Event-driven traffic network modeling method, event prediction method and system

Abstract

The application relates to an event-driven traffic network modeling method, an event prediction method and a system, and proposes a spatiotemporal interaction Hawkes process, solves the problem of event-driven dynamic network modeling, can effectively learn the influence mode of historical interaction events on subsequent events, and describes the network dynamic change process. The spatiotemporal interaction Hawkes process proposed in the application describes a series of event occurrence processes, and utilizes relevant historical events and relevant adjacent nodes including geographical adjacent nodes and semantic adjacent nodes to establish an interaction event occurrence rate model between any two nodes on the network. Model parameters are estimated by maximizing the likelihood function of the Hawkes process. The estimation result can fully reveal the complex spatiotemporal influence mode between interaction events, and realizes prediction of the number of network events at a future moment.

Description

Technical Field

[0001] This invention relates to the field of event-driven dynamic network technology, and in particular to an event-driven traffic network modeling method, event prediction method and system. Background Technology

[0002] With the increasing development of urban transportation, the number of vehicles on the road is increasing, and the road traffic conditions are becoming more and more complex. Accurate prediction of road events is of great significance for traffic flow prediction, traffic planning, and driving decisions. Event-driven dynamic networks can be used to predict road events.

[0003] Network data has been widely applied in numerous fields such as traffic demand, shared bicycles, and solar radiation. It can be viewed as a sequence of interactive events at different spatial locations; for example, a vehicle's movement between two intersections is an event with both temporal and spatial information. Due to the specific spatial structure characteristics of networks and the geographical and semantic proximity between events, each event may influence or trigger a series of subsequent events. These events are typically recorded in the form of event-based dynamic networks to predict the probability of event occurrence. Interactive events reflect the fundamental characteristics of networks and the behavior of people or objects within them; therefore, correctly establishing event-based dynamic network models is crucial.

[0004] Event-based dynamic networks are formed by nodes and edges with specific spatial network structures, where edges typically represent events. An event-based dynamic network is a spatiotemporal counting system, represented by the count of events occurring between two related nodes in the spatial domain. Despite the availability of a large amount of counting observation data for modeling research, modeling event-based dynamic networks remains very difficult, especially for transportation networks. Existing modeling methods cannot adequately describe the spatiotemporal correlation structure of the network, resulting in poor accuracy in event prediction.

[0005] First, most existing event-based methods primarily model the temporal correlation between events, neglecting spatial information. These methods ignore the essential characteristics of events as count data and the spatial correlation of the network, thus making them unsuitable for event-based dynamic network modeling. While many methods can handle the spatiotemporal correlation between count data, they still fall short in event-based dynamic network modeling because they fail to consider the directionality between two nodes, resulting in an inability to adequately describe the spatiotemporal correlation structure of the network. Second, interactive events can be driven by various real-world causes, including periodic connections between nodes and responses to preceding interactive events. However, previous studies have largely aggregated these dynamic interactions into simple edges between network nodes, completely ignoring a wealth of information such as the frequency of events, the time intervals between related events, and the motivation behind each interactive event. In event-based dynamic networks, considering the impact of relevant historical interactive events on each new interactive event is meaningful for modeling interactive event networks.

[0006] Therefore, it is necessary to propose an event-driven modeling scheme for traffic networks to predict events. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the existing technology and provide an event-driven traffic network modeling method, event prediction method and system.

[0008] The objective of this invention can be achieved through the following technical solutions:

[0009] An event-driven traffic network modeling method includes the following steps:

[0010] Obtain an event-based dynamic network, wherein the event-based dynamic network has n nodes in the time domain;

[0011] Obtain the historical event set of the dynamic network before time t. The historical event set It stores the number of events that occur between node pairs in a dynamic network;

[0012] The occurrence rate λ of events between node pairs ij at time t is described by the conditional strength function. ij (t), thus obtaining λ ij The expression for (t) is as follows:

[0013]

[0014] Where, N ij(t) represents the number of events occurring between node pairs ij within the time interval (-inf,t]. The conditional strength function describes the expected occurrence rate of node pairs ij, i,j=1,…,n at time t, based on past events.

[0015] Using a spatiotemporal interactive Hawkes process to measure the occurrence rate λ of events from node i to node j ij (t) is modeled as follows:

[0016]

[0017] in, It is the background rate basis term, μ ij (t) describes the impact of historical events in a dynamic network on events occurring between node pairs ij at time t;

[0018] For μ oj (t) is decomposed to obtain:

[0019] μ ij (t)=g oj (t)+s ij (t)+d ij (t)

[0020] Among them, g oj (t), s ij (t) and d ij (t) represents the geographic neighborhood function, semantic neighborhood function, and temporal neighborhood function of node pair ij at time t, respectively, that is, the influence of the historical events of geographic neighborhood node pair, the historical events of semantic neighborhood node pair, and the historical events of temporal neighborhood node pair on the events occurring between node pair ij at time t;

[0021] Based on historical event sets And the structure of the dynamic network determines λ ij The values ​​of each term in (t) are used to obtain the occurrence rate λ of events between node pairs ij at time t. ij The model is completed by determining the value of (t).

[0022] Furthermore, node pair ij represents the event from node i to node j, and the set of geographical neighbors of the source node i is:

[0023] H i =}h|dis(i,h)≤G}

[0024] The set of geographical neighbors of target node j is:

[0025] K j ={k|dis(k,j)≤G}

[0026] Where dis(·) is the straight-line distance between two nodes, and G is the preset neighborhood range threshold;

[0027] The geographical neighborhood node pairs of node pair ij include first-type node pairs and second-type node pairs. The first-type node pairs are node pairs hj, h∈H. i The second type of node pair is node pair ik, k∈K j ;

[0028] The expression for the geographic neighborhood function is:

[0029]

[0030] in, This represents the impact of all historical events on the first type of node pair. This represents the impact of all historical events on the second type of node pair; This represents the change in the conditional intensity function of ij caused by the excitation of hj by the node. This represents the transition change of the conditional intensity function of node pair ij generated by node pair ik; in the first type of node pair, an exponential kernel function is used. To describe the time decay property, t h n represents the time period before time t. hj (t h ) indicates that node pair hj is in time period t h The number of events occurring within the node is counted; in the second type of node pair, an exponential kernel function is used. To describe the time decay property, t k n represents the time period before time t. ik (t k ) indicates that during the time period t k The internal node counts the number of events occurring at point ik.

[0031] Furthermore, Represented as the cardinality term of the influence of node on ij Incentive terms from the influence of nodes on hj And the weight w for measuring the similarity between source node i and its geographical neighbor node h. ih The product is as follows:

[0032]

[0033] The weights h between the source node i and its geographical neighbors are defined using a distance-based softmax normalized exponential function as follows:

[0034]

[0035] Among them, the closer the geographical locations of two nodes are, the greater their weight, indicating that the similarity of the events occurring at these two nodes is higher;

[0036] Will Represented as the cardinality term of the influence of node on ij Incentive terms derived from the influence of nodes on ik And the weight w that measures the similarity between the target node j and its geographical neighbor node k. kj The product is as follows:

[0037]

[0038] The weights between the source node i and its geographical neighbors are defined using a distance-based softmax normalized exponential function as follows:

[0039]

[0040] The closer two nodes are geographically, the greater their weight, indicating a higher similarity between the events occurring at these two nodes.

[0041] Furthermore, node pair ij represents the event from node i to node j, and the semantic node set of source node i is:

[0042] M i = {m | node pair mi, where m is the source node and i is the target node}

[0043] The semantic neighborhood node pairs of node pair ij are node pairs mi, m∈M i ;

[0044] The expression for the semantic neighborhood function is:

[0045]

[0046] Among them, |M i | represents M i The number of elements in The change in the conditional intensity function transition of node pair ij caused by node pair excitation mi is represented by the exponential kernel function. To describe the time decay property, t m n represents the time period before time t. mi (t m ) indicates the time period t m The number of events that occur on node pair mi.

[0047] Furthermore, the expression for the time neighborhood function is:

[0048]

[0049] Among them, D i Represents a pre-defined set of dates, |D i | represents D i The number of elements in The cardinality of the temporal neighborhood influence of nodes in the date set on ij is represented using an exponential kernel function. To describe the time decay property, t d Indicated in D i In the date d, the time period before time t, Indicates the time period t d The internal node counts the number of events occurring at position ij.

[0050] Furthermore, the occurrence rate λ of events between node pairs ij at time t is described by a conditional strength function. ij (t), through a spatiotemporal interactive Hawkes process, λ is obtained. ij The model of (t) can be summarized as follows:

[0051]

[0052] μ ij (t)=g ij (t)+s uj (t)+d ij (t)

[0053]

[0054]

[0055]

[0056] The model parameters are rewritten in matrix form, including the background rate basis term. Influence on incentives Impact on basic terms The influence of time neighborhood is the basic term. Where i,j=1,…,n, the parameter to be estimated is Θ. (0) Θ (1) α (1) ,

[0057] Furthermore, the aforementioned set of historical events There are a total of T consecutive time periods with available count event data, and the above λ ii The log-likelihood function of the model (t) is:

[0058]

[0059] in, Using numerical approximation integration, i.e.

[0060] The method combines the expectation-maximization algorithm with the least squares estimation to estimate the parameters to be estimated, where Θ... (0) Θ (1) α (2) Parameter estimation includes the following steps:

[0061] In the expectation step of the expectation-maximization algorithm, a random variable p is defined. ij (t) represents the probability that the event occurring on node pair ij at time t is a background event. and Let p represent the probability that an event occurring at time t, corresponding to a geographical, semantic, and temporal neighborhood function, is caused by an event occurring at a previous time τ. ij (t), and Approximately expressed as:

[0062]

[0063]

[0064]

[0065]

[0066]

[0067] In the maximization step of the expectation-maximization algorithm, the number of events generated by geographical, semantic, and temporal neighborhoods is first estimated, i.e. Parameter Θ (0) Θ (1) and α (2) The update is performed by minimizing the three optimization problems resulting from the following decomposition:

[0068]

[0069]

[0070]

[0071] consider In scenarios that have already occurred in the geographical, semantic, and temporal neighborhoods, and using H ij K ij and M ij Representing the corresponding set of scenes, denoted by D ij Let represent the total historical demand of the previous year. Solve the above three optimization problems. and Updated to:

[0072]

[0073] Furthermore, regarding α (1) The specific steps for parameter estimation are: to... and After the update, the least squares method is used to estimate...

[0074] First calculate y ij (t) and x ij (t):

[0075]

[0076]

[0077] Updated to:

[0078]

[0079] Among them, X ij =[x ij (t1),…,x ij (t T )] T Y ij =[y ij (t1),…,y ij (t T )] T .

[0080] A traffic incident prediction method includes the following steps:

[0081] Obtain an event-based dynamic network, wherein the event-based dynamic network has n nodes in the time domain;

[0082] Obtain the historical event set of the dynamic network before time t. The historical event set It stores the number of events that occur between node pairs in a dynamic network;

[0083] Use the modeling method described above to predict the occurrence rate of events between any two nodes at time t in a dynamic network;

[0084] The event prediction result is obtained based on the occurrence rate of events between any two nodes at time t in the dynamic network.

[0085] A traffic incident prediction system, comprising:

[0086] A dynamic network acquisition module is used to acquire an event-based dynamic network, wherein the event-based dynamic network has n nodes in the time domain;

[0087] The historical event set acquisition module is used to acquire the historical event set of the dynamic network up to time t. The historical event set It stores the number of events that occur between node pairs in a dynamic network;

[0088] The modeling module is used to predict the occurrence rate of events between any two nodes at time t in a dynamic network using the modeling methods described above.

[0089] The output module is used to obtain the event prediction result based on the occurrence rate of events between any two nodes at time t in the dynamic network.

[0090] Compared with the prior art, the present invention has the following beneficial effects:

[0091] First, considering the influence of geographical, semantic, and temporal neighborhoods, the method proposed in this application can model the occurrence rate of network interaction events by fully characterizing spatiotemporal interactions and the directionality between two nodes. Second, by considering the impact of relevant historical interaction events on each new interaction event, the proposed spatiotemporal interaction Hawkes process contains a wealth of information, including the number of events, the time intervals between relevant events, and the impact pattern of each interaction event. Therefore, the event dynamics of the network can be effectively characterized by the occurrence rate of interaction events among all node pairs on the network. Attached Figure Description

[0092] Figure 1 This is a flowchart of the present invention;

[0093] Figure 2 This is a schematic diagram of the transportation network area;

[0094] Figure 3 for Estimated values ​​of the parameters;

[0095] Figure 4 Example of traffic network event number prediction results. Detailed Implementation

[0096] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are implemented based on the technical solutions of the present invention, providing detailed implementation methods and specific operating procedures. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them, and the scope of protection of the present invention is not limited to the following embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0097] This specification provides method operation steps as shown in the embodiments or flowcharts, but based on conventional or non-inventive labor, more or fewer operation steps may be included. The order of steps listed in the embodiments is merely one possible execution order among many and does not represent the only execution order. In actual system or server products, the method can be executed in the order shown in the embodiments or drawings, or in parallel (e.g., in a parallel processor or multi-threaded processing environment), or the execution order of steps without timing constraints can be adjusted.

[0098] Example 1:

[0099] Definition: Assume an event-based dynamic network with n spatial nodes in the time domain. Spatiotemporal event data can be represented as {i,j,t}, i,j=1,…,n, where each group {i,j,t} represents the index of the event occurring between node pairs ij at time t within a day. This invention aims to explicitly represent the occurrence rate of events through a network point process, establishing a dynamic model for the interaction events of any pair of nodes in the network. The network point process is a stochastic process that can be used with all node pairs in the network. Several counting processes N = {N ij Let N(t), i,j=1,…,n be used to represent N. ij (t) represents the number of events occurring between node pairs ij within the time interval (-inf,t]. For each N ij (t), i,j=1,…,n, given the set of historical events in the network before time t. The occurrence rate λ at time t is described by the conditional intensity function (CIF). ij (t), thus obtaining λ tj The expression for (t) is as follows:

[0100]

[0101] CIF describes the expected occurrence rate of node pairs ij, i,j=1,…,n at time t, based on past events.

[0102] (1) Spatiotemporal interaction Hawkes process

[0103] This invention, by drawing on the framework of Hawkes processes, describes the network point process λ from one node i to another node j on a network. ij The CIF of (t) is modeled as follows:

[0104]

[0105] in, It is the background rate basis term, μ ij(t) describes the impact of historical network events on the upcoming events of node pair ij. Specifically, this invention considers the interactive effects of historical network events from the perspectives of spatial correlation and temporal evolution.

[0106] For a certain day μ ij In modeling μ(t), this invention considers the interactive influence of historical network events in the geographical neighborhood, semantic neighborhood, and temporal neighborhood before that day, and will... ij (t) is decomposed into three parts

[0107] μ ij (t)=g ij (t)+s ij (t)+d ij (t), (Equation 3)

[0108] Where g ij (t), s ij (t) and d ij (t) represents the geographic neighborhood function, semantic neighborhood function, and temporal neighborhood function of node pair ij at time t, respectively.

[0109] (1.1) Geographic Neighborhood Function

[0110] A node's geographic neighbors are nodes that are geographically adjacent to it. A node and its geographic neighbors are more likely to have similar numbers of events occurring in each other. The geographic neighborhood function considers the geographic neighbors of node pair ij, i.e., nodes with the same source node and neighboring target node as the current node pair, or neighboring source nodes and the same target node, to represent the influence of all historical events with similar patterns on the node pair.

[0111]

[0112] The specific explanation of Equation 4 is as follows:

[0113] For source node i, define its set of geographical neighbor nodes as follows:

[0114] H i ={h|dis(i,h)≤G}, (Equation 5)

[0115] Where dis(·) is the straight-line distance between two nodes, and G is the threshold of the neighborhood range, which can be set empirically. Therefore, the first term of Equation 4, i.e. This indicates that there are neighboring target nodes h∈H i The impact of all historical events on the same source node j. This represents the change in the conditional intensity function of ij caused by the excitation of hj by the node. An exponential kernel function is used. To describe the time decay property, where th n represents the time period before time t. hj (t h ) indicates that node pair hj is in time period t h Count the number of events that occur within. Represented as the cardinality term of the influence of node on ij Incentive terms from the influence of nodes on hj And the weight w for measuring the similarity between source node i and its geographical neighbor node h. ih The product of is shown in the following formula:

[0116]

[0117] The weights between the source node i and its geographical neighbors are defined using a distance-based softmax normalized exponential function as follows:

[0118]

[0119] The closer two nodes are geographically, the greater their weight, indicating a higher similarity between the events occurring at these two nodes.

[0120] 2) For target node j, the definition of its geographical neighbor nodes follows the same process as that of source node i, that is:

[0121] K j ={k|dis(k,j)≤G}, Equation 8

[0122] The second term of Equation 4, namely This indicates that nodes have the same source node i and neighboring target nodes k, h∈K. j The impact of all historical events, among which t represents the change in the conditional intensity function of node pair ij caused by excitation of node pair ik. k n represents the time period before time t. ik (t k ) indicates that during the time period t k The internal node counts the events occurring at point ik. Represented as

[0123]

[0124] That is Incentive terms for the influence of nodes on ik and weight The product of w kj Measure the similarity between the target node j and its geographical neighbor node k.

[0125] (1.2) Semantic neighborhood function

[0126] For semantic neighborhood nodes, if there are a large number of events between different nodes, these nodes are called semantic neighborhood nodes, which is beneficial for timely discovery of flow relationships between different nodes. The semantic neighborhood node influence considered in this invention is the impact of the count of all historical events when node i appears as a target node in the current node pair ij. Only after node i has been a target node can it become the source node of the next target node j. Here, the set of source nodes of the node pair where node i is a target node is defined as:

[0127] M i = {m | node pair mi, where m is the source node and i is the target node}. (Equation 10)

[0128] Semantic neighborhood function s ij (t) is represented as

[0129]

[0130] |M i | represents M i The number of elements in The change in the conditional intensity function transition of node pair ij caused by node pair excitation mi is represented by the exponential kernel function. To describe the time decay property, t m n represents the time period before time t. mi (t n ) indicates that during the time period t m Count the number of events that occur on node pair mi.

[0131] (1.3) Temporal Neighborhood Function

[0132] Research has found that the number of events occurring in a network on a given day is similar to the number of events occurring in the previous few days. Therefore, this invention considers the influence of a node on the count of all historical events in the previous few days, i.e., the time neighborhood function d. ij (t) is represented as:

[0133]

[0134] Among them, D i This represents a pre-defined set of dates, specifically the set of dates on which a daily or weekly influence pattern is assumed to exist for the occurrence of events on node ij on that day. This could be the previous week or the previous few days. |D i | represents D i The number of elements in This represents the cardinality of the time neighborhood influence of nodes on ij from the previous few days, using an exponential kernel function. To describe the time decay property, t d Indicated in Di In the date d, the time period before time t, Indicates the time period t d The internal node counts the number of events occurring at position ij.

[0135] (2) Parameter estimation

[0136] Based on the detailed introduction of the spatiotemporal interactive Hawkes process in Section (1), the conditional intensity function (CIF) λ from one node u to another node j on the network... ij (t) can be summarized as follows:

[0137]

[0138] Rewrite the model parameters in matrix form, including the background rate basis term. Influence on incentives Impact on basic terms The influence of the time neighborhood (the previous day and the previous week) is the basic term. Where i,j=1,…,n. Therefore, it is necessary to estimate Θ. (0) Θ (1) α (1) , Suppose the network has T consecutive time intervals of available count event data. The log-likelihood function of the proposed Hawkes process can be expressed in the following form:

[0139]

[0140] in, Using numerical approximation integration, i.e.

[0141] The model parameters are estimated using a combination of the expectation-maximization (EM) algorithm and least squares estimation. First, given α... (1) The EM algorithm is used to maximize the log-likelihood function described above. In the expectation (E) step, a random variable p is defined. ij p(t) represents the probability that the event occurring at node pair ij at time t is a background event. ij (t), and This represents the probability that an event occurring at time t, corresponding to a geographic, semantic, and temporal neighborhood function, is caused by an event that occurred at a previous time τ.

[0142] p ij (t), and Approximately expressed as:

[0143]

[0144]

[0145]

[0146]

[0147]

[0148] In the maximization (M) step, the number of events generated by geographical, semantic, and temporal neighborhoods is first estimated, i.e. Where i,j=1,…,n. Parameter Θ (0) Θ (1) and α (2) The update is performed by minimizing the three optimization problems resulting from the following decomposition:

[0149]

[0150]

[0151]

[0152] consider In scenarios that have already occurred in the geographical, semantic, and temporal neighborhoods, and using H ij K ij and M ij Representing the corresponding set of scenes, denoted by D ij This represents the total historical demand for the previous year, which can be obtained by solving equation 16. and Updated to:

[0153]

[0154]

[0155] and After the update, the least squares method is used to estimate... First calculate y ij (t) and y ij (t):

[0156]

[0157] Then, Updated to:

[0158]

[0159] Among them, X ij =[x ij (t1),…,x ij (tT )] T Y ij =[y ij (t1),…,y ij (t T )] T .

[0160] This application also provides a method for predicting traffic incidents, including the following steps:

[0161] Obtain an event-based dynamic network, wherein the event-based dynamic network has n nodes in the time domain;

[0162] Obtain the historical event set of the dynamic network before time t. The historical event set It stores the number of events that occur between node pairs in a dynamic network;

[0163] Use the modeling method described above to predict the occurrence rate of events between any two nodes at time t in a dynamic network;

[0164] The event prediction result is obtained based on the occurrence rate of events between any two nodes at time t in a dynamic network. For example, when predicting the number of events between any two nodes based on the occurrence rate of events, a threshold can be designed. When the occurrence rate is greater than the threshold, the event is considered to have occurred and counted. Alternatively, other methods can be used to calculate the event count result.

[0165] A traffic incident prediction system, comprising:

[0166] A dynamic network acquisition module is used to acquire an event-based dynamic network, wherein the event-based dynamic network has n nodes in the time domain;

[0167] The historical event set acquisition module is used to acquire the historical event set of the dynamic network up to time t. The historical event set It stores the number of events that occur between node pairs in a dynamic network;

[0168] The modeling module is used to predict the occurrence rate of events between any two nodes at time t in a dynamic network using the modeling methods described above.

[0169] The output module is used to obtain the event prediction result based on the occurrence rate of events between any two nodes at time t in the dynamic network.

[0170] Example explanation:

[0171] To demonstrate the validity of this application, this embodiment conducts a case study using the New York yellow taxi open-source dataset. This dataset contains all yellow taxi trips from the past few years, recording the start and end dates and times, pick-up and drop-off locations, trip distance, and payment information for each trip. To illustrate the main idea of ​​the proposed method, this embodiment studies the transportation network comprised of 19 of Manhattan's busiest areas, treating a single taxi order from destination to origin as an event, and dynamically modeling the daily number of taxi orders on this transportation network. Figure 2 This is a schematic diagram of the traffic network, where each numbered region represents a node in the model, i.e., n=19. This case study collected traffic network information with the origin as the source node and the destination as the target node. For the time range, the number of orders per hour for each origin-destination (OD) node in taxi trips within a day was counted. Data from the first three weeks and the last week of February 2022 were used as training and testing data, respectively. The training and testing data contained 743,292 and 259,771 taxi orders, respectively. Accumulated on an hourly basis, the above data were converted into 181,944 counts (i.e., 24 hours × 21 days = 19). 2 (for OD nodes) and 60,648 counts (i.e., 24 hours, 7 days, 19) 2 For OD nodes).

[0172] This example considers the three nearest geographical neighbors for each node. Model parameters are estimated using Equation 14-20. The estimated values ​​of the parameters are as follows: Figure 3 As shown. This represents the background rate baseline for a given day, taking into account that almost no orders are generated after midnight. The estimated values ​​are all 0; and These represent the influencing incentive terms and the influencing base terms, and they are spatially correlated; α (2) It represents the basis term of the influence of a node in the temporal neighborhood.

[0173] The estimated parameters are substituted into the method proposed in this invention to predict the event occurrence rate, thereby calculating the counting result. To measure the accuracy of the count prediction, the prediction error is defined as the predicted count value of all nodes in the network. and the actual count value n ij (t) The average of the absolute values ​​of the differences over a day, i.e.:

[0174]

[0175] The prediction error of the present invention is 1.85. Figure 4Further examples of traffic network event number prediction results from embodiments of the present invention are shown. Subgraphs (a) and (b) represent the comparison between the predicted results and actual data for node (6, 13) on the first and second days, respectively; subgraphs (c) and (d) represent the comparison between the predicted results and actual data for node (14, 14) on the third and fourth days, respectively; and subgraphs (e) and (f) represent the comparison between the predicted results and actual data for node (0, 16) on the sixth and seventh days, respectively. The figures show that actual taxi orders do indeed exhibit a high degree of temporal correlation as expected. The method of the present invention utilizes geographical, semantic, and temporal neighborhood information obtained from training data and historical observation data of OD node pairs to enable the prediction results to accurately characterize the dynamic changes of each node in the network.

[0176] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. An event-driven traffic network modeling method, characterized in that, Includes the following steps: Obtain an event-based dynamic network, wherein the event-based dynamic network has n nodes in the time domain; Obtain the historical event set of the dynamic network before time t. The historical event set It stores the number of events that occur between node pairs in a dynamic network; The occurrence rate λij(t) of events between node pairs ij at time t is described by the conditional strength function, thus obtaining λ. ij The expression for (t) is as follows: Where, N ij (t) represents the number of events occurring between node pairs ij within the time interval (-inf, t]. The conditional strength function describes the expected occurrence rate of node pairs ij, i, j = 1, ..., n at time t, based on past events. Using a spatiotemporal interactive Hawkes process to measure the occurrence rate λ of events from node i to node j ij (t) is modeled as follows: in, It is the background rate basis term, μ ij (t) describes the impact of historical events in a dynamic network on events occurring between node pairs ij at time t; For μ ij (t) is decomposed to obtain: μ ij (t)=g ij (t)+s ij (t)+d ij (t) Among them, g ij (t), s ij (t) and d ij (t) represents the geographic neighborhood function, semantic neighborhood function, and temporal neighborhood function of node pair ij at time t, respectively, that is, the influence of the historical events of geographic neighborhood node pair, the historical events of semantic neighborhood node pair, and the historical events of temporal neighborhood node pair on the events occurring between node pair ij at time t; Based on historical event sets And the structure of the dynamic network determines λ ij The values ​​of each term in (t) are used to obtain the occurrence rate λ of events between node pairs ij at time t. ij The model is completed by determining the value of (t).

2. The event-driven traffic network modeling method according to claim 1, characterized in that, Node pair ij represents the event from node i to node j, and the set of geographic neighbors of the source node i is: H i ={h|dis(i,h)≤G} The set of geographical neighbors of target node j is: K j ={k|dis(k,j)≤G} Where dis(·) is the straight-line distance between two nodes, and G is the preset neighborhood range threshold; The geographical neighborhood node pairs of node pair ij include first-type node pairs and second-type node pairs. The first-type node pairs are node pairs hj, h∈H. i The second type of node pair is node pair ik, k∈K j ; The expression for the geographic neighborhood function is: in, This represents the impact of all historical events on the first type of node pair. This represents the impact of all historical events on the second type of node pair; This represents the change in the conditional intensity function of ij caused by the excitation of hj by the node. This represents the transition change of the conditional intensity function of node pair ij generated by node pair ik; in the first type of node pair, an exponential kernel function is used. To describe the time decay property, t h n represents the time period before time t. hj (t h ) indicates that node pair hj is in time period t h The number of events occurring within the node is counted; in the second type of node pair, an exponential kernel function is used. To describe the time decay property, t k n represents the time period before time t. ik (t k ) indicates that during the time period t k The internal node counts the number of events occurring at point ik.

3. The event-driven traffic network modeling method according to claim 2, characterized in that, Will Represented as the cardinality term of the influence of node on ij Incentive terms from the influence of nodes on hj And the weight w for measuring the similarity between source node i and its geographical neighbor node h. ih The product is as follows: The weights h between the source node i and its geographical neighbors are defined using a distance-based softmax normalized exponential function as follows: Among them, the closer the geographical locations of two nodes are, the greater their weight, indicating that the similarity of the events occurring at these two nodes is higher; Will Represented as the cardinality term of the influence of node on ij Incentive terms derived from the influence of nodes on ik And the weight w that measures the similarity between the target node j and its geographical neighbor node k. kj The product is as follows: The weights between the source node i and its geographical neighbors are defined using a distance-based softmax normalized exponential function as follows: The closer two nodes are geographically, the greater their weight, indicating a higher similarity between the events occurring at these two nodes.

4. The event-driven traffic network modeling method according to claim 3, characterized in that, Node pair ij represents the event from node i to node j, and the semantic node set of the source node i is: M i = {m | node pair mi, where m is the source node and i is the target node} The semantic neighborhood node pairs of node pair ij are node pairs mi, m∈M i ; The expression for the semantic neighborhood function is: Among them, |M i | represents M i The number of elements in The change in the conditional intensity function transition of node pair ij caused by node pair excitation mi is represented by the exponential kernel function. To describe the time decay property, t m n represents the time period before time t. mi (t m ) indicates the time period t m The number of events that occur on node pair mi.

5. The event-driven traffic network modeling method according to claim 4, characterized in that, The expression for the time neighborhood function is: Among them, D i Represents a pre-defined set of dates, |D i | represents D i The number of elements in The cardinality of the temporal neighborhood influence of nodes in the date set on ij is represented using an exponential kernel function. To describe the time decay property, t d Indicated in D i In the date d, the time period before time t, Indicates the time period t d The internal node counts the number of events occurring at position ij.

6. The event-driven traffic network modeling method according to claim 5, characterized in that, The occurrence rate λ of events between node pairs ij at time t is described by the conditional strength function. ij (t), through a spatiotemporal interactive Hawkes process, λ is obtained. ij The model of (t) can be summarized as follows: μ ij (t)=g ij (t)+s ij (t)+d ij (t) The model parameters are rewritten in matrix form, including the background rate basis term. Influence on incentives Impact on basic terms The influence of time neighborhood is the basic term. Where i, j = 1, ..., n, the parameter to be estimated is Θ. (0) Θ (1) α (1) , 7. The event-driven traffic network modeling method according to claim 6, characterized in that, The historical event set There are a total of T consecutive time periods with available count event data, and the above λ ij The log-likelihood function of the model (t) is: in, Using numerical approximation integration, i.e. The method combines the expectation-maximization algorithm with the least squares estimation to estimate the parameters to be estimated, where Θ... (0) Θ (1) α (2) Parameter estimation includes the following steps: In the expectation step of the expectation-maximization algorithm, a random variable p is defined. ij (t) represents the probability that the event occurring on node pair ij at time t is a background event. and p represents the probability that an event occurring at time t, corresponding to a geographical, semantic, and temporal neighborhood function, is caused by an event occurring at a previous time τ. ij (t), and Approximately expressed as: In the maximization step of the expectation-maximization algorithm, the number of events generated by geographical, semantic, and temporal neighborhoods is first estimated, i.e. Parameter Θ (0) Θ (1) and α (2) The update is performed by minimizing the three optimization problems resulting from the following decomposition: consider In scenarios that have already occurred in the geographical, semantic, and temporal neighborhoods, and using H ij K ij and M ij Representing the corresponding set of scenes, denoted by D ij Let represent the total historical demand of the previous year. Solve the above three optimization problems. and Updated to:

8. The event-driven traffic network modeling method according to claim 7, characterized in that, For α (1) The specific steps for parameter estimation are: to... and After the update, the least squares method is used to estimate... First calculate y ij (t) and x ij (t): Updated to: where, X ij = [x ij (t1),..., x ij (t T )] T , Y ij = [y ij (t1),..., y ij (t T )] T .

9. A traffic incident prediction method, characterized in that, Includes the following steps: Obtain an event-based dynamic network, wherein the event-based dynamic network has n nodes in the time domain; Obtain the historical event set of the dynamic network before time t. The historical event set It stores the number of events that occur between node pairs in a dynamic network; Predict the occurrence rate of events between any two nodes at time t in a dynamic network using the modeling method described in any one of claims 1-8; The event prediction result is obtained based on the occurrence rate of events between any two nodes at time t in the dynamic network.

10. A traffic incident prediction system, characterized in that, include: A dynamic network acquisition module is used to acquire an event-based dynamic network, wherein the event-based dynamic network has n nodes in the time domain; The historical event set acquisition module is used to acquire the historical event set of the dynamic network up to time t. The historical event set It stores the number of events that occur between node pairs in a dynamic network; A modeling module for predicting the occurrence rate of events between any two nodes at time t in a dynamic network using the modeling method described in any one of claims 1-8; The output module is used to obtain the event prediction result based on the occurrence rate of events between any two nodes at time t in the dynamic network.

Images