passenger flow data sequence of each rail transit station according to a preset time period to obtain multiple historical passenger flow data subsequences, and construct the multiple historical passenger flow data subsequences into a two-dimensional matrix to obtain a historical passenger flow data matrix;

[0013] The feature extraction module is used to process each historical passenger flow data matrix of each rail transit station through a convolution residual block to obtain each time feature of each rail transit station;

[0014] The passenger flow prediction module is used to obtain the passenger flow prediction value of the rail transit system based on the various time characteristics of each rail transit station through a multi-graph convolutional neural network.

[0015] Compared with the related art, the present invention provides a method for predicting passenger flow in rail transit, which uses the historical passenger flow data of the rail transit system in a real-time window for prediction. The key points are: first, converting the historical passenger flow data from a sequence form into a two-dimensional matrix form, so that the convolution residual block can extract time features from it; second, constructing a variety of different graph structures based on time features, including physical graph structures, similarity graph structures, and correlation graph structures, and using a multi-graph convolutional neural network to process a variety of graph structures, so that spatial features can be more fully extracted from historical data. Therefore, the above method can fully extract time features and spatial features from the historical passenger flow data of the rail transit system, and can more accurately predict the passenger flow of each rail transit station in the future, solving the problem of low prediction accuracy in existing rail transit passenger flow prediction methods.

[0016] Details of one or more embodiments of the present application are set forth in the following drawings and description to make other features, objects, and advantages of the present application more readily apparent. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a flow chart of the rail transit passenger flow prediction method provided in this embodiment;

[0018] Figure 2It is a structural diagram of the spatiotemporal graph neural network provided in this embodiment. DETAILED DESCRIPTION

[0019] In order to more clearly understand the purpose, technical solutions and advantages of the present application, the present application is described and illustrated below in conjunction with the accompanying drawings and embodiments.

[0020] Unless otherwise defined, the technical terms or scientific terms involved in this application shall have the general meaning understood by people with ordinary skills in the technical field to which this application belongs. The words "one", "a", "a", "the", "these" and the like in this application do not represent quantitative restrictions, and they can be singular or plural. The terms "include", "comprise", "have" and any variants thereof involved in this application are intended to cover non-exclusive inclusions; for example, a process, method and system, product or device comprising a series of steps or modules (units) is not limited to the listed steps or modules (units), but may include unlisted steps or modules (units), or may include other steps or modules (units) inherent to these processes, methods, products or devices. The words "connect", "connected", "coupled" and the like involved in this application are not limited to physical or mechanical connections, but may include electrical connections, whether directly or indirectly. The "multiple" involved in this application refers to two or more. "And / or" describes the association relationship of associated objects, indicating that there can be three relationships. For example, "A and / or B" can mean: A exists alone, A and B exist at the same time, and B exists alone. Usually, the character " / " indicates that the objects associated with each other are in an "or" relationship. The terms "first", "second", "third", etc. involved in this application are only used to distinguish similar objects and do not represent a specific ordering of the objects.

[0021] An embodiment of the present invention provides a rail transit passenger flow prediction method.

[0022] Figure 1 is a flow chart of the rail transit passenger flow prediction method provided in this embodiment. Figure 1 As shown, it includes step S110, step S120, step S130 and step S140.

[0023] Step S110, for each rail transit station in the rail transit system, multiple historical passenger flow data sequences are constructed based on the number of passengers entering the station in each time interval in the real-time window, and the t-th historical passenger flow data sequence includes the number of passengers entering the station before the t-th time interval.

[0024] During the operation of the rail transit system, the card swiping records of each rail transit station will be obtained in real time. Each card swiping record contains: the initial station, the end station, the start time, and the end time. Based on these card swiping records, the number of people entering each rail transit station in each time interval can be obtained.

[0025] Can be defined is the i-th historical passenger flow data sequence of the j-th rail transit station, where is the number of people entering the j-th rail transit station in the i-th time interval. From this definition, it can be seen that the number of people entering the station before the t-th time interval includes the number of people entering the station from the 1st to the t-th time interval.

[0026] Step S120, for each historical passenger flow data sequence of each rail transit station, segment it according to a preset time period to obtain multiple historical passenger flow data subsequences, and construct the multiple historical passenger flow data subsequences into a two-dimensional matrix to obtain a historical passenger flow data matrix.

[0027] In order to allow the convolution residual block to more fully extract the time features in the historical passenger flow data sequence, the present invention converts the historical passenger flow data sequence into a matrix form. The preset time period can be set according to the situation, for example, it can be composed of 5 or 10 time intervals, that is, a historical passenger flow data subsequence contains the number of people entering the station in 5 or 10 time intervals.

[0028] In this embodiment, constructing multiple historical passenger flow data subsequences into a two-dimensional matrix to obtain a historical passenger flow data matrix specifically includes: taking multiple historical passenger flow data subsequences as different rows in the two-dimensional matrix to obtain a historical passenger flow data matrix. Among them, the t-th historical passenger flow data matrix of the j-th rail transit station is:

[0029]

[0030] in, represents the number of people entering the j-th rail transit station in the x-th (representing the superscript f)-th time interval, T represents the number of time intervals in a single preset time period, and D represents the number of preset time periods.

[0031] The first row of the above matrix represents the number of people entering the station in the T time intervals closest to the t-th time interval, and the last row of the matrix represents the number of people entering the station in the T time intervals farthest from the t-th time interval. The above row arrangement is better than other row arrangements, and the time features in the time data can be better extracted.

[0032] Step S130, processing each historical passenger flow data matrix of each rail transit station through a convolution residual block to obtain each time feature of each rail transit station.

[0033] Specifically, the convolution residual block contains multiple layers of convolution (Conv2D), batch normalization (BN), activation function (ReLU), and residual connection (Residual Connection). The convolution kernel size can be set to m×m, according to the formula After the residual block operation, we finally get It represents the t-th time feature (multi-period feature) of the j-th rail transit station, and combines the time of all stations into a four-dimensional tensor It is the tth time characteristic of the rail transit system.

[0034] Step S140, obtaining the passenger flow prediction value of the rail transit system based on the various time characteristics of each rail transit station through a multi-graph convolutional neural network.

[0035] Specifically, based on the time characteristics of each rail transit station and the adjacency status, similarity and correlation between each rail transit station, the physical graph structure, similarity graph structure and correlation graph structure of the rail transit system are constructed; the physical graph structure, similarity graph structure and correlation graph structure of the rail transit system are processed through a multi-graph convolutional neural network to obtain the passenger flow prediction value of the rail transit system.

[0036] In this embodiment, in order to better extract the spatial features of passenger flow data, a multi-graph convolutional neural network is used to extract different spatial features in the passenger flow data.

[0037] The tth physical graph structure of the rail transit system is Among them, V t is the location set of each rail transit station in the real-time window, is the adjacency matrix between each rail transit station in the t-th time in the real-time window. Each element in the adjacency matrix represents the adjacency status between different pairs of rail transit stations. t is the t-th time feature matrix of the rail transit system in the real-time window. Each element in the t-th time feature matrix represents the t-th time feature of different rail transit stations. In the t-th time interval, if rail transit station i is adjacent to rail transit station j, then let The element in the i-th row and j-th column has the value 1, otherwise it has the value 0.

[0038] The t-th similarity graph structure of the rail transit system is Among them, V t is the location set of each rail transit station in the real-time window, is the t-th similarity matrix between each rail transit station in the real-time window. Each element in the similarity matrix represents the similarity between different pairs of rail transit stations. t is the t-th time feature matrix of the rail transit system in the real-time window, and each element in the t-th time feature matrix represents the t-th time feature of different rail transit stations.

[0039] The tth similarity matrix The build steps include a1 and a2.

[0040] Step a1: construct the t-th similarity score matrix between each rail transit station in the real-time window The diagonal elements in St are 0 (to avoid repeated calculations of graph self-loops) and the i-th row and j-th column element S(i,j) in St is:

[0041]

[0042] in, and are the t-th time features of the ith and j-th rail transit stations respectively. DTW is a dynamic time warping algorithm.

[0043] Step a2: Set the elements in the t-th similarity score matrix St that are less than the first threshold (which can be set according to actual conditions, such as 0.3) to zero, and linearly normalize the elements in different rows that are equal to and greater than the first threshold to obtain the t-th similarity matrix Linear normalization is performed row by row. For any element in a row, the sum of the elements in the row is first calculated (the elements smaller than the first threshold are set to 0 in this process), and then the ratio of each element to the sum of the elements is used as the respective normalized value.

[0044] The correlation graph structure of the rail transit system is: Among them, V t is the location set of each rail transit station in the real-time window, is the tth correlation matrix between each rail transit station in the real-time window. Each element in the correlation matrix represents the correlation between different pairs of rail transit stations. t is the t-th time feature matrix of the rail transit system in the real-time window, and each element in the t-th time feature matrix represents the t-th time feature of different rail transit stations.

[0045] The tth correlation matrix The build steps include b1 and b2.

[0046] Step b1: construct the tth correlation score matrix between each rail transit station in the real-time window The element C(i,j) in row i and column j of Ct is:

[0047]

[0048] Among them, D(i,j) is the total number of passengers from the i-th rail transit station to the j-th rail transit station before the t-th time interval in the real-time window, and N is the total number of rail transit stations. It should be noted that C(i,i) needs to be calculated because in the real world, a small number of passengers enter and exit the same station.

[0049] Step b2: Set the elements in the t-th similarity score matrix Ct that are less than the second threshold (which can be set according to actual conditions, such as 0.3) to zero, and linearly normalize the elements in different rows that are equal to or greater than the second threshold to obtain the t-th correlation matrix Linear normalization is performed row by row. For any element in a row, the sum of the elements in the row is first calculated (the elements smaller than the second threshold are set to 0 in this process), and then the ratio of each element to the sum of the elements is used as the respective normalized value.

[0050] The multi-graph convolutional neural network used in this embodiment includes: a multi-graph convolutional layer, a fusion layer, and an output layer; wherein the output layer is composed of a fully connected neural network, and the multi-graph convolutional layer includes three graph convolutional networks, which respectively process the physical graph structure, similarity graph structure, and correlation graph structure of the rail transit system, and the fusion layer fuses the outputs of the three graph convolutional networks and sends them to the output layer, and the output layer fully connects the fused features and outputs the passenger flow prediction value of the rail transit system.

[0051] In summary, the present invention provides a method for predicting passenger flow in rail transit, which uses the historical passenger flow data of the rail transit system in a real-time window for prediction. The key points are: first, converting the historical passenger flow data from a sequence form into a two-dimensional matrix form, so that the convolution residual block can extract time features from it; second, constructing a variety of different graph structures based on time features, including physical graph structures, similarity graph structures, and correlation graph structures, and using a multi-graph convolutional neural network to process a variety of graph structures, so that spatial features can be more fully extracted from historical data. Therefore, the above method can fully extract time features and spatial features from the historical passenger flow data of the rail transit system, and can more accurately predict the passenger flow of each rail transit station in the future, solving the problem of low prediction accuracy in existing rail transit passenger flow prediction methods.

[0052] In the above-mentioned rail transit passenger flow prediction method, the convolution residual block and multi-graph convolutional neural network used together constitute a spatiotemporal graph neural network, that is, the rail transit passenger flow prediction model used in the present invention. Figure 2 : is a structural diagram of the spatiotemporal graph neural network provided in this embodiment. The following is a specific description of the training process of the network model.

[0053] In the prediction process, it is necessary to make predictions based on historical passenger flow data within a real-time window. Therefore, in the training process, the historical passenger flow data within a single window (with the same width as the real-time window) needs to be used as a sample data. First, the historical passenger flow data of the rail transit system in a larger time range (larger than the window width) can be obtained, and then a sliding window (with the same width as the real-time window) can be set to intercept the historical passenger flow data in the larger time range, so as to determine each multiple sample data.

[0054] Then the sample data set D = {(X1, Y1), (X2, Y2), ..., (X m ,Y m ),...,(X L ,Y L )},(X m ,Y m ) represents the mth sample data, X m represents the set of graph structures in the mth sliding window (the input of the multi-graph convolutional neural network), Y m represents the mth label, X m = {G m,1 ,G m,2 ,...,G m,t ,...,G m,T}, G m,t represents the tth graph structure group in the mth sliding window, and They represent the t-th physical graph structure, similarity graph structure, and correlation graph structure in the m-th sliding window, respectively. and They represent the t-th adjacency matrix, similarity matrix and correlation matrix in the m-th sliding window respectively, and H m,t It represents the t-th time feature matrix in the m-th sliding window. The t-th adjacency matrix, similarity matrix, correlation matrix and time feature matrix in the m-th sliding window are all constructed based on the output of the convolutional residual block. At this time, the input of the convolutional residual block is the t-th historical passenger flow data sequence in the m-th sliding window.

[0055] X m Taking the input of a multi-graph convolutional neural network as an example, the processing process of the multi-graph convolutional neural network is:

[0056]

[0057] Among them, H p,l+1,t , H s,l+1,t and H c,l+1,t They are the convolution results of the l+1-layer physical network (graph convolution network for processing physical graph structure), the similarity network (graph convolution network for processing similarity graph structure) and the correlation network (graph convolution network for processing correlation graph structure) for the t-th physical graph structure, similarity graph structure and correlation graph structure in the m-th sliding window, σ is the activation function, I is the unit matrix, are the self-loop operation results for the t-th physical graph structure, similarity graph structure, and correlation graph structure in the m-th sliding window, respectively. are the degree matrices corresponding to the physical network, similarity network, and correlation network, respectively. p,(l) , R s,(l) , R c,(l) They are the parameters to be learned in the l-layer convolution.

[0058] The fusion layer is used to transform the local spatiotemporal features H p,l+1,t , H s,l+1,t and H c,l+1,t Perform weighted processing to obtain the fused spatiotemporal features of the historical passenger flow data of the mth sliding window λ p , s and λ c Indicates different weights.

[0059] The output layer is The dimension is compressed to obtain the prediction result O corresponding to the historical passenger flow data of the mth sliding window. m :

[0060]

[0061] Among them, W h , are two weight parameters to be learned, b h , are two bias parameters to be learned.

[0062] The gradient descent algorithm is used to train the spatiotemporal graph neural network, and the goal is to minimize the loss function. The time back propagation algorithm is used to optimize and update the parameters in the spatiotemporal graph network until the preset number of iterations is met. The training is stopped to obtain a trained spatiotemporal graph network model, which is used to predict the number of rail transit passenger flow in the future time period. The loss function Loss is:

[0063]

[0064] Among them, O m is the prediction result of the spatiotemporal graph neural network in response to the mth sample data, Y m is the mth label, L represents the total number of sample data, and α is the weight parameter.

[0065] The above is a detailed introduction to the training method of the rail transit passenger flow prediction model adopted by the present invention using an example. In order to verify the effectiveness of the rail transit passenger flow prediction method provided by the present invention, a specific experimental example is provided as follows.

[0066] 1. Dataset.

[0067] In the experiment, the AFC dataset of Hefei Rail Transit was used. The dataset mainly includes the subway card swiping data from 6 am to 11 pm from January 1 to December 31, 2020. The time granularity is 10min, 15min, and 30min, and the entry flow time series is extracted at the same time. The extraction of the entry flow time prediction sequence at the 30min granularity is shown in Table 1 below. The station numbers are sorted according to the adjacency relationship between the subway line and the station. In the experiment, the weather conditions and air quality data used are the same as the date of the card data. In order to better capture the characteristics of holiday passenger flow data to improve the prediction accuracy, 6 consecutive months of passenger flow data are used for training and learning. Taking the passenger flow data of March 2020 and August 2020 as an example, a total of 6 months of passenger flow data are included. The passenger flow data of the first week is taken to train and verify the model. The passenger flow data of the remaining week is used to test the prediction effect of the model.

[0068] Table 1. Time prediction sequence of passenger flow entering the station

[0069]

[0070] 2. Prediction model configuration.

[0071] The experiment was conducted on a desktop computer with a CoreT Mi9-10900X CPU, 64GB memory, and NVIDIA GeForce RTX3050 GPU. The model was built using PyTorch.

[0072] Hyperparameters: In order to accurately evaluate the prediction performance of the model, the parameters of MPGCN (the rail transit passenger flow prediction model adopted by the present invention) can be set as follows: the residual block contains 64 filters, and the convolution kernel size is set to 3×3. The convolutional attention network can be set to consist of 1-3 units, each of which is mainly composed of 3 self-attention mechanism units. The parameters of the 2D unit convolutional neural network are: in_channels=out_channels=1, kernel_size=3, stride=1, padding=1. The fully connected layer consists of two hidden layers and one output layer, and the number of neuron units is [2048, 1024, number of stations*1], and the activation function is the ReLU function. The fully connected network experienced by other factors consists of a hidden layer and an output layer, and the number of neuron units is [2048, number of stations*1]. The batchsize of the model is 64, and the optimizer of the model is Adam, with a learning rate of 0.0001.

[0073] 2. Baseline model.

[0074] In the experimental verification, the proposed MPGCN model is compared with the following benchmark models. All models are run on a desktop computer with an i7-8700K processor (12M cache, up to 4.7GHz), 32GB RAM, and NVIDIA GeForce GTX 3070 graphics card.

[0075] Long Short-Term Memory Network (LSTM): An LSTM approach with fully connected layers is used to model the inflow of traffic patterns.

[0076] Two-dimensional Convolutional Neural Network (CNN-2D): CNN-2D models are applied to model the inflow of traffic patterns, each of which has a CNN layer and fully connected layers. For all CNN-2D models, the kernel size is 3×3, the padding is 1, and the stride is 1.

[0077] ConvLSTM: This model is a hybrid model that combines convolution operations with LSTM and has powerful spatiotemporal modeling capabilities for time series data. This model is used to predict future passenger flows for each transportation mode.

[0078] ST-ResNet: This model uses 2D-CNN and residual connections to capture the spatiotemporal characteristics of passenger flow within the network and predict future passenger flow.

[0079] MIX-MGC: This model is a multi-graph convolution-based model with different branches that share knowledge. This model can collaboratively predict multiple traffic modes. Specifically, the model consists of two parts. The first part learns shared knowledge across tasks through regularization, and the second part learns shared knowledge through multi-linear relationships.

[0080] STGCN: A GCN-based deep learning model that models spatiotemporal features using spatial graph-convolutional layers and temporal-gated-causal-convolutional layers.

[0081] 4. Analysis of experimental results.

[0082] The experimental results are shown in Table 2. It can be seen that MPGCN has achieved accurate prediction results in rail transit short-term passenger flow prediction compared with all benchmark models, with MSE of 28.846, RMSE of 5.375, MAE of 3.025 and WMAPE of 9.8%. In addition, the MPGCN-Attention model outperforms the benchmark model in all evaluation indicators, showing its significant advantages in modeling complex spatiotemporal relationships and multi-period features. STGCN and MIX-MGC models follow closely, indicating the importance of multi-graph convolution and spatial feature modeling for passenger flow prediction. The errors of traditional models (LSTM, CNN 2D) are relatively large, reflecting their limitations in complex rail transit passenger flow prediction problems.

[0083] Table 2 Comparison of passenger flow prediction results

[0084] Model Name MSE RMSE MAE WMAPE LSTM 45.235 6.729 3.985 13.1％ CNN2D 41.874 6.472 3.812 12.6％ ConvLSTM 38.295 6.190 3.625 11.9％ ST-ResNet 35.201 5.933 3.504 11.5％ MIX-MGC 32.584 5.709 3.318 10.9％ STGCN 30.462 5.518 3.195 10.4％ MPGCN 28.846 5.375 3.025 9.8％

[0085] Furthermore, the effectiveness of the MPGCN structure is proved through ablation experiments, and some structures and frameworks of MPGCN are changed according to the principle of controlling variables. MSE, RMSE, MAE and WMAPE are used as evaluation indicators. The results are shown in Table 3.

[0086] Table 3 Ablation experiment results

[0087]

[0088]

[0089] NMP: No-Multi-Period (NMP): Do not use multi-period feature extraction (extracting time features from the historical passenger flow data matrix), and directly extract features from the historical passenger flow data sequence.

[0090] NMGCN: No-MGCN (NMGCN): does not use multi-graph convolution modules and only models spatial features based on physical graphs.

[0091] NEF: No-External Factors (NEF): Do not use external factors such as weather and air quality.

[0092] SG: Single-GCN (SG): Modeling spatial relationships based only on the physical graph.

[0093] MPO: Multi-Period Only (MPO): Temporal modeling is performed based only on multi-period features, without considering spatial features and external factors.

[0094] Advantages of the complete model (MPGCN):

[0095] Among all experimental settings, the MSE, RMSE, MAE, and WMAPE of the complete model are the lowest, indicating that the combination of multi-period features, multi-graph convolutions, convolutional attention layers, and external factor modeling significantly improves the prediction accuracy of the model.

[0096] Importance of multi-period features (No-Multi-Period): After removing the multi-period features, MSE increased by 3.636, RMSE and MAE increased by 0.334 and 0.259 respectively. This shows that multi-period features play a significant role in modeling the long-term dependency of time series.

[0097] Contribution of Multi-Graph Convolution (No-MGCN): After removing the multi-graph convolution module, the MSE increases by 5.069, indicating that the multi-graph convolution network can effectively extract multi-dimensional information in spatial relationships, while the performance of physical graph modeling alone (Single-GCN) is worse.

[0098] Influence of external factors (No-External Factors): When the external factor features are not used, the MSE increases by 2.030, indicating that the dynamic modeling of external factors such as weather and air quality has a certain contribution to the prediction results.

[0099] The shortcomings of using only single-graph convolution (Single-GCN): The MSE of single-graph convolution is 33.512, which is 5.666 higher than the complete model, indicating that modeling spatial relationships based only on physical graphs will lose important non-Euclidean relationships (such as site similarity and association).

[0100] Limitations of using only multi-period features: The MSE of using only multi-period features is 32.247, which is quite different from the complete model, indicating that the modeling of spatial features and external factors has a significant effect on improving the overall performance.

[0101] According to the above analysis of the ablation experiment, it can be known that the ablation experiment shows that multi-period feature extraction, multi-graph convolutional network and convolutional attention layer are crucial to improving the model performance. The synergistic effect of each module significantly improves the accuracy of short-term passenger flow prediction of rail transit. In addition, although the contribution of external factor modeling to the prediction results is not as significant as multi-period features and multi-graph convolution, it still plays an optimization role. Finally, through comparative experiments, it can be found that the performance of the model using only a single module (such as single-graph convolution or multi-period features) is much lower than that of the complete model, which shows that the comprehensive modeling of multi-dimensional features is the key to improving prediction performance.

[0102] To sum up, through the above technical solutions, the present invention effectively solves the shortcomings of the prior art in time feature extraction, spatial relationship modeling, external factor dynamic modeling, etc., and provides a rail transit system with an efficient and accurate short-term passenger flow prediction method, which can be widely used in rail transit operation optimization and urban traffic management.

[0103] In the embodiments of the present invention, a rail transit passenger flow prediction device is also provided, which is used to implement the above embodiments and preferred implementation modes, and will not be repeated hereafter. The terms "module", "unit", "subunit", etc. used below may be a combination of software and / or hardware that implements predetermined functions. Although the devices described in the following embodiments are preferably implemented in software, the implementation of hardware, or a combination of software and hardware, is also possible and conceivable.

[0104] The rail transit passenger flow prediction device includes:

[0105] A sequence construction module, for each rail transit station in the rail transit system, constructs multiple historical passenger flow data sequences based on the number of people entering the station in each time interval in the real-time window, and the t-th historical passenger flow data sequence includes the number of people entering the station before the t-th time interval;

[0106] A matrix construction module is used to divide each historical passenger flow data sequence of each rail transit station according to a preset time period to obtain multiple historical passenger flow data subsequences, and construct the multiple historical passenger flow data subsequences into a two-dimensional matrix to obtain a historical passenger flow data matrix;

[0107] The feature extraction module is used to process each historical passenger flow data matrix of each rail transit station through a convolution residual block to obtain each time feature of each rail transit station;

[0108] The passenger flow prediction module is used to obtain the passenger flow prediction value of the rail transit system based on the various time characteristics of each rail transit station through a multi-graph convolutional neural network.

[0109] It should be noted that the above modules can be functional modules or program modules, and can be implemented by software or hardware. For modules implemented by hardware, the above modules can be located in the same processor; or the above modules can be located in different processors in any combination.

[0110] It should be understood that the specific embodiments described herein are only used to explain the application, rather than to limit it. Based on the embodiments provided in this application, all other embodiments obtained by ordinary technicians in this field without creative work are within the protection scope of this application.

[0111] Obviously, the drawings are only some examples or embodiments of the present application. For ordinary technicians in the field, the present application can also be applied to other similar situations based on these drawings without creative work. In addition, it is understandable that although the work done in this development process may be complicated and lengthy, for ordinary technicians in the field, certain changes in design, manufacturing or production based on the technical content disclosed in this application are only conventional technical means and should not be regarded as insufficient content disclosed in this application.

Claims

1. A rail transit passenger flow prediction method, characterized in that: include: For each rail transit station in the rail transit system, multiple historical passenger flow data sequences are constructed based on the number of people entering the station in each time interval in the real-time window, and the t-th historical passenger flow data sequence includes the number of people entering the station before the t-th time interval; For each historical passenger flow data sequence of each rail transit station, segment it according to a preset time period to obtain multiple historical passenger flow data subsequences, and construct the multiple historical passenger flow data subsequences into a two-dimensional matrix to obtain a historical passenger flow data matrix; The historical passenger flow data matrix of each rail transit station is processed through the convolution residual block to obtain the time characteristics of each rail transit station; The passenger flow prediction value of the rail transit system is obtained based on the time characteristics of each rail transit station through a multi-graph convolutional neural network.

2. The rail transit passenger flow prediction method according to claim 1, characterized in that: The passenger flow prediction values ​​of the rail transit system are obtained based on the time characteristics of each rail transit station through a multi-graph convolutional neural network, including: Based on the time characteristics of each rail transit station and the adjacency status, similarity and correlation between each rail transit station, the physical graph structure, similarity graph structure and correlation graph structure of the rail transit system are constructed; The physical graph structure, similarity graph structure and correlation graph structure of the rail transit system are processed through a multi-graph convolutional neural network to obtain the passenger flow prediction value of the rail transit system.

3. The rail transit passenger flow prediction method according to claim 2, characterized in that: The historical passenger flow data matrix is ​​constructed by constructing multiple historical passenger flow data subsequences into a two-dimensional matrix, including: A plurality of historical passenger flow data subsequences are respectively used as different rows in a two-dimensional matrix to obtain a historical passenger flow data matrix.

4. The rail transit passenger flow prediction method according to claim 3, characterized in that: The t-th historical passenger flow data matrix of the j-th rail transit station is: in, represents the number of people entering the j-th rail transit station in the x-th time interval, T represents the number of time intervals in a single preset time period, and D represents the number of preset time periods.

5. The rail transit passenger flow prediction method according to claim 2, characterized in that: The tth physical graph structure of the rail transit system is Among them, V t is the location set of each rail transit station in the real-time window, is the adjacency matrix between each rail transit station in the real-time window. Each element in the adjacency matrix represents the adjacency status between different pairs of rail transit stations. t is the t-th time feature matrix of the rail transit system in the real-time window, and each element in the t-th time feature matrix represents the t-th time feature of different rail transit stations.

6. The rail transit passenger flow prediction method according to claim 2, characterized in that: The t-th similarity graph structure of the rail transit system is Among them, V t is the location set of each rail transit station in the real-time window, is the t-th similarity matrix between each rail transit station in the real-time window. Each element in the similarity matrix represents the similarity between different pairs of rail transit stations. t is the t-th time feature matrix of the rail transit system in the real-time window, and each element in the t-th time feature matrix represents the t-th time feature of different rail transit stations.

7. The rail transit passenger flow prediction method according to claim 2, characterized in that: The correlation graph structure of the rail transit system is: Among them, V t is the location set of each rail transit station in the real-time window, is the tth correlation matrix between each rail transit station in the real-time window. Each element in the correlation matrix represents the correlation between different pairs of rail transit stations. t is the t-th time feature matrix of the rail transit system in the real-time window, and each element in the t-th time feature matrix represents the t-th time feature of different rail transit stations.

8. The rail transit passenger flow prediction method according to claim 6, characterized in that: The tth similarity matrix The build steps include: Construct the t-th similarity score matrix between each rail transit station in the real-time window The diagonal elements in St are 0 and the element S(i,j) in the i-th row and j-th column of St is: in, and are the t-th time features of the i-th and j-th rail transit stations, respectively. DTW is the dynamic time warping algorithm; Set the elements in the t-th similarity score matrix St that are less than the first threshold to zero, and linearly normalize the elements in different rows that are equal to and greater than the first threshold, respectively, to obtain the t-th similarity matrix 9. The rail transit passenger flow prediction method according to claim 7, characterized in that: The tth correlation matrix The build steps include: Construct the tth correlation score matrix between each rail transit station in the real-time window The element C(i,j) in row i and column j of Ct is: Where D(i,j) is the total number of passengers from the i-th rail transit station to the j-th rail transit station before the t-th time interval in the real-time window, and N is the total number of rail transit stations; Set the elements in the t-th similarity score matrix Ct that are less than the second threshold to zero, and linearly normalize the elements in different rows that are equal to and greater than the second threshold, respectively, to obtain the t-th correlation matrix 10. A rail transit passenger flow prediction device, characterized in that: include: A sequence construction module, for each rail transit station in the rail transit system, constructs multiple historical passenger flow data sequences based on the number of people entering the station in each time interval in the real-time window, and the t-th historical passenger flow data sequence includes the number of people entering the station before the t-th time interval; A matrix construction module is used to divide each historical passenger flow data sequence of each rail transit station according to a preset time period to obtain multiple historical passenger flow data subsequences, and construct the multiple historical passenger flow data subsequences into a two-dimensional matrix to obtain a historical passenger flow data matrix; The feature extraction module is used to process each historical passenger flow data matrix of each rail transit station through a convolution residual block to obtain each time feature of each rail transit station; The passenger flow prediction module is used to obtain the passenger flow prediction value of the rail transit system based on the various time characteristics of each rail transit station through a multi-graph convolutional neural network.

Images