Prediction method and device

Abstract

The invention discloses a prediction method and device, relates to the field of data processing, and is used for solving the problem that the passenger flow volume of a set place cannot be accurately predicted. The method comprises the steps that first information is acquired, the first information comprises historical passenger flow information of a plurality of set places, non-passenger flow information of the plurality of set places and an incidence relation among the plurality of set places, and the incidence relation comprises a physical connection relation between any two of the plurality of set places; and using the first information to train a preset model, and using the trained preset model to predict the future passenger flow volume of the plurality of set places. Through the scheme, the passenger flow volume of the set place can be accurately predicted.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and in particular to a prediction method and device. Background Art

[0002] With the rapid economic growth, population increase and urban modernization, urban traffic congestion is becoming increasingly serious, and urban public transportation can alleviate traffic congestion.

[0003] However, with the rapid development of rail transit, the number of passengers has increased dramatically, and many urban subway systems are overloaded, posing huge safety hazards. If the subway passenger flow is accurately predicted, traffic congestion can be alleviated and the safety of travelers can be guaranteed. For example, by predicting the subway passenger flow, managers can formulate scientific operation plans, reasonably dispatch vehicles, improve subway operation efficiency, and provide reference information for future subway construction (such as determining the basic scale of new subway stations, platform length and crowd evacuation capacity), and can also fundamentally eliminate safety hazards in subway stations.

[0004] Therefore, how to accurately predict the passenger flow of the subway is an urgent problem that needs to be solved. Summary of the invention

[0005] The present invention provides a prediction method and device for accurately predicting the passenger flow at a set location (such as a subway).

[0006] In a first aspect, the present invention provides a prediction method, which includes: obtaining first information, the first information including historical passenger flow information of multiple set locations, non-passenger flow information of multiple set locations, and an association relationship between the multiple set locations, the association relationship including a physical connection relationship between any two of the multiple set locations, using the first information to train a preset model, and using the trained preset model to predict future passenger flow at the multiple set locations.

[0007] Through the above scheme, a topological structure is constructed according to the association relationship of the set locations. The structure includes both the passenger flow information of the set locations and the non-passenger flow information, as well as the physical connection relationship of each set location. The future passenger flow of the set locations is predicted through a preset model. In this way, the model can capture more information beyond passenger flow information, effectively improve the model prediction performance, and make the model prediction more accurate.

[0008] Optionally, the non-passenger flow information includes one or more of the following information: temperature information, weather information, time information, holiday information, and social activity information.

[0009] Through the above scheme, by integrating information such as time, weather, geographical location, temperature, etc., in the process of passenger flow prediction, not only the passenger flow information is referred to, but also other factors affecting the passenger flow are referred to, making the prediction more accurate.

[0010] Optionally, the preset model includes two channels, the two channels have the same model structure and different model parameters, and the preset model is trained using the first information, including: inputting the first information into the two channels of the preset model respectively, and obtaining two loss values ​​obtained by the two channels using their respective model parameters to process the first information, calculating the sum of the two loss values, and determining the optimal model parameters of the two channels based on the principle of minimizing the sum, and using the trained preset model to predict the future passenger flow of multiple set locations, including: using one channel of the trained preset model to predict the future passenger flow of multiple set locations.

[0011] Through the above steps, by designing the dual structure, the model can effectively learn different feature representations. By calculating the loss function, the model can constrain the two channels to each other during the training process, prompting the two channels to learn from each other and acquire features. This complementarity helps to improve the prediction performance and significantly improve the robustness and prediction accuracy of the model.

[0012] Optionally, for each channel, including a convolution module, a Fourier transform module and an encoding module, the first information is input into each channel of a preset model, and two loss values ​​obtained by the two channels using their respective model parameters to process the first information are obtained, including: inputting the first information into the convolution module of the channel, the convolution module of the channel is used to extract spatial feature information in the first information through a graph convolutional neural network; inputting the spatial feature information into the Fourier transform module of the channel, the Fourier transform module of the channel is used to filter noise data in the spatial feature information; inputting the filtered spatial feature information into the encoding module of the channel, the encoding module of the channel is used to calculate the position feature information in the first information, and using the loss function to calculate the loss value of each channel.

[0013] Through the above steps, the convolution module is used to extract spatial feature information to capture the spatial relationship between the set locations, and then the Fourier transform module is used to filter the noise data to improve the data quality and remove unnecessary interference information. Finally, the encoding module is used to calculate the position feature information to help the model understand the position-related information in the data. In this way, the spatial characteristics, data cleaning and location information are comprehensively considered, which helps to improve the accuracy and reliability of subway passenger flow prediction and improve the overall performance of the model by reducing the impact of noise data.

[0014] Optionally, the spatial feature information is input into a Fourier transform module, including: performing Fourier transform on the spatial feature information to obtain a first implicit feature, and performing inverse Fourier transform on the first implicit feature using a second weight value to obtain a second implicit feature, wherein the second implicit feature is the spatial feature information after noise is removed.

[0015] Through the above steps, implicit features can be decomposed and captured through Fourier transform and inverse transform, and the noise data in the additional information can be filtered out, retaining the features that are important for passenger flow, so that the corresponding weights of these features are larger and the weights of non-essential features are smaller.

[0016] Optionally, obtaining the first information includes: determining multiple set locations within the area to be predicted, determining multiple points of interest corresponding to each of the multiple set locations, and obtaining historical passenger flow information and non-passenger flow information of each set location and the multiple points of interest, and assigning second weight values ​​to the multiple points of interest according to the degree of influence of the multiple points of interest on the passenger flow information of each set location; wherein the multiple points of interest are other locations within the set area near each set location that affect the passenger flow of each set location, and according to the second weight values ​​of the respective points of interest corresponding to the multiple set locations, feature aggregation is performed on the historical passenger flow information and non-passenger flow information of the multiple set locations and the respective points of interest to obtain the first information, and the feature aggregation is used to indicate the spatial characteristics of the first information.

[0017] Through the above scheme, we not only pay attention to the passenger flow information and non-passenger flow information of the set location, but also pay attention to the passenger flow and non-passenger flow information of the points of interest that affect the passenger flow of the set location, such as supermarkets, shopping malls, parks, schools, scenic spots, etc. Taking into account the influence of points of interest on the passenger flow of the set location, the passenger flow of the set location is predicted more accurately. By assigning different weights to different nodes, the features of the node can be aggregated with the features of its neighboring nodes, and then the relationship between the nodes can be better captured in the graph convolutional network. After feature aggregation, the initial feature matrix of the next step of the graph convolutional network can be obtained.

[0018] In a second aspect, the present invention provides a prediction device, the device comprising:

[0019] an acquisition module, configured to acquire first information, wherein the first information includes historical passenger flow information of multiple set locations, non-passenger flow information of multiple set locations, and associations between the multiple set locations, wherein the associations include a physical connection relationship between any two of the multiple set locations;

[0020] A training module, used for training a preset model using the first information;

[0021] The prediction module is used to predict the future passenger flow of multiple set locations using the trained preset model.

[0022] In a possible implementation, the non-passenger flow information includes one or more of the following information: temperature information, weather information, time information, holiday information, and social activity information.

[0023] In one possible implementation, the preset model includes two channels, the two channels have the same model structure and different model parameters, and the training module is specifically used to: input the first information into the two channels of the preset model respectively, and obtain two loss values ​​obtained by the two channels using their respective model parameters to process the first information, calculate the sum of the two loss values, determine the optimal model parameters of the two channels based on the principle of minimizing the sum, and use the trained preset model to predict the future passenger flow of multiple set locations, including: using one channel of the trained preset model to predict the future passenger flow of multiple set locations.

[0024] In one possible implementation, for each channel, including a convolution module, a Fourier transform module and an encoding module, the training module is specifically used to: input the first information into the convolution module of the channel, and the convolution module of the channel is used to extract the spatial feature information in the first information through a graph convolutional neural network; input the spatial feature information into the Fourier transform module of the channel, and the Fourier transform module of the channel is used to filter the noise data in the spatial feature information; input the filtered spatial feature information into the encoding module of the channel, and the encoding module of the channel is used to calculate the position feature information in the first information, and use the loss function to calculate the loss value of each channel.

[0025] In one possible implementation, the training module is specifically used to: perform Fourier transform on the spatial feature information to obtain a first implicit feature, and use a second weight value to perform inverse Fourier transform on the first implicit feature to obtain a second implicit feature, where the second implicit feature is the spatial feature information after noise is removed.

[0026] In one possible implementation, the acquisition module is specifically used to: determine multiple set locations within the area to be predicted, determine multiple points of interest corresponding to each of the multiple set locations for each of the multiple set locations, and obtain historical passenger flow information and non-passenger flow information for each set location and the multiple points of interest, and assign second weight values ​​to the multiple points of interest according to the degree of influence of the multiple points of interest on the passenger flow information of each set location; wherein the multiple points of interest are other locations within the set area near each set location that affect the passenger flow of each set location, and according to the second weight values ​​of the respective points of interest corresponding to the multiple set locations, feature aggregation is performed on the historical passenger flow information and non-passenger flow information of the multiple set locations and the respective points of interest to obtain first information, and the feature aggregation is used to indicate the spatial characteristics of the first information.

[0027] In a third aspect, the present application also provides a prediction device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the method described in the various possible designs of the first aspect.

[0028] In a fourth aspect, the present application also provides a computer-readable storage medium, on which a computer program or instruction is stored. When the computer program or instruction is executed by a processor, the method described in the various possible designs of the first aspect is implemented.

[0029] In a fifth aspect, the present application also provides a computer program product, which, when executed on a computer, enables the computer to execute any of the methods described in the first aspect above.

[0030] These implementations or other implementations of the present application will be more concise and understandable in the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0032] Figure 1 A flowchart of a prediction method provided by an embodiment of the present invention;

[0033] Figure 2 A schematic diagram of constructing a subway network topology map provided by an embodiment of the present invention;

[0034] Figure 3 A schematic diagram of a process of using first information to train a preset model provided by an embodiment of the present invention;

[0035] Figure 4 A schematic diagram of the principles of Fourier transform and inverse transform provided by an embodiment of the present invention;

[0036] Figure 5 A schematic diagram of the structure of a prediction model for training a model provided by an embodiment of the present invention;

[0037] Figure 6 A schematic diagram of the structure of a preset model for passenger flow prediction provided by an embodiment of the present invention;

[0038] Figure 7 A schematic diagram of a flow chart of another prediction method provided by an embodiment of the present invention;

[0039] Figure 8 A schematic diagram of a prediction device provided by an embodiment of the present invention;

[0040] Fig. 9 A schematic diagram of another prediction device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0041] In order to make the purpose, technical solution and beneficial effects of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0042] Below, some terms in this application are explained. It should be noted that these explanations are for the convenience of understanding by those skilled in the art and do not constitute a limitation on the scope of protection claimed in this application.

[0043] 1. Graph convolutional networks (GCN)

[0044] GCN is a deep learning model specifically designed to process graph-structured data. GCN extends the concept of convolutional neural networks to graph data and learns the representation of nodes by aggregating and updating the information of nodes and their neighbors. Compared with traditional deep learning models, GCN can better handle non-Euclidean data, such as complex topological structures such as transportation networks.

[0045] 2. Activation Function

[0046] The activation function is a key component in neural networks, responsible for introducing nonlinear characteristics that enable neural networks to learn and represent complex functions. Without the activation function, each layer of the neural network is just a linear transformation and cannot capture the complex patterns and features in the data.

[0047] 3. Multi-head self-attention network

[0048] The multi-head self-attention network is a powerful mechanism in deep learning, which is widely used in natural language processing and other sequence modeling tasks, especially in the Transformer model. Its main purpose is to enable the model to focus on different parts of the input sequence when processing sequence data, so as to capture contextual information and long-term dependencies. It mainly includes the self-attention mechanism and the multi-head mechanism. Among them, the self-attention mechanism allows the model to dynamically calculate the relationship between each element of the input sequence and other elements in the sequence when processing it. The multi-head mechanism expands the self-attention into multiple parallel self-attention calculations.

[0049] Among them, the self-attention mechanism includes the query matrix Q, the key matrix K, and the value matrix V. The explanations of these three matrices are as follows:

[0050] 1. Query matrix (Q):

[0051] The query matrix is ​​generated from the input features and is used to represent the features of the current node. The feature vector of each node is transformed through a linear transformation (usually a weight matrix) to generate a query vector. The query vector is used to calculate the attention weights between nodes.

[0052] 2. Key matrix (K):

[0053] The key matrix is ​​also generated from the input features to represent the features of all neighboring nodes. The feature vector of each node is transformed into a key vector through a linear transformation. The similarity between the key vector and the query vector is used to calculate the attention weight.

[0054] 3. Value matrix (V):

[0055] The value matrix is ​​also generated from the input features to represent the actual feature information of the node. The feature vector of each node is transformed into a value vector through a linear transformation. The value vector is used when calculating the attention weighted average and is ultimately used to update the node features.

[0056] 4. Residual connection.

[0057] Residual connection is a technology widely used in deep neural networks. It is mainly used to solve the problems of gradient vanishing and gradient exploding in the training process of deep neural networks. The basic idea of ​​residual connection is to introduce a "short-circuit" path to directly add the input to the output. This method can effectively alleviate the training difficulties caused by the increase in the number of network layers.

[0058] 5. Layer normalization.

[0059] Layer normalization is an effective normalization technique. Its core idea is to normalize the activation values ​​of each layer so that they have zero mean and unit variance, which can improve the stability and convergence speed of model training. Layer normalization has become an important component of modern deep learning models and is widely used in various tasks and fields.

[0060] 6. Pairwise mutual reinforcement graph convolutional network (PMR-GCN)

[0061] PMR-GCN is a deep learning model designed specifically for graph data structures, aiming to better capture the complex relationships between nodes. The core idea of ​​PMR-GCN is to enhance the feature representation between nodes through a dual structure. This dual structure allows nodes to reinforce each other's features, thereby capturing richer graph information. Specifically, PMR-GCN usually contains two interacting sub-networks, each of which can enhance the feature representation of the other sub-network.

[0062] 7. Kullback-Leibler divergence (KL divergence)

[0063] KL divergence is an important concept in information theory, which is used to measure the difference between two probability distributions. In machine learning, KL divergence can be used to evaluate how well the model distribution fits the true distribution. The smaller the KL divergence, the closer the model distribution is to the true distribution.

[0064] 8. Fully connected layer (FC Layer)

[0065] A fully connected layer is a basic layer type in a neural network, where each neuron is connected to all neurons in the previous layer. The function of a fully connected layer is to map input features to the output space. It is usually used as the last layer for classification tasks, but it is also applicable to regression tasks.

[0066] As described in the background art, in the field of transportation, accurate prediction of subway passenger flow can alleviate traffic congestion and ensure the safety of travelers.

[0067] At present, deep learning technology is mainly used to predict passenger flow. Since traffic data is typical time series data, obtaining time features is crucial for traffic prediction, and recurrent neural network (RNN) has the ability to model time series data. Therefore, RNN has become an important part of the traffic flow prediction model.

[0068] However, the current mainstream RNN has the problem of gradient vanishing and explosion, and it is difficult to process long sequence data. Based on this, some studies consider treating traffic data as regular images and using convolutional neural networks (CNN) and geographic attributes to predict passenger flow. However, the existing CNN has limitations when dealing with traffic networks with complex topological structures and cannot fully capture spatial correlations. Therefore, in order to solve the technical problems of CNN, researchers have extended the deep learning model to graph data to use graph convolutional networks (GCN) to represent the relationship between irregular traffic data.

[0069] However, the conventional GCN model defines a smoothing filter in the Fourier domain, and captures spatial features by continuously moving the smoothing filter. However, smoothing filters have a blurring effect, that is, they can eliminate or weaken the high-frequency part while retaining the low-frequency part. This process causes the overall prediction results of the GCN model to change less. In this way, when the GCN model is used to predict traffic data, the peak of the predicted passenger flow data will be smoother, resulting in the peak of the predicted value not being obvious, thus affecting the accuracy of the prediction.

[0070] To solve the above problems, the industry has provided a solution that uses a learnable diagonal matrix to alleviate the negative impact of the smoothing filter, that is, except for the elements on the main diagonal, all other elements in the diagonal matrix are zero. In this way, the elements on the main diagonal of the diagonal matrix can be regarded as the weights of the nodes. By adjusting these weights, the importance or contribution of the nodes in the model can be affected.

[0071] By adopting a learnable diagonal matrix, the model can learn the importance of each node and adjust the node weights accordingly. However, this solution ignores many factors that affect passenger flow data, such as weather, time, social hot spots, and geographical location. For example, bad weather will reduce passenger flow; passenger flow increases during peak hours on weekdays, while it is relatively average on weekends and holidays. Major events (such as concerts, city marathons, sports events, etc.) will cause significant passenger flow peaks, and stations near large shopping malls or hospitals usually have higher passenger flow.

[0072] In addition, with the popularization of sensor technology and smart card readers, traffic data has exploded. During the data collection process, hardware or software errors may also generate noisy data. The existence of noisy data stems from a variety of factors, such as sensor failures, data collection errors, environmental changes, and emergencies, which together lead to random fluctuations in data and the emergence of outliers. If not processed, noisy data will not only reduce the prediction accuracy of the model, but also increase the complexity of the model, affect the generalization ability of the model, and thus weaken its effectiveness in practical applications. However, the above solutions provided by the industry have also failed to effectively solve the impact of noise data on the robustness of the model.

[0073] Based on this, the present invention provides a prediction method, which accurately predicts the passenger flow of a set location by using a preset model to process passenger flow information and non-passenger flow information.

[0074] The present application scheme is described in detail below in conjunction with the accompanying drawings.

[0075] See also Figure 1 , a flow chart of a prediction method is shown, which specifically includes the following steps:

[0076] Step 110: Obtain first information.

[0077] The first information includes historical passenger flow information of multiple set locations, non-passenger flow information of multiple set locations, and association relationships between multiple set locations, and the association relationship includes a physical connection relationship between any two set locations among the multiple set locations.

[0078] Optionally, the non-passenger flow information may be other information that affects passenger flow, such as one or more of temperature information, weather information, time information, holiday information, and social activity information. Thus, by integrating information such as time, weather, geographic location, and temperature, in the process of passenger flow prediction, not only the passenger flow information but also other factors that affect passenger flow are referenced, making the prediction more accurate.

[0079] Optionally, multiple set locations within the area to be predicted can also be determined, and for each of the multiple set locations, multiple points of interest corresponding to each set location can be determined, and historical passenger flow information and non-passenger flow information of each set location and the multiple points of interest can be obtained. The connection relationship between each set location and the multiple points of interest corresponding to each set location, as well as the connection relationship between the set locations can also be obtained, wherein the points of interest are other locations within the set area near the set location that affect the passenger flow of the set location.

[0080] Through the above scheme, we not only pay attention to the passenger flow and non-passenger flow information of the set location, but also pay attention to the passenger flow and non-passenger flow information of the points of interest that affect the passenger flow of the set location, such as supermarkets, shopping malls, parks, schools, scenic spots, etc. Taking into account the impact of points of interest on the passenger flow of the set location, the passenger flow of the set location is predicted more accurately.

[0081] Optionally, a set location network can be defined as G = (V, E) based on the first information of multiple set locations, where V is a node set, representing the set locations (such as subway stations) and nearby points of interest (for example: hospitals, shopping malls, parks, schools, scenic spots, etc.); E is an edge set, representing the connection relationship between nodes, that is, the connection between subway stations and subway stations, and the connection between subway stations and nearby points of interest.

[0082] Optionally, the node set V may include multiple nodes v, each node v represents a subway station or a point of interest. That is, the node set of the set location can be expressed as: V = {v i |i∈[1,N]}, where N is the number of set locations (such as subway stations).

[0083] Optionally, if the node v is a subway station, then for each subway station v∈V, its feature vector Xv can be defined to contain the following information:

[0084] Historical passenger flow x history : Indicates the passenger flow data of the site in the past period of time;

[0085] Temperature x temperature : Indicates the temperature information of the area where the station is located;

[0086] Weather conditionsx weather : Indicates the weather conditions in the area where the station is located, such as sunny, rainy, etc.;

[0087] Time x time : Indicates time information, which can be hours, days of the week, etc.;

[0088] Is it a holiday? holiday : Indicates whether the current time is a holiday;

[0089] Are there any social activities? event : Indicates whether there are any major social activities (such as concerts, exhibitions, etc.) at the current time.

[0090] Combine the feature vectors of all nodes into a feature matrix X, where the i-th row represents node v i The eigenvector of:

[0091]

[0092] Where N is the total number of nodes. For subway stations, the feature vector of each node v∈V can be expressed as a comprehensive feature vector Xv=[x history , x temperature , x weather , x time , x holiday , x event ], the comprehensive feature matrix X is:

[0093]

[0094] Optionally, if node v is an interest point, it can be understood that the interest point set V POI It can contain multiple points of interest. For each point of interest p∈V POI , its eigenvector Xp can be defined to contain the following information:

[0095] Passenger flow x POI_flow : Indicates the passenger flow of the POI in a specific time period.

[0096] Optionally, at time t, each node v i The characteristics of can be expressed as: i ∈R C , where C represents the number of features and R represents the set of real numbers, i.e., x_t i ∈R C represents a C-dimensional real vector space, where each dimension corresponds to an eigenvalue, and these eigenvalues ​​are all real numbers. For example, in the above example, the subway station v i If there are 6 features, then C is 6, and if the interest point p has 1 feature, then C is 1.

[0097] Optionally, the edge set can be represented as: E = {eij = (vi, vj) | i, j ∈ [1, N], i ≠ j}, if v i and v j There is an edge between them, that is, node v i Then and node v j If there is a physical route between them, then e ij The value of ∈E is 1.

[0098] Through the above steps, a spatiotemporal diagram of the subway network can be constructed to provide basic data support for traffic flow prediction. Figure 2 , showing a schematic diagram of constructing a subway network topology map, in which v1, v2, v3, v4, v5 and v6 are nodes of set locations, and the lines between v1, v2, v3, v4, v5 and v6 are physical connections between different nodes.

[0099] Optionally, after constructing the subway network, second weight values ​​may be assigned to multiple points of interest based on the degree of influence of the multiple points of interest on the passenger flow information of the set locations, and based on the second weight values ​​of the multiple points of interest, feature aggregation may be performed on the historical passenger flow information and non-passenger flow information of the multiple set locations and the multiple points of interest, wherein the feature aggregation is used to preset a model to determine the spatial characteristics of the first information.

[0100] It can be understood that points of interest are other places within a set area near a set location that affect the passenger flow of the set location, that is, points of interest also include other set locations adjacent to the set location. For example, if the set location is a subway station, the points of interest include not only hospitals, shopping malls, parks, schools, scenic spots, etc., but also other subway stations connected to the subway station, such as the previous subway station or the next subway station.

[0101] Optionally, the second weight value may be calculated by the following formula:

[0102]

[0103] Where a is a learnable parameter vector (learnable parameters refer to variables updated by the back-propagation algorithm during training, which are used to adjust the weights and biases of the neural network so that the neural network can better fit the training data and improve performance), W is a learnable weight matrix, ∥ is a vector concatenation operation, LeakyReLU is an activation function, and h is v and h u is the feature vector of node v and its neighbor node u.

[0104] Through the above steps, the Xv obtained in the above steps is history , x temperature , x weather , x time , x holiday , x event ] and Xp aggregate the features in a weighted summation manner through the attention mechanism. The aggregated subway station feature vector is:

[0105]

[0106] Among them, α vp is the attention weight of node v to node p.

[0107] Through the above scheme, by assigning different weights to different nodes, the features of the node can be aggregated with the features of its neighboring nodes, so that the relationship between nodes can be better captured in the graph convolutional network. After feature aggregation, X F As the initial feature matrix for the next step of graph convolutional network.

[0108] Step 120: Use the first information to train a preset model.

[0109] Optionally, in step 120, a portion of the data in step 110 can be used as a training set to train a preset model so that the model is continuously iterated and optimized, and another portion of the data can be used as a test set to evaluate the final performance of the preset model. In this way, overfitting of the model can be prevented and the trained model can be made more accurate.

[0110] Optionally, the preset model may be a graph convolutional neural network model, which may include a convolution module, a Fourier transform module and an encoding module.

[0111] See also Figure 3 , showing a schematic diagram of a process of using the first information to train a preset model, comprising the following steps:

[0112] Step 310: Input the first information into the convolution module.

[0113] By inputting the first information into the convolution module, spatial feature information in the first information can be extracted through a graph convolutional neural network.

[0114] Optionally, the first information may be graph convolved using the following formula:

[0115]

[0116] Among them, the initial feature matrix H (0) =Xv, Xv is the feature vector of node v; H l+1 is the output feature matrix of the lth layer and the input feature matrix of the l+1th layer. When l=0, H l =Xv;W l is the learnable parameter matrix of layer l; σ(·) is the nonlinear activation function, is the normalized adjacency matrix, The calculation formula is as follows:

[0117]

[0118] in, yes The degree matrix of is the adjacency matrix A plus the identity matrix I, and Λ is a learnable diagonal matrix.

[0119] Through the above steps, the normalized adjacency matrix is ​​multiplied by the feature matrix and then multiplied by the learnable parameter matrix W l , linearly transform the aggregated features. Finally, through the activation function σ, the feature representation H of the (l+1)th layer is obtained (l+1),In this way, the expressive power of the model can be enhanced.,Through the above formula, through multi-layer convolution operations, the features of neighbor nodes,layer by layer can be aggregated to extract higher-level feature representations.

[0120] Optionally, if it is a two-layer graph convolutional network, the graph convolution formula is:

[0121]

[0122] Among them, W (0) and W (1) are the learning parameter matrices of the first and second layers of the neural network, RELU is a nonlinear activation function, X ′ It is the time series data with spatial features output by the two-layer graph neural network model.

[0123] Through the above steps, the graph convolutional neural network can be used to effectively aggregate the feature information of adjacent subway stations, thereby extracting the potential spatial features in the dataset.

[0124] Step 320: Input the spatial feature information into the Fourier transform module.

[0125] The Fourier transform is a signal processing technique used to decompose a signal into its components at different frequencies. Figure 4 , shows a schematic diagram of the principle of Fourier transform and inverse transform. It can be seen from the figure that by decomposing the T×1 dimensional vector of each node through Fourier transform, the implicit information in the node features can be captured, and then these implicit features can be fused through inverse Fourier transform to obtain new features of the node. Through the process of Fourier transform and inverse transform, the model can capture important implicit features from the new features and filter out noise data.

[0126] Optionally, a weight matrix W can be introduced Fourier , which is used to control the importance of each implicit feature so that the model can better capture important features. For example, bad weather has a greater impact on passenger flow, and its corresponding weight matrix value is close to 1. For another example, the peak period on weekdays also has a great impact on passenger flow, and its corresponding weight matrix value is also close to 1.

[0127] That is, step 320 specifically includes the following steps:

[0128] Step a: Fourier transform: transform the feature vector of each node into the frequency domain to obtain the frequency domain feature representation. And use the weight matrix W Fourier Different frequency components are weighted to control the importance of each implicit feature.

[0129] Step b: Inverse Fourier transform: Convert the weighted frequency domain features back to the time domain to obtain a new feature representation of the node.

[0130] Through the above steps, since the matrix X output in step 310 ′ It contains various implicit features, which may be hidden in the data and difficult to observe directly. Through Fourier transform and inverse transform, these implicit features can be decomposed and captured, and the noise data in the additional information can be filtered out, retaining the features that are important for passenger flow, so that the corresponding weights of these features are larger and the weights of non-essential features are smaller.

[0131] Step 330: Input the filtered spatial feature information into the encoding module.

[0132] By inputting the filtered spatial feature information into the encoding module, the position feature information in the first information can be calculated.

[0133] Optionally, the following formula can be used for position encoding:

[0134]

[0135] Among them, p is the position of the data in the passenger flow data sequence, i is the dimension of the passenger flow data, and d model is the dimension of the position embedding.

[0136] The position code calculated by the above formula Contains the position information of the data in the sequence, which is used to represent the position characteristics of each data point. ′ With position encoding By adding them together, we can get new data X” containing location information and spatial features, that is:

[0137]

[0138] Through the above steps, since the multi-head self-attention network does not have the iterative operation of the recurrent neural network, it is necessary to provide the location information of the data. Position encoding can introduce position information for each position in the data sequence. By introducing position information, the model can recognize the order of input data to help the network process sequence data; combining position encoding with time series data with spatial features provides more complete information for the multi-head self-attention network to better process sequence data and improve the performance and generalization ability of the model.

[0139] Optionally, after step 330, it also includes: calculating the query matrix Q, the key matrix K and the value matrix V according to step 330, and then using these matrices to perform attention calculations, and processing data through residual connections and layer normalization operations. For details, please refer to the following steps 340 and 350.

[0140] Step 340: Calculate the passenger flow data based on the query matrix Q, the key matrix K and the value matrix V.

[0141] Specifically, the following steps are included: define three parameter matrices: W Q , W K , W V , use the above three parameter matrices to perform linear transformation on X', obtain the query matrix Q, key matrix K and value matrix V, and calculate a new passenger flow data matrix The calculation formula is as follows:

[0142]

[0143] where d k is the dimension of the input and softmax(·) is the activation function.

[0144] Step 350: Perform residual processing and normalization on the passenger flow data.

[0145] Will With X ″ Perform residual connection and layer normalization. The specific formula is as follows:

[0146]

[0147] Where LayerNormal() is a layer normalization function, and X' is the result of position encoding in step 330. Through attention calculation, the model can dynamically adjust the weight of the value according to the relationship between the query and the key to better focus on the important parts of the input data, and through residual connection and layer normalization, improve the training stability and convergence speed of the model, while retaining the information of the original features, so that the model can learn the representation of the data more effectively.

[0148] Step 360: Calculate the loss value using the loss function.

[0149] Optionally, the loss function can be a regularized loss function of KL divergence. By calculating the bidirectional KL divergence between the two output distributions (i.e., the KL divergence of A to B and the KL divergence of B to A), the models can constrain each other during the training process, prompting the two sub-models to learn from each other and acquire features. This complementarity helps to improve the prediction performance.

[0150] Through the above steps 310 to 360, the spatial feature information is first extracted through the convolution module to capture the spatial relationship between the set locations, and then the noise data is filtered through the Fourier transform module to improve the data quality and remove unnecessary interference information. Finally, the position feature information is calculated through the encoding module to help the model understand the position-related information in the data. In this way, the spatial characteristics, data cleaning and location information are comprehensively considered, which helps to improve the accuracy and reliability of subway passenger flow prediction and improve the overall performance of the model by reducing the impact of noise data.

[0151] Optionally, the preset model may include two channels, the two channels have the same model structure and different model parameters. In step 120, the first information is input into the two channels of the preset model, and two loss values ​​obtained by the two channels using their respective model parameters to process the first information are obtained, and the sum of the two loss values ​​is calculated. Based on the principle of minimizing the sum, the optimal model parameters of the two channels are determined.

[0152] For details, please refer to Figure 5 , showing the schematic diagram of the prediction model for training model, the model contains two channels, each channel contains three parts: PM-GCN, Fourier transform module, encoder module. The following introduces these three parts respectively:

[0153] 1. PM-GCN: It is used to process graph structure data and capture the relationship between subway stations. The two independent PM-GCN models in the upper and lower channels have the same structure but different parameters. This allows each model to learn different feature representations. Through these two channels, the model can generate two different output distributions. Through the above scheme, the expressiveness of the model can be enhanced, enabling it to capture richer information.

[0154] 2. Fourier transform module: used for frequency domain analysis. Combining Fourier transform with dual structure can capture implicit features in traffic data.

[0155] 3. Encoder module: used to encode the extracted features for subsequent prediction tasks.

[0156] Optionally, in order to eliminate the difference between the output distributions of the two channels, the loss function of the model can be calculated, and based on the principle of minimizing the sum of the loss functions, the parameters of the two channels can be adjusted to obtain the optimal parameters of the model.

[0157] Through the above steps, by designing the dual structure and regularization mechanism, the PMR-GCN model can effectively learn different feature representations. By combining these different feature extraction methods, the model can not only capture local graph structure information, but also understand the global time series characteristics, thereby significantly improving the robustness and prediction accuracy of the model.

[0158] Step 130: Use the trained preset model to predict future passenger flow at multiple set locations.

[0159] Optionally, in step 120, historical passenger flow information, historical non-passenger flow information and the connection relationship of nodes can be used to train a preset model, and in step 130, future non-passenger flow information and the connection relationship of nodes are input into the preset model to obtain future passenger flow information output by the preset model.

[0160] Optionally, in step 120, the optimal parameters of the two channel models during the training process are saved respectively, and they are verified using the test set. Finally, the parameters that perform best on the test set are selected as the final parameter set of the model. In this way, it can be ensured that the model has good generalization ability and can effectively predict and infer in real scenarios.

[0161] Optionally, in step 130, a channel of the trained preset model is used to predict the future passenger flow of multiple set locations. Figure 6 , showing the schematic diagram of the structure of the preset model for passenger flow prediction, the obtained optimal model parameters are loaded into the single channel model. The future passenger flow information output by the model can be obtained.

[0162] Optionally, a fully connected layer can be used to predict the data of the corresponding time in the output part of the model, that is, to predict the traffic flow data at different times in the future, for example, to predict the traffic flow in the next 15 minutes, 30 minutes, 45 minutes and 60 minutes. In this way, short-term (15 minutes) to long-term (60 minutes) prediction results can be obtained at the same time, which is helpful to formulate more comprehensive traffic scheduling and management strategies.

[0163] The above content introduces the method of predicting passenger flow. The following is another method of predicting passenger flow. Please refer to Figure 7 , a flow chart of another prediction method is shown, which specifically includes the following steps:

[0164] S1: Acquire first information and connection relationship of a set location and a point of interest.

[0165] S2: Perform feature aggregation through the second weight value.

[0166] Among them, steps S1-S2 correspond to the above-mentioned step 110 and will not be repeated here.

[0167] S3: Capturing spatial features through graph convolutional layer modules.

[0168] S4: Perform Fourier transform and inverse transform through the Fourier transform module.

[0169] S5: Position encoding is performed through the encoding module.

[0170] S6: Capturing temporal features.

[0171] Among them, steps S3-S6 correspond to the above-mentioned steps 310-350, which will not be repeated here.

[0172] S7: Capturing latent features via regularized dual structure.

[0173] S8: Obtain optimal model parameters.

[0174] Among them, steps S7-S8 correspond to the above steps 120 and Figure 5 , I will not go into details here.

[0175] S9: Output the final prediction result through the fully connected layer.

[0176] Wherein, step S9 corresponds to the above steps 130 and Figure 6 , I will not go into details here.

[0177] Through the above scheme, a topological structure is constructed according to the association relationship of the set locations. The structure includes both the passenger flow information of the set locations and the non-passenger flow information, as well as the physical connection relationship of each set location. The future passenger flow of the set locations is predicted through a preset model. In this way, the model can capture more information beyond passenger flow information, effectively improve the model prediction performance, and make the model prediction more accurate.

[0178] The above introduces the method for predicting passenger flow. The following is an example to illustrate the above method.

[0179] First, train the model, which includes the following steps:

[0180] Step 1: Get input features.

[0181] Assume that traffic flow is predicted by feature fusion, where the features include weather, passenger flow, time, holidays, and points of interest information. The input feature matrix is:

[0182]

[0183] Step 2: Get the weight matrix.

[0184] Assume that the first learnable weight matrix (learnable parameter) W is:

[0185]

[0186] Assume the initial weight is: w weather =1.0,w passenger =1.0,w time =1.0,wholiday =1.0,w POI =1.0.

[0187] Step 3: Fourier transform.

[0188] Perform Fourier transform on the feature matrix X to obtain the frequency domain feature matrix X f .

[0189] Assume X f The result is:

[0190]

[0191] Step 4: Perform weighted summation on the Fourier transformed matrix.

[0192] Apply the learnable weight matrix W to the frequency domain feature matrix X f , assuming that W is the unit matrix at the beginning, then the weighted frequency domain feature matrix X f ′ and X f The same, that is:

[0193] X f ′=X f ·W

[0194] Step 5: Inverse Fourier transform.

[0195] The weighted frequency domain feature X f ′ performs an inverse transformation to obtain the fused feature matrix X':

[0196]

[0197] Step 6: Calculate the loss function.

[0198] Step 7: Calculate the gradient of the loss function with respect to the weight matrix W and update the weight matrix according to the learning rate η.

[0199] The update formula is as follows:

[0200]

[0201] The new weight matrix will be used for weighted feature fusion in the next round of forward propagation. The model gradually learns the importance of each feature through iteration and optimizes the prediction results.

[0202] Through the above steps 1 to 7, the preset model can be trained, and then the trained preset model can be used to predict passenger flow. For details, please refer to the above step 130, which will not be repeated here.

[0203] Through the above scheme, a topological structure is constructed according to the association relationship of the set locations. The structure includes both the passenger flow information of the set locations and the non-passenger flow information, as well as the physical connection relationship of each set location. The future passenger flow of the set locations is predicted through a preset model. In this way, the model can capture more information beyond passenger flow information, effectively improve the model prediction performance, and make the model prediction more accurate.

[0204] Based on the same concept, an embodiment of the present application also provides a prediction device, which can execute the prediction method introduced in the above content.

[0205] See also Figure 8 , a schematic diagram of the structure of a prediction device provided in an embodiment of the present application is given, such as Figure 8 As shown, the prediction device comprises:

[0206] An acquisition module 801 is used to acquire first information, wherein the first information includes historical passenger flow information of multiple set locations, non-passenger flow information of multiple set locations, and associations between the multiple set locations, wherein the associations include a physical connection relationship between any two of the multiple set locations;

[0207] A training module 802, configured to train a preset model using the first information;

[0208] The prediction module 803 is used to predict the future passenger flow of multiple set locations using the trained preset model.

[0209] In a possible implementation, the non-passenger flow information includes one or more of the following information: temperature information, weather information, time information, holiday information, and social activity information.

[0210] In one possible implementation, the preset model includes two channels, and the two channels have the same model structure and different model parameters. The training module 802 is specifically used to: input the first information into the two channels of the preset model respectively, and obtain two loss values ​​obtained by the two channels using their respective model parameters to process the first information, calculate the sum of the two loss values, determine the optimal model parameters of the two channels based on the principle of minimizing the sum, and use the trained preset model to predict the future passenger flow of multiple set locations, including: using one channel of the trained preset model to predict the future passenger flow of multiple set locations.

[0211] In one possible implementation, for each channel, including a convolution module, a Fourier transform module and an encoding module, the training module 802 is specifically used to: input the first information into the convolution module of the channel, and the convolution module of the channel is used to extract the spatial feature information in the first information through a graph convolutional neural network; input the spatial feature information into the Fourier transform module of the channel, and the Fourier transform module of the channel is used to filter the noise data in the spatial feature information; input the filtered spatial feature information into the encoding module of the channel, and the encoding module of the channel is used to calculate the position feature information in the first information, and use the loss function to calculate the loss value of each channel.

[0212] In one possible implementation, the training module 802 is specifically used to: perform Fourier transform on the spatial feature information to obtain a first implicit feature, and use a second weight value to perform inverse Fourier transform on the first implicit feature to obtain a second implicit feature, where the second implicit feature is the spatial feature information after noise is removed.

[0213] In one possible implementation, the acquisition module 801 is specifically used to: determine multiple set locations within the area to be predicted, determine multiple points of interest corresponding to each of the multiple set locations for each of the multiple set locations, and obtain historical passenger flow information and non-passenger flow information for each set location and the multiple points of interest, and assign second weight values ​​to the multiple points of interest according to the degree of influence of the multiple points of interest on the passenger flow information of each set location; wherein the multiple points of interest are other locations within the set area near each set location that affect the passenger flow of each set location, and according to the second weight values ​​of the respective points of interest corresponding to the multiple set locations, feature aggregation is performed on the historical passenger flow information and non-passenger flow information of the multiple set locations and the respective points of interest to obtain first information, and the feature aggregation is used to indicate the spatial characteristics of the first information.

[0214] See also Fig. 9 , showing a schematic diagram of the structure of another prediction device provided in an embodiment of the present application, such as Fig. 9 As shown, the test device includes: a memory 901 and a processor 902, and the processor 902 is coupled to the memory 901. The memory 901 is used to store program instructions, and the processor 902 is used to call the program instructions stored in the memory 901 and execute the above prediction method according to the obtained program.

[0215] Optionally, the prediction device may further include an interface circuit 903, which may be a transceiver or an input / output interface. The input / output interface is used to input and / or output information, where output may be understood as sending and input may be understood as receiving. The processor 902 may communicate with other devices in the prediction device or other devices other than the prediction device through the interface circuit 903 to obtain information required for executing the above prediction method.

[0216] When the prediction device 900 is used to implement Figure 1 When performing the method shown, the processor 902 is used to implement the functions of the above-mentioned acquisition module 801, training module 802, and prediction module 803.

[0217] It is understandable that the processor in the embodiments of the present application may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, transistor logic devices, hardware components or any combination thereof. The general-purpose processor may be a microprocessor or any conventional processor.

[0218] The memory in the embodiments of the present application can be a random access memory, a flash memory, a read-only memory, a programmable read-only memory, an erasable programmable read-only memory, an electrically erasable programmable read-only memory, a register, a hard disk, a mobile hard disk, a compact disc read-only memory (CD-ROM) or any other form of storage medium known in the art. An exemplary storage medium is coupled to a processor so that the processor can read information from the storage medium and write information to the storage medium. The storage medium can also be a component of the processor.

[0219] Based on the same technical concept, an embodiment of the present invention further provides a computer-readable storage medium, on which a computer program or instruction is stored. When the computer program or instruction is executed by a processor, the computer executes the above prediction method.

[0220] Based on the same technical concept, an embodiment of the present invention further provides a computer-readable program product, which, when executed, enables a computer to execute the above prediction method.

[0221] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0222] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0223] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0224] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0225] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A prediction method, characterized in that: The method comprises: Acquire first information, wherein the first information includes historical passenger flow information of multiple set locations, non-passenger flow information of the multiple set locations, and associations between the multiple set locations, wherein the associations include a physical connection relationship between any two of the multiple set locations; The first information is used to train a preset model, and the trained preset model is used to predict future passenger flow at the plurality of set locations.

2. The method according to claim 1, characterized in that The non-passenger flow information includes one or more of the following information: Temperature information, weather information, time information, holiday information, and social activity information.

3. The method according to claim 1, characterized in that The preset model includes two channels, and the two channels have the same model structure and different model parameters; The using the first information to train a preset model includes: Inputting the first information into two channels of the preset model respectively, and obtaining two loss values ​​obtained by the two channels processing the first information using respective model parameters; Calculating the sum of the two loss values, and determining the optimal model parameters of the two channels based on the principle of minimizing the sum; The using the trained preset model to predict the future passenger flow of the plurality of set locations includes: A channel of the trained preset model is used to predict future passenger flow of the plurality of set locations.

4. The method according to claim 3, characterized in that For each of the channels: Includes convolution module, Fourier transform module and encoding module; Inputting the first information into each channel of the preset model, and obtaining two loss values ​​obtained by the two channels using respective model parameters to process the first information, comprising: Inputting the first information into a convolution module of the channel, wherein the convolution module of the channel is used to extract spatial feature information in the first information through a graph convolutional neural network; Inputting the spatial feature information into the Fourier transform module of the channel, the Fourier transform module of the channel is used to filter the noise data in the spatial feature information; Inputting the filtered spatial feature information into the encoding module of the channel, the encoding module of the channel is used to calculate the position feature information in the first information; The loss function is used to calculate the loss value of each channel.

5. The method according to claim 3, characterized in that The step of inputting the spatial feature information into the Fourier transform module comprises: Performing Fourier transform on the spatial feature information to obtain a first implicit feature; The first implicit feature is subjected to inverse Fourier transform using a first weight value to obtain a second implicit feature, where the second implicit feature is the spatial feature information after noise is removed.

6. The method according to claim 1, characterized in that The obtaining of the first information includes: Determine a plurality of set locations within the area to be predicted; For each of the multiple set locations, multiple points of interest corresponding to each set location are determined, and historical passenger flow information and non-passenger flow information of each set location and the multiple points of interest are obtained, and second weight values ​​are assigned to the multiple points of interest according to the degree of influence of the multiple points of interest on the passenger flow information of each set location; wherein the multiple points of interest are other locations in a set area near each set location that affect the passenger flow of each set location; According to the second weight values ​​of the respective points of interest corresponding to the plurality of set locations, feature aggregation is performed on the historical passenger flow information and non-passenger flow information of the plurality of set locations and the respective points of interest to obtain the first information, wherein the feature aggregation is used to indicate the spatial features of the first information.

7. A prediction device, characterized in that: include: an acquisition module, configured to acquire first information, wherein the first information includes historical passenger flow information of a plurality of set locations, non-passenger flow information of the plurality of set locations, and associations between the plurality of set locations, wherein the associations include a physical connection relationship between any two of the plurality of set locations; A training module, used for training a preset model using the first information; The prediction module is used to predict the future passenger flow of the plurality of set locations using the trained preset model.

8. A prediction device, characterized in that: include: A processor, wherein the processor is coupled to a memory, the memory is used to store a computer program or an instruction, and the processor is used to execute the computer program or the instruction to implement the method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that: It stores a computer program executable by a computer device, and when the program is run on the computer device, the computer device executes the steps of any method described in claims 1 to 6.

10. A computer program product, characterized in that When the method is executed on a computer, the computer is enabled to execute the method according to any one of claims 1 to 6.

Images