A crowd traffic big data prediction method for 5G needs

Abstract

The present invention belongs to the field of data analysis technology, and particularly relates to a method for predicting crowd flow big data for 5G needs, comprising dividing a to-be-predicted area into a plurality of small areas, regarding each small area as a node, obtaining node features and edge weight features between nodes; converting node features and edge weight features into vector representations, and inputting the vector representations together with node and edge position codes as input data into a graph neural network; constructing graph data through node and edge weight relationships, iteratively updating each node in the constructed graph data using a graph neural network, and obtaining a first feature vector of each node; inputting the first feature vector of each node into a graph attention network, and obtaining a second feature vector of each node; dividing data at a certain moment in the to-be-predicted area into three layers of data, namely, real-time data, periodic data, and abnormal data, processing the current layer of data through an enhanced memory adjustment network model, splicing the current layer of data with the next layer of data as the input of the next layer of enhanced memory adjustment network model, and outputting a third feature vector from the last layer of enhanced memory adjustment network model; splicing the second feature vector and the third feature vector together and inputting them into a fully connected layer, and obtaining a flow prediction for the current area; the present invention effectively improves the accuracy and efficiency of crowd flow prediction by comprehensively utilizing automated learning and advanced network models, and is particularly suitable for the big data environment of the 5G era.

Description

Technical Field

[0001] The present invention belongs to the field of data analysis technology, and in particular relates to a crowd flow big data prediction method oriented to 5G needs. Background Art

[0002] With the rapid development of mobile network technology, the amount of user data traffic has increased dramatically, which has promoted the development and commercialization of 5G technology. In particular, 5G networks can significantly improve network bandwidth and service quality through the dense deployment of small cellular networks and the application of mobile edge computing. However, how to improve the efficiency of numerous small base stations and give full play to the capabilities of edge computing has become a key issue in improving service quality.

[0003] In order to effectively manage and allocate base station resources, it is necessary to accurately predict the number of users within the coverage area of ​​the base station. Crowd flow prediction in specific functional areas (such as scenic spots, office areas, residential areas, etc.) will help to rationally deploy and optimize network resources. This prediction is not only beneficial to base station cache deployment and resource allocation, but also helps to formulate base station active / dormant switching plans, thereby improving network energy efficiency and reducing energy consumption. In addition, accurate crowd flow prediction also supports the dense deployment planning of 5G base stations, helps mobile network operators minimize capital expenditures, and provides decision-making support in urban planning, traffic management, etc. Therefore, how to accurately predict crowd flow in a specific area is an urgent problem to be solved in this field. Summary of the invention

[0004] In order to solve the problems existing in the above prior art, the present invention proposes a crowd flow big data prediction method for 5G demand, which specifically includes the following steps:

[0005] Divide the area to be predicted into multiple small areas, regard each small area as a node, and obtain the node features and edge weight features between nodes;

[0006] Convert node features and edge weight features into vector representations, and input the vector representations together with the node and edge position encodings into the graph neural network as input data;

[0007] Graph data is constructed through node and edge weight relationships, and each node in the constructed graph data is iteratively updated using a graph neural network to obtain the first eigenvector of each node;

[0008] Input the first feature vector of each node into the graph attention network to obtain the second feature vector of each node;

[0009] The data at a certain moment in the area to be predicted is divided into three layers of data: real-time data, periodic data and abnormal data. In the present invention, real-time data refers to data within a certain time range before the prediction time point. Preferably, data within a time range of 30 to 60 minutes before the time to be predicted is selected in the present invention. Periodic data refers to data that is transformed with a fixed period from historically collected data. Abnormal data refers to unconventional data. Such unconventional data may lead to a large change compared with historical trend data. In the present invention, it may refer to activities held in large public places in a certain area, including large-scale concerts and other activities. Such activities may lead to a large change in the flow of people and vehicles in the area compared with the historical data of the same period. The current layer of data is processed by an enhanced memory regulation network (EMRN) model and then spliced ​​with the next layer of data as the input of the next layer of enhanced memory regulation network model. The last layer of enhanced memory regulation network model outputs a third eigenvector.

[0010] The second eigenvector and the third eigenvector are concatenated together and input into the fully connected layer to obtain the traffic prediction of the current area.

[0011] Furthermore, the node characteristics include environmental characteristics of the area corresponding to the node, and the environmental characteristics include at least weather conditions, air quality, economic indicators, and commuting patterns in the current area. The weather conditions include at least temperature, humidity, wind speed, and precipitation. The air quality includes at least PM2.5 index, PM10 index, and NO2 index. The economic indicators include at least regional per capita GDP. The commuting patterns include working days and non-working days, holidays and non-holidays. The edge weight characteristics include at least vehicle flow and crowd flow between two nodes.

[0012] Furthermore, the node features and edge weight features are converted into vector representations through the embedding layer. The conversion process includes:

[0013] Embed(f i )=V f ×f i +b f

[0014] Embed(e ij )=V e ×e ij +b e

[0015] Among them, Embed(f i ) is the i-th node feature f i The corresponding vector representation, V f is the embedding matrix of node features, b f is the bias term corresponding to the embedding of node features; Embed(eij ) is the edge weight feature e between the i-th node and the j-th node ij The corresponding vector representation, V e is the embedding matrix of edge weight features, b e is the bias term corresponding to the embedding of edge weight features.

[0016] Furthermore, the process of position encoding the node features and edge weight features represented by the vector includes:

[0017]

[0018] Among them, PE(pos,k) represents the position code corresponding to the element with serial number pos in the vector sequence; pos represents the serial number of the data point in the vector sequence; d represents the dimension of the vector sequence; k is the index of the position code.

[0019] Furthermore, the graph data G constructed by the node and edge weight relationship is expressed as G=(V,E), where V is the set of all nodes and E is the set of edge relationships between all nodes.

[0020] Furthermore, the graph neural network is used to iteratively update each node in the constructed graph data, and the feature vector output by the last iteration is used as the first feature vector. The iterative process is:

[0021]

[0022] in, is the iterative feature vector corresponding to the i-th node at the t+1th iteration; σ(·) represents the ReLU function; W (t) is the weight matrix at the tth iteration, b (t) is the bias term at the tth iteration; c ij is a normalization constant; N(i) is the set of neighbor nodes of the i-th node; It represents the iterative feature vector corresponding to the jth node at the tth iteration. When t = 0, that is, The node features and edge weight features corresponding to the nodes and edges are converted into vector representations and concatenated with their position encodings.

[0023] Furthermore, the process of obtaining the second feature vector of the node through the graph attention network includes:

[0024] h i '=σ(∑ j∈N(i)∪{i} α ij W j )

[0025] Among them, h i' is the second eigenvector of the i-th node, σ(·) represents the ReLU function; N(i) is the set of neighbor nodes of the i-th node; α ij is the attention coefficient of the i-th node to the j-th node calculated by the attention mechanism; W is the feature conversion matrix; h j is the first eigenvector of the jth node.

[0026] Furthermore, the data at a certain moment in the area to be predicted is divided into three layers of data: real-time data, periodic data, and abnormal data. The process of processing the three layers of data using the EMRN model includes:

[0027] The real-time data realtime As the first layer of data, the real-time data is input into the EMRN model, which is expressed as:

[0028]

[0029] The output of the EMRN model based on the first layer of data With period data D cyclical After splicing, we get the input data D of the second layer. cyclical ', D cyclical 'As the input of the EMRN model, it is expressed as:

[0030]

[0031] The output of the EMRN model based on the second layer data With abnormal data D anomaly After splicing, we get the input data D of the third layer. anomaly ', D anomaly 'As the input of the EMRN model, it is expressed as:

[0032]

[0033] Will as the third eigenvector;

[0034] Where EMRN(·) represents the EMRN model.

[0035] Furthermore, the EMRN model processes data in the following ways:

[0036] i t =σ(W xi x t +W hi h t-1 +W ci c t-1 +b i )

[0037] ft =σ(W xf x t +W hf h t-1 +W cf c t-1 +b f )

[0038] o t =σ(W xo x t +W ho h t-1 +W co c t-1 +b o )

[0039] g t =tanh(W xg x t +W hg h t-1 +b g )

[0040] c t =f t ×c t-1 +i t ×g t ×tanh(W xc x t +W hc h t-1 +b c )

[0041] h t =o t ×tanh(c t )

[0042] Among them, i t represents the variable that controls the contribution of the input at time t to the state update, σ(·) represents the ReLU function, and W xi , W hi , W ci i t The weight matrix of the input vector at the current moment, the hidden state at the previous moment, and the unit state at the previous moment during the update process, x t represents the input vector at time t, h t-1 represents the hidden state at time t-1, c t-1 represents the unit state vector at time t-1, b i for i t The corresponding bias term during the update process; f t Indicates that control keeps c at time t t-1 The number of variables, W xf , Whf , W cf f t During the update process, the weight matrix of the input vector at the current moment, the hidden state at the previous moment, and the unit state at the previous moment, b f f t The corresponding bias term during the update process; o t represents the variable that controls the output from the cell state to the hidden state at time t, W xo , W ho , W co Respectively o t During the update process, the weight matrix of the input vector at the current moment, the hidden state at the previous moment, and the unit state at the previous moment, b o for o t The corresponding bias term during the update process; g t is an additional control layer variable at time t, W xg , W hg g t During the update process, the weight matrix of the input vector at the current moment and the hidden state at the previous moment, b g g t The corresponding bias term in the update process; tanh(·) is the hyperbolic tangent function, W xc , W hc c t During the update process, the weight matrix of the input vector at the current moment and the hidden state at the previous moment, b c c t The corresponding bias term during the update process; h t is the output of the EMRN model at time t.

[0043] Furthermore, the second eigenvector and the third eigenvector are concatenated and input into the fully connected layer to obtain the traffic prediction of the current area, including:

[0044] P=σ(W p O F +b p )

[0045] Where P is the final traffic prediction; σ(·) represents the ReLU function; W p is the weight matrix of the fully connected layer, b p is the bias term of the fully connected layer; O F It is the feature representation obtained by concatenating the second eigenvector and the third eigenvector.

[0046] The present invention adopts a deep spatiotemporal model to model spatiotemporal dependencies of different scales by fusing various external feature information; at the same time, the use of position coding also enhances the model's sensitivity to the temporal position in time series data, enabling the model to more accurately capture the impact of temporal changes; in addition, the present invention is particularly suitable for the big data environment of the 5G era, and can be widely used in urban planning, traffic management, public safety, commercial activity planning and other fields, helping relevant departments or enterprises to perform efficient traffic management and resource optimization in areas with high population density. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is a flow chart of the deep spatiotemporal model training process of the present invention;

[0048] Figure 2 It is a logic block diagram of the deep spatiotemporal model of the present invention. DETAILED DESCRIPTION

[0049] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0050] The present invention proposes a method for predicting crowd flow big data based on 5G demand, which specifically includes the following steps:

[0051] Divide the area to be predicted into multiple small areas, regard each small area as a node, and obtain the node features and edge weight features between nodes;

[0052] Convert node features and edge weight features into vector representations, and input the vector representations together with the node and edge position encodings into the graph neural network as input data;

[0053] Graph data is constructed through node and edge weight relationships, and each node in the constructed graph data is iteratively updated using a graph neural network to obtain the first eigenvector of each node;

[0054] Input the first feature vector of each node into the graph attention network to obtain the second feature vector of each node;

[0055] The data at a certain moment in the area to be predicted is divided into three layers of data: real-time data, periodic data, and abnormal data. The current layer of data is processed by the enhanced memory adjustment network model and then spliced ​​with the next layer of data as the input of the next layer of enhanced memory adjustment network model. The enhanced memory adjustment network model of the last layer outputs the third eigenvector;

[0056] The second eigenvector and the third eigenvector are concatenated together and input into the fully connected layer to obtain the traffic prediction of the current area.

[0057] like Figure 2 In this embodiment, the feature representation of nodes and nodes is obtained according to the environmental impact privacy, and then the feature representation of data is constructed according to the characteristics of time series data, and the two feature representations are fused through the fully connected layer to obtain the final prediction result. Next, this embodiment describes the present invention from three aspects: obtaining the feature representation of nodes and nodes, that is, the flow relationship graph features, the time series features of the data, and the fusion of the obtained features.

[0058] (I) Characteristics of flow relationship diagram

[0059] In this embodiment, the graph neural network GNN and the graph attention network GAT process the graph feature data to obtain the corresponding traffic relationship graph features.

[0060] In this embodiment, the process of obtaining the flow relationship graph includes the following steps:

[0061] 1) Obtain environmental influencing factors within the area to be predicted. Environmental influencing factors include factors within the area and factors between areas. Factors within the area are node features. In this embodiment, node features include weather conditions, air quality, economic indicators, and commuting modes in the current area. Weather conditions include at least temperature, humidity, wind speed, and precipitation. Air quality includes at least PM2.5 index, PM10 index, and NO2 index. Economic indicators include at least regional per capita GDP. Commuting modes include working days and non-working days, holidays and non-holidays. Factors between areas are edge weight features, including vehicle flow and crowd flow between two nodes.

[0062] 2) Obtain node features and edge weight features, obtain corresponding vector representations through embedding layer processing, and obtain corresponding position encodings based on the obtained vector representations;

[0063] Obtain node features and edge weight features and process them through the embedding layer to obtain the corresponding vector representation. The communication process includes:

[0064] Embed(f i )=V f ×f i +b f

[0065] Embed(e ij )=V e ×e ij +b e

[0066] Among them, Embed(f i ) is the i-th node feature fi The corresponding vector representation, V f is the embedding matrix of node features, b f is the bias term corresponding to the embedding of node features; Embed(e ij ) is the edge weight feature e between the i-th node and the j-th node ij The corresponding vector representation, V e is the embedding matrix of edge weight features, b e is the bias term corresponding to the embedding of edge weight features;

[0067] The process of position encoding the node features and edge weight features represented by the vector includes:

[0068]

[0069] Among them, PE(pos,k) represents the position code corresponding to the element with serial number pos in the vector sequence; pos represents the serial number of the data point in the vector sequence; d represents the dimension of the vector sequence; k is the index of the position code.

[0070] In particular, in this embodiment, a series of preprocessing and feature engineering processes are also applied to the data set through automatic machine learning (AutoML) technology. The preprocessing includes filling missing values, encoding non-numerical features, and generating feature cross-terms that may help improve prediction performance; the feature engineering process automatically tests a variety of feature selection algorithms and machine learning model combinations, and evaluates the prediction effect of each combination, including but not limited to decision trees, random forests, gradient boosting machines and other models, and their corresponding feature importance evaluation methods. The feature importance score is calculated during the model training process and reflects the contribution of each feature to the model prediction performance.

[0071] 3) Construct a traffic relationship graph through the acquired data, which is represented by the graph data G constructed through the node and edge weight relationship as G = (V, E), where V is the set of all nodes. In this embodiment, the i-th node v in the set i represents a specific area, carrying important characteristic information about the area, including but not limited to weather conditions, air quality, economic indicators and event data; E is the set of edge relations between all nodes, and the edge relations e in the set ij represents the i-th node v i With the jth node v j The connection relationship between the nodes is obtained; the node features, edge weight features and position encoding information of the i-th node are used as the graph data of the i-th node.

[0072] The graph data G is preprocessed using the graph neural network GNN to extract and learn the initial feature representation of the nodes, and the graph neural network GNN is used to iteratively update the features of each node through the following update rules:

[0073]

[0074] in, is the iterative feature vector corresponding to the i-th node at the t+1th iteration; σ(·) represents the ReLU function; W( t ) is the weight matrix at the tth iteration, b( t ) is the bias term at the tth iteration; c ij is a normalization constant; N(i) is the set of neighbor nodes of the i-th node; represents the iterative feature vector corresponding to the jth node at the tth iteration. When t = 0, It is the feature obtained by concatenating the node features through the embedding layer and the position encoding, that is, the graph data of the i-th node.

[0075] After being processed by the graph neural network GNN, the data flow enters the graph attention network GAT to further refine the node representation. The update rules of the graph attention network GAT are as follows:

[0076] h i '=σ(∑ j∈N(i)∪{i} α ij W j )

[0077] Among them, h i ' is the second eigenvector of the i-th node, α ij is the attention coefficient of the i-th node to the j-th node calculated by the attention mechanism; W is the feature conversion matrix; h j is the first eigenvector of the jth node.

[0078] (II) Timing characteristics

[0079] The data at a certain moment in the area to be predicted is divided into three layers of data: real-time data, periodic data, and abnormal data. The process of processing the three layers of data using the enhanced memory adjustment network model includes:

[0080] The real-time data realtime As the first layer of data, the real-time data is input into the EMRN model, which is expressed as:

[0081]

[0082] The output of the EMRN model based on the first layer of data With period data D cyclicalAfter splicing, we get the input data D of the second layer. cyclical ', D cyclical 'As the input of the EMRN model, it is expressed as:

[0083]

[0084] The output of the EMRN model based on the second layer data With abnormal data D anomaly After splicing, we get the input data D of the third layer. anomaly ', D anomaly 'As the input of the EMRN model, it is expressed as:

[0085]

[0086] Will as the third eigenvector;

[0087] Where EMRN(·) represents the EMRN model.

[0088] In this embodiment, the multidimensional time series data D is processed by the EMRN model. t , capturing long-term dependencies, t represents the temporal data of different layers, and the EMRN model is processed as follows:

[0089] i t =σ(W xi x t +W hi h t-1 +W ci c t-1 +b i )

[0090] Among them, i t The variable that controls the contribution of the input to the state update at time t, W xi , W hi , W ci i t The weight matrix of the input vector at the current moment, the hidden state at the previous moment, and the unit state at the previous moment during the update process, x t represents the input vector at time t, h t-1 represents the hidden state at time t-1, c t-1 represents the unit state vector at time t-1, b i for i t The corresponding bias term during the update process;

[0091] f t =σ(W xf x t +W hf h t-1 +Wcf c t-1 +b f )

[0092] Among them, f t Indicates that control keeps c at time t t-1 The number of variables, W xf , W hf , W cf f t During the update process, the weight matrix of the input vector at the current moment, the hidden state at the previous moment, and the unit state at the previous moment, b f f t The corresponding bias term during the update process;

[0093] o t =σ(W xo x t +W ho h t-1 +W co c t-1 +b o )

[0094] Among them, t represents the variable that controls the output from the cell state to the hidden state at time t, W xo , W ho , W co Respectively o t During the update process, the weight matrix of the input vector at the current moment, the hidden state at the previous moment, and the unit state at the previous moment, b o for o t The corresponding bias term during the update process;

[0095] g t =tanh(W xg x t +W hg h t-1 +b g )

[0096] Among them, g t is an additional control layer variable at time t, W xg , W hg g t During the update process, the weight matrix of the input vector at the current moment and the hidden state at the previous moment, b g g t The corresponding bias term in the update process; tanh(·) is the hyperbolic tangent function;

[0097] c t =f t ×c t-1 +i t×g t ×tanh(W xc x t +W hc h t-1 +b c )

[0098] Among them, W xc , W hc c t During the update process, the weight matrix of the input vector at the current moment and the hidden state at the previous moment, b c c t The corresponding bias term during the update process;

[0099] h t =o t ×tanh(c t )

[0100] Among them, h t is the output of the EMRN model at time t.

[0101] (III) Feature Fusion

[0102] The final output slices obtained from step (1) and step (2) are concatenated together to form the final weight vector O F , the final weight vector O F The process of inputting into the fully connected layer for processing includes:

[0103] P=σ(W p O F +b p )

[0104] Among them, P is the final traffic prediction; W p is the weight matrix of the fully connected layer, b p is the bias term of the fully connected layer.

[0105] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A crowd traffic big data prediction method for 5G needs, characterized in that: The specific steps include: The area to be predicted is divided into multiple small areas, each small area is regarded as a node, and node features and edge weight features between nodes are obtained; the node features include environmental features of the area corresponding to the node, and the environmental features at least include the weather conditions, air quality, economic indicators, and commuting mode of the current area; the edge weight features at least include the traffic flow and crowd flow between two nodes; Convert node features and edge weight features into vector representations, and input the vector representations together with the node and edge position encodings into the graph neural network as input data; Graph data is constructed through node and edge weight relationships, and each node in the constructed graph data is iteratively updated using a graph neural network to obtain the first eigenvector of each node. The iterative process includes: in, is the iterative feature vector corresponding to the i-th node at the t+1th iteration; σ(·) represents the ReLU function; W (t) is the weight matrix at the tth iteration, b (t) is the bias term at the tth iteration; c ij is a normalization constant; N(i) is the set of neighbor nodes of the i-th node; It represents the iterative feature vector corresponding to the jth node at the tth iteration. When t = 0, that is, It is the feature obtained by concatenating the node features through the embedding layer and position encoding; The first feature vector of each node is input into the graph attention network to obtain the second feature vector of each node, including: h i '=σ(Σ j∈N(i)∪{i} a ij Wh j ) Among them, h i ' is the second eigenvector of the i-th node, σ(·) represents the ReLU function; N(i) is the set of neighbor nodes of the i-th node; α ij is the attention coefficient of the i-th node to the j-th node calculated by the attention mechanism; W is the feature conversion matrix; h j is the first eigenvector of the jth node; The data at a certain moment in the area to be predicted is divided into three layers of data: real-time data, periodic data, and abnormal data. The current layer of data is processed by the enhanced memory adjustment network model and spliced ​​with the next layer of data as the input of the next layer of enhanced memory adjustment network model. The enhanced memory adjustment network model of the last layer outputs the third eigenvector. The process of processing the three layers of data using the enhanced memory adjustment network model includes: The real-time data realtime As the first layer of data, the real-time data is input into the EMRN model, which is expressed as: The output of the EMRN model based on the first layer of data With period data D cyclical After splicing, we get the input data D of the second layer. cyclical ', D cyclical 'As the input of the EMRN model, it is expressed as: The output of the EMRN model based on the second layer data With abnormal data D anomaly After splicing, we get the input data D of the third layer. anomaly ', D anomaly 'As the input of the EMRN model, it is expressed as: Will as the third eigenvector; Among them, EMRN(·) represents the EMRN model; The data processing process of the enhanced memory adjustment network model includes: i t =σ(W xi x t +W hi h t-1 +W ci c t-1 +b i ) f t =σ(W xf x t +W hf h t-1 +W cf c t-1 +b f ) o t =σ(W xo x t +W ho h t-1 +W co c t-1 +b o ) g t =tanh(W xg x t +W hg h t-1 +b g ) c t =f t ×c t-1 +i t ×g t ×tanh(W xc x t +W hc h t-1 +b c ) h t =o t ×tanh(c t ) Among them, i t represents the variable that controls the contribution of the input at time t to the state update, σ(·) represents the ReLU function, and W xi , W hi , W ci i t The weight matrix of the input vector at the current moment, the hidden state at the previous moment, and the unit state at the previous moment during the update process, x t represents the input vector at time t, h t-1 represents the hidden state at time t-1, c t-1 represents the unit state vector at time t-1, b i for i t The corresponding bias term during the update process; f t Indicates that control keeps c at time t t-1 The number of variables, W xf , W hf , W cf f t During the update process, the weight matrix of the input vector at the current moment, the hidden state at the previous moment, and the unit state at the previous moment, b f f t The corresponding bias term during the update process; o t represents the variable that controls the output from the cell state to the hidden state at time t, W xo , W ho , W co Respectively o t During the update process, the weight matrix of the input vector at the current moment, the hidden state at the previous moment, and the unit state at the previous moment, b o for o t The corresponding bias term during the update process; g t is an additional control layer variable at time t, W xg , W hg g t During the update process, the weight matrix of the input vector at the current moment and the hidden state at the previous moment, b g g t The corresponding bias term in the update process; tanh(·) is the hyperbolic tangent function, W xc , W hc c t During the update process, the weight matrix of the input vector at the current moment and the hidden state at the previous moment, b c c t The corresponding bias term during the update process; h t is the output of the EMRN model at time t; The second eigenvector and the third eigenvector are concatenated together and input into the fully connected layer to obtain the traffic prediction of the current area.

2. According to a method for predicting crowd flow big data for 5G demand according to claim 1, it is characterized in that: The node features and edge weight features are converted into vector representations through the embedding layer. The conversion process includes: Embed(f i )=V f ×f i +b f Embed(e ij )=V e ×e ij +b e Among them, Embed(f i ) is the i-th node feature f i The corresponding vector representation, V f is the embedding matrix of node features, b f is the bias term corresponding to the embedding of node features; Embed(e ij ) is the edge weight feature e between the i-th node and the j-th node ij The corresponding vector representation, V e is the embedding matrix of edge weight features, b e is the bias term corresponding to the embedding of edge weight features.

3. A method for predicting crowd flow big data based on 5G demand according to claim 1 or 2, characterized in that: The process of position encoding the node features and edge weight features represented by the vector includes: Among them, PE(pos,k) represents the position code corresponding to the element with serial number pos in the vector sequence; pos represents the serial number of the data point in the vector sequence; d represents the dimension of the vector sequence; k is the index of the position code.

4. The method for predicting crowd flow big data for 5G demand according to claim 1 is characterized in that: The graph data G constructed by node and edge weight relationships is expressed as G=(V,E), where V is the set of all nodes and E is the set of edge relationships between all nodes.

5. The method for predicting crowd flow big data for 5G demand according to claim 1 is characterized in that: The second eigenvector and the third eigenvector are concatenated and input into the fully connected layer to obtain the traffic prediction for the current area, including: P=σ(W p The F +b p ) Where P is the final traffic prediction; σ(·) represents the ReLU function; W p is the weight matrix of the fully connected layer, b p is the bias term of the fully connected layer; O F It is the feature representation obtained by concatenating the second eigenvector and the third eigenvector.

Images