A UPF Flow Prediction Method Based on Event Constraints and Boundary Recursion

Abstract

This invention discloses a UPF (Ultra-High Frequency) traffic prediction method based on event constraints and boundary recursion. For UPF network elements, a basic traffic prediction value is generated based on historical traffic, calendar records, and historical event states. An event coverage domain is constructed based on the intersection of candidate event intervals and future sampling intervals. The event response is jointly solved using a symmetric positive definite linear system. After mapping according to the event coverage ratio, this is combined with the basic traffic prediction value to obtain the initial traffic prediction value for the future sampling location. After the actual traffic observation value arrives, observation constraints are added and the observed response variables are eliminated. The remaining response is recursively solved to update the traffic prediction value for the location to be observed. This invention utilizes event information and newly added actual traffic observation values ​​to correct the traffic prediction results, improving the prediction accuracy and stability under event disturbances and traffic fluctuation scenarios, and providing a basis for UPF capacity assessment and resource allocation.

Description

Technical Field

[0001] This invention belongs to the field of intelligent operation and maintenance and time series prediction technology of communication networks, specifically involving a UPF traffic prediction method based on event constraints and boundary recursion. Background Technology

[0002] In the 5G core network, the UPF (User Plane Forwarding) element is responsible for forwarding user plane data. Predicting its traffic can provide a basis for capacity assessment and resource allocation. In addition to periodic changes and long-term trends, UPF traffic may also be affected by events such as holidays, planned upgrades, and maintenance. Existing methods input event information along with historical traffic and calendar records into the prediction model. However, inputting the start and end times of the event does not necessarily limit the scope of the event response in the calculation. When the start and end times of the event fall within the sampling interval, it is also necessary to consider the case where the event only covers part of the interval. At the same time, if the event responses at each location are calculated separately, without common constraints, it is difficult to consider the influence of adjacent locations and event boundaries simultaneously. After the arrival of new actual traffic observations, it is also necessary to correct the subsequent traffic prediction values ​​accordingly. Directly using local residuals for subsequent locations may amplify the impact of local fluctuations. Adjusting model parameters online may change the traffic patterns that the model has learned. To address the above problems, this invention determines the coverage location and coverage ratio based on the event interval. By jointly solving the response within the constraint domain and the boundary response, after the arrival of new observations, the traffic prediction values ​​of the locations to be observed are recursively corrected using the observation constraints. The trained network parameters and the initially generated basic traffic prediction vector within the same prediction window remain unchanged. Summary of the Invention

[0003] To address the aforementioned issues, this invention discloses a UPF traffic prediction method based on event constraints and boundary recursion. This method improves traffic prediction accuracy, enhances prediction stability under complex scenarios such as event disturbances and traffic fluctuations, and enables online traffic prediction correction driven by newly added actual traffic observations. It provides a basis for traffic prediction in UPF network element capacity assessment, resource allocation, and risk warning.

[0004] To achieve the above objectives, the technical solution of the present invention is as follows:

[0005] A UPF traffic prediction method based on event constraints and boundary recursion includes the following steps:

[0006] S1. Determine the forecast starting point and future forecast window, and clarify the scope of traffic observation records and event information that can be used for the first traffic forecast;

[0007] S2. Construct a historical joint state based on historical traffic, calendar records, and historical event states to generate a basic traffic prediction vector for the future prediction window;

[0008] S3. Filter candidate events, determine the coverage location and event coverage ratio based on the intersection of the event interval and the future sampling interval, set the response nodes and auxiliary boundary nodes in the domain, and form the event coverage domain;

[0009] S4. Generate event response parameters based on historical joint states, event attributes, and node characteristics, construct a symmetric positive definite linear system, and jointly solve for the event response.

[0010] S5. Map the event responses within the domain to future sampling locations according to the event coverage ratio, add them to the corresponding basic traffic prediction values, and output the initial traffic prediction values.

[0011] S6. After the actual flow observation value arrives, the difference between the actual flow value and the basic flow prediction value is used as the correction target. The observation constraints are constructed in combination with the coverage mapping, and the constraint accuracy is determined by the residual of the first flow prediction.

[0012] S7. Within the same prediction window, the trained network parameters and the initial generated basic flow prediction vector remain unchanged; the linear system is divided into blocks, observation constraints are added to the observed blocks, the observed response variables are eliminated, the remaining response is solved, and only the flow prediction value of the location to be observed is updated.

[0013] Further, step S1 includes the following sub-steps:

[0014] S101. Determine the prediction starting point and future prediction window: Record the sampling location corresponding to the latest observed flow rate as... Define the sampling time corresponding to that position as the prediction starting point; select the continuous time up to that time. Each sampling location constitutes a historical window, and subsequent sampling locations are then... Each sampling location serves as a future prediction window.

[0015] S102. Determine the event information boundary: The initial traffic forecast only uses event information that has been registered or confirmed before the forecast start point.

[0016] Furthermore, in step S2, generating the standardized basic flow prediction vector includes the following sub-steps:

[0017] S201. Historical Traffic Standardization: Standardize historical traffic using the logarithmic mean of traffic over the target UPF element training interval. and standard deviation Calculate the standardized flow values ​​and historical flow column vectors:

[0018] ;

[0019] in, For the first The actual flow rate at each sampling location To standardize the flow value, This is to standardize the historical traffic vector.

[0020] S202, Constructing historical location features: [The text abruptly ends here, likely due to an incomplete sentence or a missing section.] Standardized flow values ​​at each historical sampling location Calendar record coding and historical event state vector Concatenate into feature vectors ; Includes hour, day of the week, and holiday markings. This includes event type coding, event relative stages, standardized event intensity, and intensity validity marking.

[0021] S203. Generate basic flow prediction values: Divide the historical location feature sequence according to a preset length and step size, obtain the historical joint state through shared segment mapping and temporal coding, and generate a standardized basic flow prediction vector from the basic flow prediction branch:

[0022] ;

[0023] in, For the historical state of unity, for Dimensionally standardized basic flow prediction vector.

[0024] Furthermore, in step S3, filtering candidate events and constructing the event coverage domain includes the following sub-steps:

[0025] S301. Filter candidate events: Select events whose time intervals overlap with the future prediction window as candidate events; if there are no candidate events, proceed to step S503.

[0026] S302, Calculate the coverage area: [Select candidate events] The start and end boundaries used at the prediction starting point are denoted as follows: and , No. The future sampling interval is denoted as Calculate the intersection of the event interval and each future sampling interval.

[0027] S303. Determine coverage location and coverage ratio: The intersection interval, event coverage ratio, set of coverage locations, and relative event stages are as follows:

[0028] ;

[0029] in, The intervals are the intersections, where dur represents the duration of the interval. For event coverage ratio, For the set of covered locations;

[0030] S304. Constructing an augmented response node set: In Included Each coverage location is assigned a response node within the domain, with each node corresponding to a response variable for an event to be determined. The nodes are arranged by time, and auxiliary boundary nodes are set at both ends to form an augmented response node set.

[0031] S305. Construct a chain-like adjacency relationship: Connect the response nodes and auxiliary boundary nodes within each domain in chronological order to construct a node feature matrix. ; This includes relative node time, relative event stage, coverage ratio, auxiliary boundary node type, boundary source, and constraint validity. The endpoint selection matrix is ​​used to select the left and right auxiliary boundary responses. It is a symmetric positive semi-definite chained Laplace matrix.

[0032] Furthermore, in step S4, generating event response parameters and jointly solving for the event response includes the following sub-steps:

[0033] S401, Parameter Input: For candidate events In the historical state of union Event attribute vector and node feature matrix For input; It consists of event type coding, standardized event strength, and strength validity markers;

[0034] S402. Generate node parameters: Obtain the response prior center through the shared node parameter generation function. Response prior accuracy matrix Left and right endpoints constrained target vector Endpoint constraint precision matrix and chain coordination weights :

[0035] ;

[0036] S403. Construct a linear system: Using the responses of all augmented nodes as variables, incorporate the prior consistency constraint of responses, the difference constraint of responses between adjacent nodes, and the endpoint constraint into the same linear system.

[0037] The joint solution system is as follows:

[0038] ;

[0039] in, The augmented response vector is arranged in node order. It is a chain-like Laplace matrix. The endpoint selection matrix is ​​used to select the left and right auxiliary boundary responses; the coefficient matrix is ​​denoted as... The vector on the right is denoted as .

[0040] Furthermore, in step S5, obtaining the initial flow prediction value includes the following sub-steps:

[0041] S501, Mapped Event Response: For candidate events Construct an event coverage mapping matrix Based on event coverage ratio Intra-domain responses are mapped to the future. One sampling location;

[0042] S502, Combined Flow Prediction: The mapped event response is added to the standardized basic flow prediction vector, then inversely standardized, exponentially calculated, subtracted by one, and truncated to obtain the initial flow prediction value at the original flow scale.

[0043] ;

[0044] in, and All dimensional vector, for A column vector of all 1s.

[0045] S503, Branch without candidate events: When there are no candidate events, the standardized basic flow prediction vector is directly output after inverse standardization, inverse flow transformation and non-negative truncation.

[0046] Furthermore, in step S6, determining the observation boundary and constructing observation constraints includes the following sub-steps:

[0047] S601. Determine the observation boundary and node set: Within the current future prediction window, actual flow observations arrive consecutively in chronological order. The last observed coverage location is taken as the observation boundary, and the sampling locations from the first coverage location to the observation boundary form the set of observed sampling locations. The corresponding set of response nodes within the domain is denoted as The remaining domain response nodes, together with the two auxiliary boundary nodes, form the remaining augmented response node set. ;

[0048] S602. Constructing the Corrected Target: The corrected target, the observation mapping submatrix, and the initial flow prediction residual are as follows:

[0049] ;

[0050] in, To standardize the actual flow vector, This is the first time that the basic flow prediction subvector has been standardized. To correct the target, To observe the mapping submatrix, This is the response sub-vector obtained from the first joint solution. To standardize the initial flow forecast residual;

[0051] S603. Determine the observation constraint accuracy: Based on the absolute value of the initial flow prediction residual, calculate the accuracy of the first observation constraint according to the exponential decay relationship. Observation constraint accuracy for each observed location:

[0052] ;

[0053] in, For the first The observation constraint accuracy of each observed location. and These are the preset lower and upper bounds of precision, respectively. For attenuation scale parameters, The first residual vector of the initial flow prediction Each component, and the precision at each position, forms a diagonal matrix. .

[0054] Furthermore, in step S7, performing the recursive calculation of the remaining event coverage subdomains includes the following sub-steps:

[0055] S701. Divide the joint solution system into blocks: divide the coefficient matrix and the right-hand vector By node set and The corresponding response variable is divided into blocks, resulting in matrix blocks. , , , and vector partitioning , ;

[0056] S702. Add observation constraints and eliminate variables: Based on the observation mapping submatrix, correction target, and accuracy, add constraint terms to the coefficient matrix and right-hand vector of the observed blocks, and then eliminate the response variables in the observed domain to obtain the remaining system.

[0057] ;

[0058] in, , , and for The corresponding blocks, and for The corresponding blocks; This is the coefficient matrix of the remaining system.

[0059] S703. Solve for the residual response and update the flow forecast: Solve for the residual augmented response, map the intra-domain response to the observation location according to the event coverage ratio, add it to the normalized basic flow forecast, and update the flow forecast after scaling and non-negative truncation.

[0060] ;

[0061] in, For the remaining augmented response vector, This is the set of future sampling locations within the event coverage area for which actual flow observations have not yet been obtained. The event coverage mapping submatrix corresponding to the observed location and the remaining response variables; This is the first standardized basic flow prediction sub-vector for the location to be observed. The vector is a column vector of all 1s with the same dimension as the number of locations to be observed. Flow prediction values ​​will no longer be output for locations that have already been observed.

[0062] S704. Update the observation boundary: As actual flow rate observations arrive sequentially over time, advance and expand the observation boundary. Shrink Repeat steps S6 and S7 until there are no more locations to be observed within the event coverage area. Within the same window, keep the prediction starting point, window length, node structure, trained network parameters, and the first generated standardized basic traffic prediction vector unchanged.

[0063] The beneficial effects of this invention are as follows:

[0064] This invention determines the coverage location and event coverage ratio based on the intersection of the event interval and the future sampling interval, and jointly solves the intra-domain response and boundary response, ensuring that the impact of the event on traffic corresponds to the actual coverage area. After a new actual traffic observation arrives, the traffic prediction value at the observation location is updated using observation constraints, without changing the trained network parameters and the initially generated basic traffic prediction vector within the same prediction window. The observation constraint accuracy is determined based on the initial traffic prediction residual, reducing the impact of large residuals on subsequent traffic predictions. The initial traffic prediction only uses information available at the prediction starting point. This method helps improve the accuracy and stability of UPF traffic prediction in scenarios of event disturbances and traffic fluctuations, providing a basis for capacity assessment, resource allocation, and risk warning. Attached Figure Description

[0065] Figure 1 This is a flowchart illustrating the overall implementation of the present invention.

[0066] Figure 2This is the overall framework diagram of the event boundary recursive flow prediction network of the present invention. Detailed Implementation

[0067] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0068] Example 1

[0069] In this embodiment, samples are independently constructed, models are trained, and future traffic prediction values ​​are generated for a single target UPF network element. Different UPFs can use the same structure, but the model parameters are saved separately.

[0070] Inputs include actual traffic observations, event information, and calendar records. Traffic records contain UPF identifiers and sampling times, while event records contain UPF identifiers and event identifiers. Records are associated according to their UPF identifiers, and the sampling location corresponding to the event is determined based on the intersection of the event interval and the sampling interval. The specific process is as follows: Figure 1 As shown, the steps are as follows.

[0071] S1. Determine the target UPF network element, prediction starting point, and information boundary. Specific details are as follows:

[0072] S101. Determine the prediction starting point and future prediction window: Record the sampling location corresponding to the latest observed flow rate as... The sampling time corresponding to this position is defined as the prediction starting point; the length of the historical window is denoted as... The predicted future length is denoted as In this embodiment, , .

[0073] S102. Determine the event information boundary: The initial traffic forecast only uses event information that has been registered or confirmed before the forecast start point.

[0074] S2. Extract historical joint states and generate basic flow prediction values, as detailed below:

[0075] S201, Historical Flow Standardization: Standardizing historical flow values Perform the transformation using the logarithmic mean of the flow rate over the training interval of the target UPF element. and standard deviation Standardize and assemble standardized historical traffic vectors in chronological order. ,in, This indicates the index of the historical sampling location.

[0076] S202. Constructing historical location characteristics: Standardizing the flow values ​​at each historical sampling location. Calendar record coding and historical event state vector Concatenate into feature vectors The features are arranged in chronological order. Includes hour, day of the week, and holiday markings. This includes event type coding, relative event stages, standardized event intensity, and intensity validity marking.

[0077] S203. Generate basic flow prediction values: Divide the historical location feature sequence according to a preset length and step size, and obtain the historical joint state through shared segment mapping and temporal coding. The standardized basic flow prediction vector is generated from the basic flow prediction branch. .

[0078] S3. Filter candidate events and construct an event coverage area based on the event range, as detailed below:

[0079] S301. Filter candidate events: Select events whose time intervals overlap with the future prediction window as candidate events. If there are no candidate events, proceed to step S503.

[0080] S302, Calculate the coverage area: [Select candidate events] The start and end boundaries used at the prediction starting point are denoted as follows: and , No. The future sampling interval is denoted as Calculate the intersection of the event interval and each future sampling interval. .

[0081] S303. Determine coverage location and coverage ratio: Add the corresponding locations of sampling intervals with overlapping time periods to the coverage location set. The ratio of the intersection duration to the sampling interval duration is used as the event coverage ratio. The midpoint of the intersection interval is denoted as The relative stage of the event is determined based on the relative position of the midpoint within the event interval. .

[0082] S304. Constructing an augmented response node set: In Each coverage location is assigned an intra-domain response node, and each node corresponds to a response variable for an event to be determined. The nodes are arranged by time, and auxiliary boundary nodes are set at both ends. The augmented response node set consists of intra-domain response nodes and two auxiliary boundary nodes.

[0083] S305. Construct a chained adjacency relationship: Connect the response nodes and auxiliary boundary nodes within each domain in chronological order, using a chained Laplace matrix. Represent adjacency relationships and construct an endpoint selection matrix to select auxiliary boundary responses. .

[0084] S4. Generate event response parameters and jointly solve for the event response. The specific steps are as follows:

[0085] S401, Parameter Input: Based on historical joint state Event attribute vector and node feature matrix For input; It consists of event type coding, standardized event strength, and strength validity markers. It includes node relative time, event relative stage, coverage ratio, auxiliary boundary node type, boundary source, and constraint validity.

[0086] S402. Generate node parameters: Generate a response prior center through a shared node parameter generation function. Response prior accuracy matrix Endpoint constraint target vector Endpoint constraint precision matrix and chain coordination weights .

[0087] S403. Constructing a linear system: Combining the domain response and the auxiliary boundary response into an augmented response vector according to the node order. By incorporating the prior consistency constraint of response, the difference constraint of response between adjacent nodes, and the endpoint constraint into the same linear system, the following can be achieved: , , , and Constructing the coefficient matrix , combined and Construct the right-hand vector .

[0088] S5. Generate the initial flow prediction values ​​for each sampling location within the future prediction window, as detailed below:

[0089] S501, Mapping Event Response: Constructing an Event Coverage Mapping Matrix According to coverage ratio Will The domain response is mapped to the corresponding future sampling location.

[0090] S502, Combined Flow Forecast Value: Combines the mapped event response with the standardized basic flow forecast vector. Add them together, using training statistics. and After inverse standardization, exponential operation, subtraction by one, and non-negativity truncation, the initial flow prediction vector at the original flow scale is obtained. .

[0091] S503, Branch without candidate events: When there are no candidate events, directly use the standardized basic flow prediction vector. The output is obtained after inverse normalization, inverse flow transformation, and non-negative truncation.

[0092] S6. Determine the observation boundary and node set, and construct observation constraints, as detailed below:

[0093] S601. Determine the observation boundary and node set: Within the current future prediction window, actual flow observations arrive consecutively in chronological order. The last observed coverage location is taken as the observation boundary, and the sampling locations from the first coverage location to the observation boundary form the set of observed sampling locations. The corresponding set of response nodes within the domain is denoted as The remaining domain response nodes, together with the two auxiliary boundary nodes, form the remaining augmented response node set. .

[0094] S602, Constructing the correction target: with Standardized actual flow vector With standardized basic flow prediction subvector The difference is the correction target. ;from Selected from Corresponding rows and The corresponding columns yield the observation mapping submatrix. This will be the first time the response has been broadened. The components on are denoted as ,by minus through The mapped event response yields the first flow prediction residual at a standardized scale. .

[0095] S603. Determine the observation constraint accuracy: Based on The absolute value of is calculated according to the exponential decay relationship. Observation constraint accuracy of each observed location The lower bound of precision is The upper boundary is The attenuation scale parameter is Construct the observation constraint accuracy matrix based on the accuracy at each location. .

[0096] S7. Perform the recursive calculation of the remaining event coverage subdomain at the observation boundary, as detailed below:

[0097] S701. Divide the joint solution system into blocks: by node set and The corresponding response variables will be used to solve the system's coefficient matrix for the first time. Divided into , , and , the right-hand vector Divided into and .

[0098] S702, Add observation constraints and eliminate variables: According to , and To the observed blocks and By adding observation constraints and then eliminating the response variables within the observed domain, the residual system is obtained, whose coefficient matrix is ​​denoted as... The vector on the right is denoted as .

[0099] S703. Solve for the residual response and update the flow forecast: Solve for the residual system to obtain the residual augmented response vector. ,by Represents the set of future sampling locations to be observed within the event coverage area, from Select the corresponding mapping submatrix Map responses within the domain to events based on event coverage ratios. With the corresponding standardized basic flow forecast value The values ​​are added together, scaled back, and truncated to update the predicted flow rate at the location to be observed.

[0100] S704. Update the observation boundary: As actual flow observations arrive sequentially over time, advance the observation boundary, repeating steps S6 and S7 until there are no more locations to be observed within the event coverage area. Within the same window, maintain the prediction starting point, window length, node structure, trained network parameters, and the initially generated standardized basic flow prediction vector. Without changing the initial forecast window, when establishing a new forecast window, the information boundaries are redefined and a standardized basic flow forecast vector is generated. Candidate events are selected; trained network parameters can be reused.

[0101] In this embodiment, the coverage location and coverage ratio are determined based on the event interval. The event response of each coverage location is obtained by joint solution of endpoint constraint response. After the actual traffic observation value arrives in time sequence, the observation constraint is added to the solution system, the observed response variable is eliminated, and the traffic prediction value of the location to be observed is recursively corrected. The above processing helps to improve the traffic prediction accuracy of UPF network elements and enhance the prediction stability under event disturbance and traffic fluctuation scenarios.

[0102] This invention also proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. It should be noted that when the processor executes the computer program, it corresponds to the specific steps of the method provided in this invention, and has the corresponding functional modules and beneficial effects of executing the method. For technical details not described in detail in this embodiment, please refer to the method provided in this invention.

[0103] This invention also proposes a computer-readable storage medium storing a computer program. It should be noted that when the computer program is run by a processor, it corresponds to the specific steps of the method provided in this invention, and has the corresponding functional modules and beneficial effects for executing the method. For technical details not described in detail in this embodiment, please refer to the method provided in this invention.

[0104] It should be noted that the above content merely illustrates the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, various improvements and modifications can be made without departing from the principle of the present invention, and all such improvements and modifications fall within the scope of protection of the claims of the present invention.

Claims

1. A UPF flow prediction method based on event constraints and boundary recursion, characterized in that, Perform the following steps for the target UPF network element: S1. Determine the forecast starting point and future forecast window, and clarify the scope of traffic observation records and event information that can be used for the first traffic forecast; S2. Construct a historical joint state based on historical traffic, calendar records, and historical event states to generate a basic traffic prediction vector for the future prediction window; S3. Filter candidate events, determine the coverage location and event coverage ratio based on the intersection of the event interval and the future sampling interval, set the response nodes and auxiliary boundary nodes in the domain, and form the event coverage domain; S4. Generate event response parameters based on historical joint states, event attributes, and node characteristics, construct a symmetric positive definite linear system, and jointly solve for the event response. S5. Map the event responses within the domain to future sampling locations according to the event coverage ratio, add them to the corresponding basic traffic prediction values, and output the initial traffic prediction values. S6. After the actual flow observation value arrives, the difference between the actual flow value and the basic flow prediction value is used as the correction target. The observation constraints are constructed in combination with the coverage mapping, and the constraint accuracy is determined by the residual of the first flow prediction. S7. Within the same prediction window, the trained network parameters and the initially generated basic traffic prediction vector remain unchanged. The linear system is divided into blocks. After adding observation constraints to the observed blocks, the observed response variables are eliminated, the remaining response is solved, and only the flow prediction values ​​at the locations to be observed are updated. When there are no candidate events, output the basic flow prediction value and do not execute steps S4 to S7; when there are candidate events, as the actual flow observation value gradually arrives, repeat steps S6 and S7 until there are no locations to be observed within the event coverage area.

2. The method according to claim 1, characterized in that, Step S1 includes the following sub-steps: S101. Determine the prediction starting point and future prediction window: Define the sampling time corresponding to the latest observed flow rate as the prediction starting point, and select the continuous flow rate up to that time. Each sampling location constitutes a historical window, and subsequent sampling locations are then... Each sampling location serves as a future prediction window; S102. Determine the event information boundary: The initial traffic forecast only uses event information that has been registered or confirmed before the forecast start point.

3. The method according to claim 1, characterized in that, In step S2, generating the standardized basic flow prediction vector corresponding to the future prediction window includes the following sub-steps: S201, Historical Traffic Standardization: Perform a transformation on historical traffic and calculate standardized traffic values ​​using the logarithmic traffic mean and standard deviation of the target UPF element training interval; S202. Constructing historical location features: Concatenate the standardized flow values, calendar record codes, and historical event state vectors of each historical sampling location to form a historical location feature sequence in chronological order; S203. Generate basic flow prediction values: Input the historical location feature sequence into the historical state encoder to obtain the historical joint state. Generated from the basic flow prediction branch 3D Standardized Basic Flow Prediction Vector .

4. The method according to claim 1, characterized in that, Step S3, constructing the event coverage domain includes the following sub-steps: S301. Filtering candidate events: Events whose time intervals overlap with the future prediction window are selected as candidate events; S302, Calculate the coverage area: [Select candidate events] The start and end boundaries used at the prediction starting point are denoted as follows: and , No. The future sampling interval is denoted as Calculate the intersection of the event interval and each future sampling interval; S303. Determine the coverage location and coverage ratio: Use the ratio of the intersection duration to the sampling interval duration as the event coverage ratio, and add locations with an event coverage ratio greater than zero to the coverage location set. ； in, The intervals are the intersections, where dur represents the duration of the interval. For event coverage ratio, For the set of covered locations; S304. Constructing an augmented response node set: In Each coverage location is assigned an intra-domain response node, and each node corresponds to a response variable for an event to be determined. The nodes are arranged by time, and auxiliary boundary nodes are set at both ends to form an augmented response node set. S305. Construct a chain-like adjacency relationship: Connect the response nodes and auxiliary boundary nodes in each domain in chronological order.

5. The method according to claim 1, characterized in that, Step S4, generating the event response parameters and jointly solving the event response, includes the following sub-steps: S401, Parameter Input: For candidate events In the historical state of union Event attribute vector and node feature matrix For input; It consists of event type coding, standardized event strength, and strength validity markers. It includes relative node time, relative event stage, coverage ratio, and boundary characteristics; S402. Generate node parameters: Obtain the response prior center through the shared node parameter generation function. Response prior accuracy matrix Left and right endpoints constrain the target vector Endpoint constraint precision matrix and chained coordination weights ; S403. Constructing a linear system: Incorporating prior consistency constraints on responses, differences in responses between adjacent nodes, and endpoint constraints into the same linear system: ； in, The augmented response vector is arranged in node order. It is a chain-like Laplace matrix. The endpoint selection matrix is ​​used to select the left and right auxiliary boundary responses; the coefficient matrix is ​​denoted as... The vector on the right is denoted as .

6. The method according to claim 1, characterized in that, Step S5, obtaining the initial traffic prediction value includes the following sub-steps: S501, Mapped Event Response: For candidate events Construct an event coverage mapping matrix Based on event coverage ratio Intra-domain responses are mapped to the future. One sampling location; S502, Combined Flow Forecast Value: Combines the mapped event response with the standardized basic flow forecast vector. The sums are then processed through inverse standardization, inverse flow transformation, and non-negative truncation to obtain the initial flow prediction value. S503, Branch without candidate events: When there are no candidate events, the standardized basic flow prediction vector is directly output after inverse standardization, inverse flow transformation and non-negative truncation.

7. The method according to claim 1, characterized in that, In step S6, constructing the observation constraints includes the following sub-steps: S601. Determine the observation boundary and node set: Within the current future prediction window, actual flow observations arrive consecutively in chronological order. The last observed coverage location is taken as the observation boundary, and the sampling locations from the first coverage location to the observation boundary form the set of observed sampling locations. The corresponding set of response nodes within the domain is denoted as The remaining domain response nodes, together with the two auxiliary boundary nodes, form the remaining augmented response node set. ; S602. Constructing the Corrected Target: The corrected target, the observation mapping submatrix, and the initial flow prediction residual are as follows: ； in, To standardize the actual flow vector, This is the first time that a basic flow prediction subvector has been standardized. To correct the target, To observe the mapping submatrix, This is the response sub-vector obtained from the first joint solution. To standardize the initial flow forecast residual; S603. Determine the observation constraint accuracy: Based on the absolute value of the initial flow prediction residual, calculate the observation constraint accuracy for each observed location according to the exponential decay relationship; construct the observation constraint accuracy matrix based on the accuracy of each location. .

8. The method according to claim 7, characterized in that, Step S7, performing the recursive calculation of the remaining event coverage subdomains, includes the following sub-steps: S701. Divide the joint solution system into blocks: divide the coefficient matrix and the right-hand vector By node set and The corresponding response variable is divided into blocks, resulting in matrix blocks. , , , and vector partitioning , ; S702, Add observation constraints and eliminate variables: , and The observation constraints are added to the observed blocks, and then the response variables within the observed domain are eliminated to obtain the remaining event-covered subdomain system: ； in, Here is the coefficient matrix of the remaining system. This is the right-hand vector of the remaining system; S703, Solve for remaining responses and update traffic predictions: Solve for remaining event coverage subdomain system: ； in, For the remaining augmented response vector, the in-domain response is mapped to the future sampling location to be observed within the event coverage domain according to the event coverage ratio, and added to the corresponding standardized basic flow prediction value. After inverse standardization, inverse flow transformation and non-negative truncation, the flow prediction value of the location to be observed is updated. S704. Update the observation boundary: Within the same window, keep the trained network parameters and the initially generated standardized basic traffic prediction vector unchanged; as actual traffic observations arrive continuously, advance the observation boundary and expand it. Shrink Repeat steps S6 and S7 until there are no more locations to be observed within the event coverage area.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the UPF traffic prediction method based on event constraints and boundary recursion as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the UPF traffic prediction method based on event constraints and boundary recursion as described in any one of claims 1 to 8.

Images