Traffic hub passenger flow prediction method considering multiple collection and distribution modes

Abstract

The invention relates to the field of artificial intelligence, in particular to a traffic hub passenger flow prediction method considering multiple collection and distribution modes, which comprises the following steps: collecting multi-mode passenger flow data and external factor data, and preprocessing the data to obtain original data; constructing a prediction model, wherein the prediction model comprises a multi-channel time sequence enhancement module, a space-time coding and channel fusion module and an optimization and prediction module; and inputting the multi-mode passenger flow data and the external factor data into the trained prediction model to obtain passenger flow distribution prediction values of multiple traffic modes in the traffic hub. The method not only pays attention to the dynamic space-time dependence of multiple modes, but also fuses external factors such as holidays, festivals, weather conditions and time information, and can perform refined prediction on the passenger flow of multiple transportation modes in the transportation hub.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence, and in particular to a method for predicting passenger flow at a transportation hub taking into account multiple distribution modes. Background Art

[0002] In recent years, with the rapid development of global urbanization and economy, the passenger flow of large integrated transportation hubs has increased significantly. With the vigorous development of intelligent transportation systems, passenger travel needs have become increasingly diversified. Due to the complex changes in passenger flow, the various distribution modes of hubs have brought unique challenges and opportunities to transportation planning and management. Accurate passenger flow prediction can provide an important basic reference for optimizing operation strategies and ensuring efficient resource allocation. However, obtaining accurate, refined and instructive prediction results remains a challenging task. Convolutional neural networks (CNNs) convert passenger flow data into image representations to extract temporal features. The use of convolutional layers can greatly improve parallel computing capabilities and efficiency. However, relatively simple models may not be able to fully capture nonlinear relationships and complex patterns in temporal serial data. The self-attention mechanism enables Transformer to consider all elements in the series and create various correlations, which has attracted great attention from researchers. However, the window size directly limits the contextual dependencies in long-sequence processing, affecting training and reasoning. Some studies divide transportation centers into different areas to predict passenger flow distribution and extract spatial features based on graph convolutional neural networks (GCNs). In addition, there are studies that consider multi-source data, including mobile phone data, special events, weather conditions, and smart card data for comprehensive prediction.

[0003] However, most existing studies only consider the changes in overall passenger flow within a transportation hub from a temporal perspective, while ignoring the impact of other surrounding transportation modes. Passengers arriving at or leaving a railway station often use multiple modes of transportation, so ignoring the interactions between them will result in the inability to identify specific transportation modes. Multimodal refined predictions enable managers to understand the specific needs of various transportation modes, thereby improving service quality and effectively identifying potential safety hazards. However, when considering the interactions of multiple modes, the goal becomes to effectively model their complex spatiotemporal correlations, and current research still has the following problems: In the spatial dimension, passenger flow exhibits significant interdependence and transitivity between different transportation entities. Changes in traffic status will change the connection between different transportation modes. Some studies use CNN to capture local spatial correlations. A significant limitation is that the irregularity of real-world traffic structure is ignored, resulting in the loss of key topological information. A better way to avoid this problem is to construct the transportation network as a graph. Specifically, GCN usually uses an adjacency matrix or a Laplacian matrix to describe the graph structure. However, the various transportation modes around the transportation hub are unevenly distributed, and their spatial associations change dynamically over time in a complex manner. Traffic graphs using static or adaptive adjacency matrices cannot dynamically reflect the real spatial relationship at every moment. Therefore, it is of practical significance to construct dynamic traffic graphs driven by data. In the time dimension, traffic conditions fluctuate periodically and variably. Therefore, multimodal passenger flow data exhibit significant volatility and continuous interactions. Unlike the perspective of focusing only on overall passenger flow, fine prediction is more challenging, but more detailed passenger flow distribution prediction can be achieved. Deep learning models can learn regular time patterns through continuous training. However, special situations such as holidays require special treatment. Some studies incorporate external factors into model training to adapt to abnormal situations. Factors such as weather, road conditions, and holidays can be integrated into the network together with time data for joint learning. The time interval of passenger flow also affects the difficulty of capturing time features. Within a certain range, larger intervals will introduce more volatility and lose more details, making accurate predictions in the future more challenging. Especially at railway stations, the strong volatility and data spikes of multimodal passenger flow bring difficulties to prediction. And the management department has extremely high requirements for accuracy during peak hours in order to take corresponding emergency measures. A comprehensive time feature enhancement solution needs to be specially designed to better serve multimodal prediction.

[0004] Therefore, how to accurately predict passenger flow at transportation hubs with multiple distribution modes is an urgent problem to be solved. Summary of the invention

[0005] The present application provides a method for predicting passenger flow at a transportation hub taking into account a variety of distribution modes, which can accurately predict the passenger flow of various modes of transportation within a transportation hub.

[0006] In order to achieve the above object, the present invention adopts the following technical scheme:

[0007] In a first aspect, the present invention provides a method for predicting passenger flow at a transportation hub taking into account a variety of distribution modes, the method comprising:

[0008] Collect multi-mode passenger flow data and external factor data, pre-process the data, and obtain raw data;

[0009] Constructing a prediction model, the prediction model includes a multi-channel timing enhancement module, a spatiotemporal coding and channel fusion module, and an optimization and prediction module;

[0010] The multi-channel timing enhancement module enhances the timing characteristics of the original data through multiple channels;

[0011] In the spatiotemporal coding and channel fusion module, a spatiotemporal dynamic graph convolutional recurrent network is obtained based on a gated recurrent unit and a dynamic graph convolutional structure, wherein the matrix product in the gated recurrent unit is replaced by a dynamic graph convolutional operation, and then the original data enhanced by the temporal features is input into the spatiotemporal dynamic graph convolutional recurrent network to obtain coded data, and then the coded data is fused by an adaptive channel attention mechanism to obtain fused data;

[0012] The optimization and prediction module converts the external factor data into a continuous vector representation, designs the external information encoding corresponding to the passenger flow data of each traffic mode, uses position encoding to represent the order of the fused data, obtains auxiliary prediction data, defines each transportation hub in the fused data as a multi-modal transportation graph network, abstracts each traffic flow as a node, combines the auxiliary prediction data, and projects the feature representation of each node in the time dimension and feature dimension through the self-attention mechanism to obtain the passenger flow distribution prediction values ​​of multiple transportation modes;

[0013] Based on the multi-modal passenger flow data and the external factor data, training the prediction model based on a preset loss function and training settings;

[0014] The multi-mode passenger flow data and external factor data are input into the trained prediction model to obtain the passenger flow distribution prediction values ​​of various transportation modes in the transportation hub.

[0015] In a preferred example of the present application, it can be further configured that the multi-channel timing enhancement module enhances the timing characteristics of the original data through multiple channels, including:

[0016] The multi-channel time series enhancement module uses a sequence decomposition channel, a history enhancement channel and a peak enhancement channel to process the original data to obtain original data with enhanced time series features.

[0017] In a preferred example of the present application, it can be further configured to include:

[0018] The sequence decomposition channel decomposes the signal serially into a series of intrinsic mode functions through empirical mode decomposition, each intrinsic mode function represents an oscillation component in a different frequency range, and then integrates each decomposition result through the CEEMDAN algorithm to obtain a decomposition result;

[0019] The history enhancement channel performs history enhancement on the original data based on dilated causal convolution;

[0020] The peak enhancement channel captures peak information on different time scales in the original data by combining different MaxPooling 1D convolution kernels.

[0021] In a preferred example of the present application, it can be further configured that the raw data enhanced by the time series feature is input into the spatiotemporal dynamic graph convolutional recurrent network to obtain the encoded data, including:

[0022] After the original data enhanced by the time series features is input into the spatiotemporal dynamic graph convolutional recurrent network, the spatiotemporal dynamic graph convolutional recurrent network extracts the spatiotemporal association of each node from the time series data at each moment, and maps it into a dynamic adjacency matrix;

[0023] Normalizing the generated dynamic adjacency matrix to obtain a dynamic convolution kernel;

[0024] Based on the dynamic convolution kernel, the transmission of interactive information between each node is completed through the graph convolution operation to obtain the encoded data.

[0025] In a preferred example of the present application, it can be further configured that the adaptive channel attention mechanism is then used to fuse the encoded data to obtain fused data, including:

[0026] At each moment, the multi-channel spatiotemporal codes of each traffic flow in the encoded data are stacked into one code, and then the code is transformed into a shape for matrix operation, and the matrix multiplication result of the SoftMax layer is used to obtain a channel attention map;

[0027] Use an average pooling layer to extract the overall channel weight;

[0028] The channel attention map is used to weight-enhance the multi-channel spatiotemporal coding, and the overall weight is used to fuse the multi-channel coding to obtain fused data.

[0029] In a preferred example of the present application, it can be further configured that the external factor data includes:

[0030] Holidays, weather conditions and time information.

[0031] In a second aspect, the present application provides a transportation hub passenger flow prediction device considering multiple distribution modes, the device comprising:

[0032] The data collection module is used to collect multi-mode passenger flow data and external factor data, pre-process the data, and obtain the original data;

[0033] A model building module is used to construct a prediction model, which includes a multi-channel time series enhancement module, a spatiotemporal coding and channel fusion module, and an optimization and prediction module; the multi-channel time series enhancement module performs time series feature enhancement on the original data through multiple channels; in the spatiotemporal coding and channel fusion module, a spatiotemporal dynamic graph convolution recurrent network is constructed based on a gated recurrent unit and a dynamic graph convolution, wherein the matrix product in the gated recurrent unit is replaced by a dynamic graph convolution operation, and then the original data after the time series feature enhancement is input into the spatiotemporal dynamic graph convolution recurrent network to obtain encoded data, and then the encoded data is fused by an adaptive channel attention mechanism to obtain fused data; the optimization and prediction module converts the external factor data into a continuous vector representation, designs the external information encoding corresponding to the passenger flow data of each traffic mode, uses position encoding to represent the order of the fused data, obtains auxiliary prediction data, defines each transportation hub in the fused data as a multi-mode transportation graph network, and abstracts each traffic flow as a node. Combined with the auxiliary prediction data, the feature representation of each node in the time dimension and feature dimension is projected by the self-attention mechanism to obtain the passenger flow distribution prediction values ​​of multiple transportation modes;

[0034] A model training module, used for training the prediction model based on the multi-modal passenger flow data and external factor data, based on a preset loss function and training settings;

[0035] The prediction module is used to input the multi-mode passenger flow data and external factor data into the trained prediction model to obtain the passenger flow distribution prediction values ​​of various transportation modes in the transportation hub.

[0036] In a third aspect, the present application provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the processor implements the steps of a method for predicting passenger flow at a transportation hub taking into account a variety of distribution modes as described in any one of the above items.

[0037] In a fourth aspect, the present application provides a computer-readable storage medium having a program stored thereon, wherein when the program is executed by a processor, a method for predicting passenger flow at a transportation hub taking into account a variety of distribution modes as described in any one of the above items is implemented.

[0038] In a fifth aspect, the present application provides a computer program product, comprising computer instructions, which, when executed by a processor, implement the steps of a method for predicting passenger flow at a transportation hub taking into account a variety of collection and distribution modes as described in any of the above items.

[0039] In summary, compared with the prior art, the technical solution provided in the embodiment of the present application has at least the following beneficial effects:

[0040] The present application provides a method for predicting passenger flow at a transportation hub that takes into account multiple distribution modes. It not only focuses on the dynamic spatiotemporal dependencies of multiple modes, but also integrates external factors such as holidays, weather conditions, and time information. It can make refined predictions of passenger flow of multiple modes of transportation in the transportation hub, rather than just predicting the overall flow. This unique perspective provides stronger practical value for significantly improving collaborative management under abnormal conditions. The integrated time enhancement strategy of the present application using signal decomposition and convolution processing technology can effectively deal with the problems of strong volatility and insufficient data spike processing capabilities in time series data. Compared with existing methods, better time series feature representation can be obtained to serve downstream prediction tasks. The spatiotemporal dynamic graph convolutional recurrent network is data-driven, can comprehensively capture global and local spatiotemporal patterns, and simultaneously complete efficient encoding of spatiotemporal features. Compared with existing methods, it can break through the inherent limitations of static feature extraction. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 A flowchart of a method for predicting passenger flow at a transportation hub taking into account multiple distribution modes is provided for one embodiment of the present application.

[0042] Figure 2 A spatial dynamic correlation diagram of a transportation hub for a transportation hub passenger flow prediction method taking into account multiple distribution modes is provided as an embodiment of the present application.

[0043] Figure 3 A prediction model structure diagram of a method for predicting passenger flow at a transportation hub taking into account multiple distribution modes provided in one embodiment of the present application.

[0044] Figure 4 A diagram of the construction of a multi-head self-attention module for a method for predicting passenger flow at a transportation hub taking into account a variety of distribution patterns, provided as an embodiment of the present application.

[0045] Figure 5 A temporal convolutional neural network diagram composed of dilated causal convolutions, providing a method for predicting passenger flow at a transportation hub taking into account multiple distribution patterns, provided as an embodiment of the present application.

[0046] Figure 6A spatiotemporal dynamic graph convolutional recurrent network diagram of a method for predicting passenger flow at a transportation hub taking into account multiple distribution patterns is provided as an embodiment of the present application.

[0047] Figure 7 A flowchart of a method for predicting passenger flow at a transportation hub taking into account a variety of distribution patterns, which adopts an adaptive channel attention mechanism to fuse the encoded data, is provided as an embodiment of the present application.

[0048] Figure 8 A schematic diagram of a loss function of a method for predicting passenger flow at a transportation hub taking into account multiple distribution modes is provided as an embodiment of the present application.

[0049] Fig. 9 A structural diagram of a transportation hub passenger flow prediction device that takes into account multiple distribution modes is provided for one embodiment of the present application. DETAILED DESCRIPTION

[0050] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0051] In one embodiment of the present application, a method for predicting passenger flow at a transportation hub taking into account multiple distribution modes is provided. Figure 1 As shown, the method includes:

[0052] S100: Collect multi-mode passenger flow data and external factor data, pre-process the data, and obtain original data;

[0053] Specifically, the multi-modal passenger flow data includes passenger flow data of passengers arriving or leaving the transportation hub using multiple modes of transportation, such as Figure 2 As shown, the transportation hub can be a railway station, and the means of arrival or departure can be taxi, private car, etc. The data preprocessing includes abnormal data cleaning and data normalization. The external factor data is the factors that affect the passenger flow of the transportation hub, which can be factors such as weather, road conditions and holidays.

[0054] Furthermore, the external factor data includes holidays, weather conditions, temperature and time information.

[0055] S200: construct a prediction model, the prediction model includes a multi-channel time series enhancement module, a spatiotemporal coding and channel fusion module and an optimization and prediction module; the multi-channel time series enhancement module performs time series feature enhancement on the original data through multiple channels; in the spatiotemporal coding and channel fusion module, a spatiotemporal dynamic graph convolutional recurrent network is constructed based on gated recurrent units and dynamic graph convolution, wherein the matrix product in the gated recurrent unit is replaced by a dynamic graph convolution operation, and then the original data after time series feature enhancement is input into the spatiotemporal dynamic graph convolutional recurrent network to obtain encoded data, and then the encoded data is fused by an adaptive channel attention mechanism to obtain fused data; the optimization and prediction module converts the external factor data into a continuous vector representation, designs the external information encoding corresponding to the passenger flow data of each traffic mode, uses position encoding to represent the order of the fused data, obtains auxiliary prediction data, defines each transportation hub in the fused data as a multi-mode transportation graph network, and abstracts each traffic flow as a node, and combines the auxiliary prediction data, and projects the feature representation of each node in the time dimension and feature dimension through the self-attention mechanism to obtain the passenger flow distribution prediction values ​​of multiple transportation modes;

[0056] Specifically, the overall network structure is as follows: Figure 3 As shown, the processing steps of the original data are:

[0057] The first step attempts to enhance temporal features over multiple channels, including serial decomposition to reveal complex temporal patterns, history enhancement to mitigate data volatility, and peak enhancement to emphasize abnormal data.

[0058] In the second step, the spatiotemporal dependencies of each channel are synchronously encoded through a data-driven spatiotemporal dynamic graph convolutional recurrent network (STDGCRN). Then, an adaptive channel attention mechanism is used to fuse the multi-channel encodings and obtain optimized feature representations according to different prediction tasks.

[0059] The third step is to use the self-attention mechanism and combine it with external factors to strengthen the guiding role of external priors, and finally map it into the predicted values ​​of passenger flow distribution for multiple modes of transportation. Among them, each transportation hub can be defined as a multimodal transportation graph network, that is, G = (V, E, A). V is a set of N nodes, and E is a set of connected edges. Each mode of transportation can generate passenger flow data in up to two directions (i.e., arrival and departure). Each traffic flow is abstracted as a node. Figure 3 The adjacency matrix in represents the correlation between each traffic passenger flow and is an N×N real symmetric matrix.

[0060] Given a multivariate traffic flow data X with a fixed history length P =[x t-P+1 ,…,x t-1 ,xt ]∈R P×N×C , by learning a complex prediction function, aiming to predict the future fixed-length multivariate traffic flow data Y Q =[y t+1 ,…,y t+Q ]∈R Q ×N×C , where P represents the size of the history window and Q represents the size of the prediction window.

[0061] S300: Based on the multi-mode passenger flow data and the external factor data, the prediction model is trained based on a preset loss function and training settings;

[0062] S400: Inputting the multi-mode passenger flow data and external factor data into the trained prediction model to obtain passenger flow distribution prediction values ​​of multiple transportation modes in the transportation hub.

[0063] Specifically, the text information of various external factors including holidays, weather conditions and time information is first numerically numbered and converted into numerical types. Then, the embedding technology is used to convert the data into a continuous vector representation.

[0064] Under the same external environment, each type of traffic flow data presents different spatiotemporal characteristics, so it is necessary to design a unique external information encoding for each type of traffic flow data. P ∈R P×N×F Through this encoding, the model will learn the unique association between external information and each type of traffic flow data as much as possible, thereby improving the accuracy of the prediction as much as possible.

[0065] In addition, in the Transformer model, positional encoding is introduced to solve the problem that the self-attention mechanism itself does not contain position information. When the Transformer processes sequence data, the self-attention mechanism will treat them as equivalent regardless of how the sequence order changes, which is obviously not in line with the actual situation. Since the self-attention mechanism is used, positional encoding is also required to represent the sequence. Positional encoding is generally composed of sine and cosine functions, and is represented by formulas 24 and 25, where t represents the position of the sequence moment, i represents the specific dimension, and F represents the length of the embedding vector.

[0066]

[0067] In this way, for each position in the input sequence, an embedding vector of length F corresponding to its position information can be obtained. This position encoding method of sine and cosine functions can provide the model with position information of different frequencies, which helps the model capture long-distance dependencies in the sequence. Secondly, due to the periodicity of sine and cosine functions, this position encoding method also has a certain robustness, that is, when the sequence length changes, the model can still handle it well.

[0068] It should be noted that this positional encoding does not change during model training and only represents a fixed positional encoding.

[0069]

[0070] As shown in Formula 25, all the codes are added together to get the final spatiotemporal feature representation Emb P ∈R P×N×F , and use it as the input of the time series attention enhancement layer. The fusion process allows the model to learn the impact of external factors on time series prediction while maintaining the model's sensitivity to the order of time series data.

[0071] Given sequence data, it can be mapped to query vector Q, key vector K and value vector V. The self-attention mechanism is expressed as shown in Formula 26:

[0072]

[0073] For each traffic flow representation, a self-attention layer is used to perform temporal attention enhancement operation. i Input (Q i , K i and V i ) use the feature encoding obtained previously that combines external information and position information is represented by, where i = 0, 1, ..., N-1 represents the sequence number of the multivariate traffic flow, d k For Emb P Embedding dimension F. After the temporal enhancement processing of the self-attention layer, the output result It can be expressed by formula 27:

[0074]

[0075] like Figure 4As shown in the figure, the multi-head self-attention block consists of layer normalization, node-level multi-head self-attention layer, random dropout layer, and residual connection layer. Layer normalization is applied to stabilize the network layer data and accelerate training convergence. The model residually connects (i.e., adds) the output of the self-attention module to the original input, which helps the model learn the identity mapping and accelerates the training process. The random dropout layer is used to prevent overfitting and enhance the generalization ability of the model. The final representation after attention enhancement is

[0076] Finally, the feature representations of each node in the time dimension and feature dimension are projected to obtain the predicted data

[0077]

[0078] In this embodiment, not only the dynamic spatiotemporal dependencies of multiple modes are emphasized, but also external factors such as holidays, weather conditions and time information are integrated. It is possible to make refined predictions of passenger flows of multiple modes of transportation in transportation hubs, rather than just predicting the overall flow. This unique perspective provides stronger practical value for significantly improving collaborative management under abnormal circumstances. The integrated time enhancement strategy of the present application using signal decomposition and convolution processing technology can effectively deal with the problems of strong volatility and insufficient data spike processing capabilities in time series data. Compared with existing methods, better time series feature representation can be obtained to serve downstream prediction tasks. The spatiotemporal dynamic graph convolutional recurrent network is data-driven, can comprehensively capture global and local spatiotemporal patterns, and simultaneously complete efficient encoding of spatiotemporal features. Compared with existing methods, it can break through the inherent limitations of static feature extraction.

[0079] In some embodiments, the multi-channel timing enhancement module performs timing feature enhancement on the original data through multiple channels, including:

[0080] The multi-channel time series enhancement module uses a sequence decomposition channel, a history enhancement channel and a peak enhancement channel to process the original data to obtain original data with enhanced time series features.

[0081] The sequence decomposition channel decomposes the signal serially into a series of intrinsic mode functions through empirical mode decomposition, each intrinsic mode function represents an oscillation component in a different frequency range, and then integrates each decomposition result through the CEEMDAN algorithm to obtain a decomposition result;

[0082] The history enhancement channel performs history enhancement on the original data based on dilated causal convolution;

[0083] The peak enhancement channel captures peak information on different time scales in the original data by combining different MaxPooling 1D convolution kernels.

[0084] In specific implementation, the original data X P Through a comprehensive timing enhancement solution (i.e., serial decomposition, history enhancement, and peak amplification), it is processed into Dp, Bp, and Mp to cope with strong volatility, data sparsity, and data peaks, respectively.

[0085] The specific steps of processing the original data by the sequence decomposition channel are:

[0086] Empirical Mode Decomposition (EMD) can adaptively decompose the signal serially into a series of Intrinsic Mode Functions (IMFs). Each IMF represents an oscillating component in a different frequency range.

[0087] Fully Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) introduces adaptive noise to address the sensitivity of EMD to noise when dealing with non-stationary signals, thereby improving the modal decomposition. For each time series, CEEMDAN is applied to capture different temporal patterns and mitigate the volatility of the data.

[0088] In a given one-dimensional sequence s o (t), add I-order Gaussian white noise Then perform EMD decomposition. For the first component Find the average value and get the corresponding first residual component As shown in Formula 1 and Formula 2. Among them, Represents the signal-to-noise ratio.

[0089]

[0090] Repeat the above steps for the obtained residual signal until the last is a monotonic function or a constant, where u represents the final number of decompositions. For multimodal passenger flow, the maximum number of decomposed IMFs is m. Zero padding is then used to ensure consistency and facilitate subsequent network model processing. The final result can be expressed as Formula 3.

[0091]

[0092] Finally, the N passenger flows decomposed by CEEMDAN and their respective components are recombined into new data, that is, Among them C dRepresents the number of feature channels after decomposition. The CEEMDAN algorithm integrates each decomposition result of EMD to obtain a more complete and robust decomposition result. This completeness can better capture the characteristics of time series and improve the adaptability to nonlinearity and non-stationarity.

[0093] The specific steps of the history enhancement channel processing the original data are:

[0094] Dilated causal convolution is a special convolution operation that aims to simultaneously increase the receptive field of the model and ensure temporal causality, so that dependencies over a longer time range can be captured. Dilated convolution introduces a dilation factor into the convolution kernel, allowing the convolution kernel to extract features within a larger receptive field, effectively increasing the range of information that the model can observe. Causal convolution is a strict time-constrained model that ensures that each convolution kernel can only depend on its previous input data, thereby avoiding information leakage and enabling the network to better cope with the processing needs of time series data. The temporal convolutional neural network composed of dilated causal convolution is as follows: Figure 5 As shown. In a given one-dimensional sequence s o (t), the dilated causal convolution is expressed as shown in Formula 4, d represents the dilation factor, k T Indicates the size of the convolution kernel.

[0095]

[0096] Each dilated causal convolutional block can be expressed as Formula 5. The combination of two dilated causal convolutional blocks and residual connections forms a TCN block. Combining TCN blocks with different receptive fields, the information transfer capability of long sequences is further improved. Finally, for X P Each traffic flow time series data is processed using multiple TCN Blocks with different parameters. Among them C b It represents the number of feature channels after historical enhancement, so as to achieve the purpose of historical enhancement of long-sequence data and solve the data sparsity problem in the original data.

[0097] DCBlock(s o (t))=Dropout(Relu(WeightNorm(DConv(s) o (t))))); (5)

[0098] The specific steps of processing the raw data by the peak enhancement channel are as follows:

[0099] Combining different MaxPooling 1D convolution kernels can capture peak information at different time scales. Smaller kernels can capture finer temporal patterns, while larger kernels focus more on the overall structure. The whole process is expressed as shown in Formula 6. M Indicates the size of the convolution kernel. P Each traffic flow time series data is subjected to peak enhancement operation to obtain Among them C m Indicates the number of feature channels after peak enhancement.

[0100]

[0101] In some embodiments, the step of inputting the raw data enhanced by the temporal features into the spatiotemporal dynamic graph convolutional recurrent network to obtain the encoded data includes:

[0102] After the original data enhanced by the time series features is input into the spatiotemporal dynamic graph convolutional recurrent network, the spatiotemporal dynamic graph convolutional recurrent network extracts the spatiotemporal association of each node from the time series data at each moment, and maps it into a dynamic adjacency matrix;

[0103] Normalizing the generated dynamic adjacency matrix to obtain a dynamic convolution kernel;

[0104] Based on the dynamic convolution kernel, the transmission of interactive information between each node is completed through the graph convolution operation to obtain the encoded data.

[0105] Furthermore, the adaptive channel attention mechanism is then used to fuse the encoded data to obtain fused data, including:

[0106] At each moment, the multi-channel spatiotemporal codes of each traffic flow in the encoded data are stacked into one code, and then the code is transformed into a shape for matrix operation, and the matrix multiplication result of the SoftMax layer is used to obtain a channel attention map;

[0107] Use an average pooling layer to extract the overall channel weight;

[0108] The channel attention map is used to weight-enhance the multi-channel spatiotemporal coding, and the overall weight is used to fuse the multi-channel coding to obtain fused data.

[0109] In the specific implementation, in order to better synchronize and integrate spatiotemporal features, a spatial-temporal dynamic graph convolutional recurrent network (STDGCRN) was constructed based on data-driven dynamic graph convolution to encode the data of three channels respectively. This network can fully process global and local spatiotemporal associations. Then an adaptive channel attention mechanism is used to better fuse the encoding.

[0110] The STDGCRN node includes the following process when processing data:

[0111] Dynamic graph convolution kernels can be dynamically learned from the node features and structural information of the graph to adapt to changes in the graph structure and better process various complex graph data. Figure 6 As shown, STDGCRN extracts the spatiotemporal correlation of each node from the time series data at each moment and maps it into a dynamic adjacency matrix.

[0112] The input of the current time step t is passed through the MLP layer to extract the dynamic signal, as shown in Formula 7. Where, F t ∈R N×L represents the dynamic information after filtering, and L represents the channel dimension of the dynamic information.

[0113] F t =MLP(x t ); (7)

[0114] Construct a learnable embedding representation E for each node t ∈R N×L , used to represent the spatiotemporal pattern of each node, and then in F t and E t Perform an element-wise multiplication operation on to generate a dynamic graph embedding As shown in Formula 8:

[0115]

[0116] The spatial dependence can be explained by and In order to meet the requirements of Chebyshev polynomials, the generated dynamic matrix is ​​normalized, as shown in Formula 9. Among them, represents the dynamic adjacency matrix at time t, Represents the dynamic degree matrix at time t.

[0117]

[0118] The calculation expression of the GCN-based dynamic graph convolution kernel is shown in Formula 10, and its simplified form is shown in Formula 11:

[0119]

[0120] It represents the graph embedding representation after dynamic graph convolution at each moment, and F represents the embedding dimension size.

[0121] The model dynamically adapts according to the spatiotemporal characteristics of each time step in the traffic flow data, and then obtains the dynamic convolution kernel after normalization, and completes the effective transmission of interactive information between each node through graph convolution operation.

[0122] Traffic prediction involves complex temporal correlation and spatial association. The traditional GRU can effectively capture temporal features, and on this basis, the dynamic graph convolution operation is used to replace the matrix product in the GRU.

[0123] First, calculate the reset gate. Input x at the current moment t and the output h of the hidden layer at the previous moment t-1 After the dynamic graph convolution transformation, it is connected to Sigmoid (represented by σ), and its output value is Between is used to select how much information to leave, as shown in Formula 12:

[0124]

[0125] The update gate is then calculated. The calculation process is similar to the reset gate, where It is used to select how much information to use for updating, as shown in Formula 13:

[0126]

[0127] Next, the current memory content is calculated. A reset gate is used to store relevant information from the past and combine it with the current input information, as shown in Formula 14:

[0128]

[0129] The final memory of the current time step t is determined by the update gate to determine the distribution of information, as shown in Formula 15. Among them, the embedding dimension of the hidden layer representation at the current time t is determined by the dynamic graph convolution kernel To control, that is

[0130]

[0131] As shown in formula 16-18, the previously obtained multi-channel timing data (D p , B p and Mp ) are encoded using a STDGCRN respectively, and the final multi-channel spatiotemporal representation is obtained ( and ). This shows that each STDGCRN channel will learn and update parameters according to the distribution characteristics of its own data.

[0132]

[0133] The adaptive channel attention mechanism is used to fuse the encoded data to obtain a fused data channel attention mechanism, including the following process:

[0134] A channel attention mechanism is designed to perform channel fusion on multi-channel spatiotemporal feature encodings, which allows the model to focus on important information in the channel dimension during feature extraction, thereby improving the representation ability of features. In addition, for different downstream prediction tasks, the model pays different attention to different channel encodings, which can better help the model make decisions and enhance the robustness of the model. For example, for long-term predictions, the channel attention mechanism may pay attention to the historically enhanced encoded data based on more specific weights.

[0135] At each time t, the multi-channel spatiotemporal codes of each traffic flow are stacked into one code, emb t ∈R 3×N×F . Then emb t Transform the shape to get To perform matrix operations, apply the SoftMax layer to and The matrix multiplication result is used to obtain the channel attention map A t ∈R 3×3 To express the channel correlation, as shown in Equation 19 and Equation 20:

[0136]

[0137] Among them, α represents the scale factor, which is a constant. Its function is to scale the input tensor so that its value range is suitable for the input of the SoftMax function. Through this scaling, the output of the SoftMax layer can be better controlled, making the distribution of attention weights more stable and stable, which helps to improve the training stability and performance of the model.

[0138] Use the average pooling layer to extract the overall channel weight W t ∈R 3 , as shown in formula 21:

[0139] W t =SoftMax(MLP(AvgPool(embt )));(twenty one)

[0140] Using channel attention map A t The multi-channel spatiotemporal coding is weighted to obtain As shown in formula 22. Adjust the shape to get Then use the overall weight W t Perform multi-channel coding fusion to obtain the final coding representation As shown in Formula 23:

[0141]

[0142] The schematic diagram of channel attention encoding fusion is as follows Figure 7 As shown in the figure, after the encoding fusion of multi-channel attention, the model can adaptively focus on the important features on different channels and perform weighted fusion according to the importance of these features. This mechanism significantly improves the representation ability of features, allowing the model to more effectively capture and utilize key information in the data when processing complex tasks. By increasing the sensitivity of the model to key features, the robustness of the model is improved. In addition, the channel attention mechanism can also help the model achieve better sharing and migration between multiple prediction tasks. Finally, the spatiotemporal encoding is obtained as

[0143] In some embodiments, based on the multi-modal passenger flow data and the external factor data, the prediction model is trained based on a preset loss function and training settings, including:

[0144] In the specific implementation, in order to improve the prediction accuracy of peak data, a new loss function called Enhanced Peak Exponential Loss (EPEL) is designed, as shown in Formula 28. Among them, p and q are adjustment terms that control the sensitivity of EPEL to different errors. After normalization, the data is scaled to [0,1]. Therefore, the data is mapped to [0,∞] in an exponentially amplified manner. The error is further amplified by adjusting p and q, and a higher penalty is imposed on the error of the peak data. As shown in Figure 8 As shown, the exponential function and the different exponential terms make it more sensitive to larger errors, especially for peak values.

[0145]

[0146] In the actual training process of the network, 12 2-layer TCN blocks and 6 1-layer MaxPooling blocks are used. The convolution kernel size increases linearly with the number of layers in both blocks. The number of CEEMDAN decompositions is determined by the complexity of the data itself. Each STDGCRN is a single layer and encodes the data into a spatiotemporal representation of size 128. Finally, a 2-layer self-attention block is used to enhance the temporal features. The AdamW optimizer is used with a batch size of 128 and a weight decay set to 0.0001. The duration of early stopping is 6 seconds. The initial learning rate is set to 10 -4 The parameters in the EPEL function are set to p=2 and q=1.

[0147] This application also provides a transportation hub passenger flow prediction device that takes into account multiple distribution modes. Fig. 9 As shown, the device comprises:

[0148] The data collection module 100 is used to collect multi-mode passenger flow data and external factor data, and pre-process the data to obtain raw data;

[0149] The model building module 200 is used to construct a prediction model, which includes a multi-channel timing enhancement module, a spatiotemporal coding and channel fusion module, and an optimization and prediction module; the multi-channel timing enhancement module enhances the timing characteristics of the original data through multiple channels; in the spatiotemporal coding and channel fusion module, a spatiotemporal dynamic graph convolutional recurrent network is obtained based on a gated recurrent unit and a dynamic graph convolution operation, wherein the matrix product in the gated recurrent unit is replaced by a dynamic graph convolution operation, and then the original data after the timing characteristics are enhanced is input into the spatiotemporal dynamic graph convolutional recurrent network to obtain the encoded data, and then the adaptive The channel attention mechanism corresponding to the encoding data is used to fuse the encoded data to obtain fused data; the optimization and prediction module converts the external factor data into a continuous vector representation, designs the external information encoding corresponding to the passenger flow data of each traffic mode, uses position encoding to represent the order of the fused data, obtains auxiliary prediction data, defines each transportation hub in the fused data as a multi-modal traffic graph network, abstracts each traffic flow as a node, and combines the auxiliary prediction data to project the feature representation of each node in the time dimension and feature dimension through the self-attention mechanism to obtain the passenger flow distribution prediction values ​​of multiple transportation modes;

[0150] A model training module 300, for training the prediction model based on the multi-modal passenger flow data and external factor data, based on a preset loss function and training settings;

[0151] The prediction module 400 is used to input the multi-mode passenger flow data and external factor data into the trained prediction model to obtain the passenger flow distribution prediction values ​​of various transportation modes in the transportation hub.

[0152] The functional implementation of each module in the above-mentioned transportation hub passenger flow prediction device considering multiple distribution modes corresponds to the steps in the above-mentioned transportation hub passenger flow prediction method embodiment considering multiple distribution modes, and its functions and implementation process will not be repeated here one by one.

[0153] The present application also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the steps of the method for predicting passenger flow at a transportation hub taking into account a plurality of distribution modes as described in any of the above embodiments are implemented.

[0154] The present application also provides a computer-readable storage medium, on which a program is stored, wherein the computer-readable storage medium refers to a carrier for storing data, which may include but is not limited to a floppy disk, an optical disk, a hard disk, a flash memory, a USB flash drive, and / or a memory stick, etc., and the computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The working process, working details, and technical effects of the computer-readable storage medium provided in this embodiment can be found in the above embodiment of a method for predicting passenger flow at a transportation hub considering multiple distribution modes, which will not be repeated here.

[0155] The application also provides a computer program product, including computer instructions, which, when executed by a processor, implement the steps of the method for predicting passenger flow at a transportation hub taking into account multiple collection and distribution modes as described in any of the above embodiments.

[0156] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0157] The technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, all possible combinations of the technical features in the above embodiments are not described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification. The above-mentioned embodiments only express several implementation methods of the present application, and their descriptions are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that for ordinary technicians in this field, without departing from the concept of the present application, several deformations and improvements can be made, which all belong to the protection scope of the present application. Therefore, the protection scope of the patent of this application shall be based on the attached claims.

Claims

1. A method for predicting passenger flow at a transportation hub considering multiple distribution modes, characterized in that: include: Collect multi-mode passenger flow data and external factor data, pre-process the data, and obtain raw data; Constructing a prediction model, the prediction model includes a multi-channel timing enhancement module, a spatiotemporal coding and channel fusion module, and an optimization and prediction module; The multi-channel timing enhancement module enhances the timing characteristics of the original data through multiple channels; In the spatiotemporal coding and channel fusion module, a spatiotemporal dynamic graph convolutional recurrent network is obtained based on a gated recurrent unit and a dynamic graph convolutional structure, wherein the matrix product in the gated recurrent unit is replaced by a dynamic graph convolutional operation, and then the original data enhanced by the temporal features is input into the spatiotemporal dynamic graph convolutional recurrent network to obtain coded data, and then the coded data is fused by an adaptive channel attention mechanism to obtain fused data; The optimization and prediction module converts the external factor data into a continuous vector representation, designs the external information encoding corresponding to the passenger flow data of each traffic mode, uses position encoding to represent the order of the fused data, obtains auxiliary prediction data, defines each transportation hub in the fused data as a multi-modal transportation graph network, abstracts each traffic flow as a node, combines the auxiliary prediction data, and projects the feature representation of each node in the time dimension and feature dimension through the self-attention mechanism to obtain the passenger flow distribution prediction values ​​of multiple transportation modes; Based on the multi-modal passenger flow data and the external factor data, training the prediction model based on a preset loss function and training settings; The multi-mode passenger flow data and external factor data are input into the trained prediction model to obtain the passenger flow distribution prediction values ​​of various transportation modes in the transportation hub.

2. The method for predicting passenger flow at a transportation hub considering multiple distribution modes according to claim 1, characterized in that: The multi-channel timing enhancement module performs timing feature enhancement on the original data through multiple channels, including: The multi-channel time series enhancement module uses a sequence decomposition channel, a history enhancement channel and a peak enhancement channel to process the original data to obtain original data with enhanced time series features.

3. The method for predicting passenger flow at a transportation hub considering multiple distribution modes according to claim 2, characterized in that: include: The sequence decomposition channel decomposes the signal serially into a series of intrinsic mode functions through empirical mode decomposition, each intrinsic mode function represents an oscillation component in a different frequency range, and then integrates each decomposition result through the CEEMDAN algorithm to obtain a decomposition result; The history enhancement channel performs history enhancement on the original data based on dilated causal convolution; The peak enhancement channel captures peak information on different time scales in the original data by combining different MaxPooling 1D convolution kernels.

4. The method for predicting passenger flow at a transportation hub considering multiple distribution modes according to claim 3 is characterized in that: The raw data enhanced by the temporal features is input into the spatiotemporal dynamic graph convolutional recurrent network to obtain encoded data, including: After the original data enhanced by the time series features is input into the spatiotemporal dynamic graph convolutional recurrent network, the spatiotemporal dynamic graph convolutional recurrent network extracts the spatiotemporal association of each node from the time series data at each moment, and maps it into a dynamic adjacency matrix; Normalizing the generated dynamic adjacency matrix to obtain a dynamic convolution kernel; Based on the dynamic convolution kernel, the transmission of interactive information between each node is completed through the graph convolution operation to obtain the encoded data.

5. The method for predicting passenger flow at a transportation hub considering multiple distribution modes according to claim 1, characterized in that: Then, the encoded data is fused using an adaptive channel attention mechanism to obtain fused data, including: At each moment, the multi-channel spatiotemporal codes of each traffic flow in the encoded data are stacked into one code, and then the code is transformed into a shape for matrix operation, and the matrix multiplication result of the SoftMax layer is used to obtain a channel attention map; Use an average pooling layer to extract the overall channel weight; The channel attention map is used to weight-enhance the multi-channel spatiotemporal coding, and the overall weight is used to fuse the multi-channel coding to obtain fused data.

6. The method for predicting passenger flow at a transportation hub considering multiple distribution modes according to claim 1, characterized in that: The external factor data include: Holidays, weather conditions and time information.

7. A passenger flow prediction device for a transportation hub considering multiple distribution modes, characterized in that: include: The data collection module is used to collect multi-mode passenger flow data and external factor data, pre-process the data, and obtain the original data; A model building module is used to construct a prediction model, wherein the prediction model includes a multi-channel time series enhancement module, a spatiotemporal coding and channel fusion module, and an optimization and prediction module; the multi-channel time series enhancement module enhances the time series features of the original data through multiple channels; In the spatiotemporal coding and channel fusion module, a spatiotemporal dynamic graph convolutional recurrent network is obtained based on the gated recurrent unit and the dynamic graph convolution construction, wherein the matrix product in the gated recurrent unit is replaced by the dynamic graph convolution operation, and then the original data enhanced by the time series feature is input into the spatiotemporal dynamic graph convolutional recurrent network to obtain the coded data, and then the coded data is fused by an adaptive channel attention mechanism to obtain the fused data; the optimization and prediction module converts the external factor data into a continuous vector representation, designs the external information encoding corresponding to the passenger flow data of each traffic mode, uses position encoding to represent the order of the fused data, obtains auxiliary prediction data, defines each traffic hub in the fused data as a multi-modal traffic graph network, and abstracts each traffic flow as a node. Combined with the auxiliary prediction data, the feature representation of each node in the time dimension and feature dimension is projected by the self-attention mechanism to obtain the passenger flow distribution prediction values ​​of multiple traffic modes; A model training module, used for training the prediction model based on the multi-modal passenger flow data and external factor data, based on a preset loss function and training settings; The prediction module is used to input the multi-mode passenger flow data and external factor data into the trained prediction model to obtain the passenger flow distribution prediction values ​​of various transportation modes in the transportation hub.

8. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method for predicting passenger flow at a transportation hub considering multiple collection and distribution modes as described in any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a program, wherein when the program is executed by a processor, the method for predicting passenger flow at a transportation hub considering multiple collection and distribution modes as described in any one of claims 1 to 6 is implemented.

10. A computer program product comprising computer instructions, characterized in that When executed by a processor, the computer instructions implement the steps of the method for predicting passenger flow at a transportation hub taking into account multiple collection and distribution modes as described in claims 1 to 6.

Images