Subway short-time pull-in passenger flow prediction method based on double-branch cross attention

Abstract

The invention discloses a subway short-time pull-in passenger flow prediction method based on double-branch cross attention, and the method comprises the steps: preprocessing the card swiping data of an AFC, setting a time granularity, extracting a pull-in passenger flow sequence and a pull-out passenger flow sequence based on the granularity, and dividing the obtained sequences into a training set, a verification set and a test set; analyzing in-and-out passenger flow characteristics on the training set, determining a historical time step interval in which out-of-station passenger flow is influenced and an input in-and-out passenger flow sequence length, constructing training, verification and test samples, and determining an input dimension of the in-station passenger flow and an input dimension of the out-of-station passenger flow; constructing a neural network model for subway pull-in passenger flow prediction, and training to obtain an optimal model; and inputting a test sample into the optimal neural network model to obtain a prediction result. According to the method, the problem that spatial relation modeling is performed by excessively depending on a predefined station association diagram in the prior art is effectively solved, the problem that the in-and-out station passenger flow interaction relation between the same station and different stations cannot be fully modeled is effectively solved, and the prediction accuracy is remarkably improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent transportation, and relates to a subway short-time inbound passenger flow prediction method based on a double-branch cross attention. BACKGROUND

[0002] With the acceleration of urbanization, urban subways, as a public transportation mode with sufficient capacity and high energy efficiency, are rapidly developing worldwide. The subway network is expanding in scale and operating frequency is continuously improving, playing an important role in relieving road congestion and improving urban operation efficiency. At the same time, the dynamic characteristics of subway passenger flow are increasingly complex, especially in the morning and evening peak, holidays and emergency scenarios. The passenger flow fluctuation not only has obvious periodicity, but also shows high uncertainty. Accurate prediction of subway passenger flow, especially short-time inbound passenger flow, is not only crucial for passenger transport organization and facility configuration, but also provides important support for the operation management department to develop scheduling plans, optimize operation capacity allocation and ensure passenger travel safety.

[0003] Existing subway passenger flow prediction methods have experienced the development and evolution from traditional mathematical statistical models to machine learning methods, and then to deep learning models in recent years. Early researches mostly adopt mathematical statistical models, such as autoregressive moving average model (ARIMA), Kalman filter, etc. Such models have strong interpretability and perform relatively stably under small-scale and relatively stable data conditions. However, subway passenger flow usually has significant nonlinear characteristics and multi-scale periodicity, and traditional statistical models are difficult to fully describe its variation law, so the prediction performance is greatly limited in actual application. With the development of machine learning methods, support vector machine (SVM), support vector regression (SVR), random forest (RF), extreme gradient boosting decision tree (XGBoost), K nearest neighbor (KNN), etc. are gradually applied to subway passenger flow prediction. These methods construct nonlinear mapping in high-dimensional space to a certain extent, break through the dependence of traditional methods on linear hypothesis, and improve prediction accuracy with flexible feature combination. However, they generally rely on manual feature engineering and are insufficient in describing complex spatio-temporal dependence relationship, making it difficult to cope with the high dynamicity and uncertainty of subway passenger flow.

[0004] In recent years, deep learning methods have gradually become an important research direction for metro passenger flow prediction due to their strong feature extraction and representation learning capabilities. Among them, recurrent neural networks (RNN), its variants long short-term memory networks (LSTM), and gated recurrent units (GRU) can effectively model long-term temporal dependencies. Convolutional neural networks (CNN) rely on local receptive fields and parameter sharing mechanisms, making them more advantageous in capturing short-term patterns. However, these time series models mainly focus on the temporal variation of station passenger flow and cannot effectively capture the spatial dependence between station passenger flows. With the rise of graph neural networks (GNN), graph convolutional networks (GCN), gated graph convolutional networks (GGCN), and graph attention networks (GAT) have been widely applied to metro passenger flow prediction to capture complex topological relationships and passenger flow interaction patterns between stations.

[0005] On this basis, spatio-temporal fusion models that can simultaneously capture time evolution rules and spatial dependence features have gradually become the mainstream research direction. For example, by combining graph convolutional networks (GCN) with recurrent neural networks (such as GRU and LSTM), we can achieve joint modeling of spatial dependence and temporal dependence. The combination of convolutional neural networks (CNN) and residual networks (ResNet) enhances the ability to express complex spatio-temporal patterns. On this basis, more complex hybrid architectures have emerged, such as the introduction of three-dimensional convolution (3D CNN) to capture dynamic spatial features or the construction of enhanced GCN through physical and virtual adjacency matrices, combined with an encoder-decoder structure to improve prediction accuracy. With the evolution of models, researchers have further explored the joint application of graph convolutional networks (CNN), gated graph convolutional networks (GGCN), and graph attention networks (GAT) with new time series models TimesNet, achieving higher accuracy in short-term prediction tasks. At the same time, attention mechanisms have gradually become a key modeling tool. With the advent of the Transformer architecture, attention mechanisms have been widely introduced into passenger flow prediction models and combined with GCN, GRU, or CNN to enhance spatio-temporal dependence modeling and improve prediction accuracy.

[0006] Although some progress has been made in existing research, there are still several problems to be solved. First, when modeling the relationship between inter-station passenger flow, most studies still rely on pre-defined station association graph to extract spatial features. However, a single graph structure is difficult to fully capture the complex dependence between stations. Although the introduction of multi-graph structure enriches the representation of spatial relationship to some extent, the increased model complexity is usually not proportional to the improvement of prediction accuracy. In addition, whether it is a single graph or a multi-graph, it essentially assumes a fixed station association pattern and strength, which is difficult to reflect the dynamic changes of passenger flow relationship. In contrast, dynamic graph structure has higher flexibility and can better adapt to time-varying characteristics, but its construction quality is easily affected by the completeness and noise level of passenger flow data. When there is missing or sparsity in the data, the dynamic graph may weaken the model stability and generalization ability.

[0007] Second, the internal relationship between inbound and outbound passenger flow at the same station and the interaction relationship between inbound and outbound passenger flow at different stations have not been fully explored. For the inbound and outbound passenger flow at the same station, existing methods mostly treat it as a double-channel input, simply concatenating and maintaining the time sequence correspondence at the feature level, without explicitly depicting the influence of historical outbound passenger flow on future inbound passenger flow. For example, in a typical commuting scenario, passengers often leave at the work site during the morning rush hour and arrive at the same station during the evening rush hour. For another example, if a passenger leaves at the residential area in the evening of the previous day, he or she will usually arrive at the same station in the morning of the next day. These rules show that there is a significant time-dependent relationship between inbound and outbound passenger flow at the same station, and if this relationship is ignored, the prediction accuracy will be limited. For the interaction relationship between inbound and outbound passenger flow at different stations, existing methods mainly rely on time-step-based OD (Origin-Destination) data. However, OD data can only reflect the corresponding relationship between the origin and destination of a passenger's trip, and cannot describe more complex cross-station interaction patterns. For example, in a business center area, if two stations are close to each other, passengers may leave at one station and arrive at another station after some activities, and such cross-station flow relationship cannot be captured by OD data. In addition, OD data itself has inherent limitations. OD data is generated based on single complete trip records, and can only reflect short-term origin-destination spatial dependence, but cannot reflect long-term spatio-temporal evolution characteristics of inbound and outbound passenger flow. SUMMARY

[0008] The present invention aims to provide a subway short-term inbound passenger flow prediction method based on double-branch cross-attention to solve the problem that existing technology relies too much on pre-defined station association graph for spatial relationship modeling, and the problem that it fails to fully model the interaction relationship between inbound and outbound passenger flow at the same station and at different stations.

[0009] The technical solution adopted by the present invention is a subway short-term inbound passenger flow prediction method based on double-branch cross-attention, which is performed according to the following steps:

[0010] S1, preprocessing the card swiping data of the subway automatic fare collection system, setting a time granularity and extracting the inbound and outbound passenger flow sequences based on the granularity, dividing the obtained sequences into a training set, a validation set and a test set;

[0011] S2, analyzing the inbound and outbound passenger flow characteristics on the training set, determining the historical time step interval that affects the outbound passenger flow, inputting the inbound and outbound passenger flow sequence length, and constructing training samples, validation samples and test samples based thereon, and determining the input dimension of the inbound passenger flow and the input dimension of the outbound passenger flow;

[0012] S3, constructing a neural network model for subway inbound passenger flow prediction based on a double-branch time series network and an interaction attention mechanism, and obtaining the best neural network model through training;

[0013] S4, inputting the test sample into the best neural network model to obtain the inbound passenger flow prediction result for a future time step.

[0014] Further, the specific steps of S1 are as follows:

[0015] S1.1: Obtain the original card swiping data of the target subway network, clean the original data, aggregate the cleaned data according to a predetermined time granularity, and generate inbound passenger flow time series of all stations outbound passenger flow time series , as shown in equations (1) and (2):

[0016] (1)

[0017] wherein: is the number of stations, is the total number of time steps contained in the data set after aggregation according to the predetermined time granularity, is the inbound passenger flow time series of all stations, is a real number set, is a real number set with a dimension of is the inbound passenger flow of all stations at time step ;

[0018] (2)

[0019] wherein: is the outbound passenger flow time series of all stations, is the outbound passenger flow vector of all stations at time step ;

[0020] S1.2: Divide the inbound and outbound passenger flow time series of all stations into a training set , a validation set​ and test set .

[0021] Furthermore, the specific steps of S2 are as follows:

[0022] S2.1: Calculate the total passenger flow of all stations within the training set time step range, and plot the curve of its change over time to identify the daily cycle pattern of passenger flow. Plot the curve of passenger flow of representative stations within the training set time step range to identify the time lag pattern between passenger flow of each station.

[0023] S2.2: Recording Stations The inbound sequence on the training set is Site The outbound sequence is , will the site Remove the pre-lag step size from the inbound sequence Data, obtain site Cut-off passenger flow sequence , will the site Remove the trailing lag step from the outbound sequence Data, obtain site The outbound passenger flow sequence after truncation ,calculate and Pearson correlation coefficient ,based on To determine the correlation between historical outbound passenger flow and future inbound passenger flow, Pearson correlation coefficients are calculated sequentially at multiple lag steps. The correlation between inbound and outbound passenger flow at each station under different lag steps is analyzed. Based on the periodicity of the correlation curve, the historical time interval in which outbound passenger flow has an impact is determined. ;

[0024] S2.3: Input the length of the passenger flow sequence entering the station. Set as the first preset value, and input the length of the outbound passenger flow sequence. Set to the second preset value to satisfy , This is the termination value. Given the historical time step interval length, and based on the set sequence length, construct training, validation, and test samples, and determine the data batch size. Set as the third preset value, and determine the input dimensions of the inbound passenger flow based on the set batch size and sequence length. , For the data batch size, To input the length of the passenger flow sequence entering the station, The number of stations and the input dimension of outbound passenger flow. , is inputted.

[0025] Further, the model in S3 includes a feature extraction layer, a time sequence representation layer, a cross-branch interaction layer, and a fusion and prediction layer.

[0026] Further, the feature extraction layer, the time sequence representation layer, the cross-branch interaction layer, and the fusion and prediction layer in the model in S3 have the following specific structures and data processing methods:

[0027] The feature extraction layer includes a main branch and a sub-branch arranged in parallel, and both the main branch and the sub-branch use a time sequence module of TimesNet structure for feature extraction.

[0028] The time sequence representation layer includes a query generation module connected with the main branch, which is used to perform tensor quantization scaling and dimension expansion on the features output by the main branch to generate a Query vector with a dimension of , where is the embedding space dimension.

[0029] The time sequence representation layer also includes a key-value pair generation module connected with the sub-branch, which is used to generate a Key vector and a Value vector with dimensions of respectively through linear projection, where is the length of the historical time step interval .

[0030] The Query, Key, and Value vectors are used to perform cross-branch attention calculation between the main branch and the sub-branch.

[0031] The cross-branch interaction layer includes a multi-head attention module, which is used to receive the Query vector from the main branch, and the Key vector and the Value vector from the sub-branch, calculate the correlation between the Query vector and the Key vector to generate attention weights, and then use the attention weights to perform weighted summation on the Value vector.

[0032] The fusion and prediction layer is used to receive the original feature representation of the main branch and the cross-attention feature output by the cross-branch interaction layer , perform weighted summation to generate a fused feature representation , and then output the inbound passenger flow prediction results of each station through linear projection.

[0033] Further, the time sequence module in the feature extraction layer in S3 processes data according to the following steps: ​

[0034] S3.1: Z-score standardization is performed on the input inbound and outbound passenger flow samples along the time dimension, and the processed data is mapped to a preset dimension embedding space through a data embedding operation;

[0035] S3.2: A linear projection is used to generate a predicted time step representation, and the input sequence length is extended;

[0036] S3.3: The extended sequence is input into a multi-layer TimesBlock module for time series modeling;

[0037] S3.4: Layer normalization is performed after each TimesBlock module, so that the main branch and the auxiliary branch obtain representation sequences under the extended time series dimension;

[0038] S3.5: The main branch outputs a representation corresponding to the predicted target time step , and the auxiliary branch outputs a representation corresponding to the historical time step interval .

[0039] Further, the method for obtaining the best neural network model by training in S3 is as follows:

[0040] The training samples are input into the model in batches according to a preset batch size, and the model parameters are continuously optimized by the back propagation algorithm according to the set model structure and training hyperparameters, so that the loss function is continuously reduced. When the early stopping condition is met, the training is terminated, and the best neural network model is obtained.

[0041] Further, the specific steps of S4 are as follows: The test set samples are input into the trained neural network model in batches according to a preset batch size, and the inbound passenger flow prediction results of each station at a future time step are output. The prediction results are de-standardized to restore them to the actual passenger flow scale, and the model prediction accuracy and generalization performance are evaluated based on the real data of the test set.

[0042] The beneficial effects of the embodiments of the application are as follows:

[0043] 1) The DBCA-TN (Dual-Branch Cross-Attention Temporal Network, DBCA-TN) proposed in the present application adopts a main / auxiliary dual-branch structure. The main branch is used to extract inbound time series features of the target prediction time step, and the auxiliary branch is used to extract outbound time series features in the prior historical time step interval. The cross-branch multi-head attention is used to simultaneously depict the global spatio-temporal interaction relationship of inbound and outbound passenger flow across stations in the subway network, as well as the influence of historical outbound passenger flow in a single station on future inbound passenger flow. This design can explicitly strengthen the representation of the influence of historical outbound passenger flow on inbound passenger flow, thereby significantly improving the prediction accuracy.

[0044] 2) The DBCA-TN proposed in the present application automatically learns the potential dependency relationship of inter-station passenger flow in the embedding space through multi-head attention, without the need to pre-construct a graph structure and additional spatial modules, thereby effectively controlling the model parameter size while ensuring prediction accuracy, and exhibiting obvious advantages in terms of computational efficiency and robustness.

[0045] 3) The DBCA-TN proposed in the present application does not rely on track adjacency graphs, OD data, external time labels or weather information, and only uses the inbound and outbound passenger flow sequence as input, thereby achieving high-precision short-term inbound passenger flow prediction. Even when some station data is missing, it still maintains strong adaptability and prediction ability.

[0046] 4) The DBCA-TN proposed in the present application uses a learnable fusion coefficient α to dynamically adjust the information injection intensity of the main and auxiliary branches, thereby suppressing noise interference to some extent and enhancing the robustness of the model.

[0047] 5) The experimental results on the public data set show that the average prediction error of the DBCA-TN proposed in the present application is significantly lower than that of existing mainstream models, and it is suitable for short-term passenger flow prediction tasks in large-scale urban rail transit. BRIEF DESCRIPTION OF DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0049] Figure 1 The DBCA-TN network framework proposed in the present application;

[0050] Figure 2 The total inbound and outbound passenger flow graph of Hangzhou Metro from January 1, 2019 to 25:30 on January 25, 2019;

[0051] Figure 3 The total outbound passenger flow graph of station 0;

[0052] Figure 4 The total outbound passenger flow graph of station 7;

[0053] Figure 5 The total outbound passenger flow graph of station 9;

[0054] Figure 6 The total outbound passenger flow graph of station 15;

[0055] Figure 7Pearson correlation coefficient curves of in-out station passenger flow sequences of stations 0, 7, 9 and 15; wherein (a) is station 0, (b) is station 7, (c) is station 9, and (d) is station 15;

[0056] Figure 8 It is a flow chart of a subway in-station passenger flow prediction method based on a double-branch cross attention. DETAILED DESCRIPTION

[0057] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the present application.

[0058] The improvement of the present application lies in that the present application introduces a double-branch time sequence network and an in-out station cross attention mechanism, fully excavates the potential correlation between in-station passenger flow and out-station passenger flow, realizes explicit injection of out-station passenger flow, thereby enhancing the interpretability of the model, improving the in-station passenger flow prediction accuracy, taking into account the training efficiency, reducing the dependence on data integrity and external information, and significantly improving the generalization ability and practical application value of the model.

[0059] As shown in Figure 8 , the present application proposes a subway short-time in-station passenger flow prediction method based on a double-branch cross attention, which specifically performs the following steps:

[0060] S1, preprocessing the card swiping data of the subway automatic fare collection system (AFC), setting a time granularity and extracting in-station and out-station passenger flow sequences based on the granularity, and dividing the obtained sequences into a training set, a validation set and a test set;

[0061] The specific steps of S1 are as follows:

[0062] S1.1: Obtain the original card swiping data of the target subway network, clean the original data, aggregate the cleaned data according to the predetermined time granularity, and generate all-station in-station passenger flow time sequences , out-station passenger flow time sequences , as shown in formulas (1) and (2):

[0063] (1)

[0064] Wherein: is the number of stations, is the total number of time steps contained in the data set after aggregation according to the predetermined time granularity, is the all-station in-station passenger flow time sequence, For the set of real numbers, For dimension The set of real numbers, In time step Passenger flow at all stations at that time.

[0065] (2)

[0066] in: For the time series of outbound passenger flow at all stations, In time step At that time, the outbound passenger flow vector of all stations;

[0067] S1.2: Divide the time series of passenger flow entering and exiting all stations into a training set according to time sequence. Validation set and test set .

[0068] S2. Analyze the characteristics of passenger flow in and out of the training set, determine the historical time interval of the impact of passenger flow out of the station and the length of the input passenger flow sequence, and construct training samples, validation samples and test samples accordingly to determine the input dimension of passenger flow in and passenger flow out of the station.

[0069] The specific steps for S2 are as follows:

[0070] S2.1: Calculate the total passenger flow of all stations within the training set time step range, and plot the curve of its change over time to identify the daily cycle pattern of passenger flow. Plot the curve of passenger flow of representative stations within the training set time step range to identify the time lag pattern between passenger flow of each station.

[0071] S2.2: Recording Stations The inbound sequence on the training set is Site The outbound sequence is For a given lag step size , will the site Remove the pre-lag step size from the inbound sequence Data, obtain site Cut-off passenger flow sequence , will the site Remove the trailing lag step from the outbound sequence Data, obtain site The outbound passenger flow sequence after truncation ,calculate and Pearson correlation coefficient .

[0072] In the detailed description of the application: the station The inbound sequence on the training set is , the outbound sequence of the station is For a given lag step , the inbound sequence of the station is removed from the front data, and the truncated inbound passenger flow sequence of the station is obtained , as shown in equation (3):

[0073] (3)

[0074] Wherein: is the inbound passenger flow of the station at time step , T is the total number of time steps contained in the training set, is the lag step, is the length of the truncated sequence. The outbound sequence of the station

[0075] is removed from the end data, and the truncated outbound passenger flow sequence of the station is obtained , as shown in equation (4):

[0076] (4) Wherein:

[0077] is the outbound passenger flow of the station at time step 0.

[0078] and correspond to the original sequence in time, separated by time steps, and both have a length of , and the Pearson correlation coefficient between and is calculated , as shown in equation (5)

[0079] (5)

[0080] Wherein: is the Pearson correlation coefficient between and , Cov represents the covariance between and , and represents standard deviation express standard deviation express The mean, express The mean, For the site At time step The number of passengers entering the station, For the site At time step The number of passengers exiting the station The total number of time steps contained in the training set;

[0081] based on The value is used to determine the correlation between historical outbound passenger flow and future inbound passenger flow. The lag step starts from 0 and is used to calculate the Pearson correlation coefficient at multiple lag steps. The correlation between inbound and outbound passenger flow at each station under different lag steps is analyzed. Based on the periodicity of the correlation curve, the historical time interval in which outbound passenger flow has an impact is determined. The initial value of the lag step is set to 0, which aims to fully utilize the latest exit passenger flow information when predicting inbound passenger flow, while the termination value is... The magnitude of the value determines the validity of the exit passenger flow information within that time interval.

[0082] S2.3: Input the length of the passenger flow sequence entering the station. Set as the first preset value, and input the length of the outbound passenger flow sequence. To be set as the second preset value, it must meet the following requirements. , This is the termination value. The length of the historical time step interval is set to ensure that the input sequence length can cover the historical time step interval. Based on the set sequence length, training, validation, and test samples are constructed, and the data batch size is determined. Set as the third preset value, and determine the input dimensions of the inbound passenger flow based on the set batch size and sequence length. , For the data batch size, To input the length of the passenger flow sequence entering the station, The number of stations and the input dimension of outbound passenger flow. , The input is for long passenger flow sequences at the station.

[0083] S3. Construct a neural network model for predicting subway passenger flow based on a bi-branch temporal network and interactive attention mechanism (e.g., Figure 1 As shown, the optimal neural network model is obtained through training;

[0084] The model in S3 comprises a feature extraction layer, a time sequence representation layer, a cross-branch interaction layer, and a fusion and prediction layer.

[0085] The feature extraction layer, the time sequence representation layer, the cross-branch interaction layer, and the fusion and prediction layer in the model in S3 have the following specific structures and data processing methods:

[0086] The feature extraction layer comprises a main branch and a secondary branch arranged in parallel, and both the main branch and the secondary branch use a time sequence module of a TimesNet structure for feature extraction.

[0087] The time sequence representation layer comprises a query generation module connected with the main branch, which is configured to perform tensor quantization scaling and dimension expansion on the features output by the main branch to generate a Query vector with a dimension of wherein d is an embedding space dimension. The time sequence representation layer further comprises a key-value pair generation module connected with the secondary branch, which is configured to perform linear projection on the features output by the secondary branch to generate a Key vector with a dimension of

[0088] and a Value vector with a dimension of wherein L is the length of a historical time step interval. The Query vector, the Key vector, and the Value vector are used to perform cross-branch attention calculation between the main branch and the secondary branch. The cross-branch interaction layer comprises a multi-head attention module, which is configured to receive the Query vector from the main branch, the Key vector and the Value vector from the secondary branch, calculate the correlation between the Query vector and the Key vector to generate attention weights, and then perform weighted summation on the Value vector using the attention weights.

[0089] The multi-head attention module can calculate the correlation in different subspaces in parallel, thereby highlighting the outgoing features that contribute most to the prediction task while suppressing redundant information. Each attention head

[0090] comprises an independent first parameter matrix Wq , a second parameter matrix Wk

[0091] , and a third parameter matrix Wv , which are used to map the original d-dimensional embedding to a d-dimensional subspace. ​​​​​​​​​The calculated correlation matrix focuses on different types of spatio-temporal dependencies, formula (6) represents the calculation of a single attention head, formula (7) represents the scaled dot-product attention, and formula (8) represents the attention output.

[0092] (6)

[0093] wherein: is a query matrix, is a key matrix, is a value matrix, is a first parameter matrix, is a second parameter matrix, is a third parameter matrix.

[0094] (7)

[0095] wherein: is the number of attention heads, is a scaling factor, and Softmax(·) represents an activation function.

[0096] The results of all attention heads are spliced and projected to a dimension of by a projection matrix to obtain the attention output , .

[0097] (8)

[0098] wherein: is a projection matrix.

[0099] The projection matrix maps the spliced results of multi-head attention to a unified representation space, facilitating subsequent feature fusion.

[0100] The fusion and prediction layer is used to receive the original feature representation of the main branch and the cross-attention feature output by the cross-branch interaction layer , perform weighted summation to generate the fused feature representation , and then output the station entry passenger flow prediction results through linear projection.

[0101] The time sequence module in the feature extraction layer in S3 processes data according to the following steps: the feature extraction layer includes a main branch and a secondary branch arranged in parallel, each branch is composed of multiple TimesBlock modules connected in series, and a time sequence module in the style of TimesNet is used to perform feature extraction.

[0102] The specific process is as follows: first, the input inbound and outbound passenger flow samples are subjected to Z-score standardization processing along the time dimension, and normalization operation is completed, and then the samples are mapped to an embedding space with a dimension of by data embedding. Then, additional prediction time step representations are generated by linear projection, so that the input sequence length is extended from the original to and to . The expanded sequence is input into a multi-layer TimesBlock module for time series modeling.

[0103] In each TimesBlock module, the input sequence is first converted to the frequency domain by fast Fourier transform (FFT). The frequency components are sorted according to their spectral energy, and the top k components with the highest energy are selected to capture the main periodic patterns. Then, local time windows are divided according to these periodic information, and multi-scale time series features are extracted by Inception-style convolution modules. The output results corresponding to different periods are weighted and fused, and the fused results are added to the original input through residual connection to enhance information retention and improve the stability of gradient propagation. Layer normalization (LayerNorm) is used after each TimesBlock to maintain the stability of the feature distribution.

[0104] Finally, the main branch and the auxiliary branch obtain the representation sequences under the expanded time series dimension, i.e. and where the main branch generates the representation corresponding to the prediction target time step, and the auxiliary branch generates the representation corresponding to the historical time step interval.

[0105] The method for obtaining the optimal neural network model by training in S3 is as follows: the specific method for training the constructed model is to input the training set samples into the model in batches according to the preset batch size, continuously optimize the model parameters by the back propagation algorithm according to the set model structure and training hyperparameters, so that the loss function continuously decreases, and the validation set error changes are monitored in real time during the training process, and the training is terminated when the preset early stopping condition is met, and after the training is completed, the model parameter configuration with the optimal validation set performance is saved for prediction.

[0106] In the specific embodiments of the present application, the prediction results are measured by RMSE (Root Mean Square Error), MAE (Mean Absolute Error), and WMAPE (Weighted Mean Absolute Percentage Error). the error of the true value The RMSE, MAE and WMAPE calculation formulas are shown in formulas (9)-(11) respectively:

[0107] (9)

[0108] Wherein: N represents the number of samples.

[0109] (10)

[0110] (11)

[0111] S4, output the test sample to the best neural network model to obtain the inbound passenger flow prediction result at the next time step.

[0112] The specific steps of S4 are:

[0113] The test set sample is input into the trained neural network model in batches according to the preset batch size, and the inbound passenger flow prediction result of each station at the next time step is output. The prediction result is subjected to inverse standardization processing to restore the actual passenger flow scale, and the real passenger flow data of the test set is used to test the prediction accuracy and generalization performance of the model.

[0114] Embodiment 1

[0115] As shown in Figure 1 To evaluate and verify the performance of the subway inbound passenger flow prediction model DBCA-TN proposed in the application, the real AFC card data of Hangzhou City, China from January 1, 2019 to January 25, 2019, 5:30-23:30 is used as the experimental data set in Embodiment 1 of the application, covering 81 subway stations of lines 1, 2, 4 and 9, and the station number is 0-80. The original card data includes card time, device number, entry and exit state, user ID, line and station information.

[0116] Due to random errors in the collection process, the original data needs to be cleaned in Embodiment 1 of the application. The cleaning rules include: 1) remove the card records outside the operating time; 2) delete incomplete records (only inbound or outbound records); 3) remove logical error records (such as inbound time later than outbound time). After cleaning, a total of 29,159,711 valid data are obtained.

[0117] Embodiment 1 of the application aggregates the passenger flow data with a time granularity of 15 minutes to obtain the inbound and outbound passenger flow time series of 81 stations. The daily operation time of the subway is 18 hours, corresponding to 72 time steps, and 25 days correspond to 1800 time steps, numbered 0-1799, and the obtained inbound and outbound passenger flow sequence data has a dimension of where , . The inbound passenger flow time series is denoted as: , the outbound passenger flow time series is denoted as: , is the number of stations, is the total number of time steps contained in the dataset after aggregation according to the predetermined time granularity, and is the inbound passenger flow time series of all stations, is a real number set, is a real number set with dimension .

[0118] The dataset is divided into a training set (T: time steps 0-1295), a validation set (V: time steps 1296-1439), and a test set (Test: time steps 1440-1799), corresponding to January 1-18, 19-20, and 21-25, 2019, respectively. This division avoids information leakage during training and ensures the objectivity of model performance evaluation.

[0119] The inbound passenger flow sequence input length is and the outbound passenger flow sequence input length is to construct training, validation, and test samples on the training set, validation set, and test set, respectively. The inbound passenger flow sequence sample is as shown in equation (12):

[0120] (12)

[0121] where: is the inbound passenger flow sequence input length.

[0122] The inbound passenger flow sequence sample corresponding outbound passenger flow sequence sample is as shown in equation (13):

[0123] (13)

[0124] where: is the outbound passenger flow sequence input length.

[0125] When constructing samples, it is necessary to ensure that if , then ; if , then so as to use the latest outbound passenger flow information as input while avoiding leakage of future outbound passenger flow data.

[0126] The inbound passenger flow sequence sample​​​ corresponding true value As shown in formula (14):

[0127] (14)

[0128] wherein: is the in-station passenger flow value at time step.

[0129] The target equation is defined as shown in formula (15):

[0130] (15)

[0131] wherein: represents the DBCA-TN model constructed by the application, Θ is a model parameter, is the prediction value of the i-th sample, and the prediction result the difference between the true value reflects the prediction accuracy of the model.

[0132] To avoid the model from learning the in-out station passenger flow law of the validation set and the test set in advance in the training process, thereby causing the prediction accuracy to be high, the embodiment 1 of the application only analyzes the in-out station passenger flow characteristics of the training set, and the total in-out station passenger flow of the 81 stations in the training set at the 0-1295 time steps is counted, and the corresponding curve is drawn. Figure 2 ). Figure 2 It is shown that the passenger flow of the Hangzhou subway presents a significant daily cycle law: the early and late peaks appear on weekdays and weekends, but the peak value on weekends is obviously lower than that on weekdays; the passenger flow on the New Year's Day presents a single-peak mode. The in-station and out-station total passenger flow curves are highly coincident, indicating that the overall change law is consistent.

[0133] Further, four representative stations (0, 7, 9, 15) with different passenger flow scales are selected, and the in-out station passenger flow curves of the stations on the training set are drawn. Figure 3 Figure 6 ). Figure 3 It is shown that the in-station passenger flow peak of the station 0 is obviously higher than the out-station passenger flow, Figure 4 it is shown that the out-station passenger flow peak of the station 7 is obviously higher than the in-station passenger flow, Figure 5 it is shown that the in-out station passenger flow of the station 9 is relatively close, Figure 6 it is shown that the out-station passenger flow of the station 15 is relatively close. Combined with Figure 3 Figure 6 Although the passenger flow scales of different stations are different, the in-out station passenger flow curves of the same station are highly consistent in the overall trend, and there is an obvious time lag. This feature shows that there is a significant time-dependent relationship between the in-out station passenger flow of the same station. ​​​

[0134] For the missing problem of the inbound and outbound passenger flow of the station 54, the embodiment 1 of the application selects to reserve the station to maintain the connectivity of the adjacent graph and avoid destroying the overall network structure.

[0135] The single round trip of the subway passenger is usually completed within a day, but in order to more fully reveal the potential influence law of the outbound passenger flow on the inbound passenger flow, the value range of the lag step is set to cover the subway operation time length of two days, that is, the integer values are sequentially taken in the interval [0, 143]. The station The inbound passenger flow sequence of the station is truncated backward by values, and the truncated inbound passenger flow sequence of the station is as shown in formula (3). The outbound passenger flow sequence of the station is truncated forward by values, and the truncated outbound passenger flow sequence of the station is as shown in formula (4). The Pearson correlation coefficient between the inbound passenger flow sequence of the station and the outbound passenger flow sequence of the station

[0136] is calculated as shown in formula (5). Figure 7 Taking the stations 0, 7, 9 and 15 as examples, the correlation curves of the inbound passenger flow sequence of the station and the outbound passenger flow sequences of the four stations at 144 values are drawn. Figure 7 (a) is the change trend of the correlation between the inbound passenger flow of the station 0 and the outbound passenger flows of the stations 0, 7, 9 and 15 with values, Figure 7 (b) is the change trend of the correlation between the inbound passenger flow of the station 7 and the outbound passenger flows of the stations 0, 7, 9 and 15 with values, Figure 7 (c) is the change trend of the correlation between the inbound passenger flow of the station 7 and the outbound passenger flows of the stations 0, 7, 9 and 15 with values, Figure 7 (d) is the change trend of the correlation between the inbound passenger flow of the station 7 and the outbound passenger flows of the stations 0, 7, 9 and 15 with values. Figure 7 It is shown that the correlation curves of the inbound and outbound passenger flows of each station and four stations are different in shape, but all show obvious periodic characteristics, and the period is about 72 steps, which is basically consistent with the number of steps of the single day operation time of the subway. The periodicity shows that the inbound and outbound passenger flows of the same station follow the daily round trip rule, and the inbound and outbound passenger flows of different stations reflect the space-time dependence within the single day operation time. Based on this periodic characteristic, the historical time step interval is set to [0, 71] to cover the single day operation period of the subway, and the latest and effective historical outbound passenger flow information is provided for the inbound passenger flow prediction.

[0137] To fully capture the characteristics of the day cycle, the input inbound passenger flow sequence length of the main branch of the DBCA-TN model constructed in Embodiment 1 of the present application is set to According to the determined historical time step interval, the input outbound passenger flow sequence length of the auxiliary branch is to cover the entire historical time step interval and correspond to the length of the subway single-day operation, and finally the sample dimensions of the training set, the validation set and the test set are X_train_in: (1224, 81, 72), X_train_out: (1224, 81, 72), X_val_in: (72, 81, 72), X_val_out: (72, 81, 72), X_test_in: (288, 81, 72), X_test_out: (288, 81, 72), and the corresponding target tensor dimensions are Y_train: (1224, 81), Y_val: (72, 81), Y_test: (288, 81).

[0138] In Embodiment 1 of the present application, the DBCA-TN model constructed by the present application adopts a double-path architecture with a main branch and an auxiliary branch in parallel, and both branches are stacked with two layers of TimesBlock modules, each of which integrates an Inception convolution block. Each Inception convolution block in TimesBlock contains 3 convolution kernels, and the key cycle number of the model is set to 2. In the parameter configuration of the main branch, the embedding vector dimension of each time step in the main branch is 128, and the intermediate convolution channel dimension is 64. The embedding dimension of the auxiliary branch is also set to 128, but the intermediate convolution channel dimension is 128. To improve the generalization performance of the model, Dropout is integrated into both branches, and the dropout rate is uniformly set to 0.1.

[0139] After the time feature extraction of the outbound passenger flow sequence, the time dimension changes from to To obtain the sequence features corresponding to the historical time step interval [0, 71], the forward time step clipping parameter of the auxiliary branch is set to the forward time step clipping number The cross-branch attention module adopts a four-head attention mechanism, and the main branch output and the attention output are weighted by a learnable fusion weight The initial value of the fusion weight is set to 0.6. To further enhance feature interaction, the main branch output feature is multiplied by a scaling coefficient , and the initial value of the scaling coefficient is set to 1.11.

[0140] During model training, the optimizer used was AdamW, and the weight decay coefficient was set to 1e-05.

[0141] In the model training process of this invention, a hierarchical learning rate strategy is adopted to achieve differentiated training for different modules. The initial learning rate of the feature extraction layer is lr=0.001, and the initial learning rate of the cross-branch interaction layer is increased to 2.7 times the initial learning rate of the feature extraction layer. The learnable parameters... and The initial learning rate is increased to 3.1 times the base initial learning rate of the feature extraction layer, thereby accelerating the convergence speed. Mean squared error (MSE) is used as the loss function during training, as shown in Equation 16.

[0142] (16)

[0143] In this invention, both the model parameters and training configuration parameters are determined based on a combination of theoretical analysis and experimental verification. These parameters are not arbitrarily set based on experience, but rather are reasonable values ​​gradually determined through theoretical analysis and experimental optimization, possessing clear design basis and stable performance.

[0144] To verify the effectiveness of the DBCA-TN model of this invention, nine representative benchmark models were selected for comparison. The benchmark models and their main parameter settings are as follows:

[0145] SVR (Support Vector Regression): As an early regression model applied to traffic prediction, it uses a radial basis function (RBF) kernel function. Its main parameters are: ε = 0.005 and regularization coefficient C = 3. The length of the passenger flow input sequence is 4. It should be noted that the parameter size of SVR depends on the number of support vectors generated during training, rather than the fixed parameters of the neural network structure. The final parameter size can only be determined after training is complete.

[0146] Two-layer MLP: Consists of two fully connected network layers, with ReLU activation functions inserted between each layer, and LayerNorm normalization applied to the output of each layer. Key parameters are: 128 feature dimensions for the first layer, 64 for the second layer, Dropout ratio of 0.2, length of the inbound passenger flow input sequence of 36, and batch size of 32.

[0147] ST-GCN: A prediction model based on spatiotemporal graph convolution, which utilizes convolution to achieve efficient modeling while considering spatial correlations. The model parameter configuration remains consistent with the original paper.

[0148] ResLSTM: a hybrid architecture combining ResNet, GCN and Attention-LSTM, where ResNet captures deep spatial features, GCN extracts topological information, and Attention-LSTM models temporal dependencies. The original design includes four branches: inflow, outflow, topology, and weather conditions and air quality. In this experiment, only weather data is used in the fourth branch, and other parameters remain the same as the original.

[0149] Conv-GCN: combines multi-graph GCN and three-dimensional convolutional network (3D CNN) to capture spatial and temporal dependencies of multiple traffic patterns (real-time, daily, and weekly). The model parameters remain the same as the original.

[0150] PVCGN (Physical-Virtual Combined Graph Network): constructs a physical and virtual combined adjacency matrix to enhance graph convolution, improves the gating recurrent unit and stacks it to form an encoder-decoder structure for prediction. Unlike the original, this implementation only predicts inbound passenger flow, and the parameter settings remain the same as the original.

[0151] TimesNet: explicitly models multi-scale periodicity through Fourier spectrum decomposition and applies multi-scale convolution (TimesBlock) on a two-dimensional expanded matrix to capture local patterns. Main parameters: Inception kernel number 3, TimesBlock stack layer number 2, key period number 2, feature dimension 128, convolution dimension 64, Dropout ratio 0.1, inbound passenger flow sequence input length 72, batch size 32.

[0152] Dual-channel TimesNet: parameter settings are consistent with TimesNet, but both inbound and outbound sequences are input, and only inbound passenger flow prediction values are output.

[0153] GCN+Dual-channel TimesNet: introduces a double-layer GCN to extract spatial features based on Dual-channel TimesNet, uses ReLU activation between the two layers of GCN, and applies LayerNorm and Dropout (ratio 0.1) after each layer. Input length and batch size remain the same as TimesNet.

[0154] For all deep learning-based benchmark models, the early stopping strategy (patience=10) is used during training to prevent overfitting.

[0155] The experimental results are shown in Table 1:

[0156] Table 1: Prediction accuracy and training efficiency of benchmark models and DBCA-TN on Hangzhou subway data set

[0157] Although ST-GCN, ResLSTM and Conv-GCN achieve high prediction accuracy on many real large-scale metro datasets, as shown in Table 1, the prediction performance of the four deep learning models, double-layer MLP, ST-GCN, ResLSTM and Conv-GCN, is lower than that of the traditional machine learning model SVR on the Hangzhou metro dataset. The main reason for the poor performance of double-layer MLP is its limited time series modeling capability; while the failure of ST-GCN, ResLSTM and Conv-GCN is due to their high dependence on the track adjacency matrix. Although the track adjacency matrix of this dataset is complete, the station 54 is missing the in-out station passenger flow data, resulting in its node features being zeroed, introducing a lot of noise when GCN extracts spatial features, and weakening the generalization ability of the model. This result shows that the combination model based on spatio-temporal graph convolution still has the problem of insufficient robustness in the scenario of missing part of the node. The PVCGN model is slightly better than SVR in prediction accuracy, but its structure is complex, the parameter quantity is large, and the training efficiency is low, which is difficult to meet the demand of large-scale practical application. In addition, compared with the simple model, this kind of complex model does not significantly improve the prediction accuracy, which shows that the deeper the learning prediction model is, the better the performance is. Double-channel TimesNet extends single-channel TimesNet to a double-branch structure that simultaneously inputs in-out station passenger flow, but only performs simple splicing at the feature layer, and fails to explicitly model the causal effect of out-of-station flow on future in-station flow, so it only slightly improves the RMSE compared with TimesNet. GCN+double-channel TimesNet adds GCN to extract spatial features based on double-channel TimesNet, but under the condition of missing node features, the introduction of GCN leads to a decrease in prediction performance.

[0158] The DBCA-TN model proposed in the application is verified by experiments, and the performance indicators are as follows: the root mean square error (RMSE) is 33.6581, the mean absolute error (MAE) is 20.8373, and the weighted average absolute percentage error (WMAPE) is 9.6946%. The model parameter quantity is 3,637,190, and the average time consumption per training cycle is 1.8835. The experimental results show that the model achieves the expected effect in terms of prediction accuracy and computational efficiency.

[0159] In addition, compared with the best-performing TimesNet in the benchmark model, the DBCA-TN model proposed in the application achieves significant improvement in all evaluation indicators: the RMSE is reduced by an average of 9.57%, the MAE is reduced by an average of 12.00%, and the WMAPE is reduced by an average of 11.67%. The results show that DBCA-TN can significantly improve the prediction accuracy of metro short-term in-station passenger flow without external information such as track adjacency graph, OD data, weather data and time label.

[0160] The above merely provides the preferred embodiments of the application, and not intended to limit the protection scope of the application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the application shall fall within the protection scope of the application.

Claims

1. A method for predicting short-term passenger flow in subway stations based on bi-branch cross-attention, characterized in that, Follow these steps: S1. Preprocess the card swiping data of the subway automatic fare collection system, set the time granularity and extract the passenger flow sequence of entering and exiting the station based on the granularity, and divide the obtained sequence into training set, validation set and test set. S2. Analyze the characteristics of passenger flow entering and exiting the station on the training set, determine the historical time interval of the impact of passenger flow exiting the station, input the length of passenger flow sequence entering and exiting the station, and construct training samples, validation samples and test samples accordingly, and determine the input dimension of passenger flow entering the station and the input dimension of passenger flow exiting the station. S3. Construct a neural network model for predicting subway passenger flow based on a dual-branch temporal network and interactive attention mechanism, and obtain the optimal neural network model through training. S4. Input the test samples into the optimal neural network model to obtain the passenger flow prediction results for the next time step.

2. The method for predicting short-term subway passenger flow based on bi-branch cross-attention as described in claim 1, characterized in that, The specific steps for S1 are as follows: S1.1: Obtain the raw card swiping data of the target subway network, clean the raw data, aggregate the cleaned data according to the predetermined time granularity, and generate the time series of passenger flow entering all stations. Time series of outbound passenger flow As shown in equations (1) to (2): （1） in: For the number of stations, This represents the total number of time steps contained in the dataset after it has been aggregated according to a predetermined time granularity. For the time series of passenger flow entering all stations, For the set of real numbers, For dimension The set of real numbers, In time step At that time, the passenger flow entering all stations; （2） in: For the time series of outbound passenger flow at all stations, In time step At that time, the outbound passenger flow vector of all stations; S1.2: Divide the time series of passenger flow entering and exiting all stations into a training set according to time sequence. Validation set and test set .

3. The method for predicting short-term subway passenger flow based on bi-branch cross-attention as described in claim 1, characterized in that, The specific steps for S2 are as follows: S2.1: Calculate the total passenger flow of all stations within the training set time step range, and plot the curve of its change over time to identify the daily cycle pattern of passenger flow. Plot the curve of passenger flow of representative stations within the training set time step range to identify the time lag pattern between passenger flow of each station. S2.2: Recording Stations The inbound sequence on the training set is Site The outbound sequence is , will the site Remove the pre-lag step size from the inbound sequence Data, obtain site Cut-off passenger flow sequence , will the site Remove the trailing lag step from the outbound sequence Data, obtain site The outbound passenger flow sequence after truncation ,calculate and Pearson correlation coefficient ,based on To determine the correlation between historical outbound passenger flow and future inbound passenger flow, Pearson correlation coefficients are calculated sequentially at multiple lag steps. The correlation between inbound and outbound passenger flow at each station under different lag steps is analyzed. Based on the periodicity of the correlation curve, the historical time interval in which outbound passenger flow has an impact is determined. ; S2.3: Input the length of the passenger flow sequence entering the station. Set as the first preset value, and input the length of the outbound passenger flow sequence. Set to the second preset value to satisfy , This is the termination value. Given the historical time step interval length, and based on the set sequence length, construct training, validation, and test samples, and determine the data batch size. Set as the third preset value, and determine the input dimensions of the inbound passenger flow based on the set batch size and sequence length. , For the data batch size, To input the length of the passenger flow sequence entering the station, The number of stations and the input dimension of outbound passenger flow. , Input the length of the outbound passenger flow sequence.

4. The method for predicting short-term subway passenger flow based on bi-branch cross-attention as described in claim 1, characterized in that, The model described in S3 includes a feature extraction layer, a temporal representation layer, a cross-branch interaction layer, and a fusion and prediction layer.

5. The method for predicting short-term subway passenger flow based on bi-branch cross-attention as described in claim 4, characterized in that, The specific structure and data processing methods of the feature extraction layer, temporal representation layer, cross-branch interaction layer, and fusion and prediction layer in the model described in S3 are as follows: The feature extraction layer includes a main branch and a sub-branch set in parallel, and both the main branch and the sub-branch use a time-series module with a TimesNet structure for feature extraction. The time-series representation layer includes a query generation module, which is connected to the main branch and is used to generate features output by the main branch. Perform tensor quantization scaling and dimensional expansion to generate a dimension of The query vector, where To embed spatial dimensions; The timing representation layer also includes a key-value pair generation module, which is connected to the sub-branch and is used to generate the features output by the sub-branch. By using linear projection, dimensions of [dimensions to be filled in] are generated respectively. The key vector and value vector, where Historical time interval Length; The Query, Key, and Value vectors are used to perform cross-branch attention calculations between the main branch and the sub-branch; The cross-branch interaction layer includes a multi-head attention module, which receives a query vector from the main branch and a key vector and a value vector from the sub-branch, calculates the correlation between the query vector and the key vector, generates attention weights, and then uses the attention weights to perform a weighted summation on the value vector. The fusion and prediction layer is used to receive the original feature representation of the main branch. Cross-attention features with the output of the cross-branch interaction layer Then, a weighted sum is performed to generate the fused feature representation. Then, the predicted passenger flow for each station is output through linear projection.

6. The method for predicting short-term subway passenger flow based on bi-branch cross-attention as described in claim 5, characterized in that, The temporal module in the feature extraction layer described in S3 processes the data according to the following steps: S3.1: Perform Z-score standardization on the input passenger flow samples along the time dimension, and map the processed data to the preset dimension embedding space through data embedding operation; S3.2: Linear projection is used to generate the prediction time step representation, thus extending the length of the input sequence; S3.3: Input the expanded sequence into the multi-layer TimesBlock module for time series modeling; S3.4: Perform layer normalization processing after each TimesBlock module so that the main branch and the sub-branch obtain the representation sequence under the extended time-series dimension respectively; S3.5: The main branch outputs the representation corresponding to the predicted target time step. The secondary branch outputs the representation corresponding to the historical time step interval. .

7. The method for predicting short-term subway passenger flow based on bi-branch cross-attention as described in claim 1, characterized in that, The method for obtaining the optimal neural network model through training, as described in S3, is as follows: Training samples are input into the model in batches according to a preset batch size. Based on the set model structure and training hyperparameters, the model parameters are continuously optimized through the backpropagation algorithm to continuously reduce the loss function. Training is terminated when the preset early stopping condition is met, and the optimal neural network model is obtained.

8. The method for predicting short-term subway passenger flow based on bi-branch cross-attention as described in claim 1, characterized in that, The specific steps of S4 are as follows: input the test set samples into the pre-trained neural network model in batches according to the preset batch size, output the passenger flow prediction results of each station in the next time step, perform de-standardization on the prediction results to restore them to the actual passenger flow scale, and evaluate the prediction accuracy and generalization performance of the model based on the real data of the test set.

Images