Road intersection congestion prediction method based on time point process neural network model

Abstract

The present invention discloses a method for predicting road intersection congestion based on a time point process neural network model. It includes the following steps: first, in the spatial association modeling process, a spatial association module is constructed by fusing the spatial regional congestion change patterns of multiple intersections into a single intersection level; second, in the dual-granularity temporal association modeling process, the temporal granularity of congestion events is captured through the time point process, and further integrated with the gated recurrent network unit to construct a new neural point process gated recurrent unit. Through the new neural point process gated recurrent unit, congestion is modeled at different time granularities to obtain a dual-granularity temporal association module; finally, a spatiotemporal point process neural network model is established based on the sequence architecture and spatiotemporal association module, and parameter optimization is performed to achieve multi-step prediction of congestion events. The present invention has the advantages of high prediction accuracy for refined urban intersection lane-to-lane and signal cycle congestion event prediction.

Description

Technical Field

[0001] The present invention relates to the field of urban traffic prediction, and more specifically to a road intersection congestion prediction method based on a time point process neural network model. Background Art

[0002] Congestion is a complete non-rigid spatiotemporal event, and its development and changes are different from the short-term regression prediction of traditional traffic parameters. At present, most studies on congestion prediction focus on free highways or intersection environments without signal control. Although there are some studies on urban congestion prediction, there are not many studies on the refined prediction of congestion events at the lane level and signal cycle level of urban signal-controlled intersections. Compared with highways, fine-grained traffic congestion prediction at urban intersections is of great significance in many advanced intelligent transportation system applications. Despite the continuous advancement of traffic congestion prediction technology, traffic congestion prediction at fine-grained signal intersections remains a challenging task. The difficulties mainly lie in the following two aspects:

[0003] There are two problems in the refined spatial correlation modeling: first, the spatiotemporal evolution of the congestion morphology of refined signal-controlled intersections is affected by the current signal control strategy; in addition, affected by the complex topology between the intersection and its upstream and downstream, the congestion pattern of the local intersection will be affected by the overall traffic pattern at the regional level. The spatiotemporal evolution process of the congestion morphology is not like the linear distribution of the highway, but presents a planar diffusion, and the traffic congestion pattern presents a complex spatial dynamic correlation. Therefore, it is difficult to accurately describe the dynamic correlation of congestion between the lane spaces of the intersection. However, existing research mainly focuses on intersection modeling, or modeling road networks through intersections, and rarely considers the spatial dependence of regional road network structures on local intersections. Secondly, considering that local signal-controlled intersections are also affected by external signal control strategies, different signal control strategies determine the connection strength between lanes and the connectivity of traffic flows. However, most existing methods rarely consider the impact of external signal control strategies on the lane-level spatial modeling of intersections. In terms of refined temporal modeling. First, such as Figure 2 As shown, the occurrence of congestion events is a "jumping" timestamp, a discrete sequence distributed within continuous time, which is inconsistent with the temporal granularity of the "continuous" traffic flow parameter sequence. Therefore, it is difficult to uniformly express the two time granularities. It is necessary to establish a reasonable relationship between the "jumping" time granularity of congestion events and the "continuous" time granularity of traffic flow parameters.

[0004] Therefore, in the dual-granularity modeling of traffic congestion time, not only the signal cycle time granularity of traffic flow parameters but also the time granularity of the traffic congestion event itself should be considered.

[0005] Existing urban intersection congestion prediction models have three drawbacks when it comes to accurately predicting congestion events between lanes and within signal cycles at urban intersections. 1) Their level of refinement is insufficient, failing to consider the characteristics of signal-controlled intersections. Modeling spatiotemporal correlations at the signal cycle and lane levels results in poor accuracy. 2) Model fusion is weak. Most deep learning prediction models are purely data-driven and lack integration with traditional traffic models, resulting in slow model convergence and poor accuracy. 3) Research examining congestion as an event-based problem lacks research, resulting in a failure to establish a reasonable prior structure for modeling the spatiotemporal correlations of congestion events and extracting spatiotemporal information about congestion.

[0006] Therefore, it is necessary to develop a method for high-precision prediction of congestion events between lanes and within signal cycles at refined urban intersections. Summary of the Invention

[0007] The purpose of the present invention is to provide a road intersection congestion prediction method based on a time point process neural network model, which can predict congestion events between lanes and within signal cycles of refined urban intersections with high prediction accuracy.

[0008] To achieve the above objectives, the technical solution of the present invention is: a road intersection congestion prediction method based on a time point process neural network model, characterized by: based on time point process theory and deep learning methods, using a spatio-temporal neural point process (STNPP) neural network model to predict the full life cycle of the occurrence, development and dissipation of intersection lane-level congestion events at the fine time scale of the signal light cycle.

[0009] The specific method is:

[0010] First, during spatial correlation modeling, the spatial regional congestion change patterns of multiple intersections are integrated into a single intersection level to construct a spatial correlation module. This assists in fine-grained congestion prediction at the intersection lane level and at the signal cycle level, achieving dynamic correlation at the local intersection and regional road network levels. Furthermore, when modeling spatial correlation at the signal-controlled intersection level, the signal cycle changes and the graph convolutional network of the local intersection are integrated to capture the dynamic spatial correlation at the lane level of the local intersection under a fixed signal cycle.

[0011] Secondly, in the dual-granularity temporal association modeling process, the occurrence of congestion events is a "jumping" timestamp, which is a discrete distribution in continuous time and inconsistent with the continuous signal cycle time granularity of traffic flow. Therefore, the temporal granularity of congestion events is captured through the time point process, and further integrated with the gated recurrent network unit (GRU) to construct a new neural point process gated recurrent unit (NPPGRU). Congestion is modeled at different time granularities, resulting in a dual-granularity temporal association module. The spatiotemporal association module is constructed by combining the spatial association module and the dual-granularity temporal association module.

[0012] Finally, based on the sequence architecture and spatiotemporal association module, a spatiotemporal point process neural network model (i.e., intersection congestion event prediction model) is established, and parameter optimization is performed to achieve multi-step prediction of congestion events.

[0013] In the above technical solution, the specific prediction method of the road intersection congestion prediction method based on the time point process neural network model includes the following steps:

[0014] Step 1: Organize basic information;

[0015] Step 2: Define congestion events;

[0016] Step 3: Spatial correlation modeling based on intersection graph structure;

[0017] Step 4: Temporal correlation modeling of integrated time point processes;

[0018] Step 5: Constructing an intersection congestion event prediction model based on the sequence-to-sequence spatiotemporal correlation module;

[0019] Step 6: Loss function representation of the model;

[0020] Step 7: Output the prediction results.

[0021] In the above technical solution, in step 1, the basic data includes traffic speed data, geometric data of the detection road network and signal timing data of all intersections in the study area;

[0022] Traffic speed data is the average of the instantaneous speed of lane-level floating vehicles within each signal cycle. For example, speed and signal timing data for 30 days from December 1, 2018 to December 31, 2018 were extracted from a certain city's traffic data set. The maximum number of events in a day was 4235, and the minimum number of events was 2354.

[0023] The geometric data of the detected road network describes the number of lanes, the geometry of each lane, the topological connections between lanes, and the turning restrictions at each intersection;

[0024] The signal timing data for all intersections in the study area include cycle length and effective green time. Figure 1 As shown in Figure 1, the road network under study consists of 18 intersections and 154 interconnected lanes.

[0025] In the above technical solution, in step 2, a congestion event represents a set of congestion occurrence times and values ​​at a specific intersection. According to the city's speed limit regulations for specific road intersections, a traffic state where the speed is less than the standard speed at a certain moment is defined as a congestion event, and the timestamp (time) and speed (value) of the event are used as the basic information of the event.

[0026] Congestion event sequence: Based on the definition of congestion events, congestion event sequence refers to the set of congestion events according to the time of occurrence, expressed as e s ={(t1, m1), (t2, m2),…, (t n , m n )}, n represents the random variable of the number of times the event occurs, where 0 <t1<t2<…<t n The time t when the congestion event occurs i ∈(0, T], s represents a fixed spatial position, for example, a fixed lane position at an intersection. To further illustrate the difference between congestion event sequences and urban traffic flows, as shown in Figure 2 As shown in Figure 3, the occurrence timestamps of asynchronous congestion event sequences are discretely distributed within equal time intervals of the continuously changing traffic flow. The time granularity of the two is different. Generally, the time change scale of events is irregular and larger than the time change granularity of traffic flow.

[0027] The specific steps for defining a congestion event are as follows:

[0028] 1) The set of congestion events of the jth lane at the i-th intersection is counted as Wherein, if no congestion event occurs at a certain time, it represents {0, 0}, t c ∈(0, P] represents the timestamp of the cth congestion event; Indicates the congestion event value information (congestion level, speed and other traffic status characteristics, which refers to speed in this invention); m j ∈[1, 2, …, K], K∈Z + represents the cumulative number of congestion events that occurred in the historical time interval (0, P]; U represents the union;

[0029] 2) Define the historical interval (0, P] signal cycles, and count the set of all congestion events for M lanes of a fixed intersection i as

[0030] Among them, if no congestion event occurs at a certain time, it represents {0, 0}, t c∈(0, P] represents the timestamp of the cth congestion event; therefore, the input historical time feature is P event timestamps t c / 0 constitutes a continuous time series; m j ∈[1, 2, …, K], K∈Z + represents the cumulative number of congestion events that occurred within the historical time interval (0, P]; therefore, within the historical interval (0, P] signal cycles, for a fixed intersection i with M lanes, the set of all congestion events is:

[0031] In the above technical solution, in step 3, the intersection congestion event prediction model accurately predicts the occurrence time and value of traffic congestion events at the signal cycle level and lane space level of the signal-controlled intersection; Figure 3 The basic structure of the intersection congestion event prediction model is presented;

[0032] The intersection congestion event prediction model mainly consists of two parts, including the spatial correlation module and the dual-granularity temporal correlation module;

[0033] Among them, in the spatial correlation modeling stage, the intersection congestion problem is affected by the combined influence of internal and external factors. The internal factors include the congestion event feature set E 0→P and traffic flow characteristics χ g The external factors are mainly affected by the dynamic changes of signal control strategies and the impact of regional road network congestion change patterns on local intersections; in formulas (2) and (14), E 0→P and χ g To calculate spatial association;

[0034] The global regional road network spatial correlation modeling method under the influence of external factors is as follows: among the external spatial dependence influences, the congestion pattern of local intersections will be affected by the overall regional traffic pattern, and the traffic congestion pattern will show complex spatial dynamic correlation;

[0035] Taking advantage of the graph convolutional network modeling graph structure, for each signal cycle, the intersection is used as the basic unit of the regional road network graph structure, and the graph convolutional network is used to input N intersections of the regional road network graph structure. Modeling spatial dependencies, we can obtain the global spatial correlation formula between regional intersections:

[0036]

[0037] in, and are the l and l+1 layers of the regional road network graph convolutional layer input data blocks; U g is an orthogonal matrix; g(·) is a polynomial kernel function acting on the diagonal matrix Λg ∈R N×N ; For the regional road network graph G g , the graph Laplace matrix L can be decomposed into U g g(Λ g )(U g ) T ;

[0038] The spatial structure information between the embedded regional road network intersections is obtained through the global spatial correlation formula between regional intersections, and the historical P signal cycles of the regional road network are calculated to obtain

[0039] Through index query, the index of target intersection n is selected from N intersections, which is characterized by Further obtain the regional road network impact characteristics of P signal cycles of all lanes at the intersection As the temporal and spatial correlation input of local intersections, it assists in modeling the impact of internal factors on congestion;

[0040] Considering that intersection congestion is affected by external factors of signal control strategies, different signal control strategies determine the connection strength between upstream and downstream lanes of the intersection and the connectivity of traffic flow; in order to take into account the external factors affecting congestion, combined with the application environment of signal-controlled intersections, the lane-level external signal control strategy changes are modeled as spatial associations. Studies have shown that by modeling external factors as adaptive matrices, the dynamic spatial dependencies hidden in some traffic tasks can be significantly captured. Therefore, the specific modeling method is: given a specific intersection with M lanes in a traffic network, through the dynamic changes of signal control strategies, an external signal control perception transfer matrix A∈R is constructed. M×M , expressing the correlation strength between spatial lanes in each signal cycle time slice, and further superimposing the signal control perception transfer matrix in the traffic history P time period to form A s ∈R M×M×P ;

[0041] Among them, the matrix A s ∈R M×M×P At different time slices, the correlation strength between lanes is quantified by the different sizes of signal periods between upstream and downstream lanes.

[0042] In the above technical solution, for local intersections, it is difficult to establish an explicit expression model due to the influence of internal factors on lane-level spatial modeling; therefore, the method of spatial correlation modeling based on intersection graph structure also includes: using the attention mechanism to integrate the congestion internal factor event feature set E 0→P and traffic flow characteristics χ lCombine and calculate the weights between lanes. Different lanes are assigned different weights. The weights are calculated by combining the combined effects of internal factors to express the joint impact on the current congested lanes. The specific method is as follows;

[0043] Assume that there is traffic congestion observation data of historical P signal cycles, and the traffic flow of M lanes at the intersection is l and event E 0→P As the input of the neural network layer, the input weight matrix is ​​calculated in each signal cycle:

[0044]

[0045] Among them: σ represents the activation function; “|” represents the connector, b l ∈R M×M Network learning parameters; the input features of congestion events are event time and scalar value;

[0046] For the i-th lane where congestion occurs, its spatial correlation α with adjacent lanes and upstream and downstream lanes i，j According to A l Calculation yields:

[0047]

[0048] Where j is the index of the lane associated with the i-th lane, α i，j Indicates the spatial correlation strength between lanes j and i in the intersection;

[0049] The geometric topological correlation between lanes at local signal-controlled intersections is affected. This spatial physical correlation is crucial for information transfer between congested lanes. Similar to the regional road network graph structure, a graph convolutional network is used to model the local intersection graph structure. The aforementioned attention mechanism and signal-controlled perception transfer matrix are integrated into the traditional graph convolutional network to adaptively capture the dynamic spatial dependency between the current intersection lane and the upstream and downstream intersection lanes. Within each signal cycle slice, a graph convolution operator is used to capture the spatial correlation between intersection lanes. The formula is as follows:

[0050] X l+1 =ReLU(Ug(Λ)U T x l ) (4)

[0051] Among them, X l+1 and X l are the l and l+1 layers of the graph convolutional network layer input data blocks, (Note: Historical event data is used as spatial attention input to calculate the spatial impact between lanes, but it is not used as the input layer of spatial convolution). Represents the global regional road network influence at time t; U is an orthogonal matrix; In order to capture the dynamic lane-level spatial correlation between adjacent intersections, it is proposed to calculate the spatial attention weight and the signal control perception transfer matrix A∈R M×M Integrating into the above graph convolution, Equation (4) is rewritten as:

[0052]

[0053] Here, ⊙ is the Hadamard product operator, or point-by-point multiplication. This integrated design offers two benefits: 1) By explicitly accounting for the impact of external signal control strategies on spatial modeling, it improves the graph convolutional network's sensitivity to spatial correlations between upstream and downstream lanes; 2) it enables the model to measure the interaction between upstream and downstream lanes at an intersection, enhancing its interpretability.

[0054] In the above technical solution, the time correlation modeling of the integrated time point process (i.e., the dual time granularity correlation modeling of urban signal-controlled intersections) is carried out. The specific method is as follows:

[0055] After capturing the dynamic lane-level correlation between adjacent intersections in space, temporal modeling focuses on a specific lane to study the temporal correlation between congestion in the current observation cycle and its adjacent signal cycles. Considering that traffic congestion is affected by multiple time granularities, the traffic time series is divided into two time granularities: a "jumping" irregular time granularity based on the interval between congestion events and a "continuous" regular signal cycle time granularity for traffic flow parameters.

[0056] For a specific (i.e., a given) intersection n, the mth lane, given a set of discrete event lists that occurred within the historical time P

[0057] Among them, t c ∈(0,P] represents the timestamp when the c-th congestion event is obtained, and it is assumed that the first observation is made at timestamp 0;

[0058] Introducing a mask vector To express t c Whether a congestion event occurs at a certain time, its value is 1 / 0 respectively, which is expressed by the following formula:

[0059]

[0060]

[0061] Where: t cIndicates the timestamp of the congestion event; represents the mask vector; Indicates the time interval between adjacent consecutive congestion events;

[0062] In order to capture the temporal correlation patterns of congestion events at dual time granularity, this paper designs an adaptively learnable prior intensity function model and integrates the model into the traditional gated recurrent network (GRU). The formula is:

[0063]

[0064] Among them, [″|″] represents the connection symbol, are model learning parameters; represents the cumulative vector of congestion events;

[0065] The selection of the external function λ(·) needs to consider two key criteria: 1) the intensity function needs to be positive; 2) as mentioned above, the intensity function should vary with the time interval and the cumulative number of congestion events.

[0066] After obtaining the results (i.e., obtaining the temporal correlation pattern of congestion events at dual time granularity), in order to model the temporal correlation between historical congestion events and the cumulative number of congestion events, we multiply the elements by λ(t c ) to update the temporal features and hidden states of the original gated recurrent unit input It is then used as a priori model and combined with the GRU model to form the Neural Point Process Gated Recurrent Unit, NPPGRU;

[0067] NPPGRU is updated as follows:

[0068]

[0069]

[0070]

[0071]

[0072]

[0073] Among them, u represents the update gate, r represents the reset gate, Candidate hidden variables, σ activation function (sigmoid), W r , W u ,U r , U u , U and Learning weight parameter, b r , b u and Indicates the bias term. [⊙] indicates dot product, and [*] indicates matrix product. Figure 4 As shown, the present invention compares the similarities and differences between the neural point process gated recurrent unit and the traditional gated recurrent unit;

[0074] The mask vector based on the congestion event Directly input the neural point process gated recurrent unit model to assist in the temporal correlation modeling of events and obtain the temporal correlation model.

[0075] In the above technical solution, in step 6, the loss function of the model is expressed (the model of the present invention refers to the spatial association module and the dual-time module part), and the specific method is:

[0076] Given a dataset containing historical, regional and intersection-level χ g , χ l A set of time-irregular traffic states, a discrete list of historical and future time-irregular congestion events t c ∈(0, P] and By setting k c =1 means congestion occurs at the current moment, otherwise it is 0 (i.e. if congestion does not occur at the current moment, Kc is 0); by maximizing The joint log-likelihood estimation updates the model parameters, and the loss function of the model is expressed as:

[0077]

[0078] In the above technical solution, the time association module is based on the sequence-to-sequence framework structure. Through recursive calculation, it learns the encoding representation of historical congestion events and generates the output of future multi-step congestion events by selecting the optimal parameter θ (Formula (14)) with the maximum conditional probability; Figure 5 As shown in the figure, during the process of learning congestion patterns with a temporal association model, event and traffic flow data at multiple historical moments are input into stacked NPPGRU units to achieve information encoding at dual time granularity. In the decoding prediction stage, the output of the previous NPPGRU unit serves as the input of the next unit to calculate whether congestion will occur and the congestion value at multiple future moments.

[0079] The present invention has the following advantages:

[0080] The present invention first models spatial associations by fusing the spatial regional congestion change patterns of multiple intersections to a single intersection level, and assisting in fine-tuning congestion predictions between lanes at intersections and at a periodic granularity, thereby resolving dynamic associations at local intersection and regional road network levels. Then, when modeling spatial associations at the signal-controlled intersection level, the signal periodic changes and the graph convolutional network of the local intersection are integrated to capture the dynamic spatial associations at the lane-level granularity of the local intersection under a fixed signal period. Finally, in the temporal association modeling process, considering that the occurrence of congestion events is a "jumping" timestamp, which is a discrete distribution in continuous time and inconsistent with the continuous signal period time granularity of the traffic flow, the temporal granularity of the congestion event is captured through a time point process model, and further integrated with a gated recurrent network unit, a new neural point process gated recurrent unit is proposed to model congestion at dual time granularities. In order to consider the spatial impact of remote intersections on locally congested intersections, the present invention incorporates changes in intersection congestion patterns within the region into local intersection modeling, thereby improving prediction accuracy.

[0081] The present invention develops an adaptive matrix that can flexibly perceive signal period changes to model dynamic spatial correlations in congestion. It can significantly capture the dynamic spatial dependencies hidden in some traffic tasks and improve prediction accuracy.

[0082] The present invention develops a new dual-time granularity model that integrates continuous traffic flow and discrete congestion event information. The continuous time granularity information is integrated into the discrete event granularity through a recurrent neural network. In this way, the congestion prediction model can not only process the "macro" trend of congestion events at the macro time granularity, but also capture more details and changes between congestion events at the "micro" signal cycle time granularity, thereby improving the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Figure 1 This is the road network structure diagram in the experiment of the present invention.

[0084] Figure 2 This is the temporal dual-scale distribution of a series of congestion events within one day at a fixed location in the present invention.

[0085] Figure 3 This is the STNPP model for refined congestion prediction at signal-controlled intersections proposed in this invention.

[0086] Figure 4 A comparison diagram of the neural point process gated recurrent unit in the present invention and the traditional gated recurrent unit.

[0087] Figure 5 This is a diagram of the neural network structure based on the dual-time granularity neural point process gated recurrent unit (NPPGRU) of the present invention.

[0088] Figure 6This is a comparison chart of the congestion prediction results for the northbound lane of the #767 intersection in an embodiment of the present invention.

[0089] exist Figure 1 In the figure, circles represent intersections, and five-pointed stars represent selected intersections where the present invention was experimented (i.e., six multi-lane intersection locations to be evaluated); numbers 1044, 1043, 1042, 771, 768, and 767 represent the numbers of the intersections.

[0090] exist Figure 2 In the figure, the height of the arrow-shaped bar represents the congestion index; the horizontal axis represents the cycle.

[0091] exist Figure 3 In, F g , F l = represents the number of traffic state characteristics for the region and intersection, respectively; T represents the number of signal cycles; N and M represent the number of intersections and lanes in the region, respectively. External factors refer to the spatial influence of the regional road network and the influence of signal cycle variations; internal factors refer to the characteristics of congestion events between lanes and between cycles at the intersection and the influence of traffic flow variations. In the output, solid dots represent lanes, and dots with crosses represent the current predicted lane results. Together, they form the local intersection graph structure. Congestion event prediction outputs the congestion event value (speed) and event time category (whether congestion occurred in each signal cycle) for multiple future signal cycles.

[0092] Figure 4 Figure (a) shows the neural point process gated recurrent unit (NPPGRU unit diagram); Figure (b) shows the traditional gated recurrent unit (GRU unit).

[0093] Figure 4 This paper demonstrates the differences between the NPPGRU and traditional GRU units. Compared to traditional GRU units, the present invention incorporates modified NPPGRU units based on the time-point process of congestion events. The integrated NPPGRU units not only leverage the original GRU unit's ability to model temporal nonlinearity, but also improve the predictive power of traditional time-point process models. This modeling approach allows the model to not only capture the global trends of congestion events at the macro-event time granularity, but also capture more details and variations of congestion at the micro-traffic flow time granularity.

[0094] exist Figure 5 In the example, the input discrete event time series set {(t1, y1), (t2, y2), ..., (t p ,y p )} and the continuous regular traffic flow parameter sequence [X1, X2, ..., X p]; After stacking NPPGRU encoding, the implicit history sequence representation C is obtained. Next, by decoding the sequence, the scale value and time series [Y1, Y2, ..., Y m ],Y i =(t i ,y i ); Y0, C, ..., Ym-1, C express congestion events, which is a tuple representation;

[0095] exist Figure 5 In the figure, S1 represents the input layer; S2 represents the output layer; S3 represents the encoding layer; and S4 represents the decoding layer.

[0096] Figure 6 Figure (a) is the STNPP prediction result diagram; Figure (b) is the NP-STNPP (indicates the removal of the modifications on the time association modeling GRU, that is, the weakened STNPP model) prediction result diagram; Figure (c) is the ASTGCN (existing method model) prediction result diagram. DETAILED DESCRIPTION

[0097] The following detailed description of the embodiments of the present invention is given in conjunction with the accompanying drawings, which do not limit the present invention but are merely examples. The description makes the advantages of the present invention clearer and easier to understand.

[0098] Example

[0099] The present invention is now described in detail by taking the application of the present invention to a certain road intersection for congestion prediction of a northbound lane of a #767 intersection as an example. The present invention also has a guiding role in applying the present invention to congestion prediction at other road intersections.

[0100] This example uses the method of the present invention to predict congestion in the northbound lanes of a #767 intersection. The specific method is as follows: refined spatiotemporal correlation modeling at the signal cycle and lane levels of urban intersections, specifically including spatial correlation modeling and temporal correlation modeling. In spatial correlation modeling, intersection lanes are treated as nodes in a graph, and all spatial lanes of a single intersection are modeled. Then, by embedding spatial structure information, temporal modeling is performed for each node (i.e., temporal correlation modeling with dual temporal granularity). Finally, refined congestion event predictions are obtained for all lanes and signal cycle levels of the intersection:

[0101] The input data includes traffic data collected from regional road networks and local intersections.

[0102] For the regional road network graph G g =(V g , E g , A g), the traffic characteristics of all the incoming sections of each intersection are aggregated every 10 minutes, that is,

[0103] Among them, P is the data acquisition signal period, F g is the aggregated traffic feature. g =[v avg，， v min , v max ] Whether it is at the peak period 0 / 1, v avg ,,v min , v max Indicates average, minimum, and maximum speeds.

[0104] At the intersection level, define the local intersection graph G l =(V l , E l , A l ), which captures the local topological relationship between lanes at a specific intersection. Each graph node corresponds to a lane spatial location. represents the lane-level local time series traffic status data, F l =[v i,avg ,,v i,min , v i,max ,TP], where v i,avg , v i，min , v i,max are the average speed, minimum speed and maximum speed of the i-th intersection point in the P period respectively.

[0105] Training STNPP mainly consists of the following steps:

[0106] Input: intersection graph structure G l ∈R M×M ; Regional road network graph structure G g ∈R N×N ;Training traffic flow parameter feature set {χ g , χ l} and congestion event collection Batch size b, number of iterations K, number of input history and predicted signal cycles P, T p .

[0107] Output: STNPP model parameters θ

[0108] / / Forward propagation calculation

[0109] Random initialization θ

[0110] For n=1…K do:

[0111] Training sample input: Each time input b traffic flow spatiotemporal features {(G g , χ g ), (G l , x l )} and congestion event set

[0112] According to formula (4.5), spatial association single-layer GCN modeling event spatial association:

[0113] By modeling event-time correlation:

[0114] Model output: Predict whether a multi-step time congestion event will occur and the value:

[0115] Calculate the congestion event time and calibration error loss using the formula:

[0116]

[0117] / / Back propagation calculation:

[0118] Perform Adam backpropagation to update the gradient

[0119] Update the model weight parameters through gradient:

[0120] Stop iteration.

[0121] This embodiment further demonstrates the prediction results of the refined lane-level congestion value, and only shows the methods with better prediction results. Figure 6 As shown in the figure, we can find that: 1) according to the standard speed regulations for urban traffic, some lanes at the intersection are currently in a slow-moving and severely congested state; 2) even when the traffic speed changes drastically, STNPP can still obtain prediction results that are very close to the true value and obtain the smallest residual; this is mainly because the NPPGRU unit can capture the historical impact of congestion and the cumulative number of triggers in time; therefore, it pays more attention to locations with drastic changes, can flexibly respond to real traffic changes, and is not easy to produce smooth prediction results; STNPP spatially utilizes the influence of traffic lights between lanes and considers upstream and downstream associations, and uses dual-granularity modeling of continuous traffic flow and the time impact of "jump" events in time, which can obtain smaller residuals (predicted value minus actual value) than other baseline models, and the residual distribution of STNPP is obviously centered on zero, without serious tilt.

[0122] Verification test

[0123] The prediction accuracy of the present invention is further illustrated by the following verification.

[0124] (1) Comparison of single lane congestion prediction accuracy

[0125] The present invention first evaluates the prediction performance of STNPP in two single-lane experiments (only one lane is tested each time, and the congestion period is not distinguished between peak and flat peaks), and selects three signal steps of 1 / 3 / 5 to predict the future, and shows them in Tables 1 and

[0126] The results are shown in Table 2. First, the results show that as the prediction step size increases, the prediction performance of all models shows a downward trend. However, the STNPP model always obtains the lowest MAE, RMSE, and MAPE for single-lane congestion prediction at intersections. The predicted MAE / RMSE / MAPE for the southbound lane in the next 5 cycles are 2.60 / 3.51 / 8.44, and for the northbound lane: 2.29 / 2.97 / 6.51, which proves that the design of spatiotemporal correlation modeling in the STNPP algorithm is more consistent with the distribution pattern of real congestion than other models. Secondly, the performance of deep learning-based spatiotemporal prediction models (GCN+ATT, ASTGCN, S2S-GRU, NP-STNPP, STNPP) is better than that of machine learning SVR in most cases, which proves that deep learning models have a stronger ability to capture spatiotemporal features. However, in

[0127] The prediction performance of GCN+ATT in the northbound lane report results in Table 2 is not as good as SVR. This is mainly due to the strong temporal correlation in congestion events. Ignoring the temporal features and only modeling the space cannot fit the congestion distribution well. Finally, ASTGCN, as the suboptimal model, is close to the variant model NP-STNPP of the present invention in performance, and is better than other congestion prediction models. GCN+ATT, S2S-GRU and NP-STNPP achieved the worst prediction results, mainly because they did not model the temporal and spatial associations at the same time, and did not consider the influence of multiple time granularity factors in congestion events. In summary, the STNPP model of the present invention has a stronger ability to capture spatiotemporal features, can better fit the congestion distribution, and simultaneously models the temporal and spatial associations, and considers the influence of multiple time granularity factors in congestion events, with the highest prediction accuracy.

[0128] Table 1 Performance comparison of southbound driving models

[0129]

[0130] Table 2 Performance comparison of northbound driving models

[0131]

[0132] (2) Migration capacity congestion prediction test at similar structure intersections

[0133] To verify the proposed method's ability to generalize and transfer traffic between similar intersections, we trained the model at intersection #767. The trained model was then transferred to intersections #771 and #1041, which have similar structures to intersection #767, to test the congestion model's migration prediction capabilities. The results in Table 3 show that the STNPP model achieves good migration prediction capabilities between structurally similar intersections, achieving satisfactory MAE, RMSE, and MAPE for the next 1, 3, and 5 signal cycles.

[0134] Among them, the #771 intersection achieved good migration application, which is mainly reflected in the following two points: 1) A high prediction accuracy was achieved; the single-step prediction MAPE of lanes traveling in different directions was: 4.68 / 6.50; 2) The prediction step size and accuracy were significantly negatively correlated; it can be seen that the accuracy of the congestion prediction results gradually decreased with the increase of the prediction step size; Based on these two evaluation criteria, the accuracy of the migration prediction results at the #1041 intersection was relatively poor, and the accuracy of the single-step prediction results was not higher than the 3-step and 5-step prediction accuracies. In fact, as the prediction step size increased, the accuracy showed an upward trend, indicating that the model's migration ability was poor.

[0135] Therefore, based on the above results, it is found that the spatiotemporal patterns of congestion changes at different intersections are different, and the generalization performance of the model varies greatly. In particular, the mean absolute error (MAE) in the single-step prediction of intersection #771 is: 1.43-1.93, which is less than 2km / h, which is considered insignificant in traffic management and control practice. This is mainly because in terms of spatial location, intersection #771 and intersection #767 for model training are located upstream and downstream of the same road section, so the spatiotemporal changes of traffic patterns are similar. However, intersection #1044 is far away from #767 and is located on a different road section, resulting in a large difference in the spatiotemporal patterns of traffic congestion. Although the structure is similar, the error generated by migration prediction is large. Although the performance is not stable enough in the migration prediction of some intersections, the overall prediction accuracy of intersections with similar layouts is good, proving that the method proposed in the present invention can meet the purpose of migration prediction between similar intersections.

[0136] Table 3. STNPP model migration capability test at intersections #771 and #1044

[0137]

[0138] Other parts not described belong to the prior art.

Claims

1. A road intersection congestion prediction method based on a time point process neural network model, characterized by: Based on the time point process theory and deep learning methods, the time point process neural network model is used to predict the full life cycle of the occurrence, development and dissipation of lane-level congestion events at intersections at the fine time scale of the signal light cycle. The specific method is: First, during spatial correlation modeling, the spatial regional congestion change patterns of multiple intersections are integrated into a single intersection level to construct a spatial correlation module. This assists in fine-grained congestion prediction at the intersection lane level and at the signal cycle level, achieving dynamic correlation at the local intersection and regional road network levels. Furthermore, when modeling spatial correlation at the signal-controlled intersection level, the signal cycle changes and the graph convolutional network of the local intersection are integrated to capture the dynamic spatial correlation at the lane level of the local intersection under a fixed signal cycle. Secondly, in the dual-granularity temporal correlation modeling process, the temporal granularity of congestion events is captured through the time point process. This is further integrated with the gated recurrent network unit to construct a new neural point process gated recurrent unit. This new neural point process gated recurrent unit is used to model congestion at different time granularities, resulting in a dual-granularity temporal correlation module. The spatiotemporal correlation module is constructed by the spatial correlation module and the dual-granularity temporal correlation module; Finally, a time point process neural network model is established based on the sequence-to-sequence architecture and spatiotemporal correlation module, and parameter optimization is performed to achieve multi-step prediction of congestion events. The specific prediction method includes the following steps: Step 1: Organize basic information; Step 2: Define congestion events; Step 3: Spatial correlation modeling based on intersection graph structure; Step 4: Temporal correlation modeling of integrated time point processes; Step 5: Build a time point process neural network model based on the sequence-to-sequence architecture and spatiotemporal association module; Step 6: Loss function representation of the model; The loss function of the model is expressed as follows: Given a region-level traffic flow feature χ g , and intersection-level traffic flow characteristics χ l A set of temporal irregular traffic states, a discrete set of historical and future time internal factor event characteristics and By setting k c =1 means congestion occurs at the current moment, if there is no congestion at the current moment, then k c is 0; by maximizing The joint log-likelihood estimation updates the model parameters, and the loss function of the model is expressed as: Among them, t c ∈(0,P] represents the timestamp of the cth congestion event; Indicates the value information of this congestion event; m j ∈[1,2,…,K],K∈Z + It represents the cumulative number of congestion events that occurred within the historical time interval (0, p] signal cycles; P represents the historical time; Step 7: Output prediction results; The time association module is based on a sequence-to-sequence framework. Through recursive calculation, it learns the encoding representation of historical congestion events, selects the optimal parameter θ with the maximum conditional probability through formula (14), and generates the output of future multi-step congestion events. In the process of learning congestion patterns using the temporal correlation model, event and traffic flow data at multiple historical moments are input into the stacked NPPGRU units to achieve information encoding at dual time granularity. In the decoding prediction stage, the output of the previous NPPGRU unit will be used as the input of the next unit to calculate whether congestion will occur and the congestion value in the next multiple steps.

2. The method for predicting road intersection congestion based on a time point process neural network model according to claim 1, characterized in that: In step 1, the basic data include traffic speed data, geometric data of the detection road network and signal timing data of all intersections in the study area; Traffic speed data is the average of the instantaneous speeds of lane-level floating vehicles in each signal cycle; The geometric data of the detected road network describes the number of lanes, the geometry of each lane, the topological connections between lanes, and the turning restrictions at each intersection; The signal timing data for all intersections in the study area include cycle length and effective green time.

3. The method for predicting road intersection congestion based on a time point process neural network model according to claim 2, characterized in that: In step 2, a congestion event represents a collection of congestion occurrence times and values ​​at a specific intersection. Based on the city's speed limit regulations for specific road intersections, a traffic state where the speed is less than the standard speed at a certain moment is defined as a congestion event. The timestamp and speed of the event are used as the basic information of the event. The specific steps for defining a congestion event are as follows: 1) The set of congestion events of the jth lane at the i-th intersection is counted as For a certain moment when no congestion event occurs, it means { 0,0 }; 2) Define the historical interval (0, P] signal cycles, and for a fixed intersection M lanes, the set of all internal factor event features is counted as Among them, if no congestion event occurs at a certain moment, it represents {0,0}; Among them: U represents the union; Input history P event timestamps t c or 0.

4. The method for predicting road intersection congestion based on a time point process neural network model according to claim 3, characterized in that: In step 3, spatial correlation modeling is performed based on the intersection graph structure, specifically: In the spatial correlation modeling stage, intersection congestion is affected by the combined effects of internal and external factors. Internal factors include the internal factor event feature set E 0→P and regional traffic flow characteristics χ g External factors are affected by the dynamic changes in signal control strategies and the changing patterns of regional road network congestion on local intersections; Global regional road network spatial correlation modeling under the influence of external factors: The graph convolutional network is used to model the graph structure. For each signal cycle, the intersection is used as the basic unit of the regional road network graph structure. The graph convolutional network is used to input N intersections of the regional road network graph structure. Modeling spatial dependencies, we can obtain the global spatial correlation formula between regional intersections: in, and are the l+1 and l layers of the regional road network graph convolutional layer input data blocks; U g is an orthogonal matrix; g(·) is a polynomial kernel function acting on the diagonal matrix Λ g ∈R N×N ; For the regional road network graph G g , its graph Laplacian matrix L is decomposed into U g g(Λ g )(U g ) T ; The spatial structure information between the embedded regional road network intersections is obtained through the global spatial correlation formula between regional intersections, and the historical time P signal cycles of the regional road network are calculated to obtain Through index query, the index of target intersection n is selected from N intersections, which is characterized by Further obtain the regional road network impact characteristics of the historical time P signal cycles of all lanes at the intersection As the temporal and spatial correlation input of local intersections, it assists in modeling the impact of internal factors on congestion; Considering that intersection congestion is affected by external factors of the signal control strategy, different signal control strategies determine the connection strength between upstream and downstream lanes of the intersection and the connectivity of traffic flow. To account for external factors affecting congestion, the lane-level external signal control strategy changes are modeled spatially in relation to the application environment of signal-controlled intersections. The specific modeling method is as follows: Given a specific intersection with M lanes in a traffic network, the external signal control perception transfer matrix A∈R is constructed through the dynamic change of signal control strategy. M×M , expressing the correlation strength between spatial lanes on each signal cycle time slice, and further superimposing the signal control perception transfer matrix in the historical time period P to form A s ∈R M×M×P ; Among them, the matrix A s ∈R M×M×P At different time slices, the correlation strength between lanes is quantified by the different sizes of signal periods between upstream and downstream lanes.

5. The method for predicting road intersection congestion based on a time point process neural network model according to claim 4, characterized in that: The method of spatial correlation modeling based on intersection graph structure also includes: using the attention mechanism to collect the congestion internal factor event feature set E 0→P and intersection-level traffic flow characteristics χ l Combine and calculate the weights between lanes. Different lanes are assigned different weights. The weights are calculated by combining the combined effects of internal factors to express the joint impact on the current congested lanes. The specific method is as follows; Assume that there is traffic congestion observation data of P signal cycles in history, and the traffic flow of M lanes at the intersection is l and event E 0→P As the input of the neural network layer, the input weight matrix is ​​calculated in each signal cycle: Among them: σ represents the activation function; "|" represents the connector, b l ∈R M×M represents the network learning parameters; the input features of the congestion event are event time and scalar value; For the sth lane where congestion occurs, its spatial correlation α with adjacent lanes and upstream and downstream lanes S,j According to A l Calculation yields: Where j is the index associated with the sth lane, α S,j Represents the spatial correlation strength between the jth lane and the sth lane in the intersection; In each signal cycle slice, the graph convolution operator is used to capture the spatial correlation between intersection lanes. The formula is as follows: X l+1 =ReLU(Ug(Λ)U T X l ) (4) Among them, X l+1 and X l are the l+1 and L layers of the graph convolutional network layer input data blocks, Represents the global regional road network influence at time t; U is an orthogonal matrix; In order to capture the dynamic lane-level spatial correlation between adjacent intersections, it is proposed to calculate the spatial attention weight and the signal control perception transfer matrix A∈R M×M Integrating into the above graph convolution, Equation (4) is rewritten as: Among them, ⊙ is the Hadamard product operator, that is, point-by-point multiplication.

6. The method for predicting road intersection congestion based on a time point process neural network model according to claim 5, characterized in that: The time correlation modeling of the integrated time point process is as follows: Considering that traffic congestion is affected by multiple time granularities, the traffic time series is divided into two time granularities: the irregular time granularity based on the interval of congestion events and the traffic flow parameter time granularity based on the continuous regular signal cycle; For a given lane m of an intersection n, a set of discrete event lists that occurred within the historical time P is given Among them, t c ∈(0,P] represents the timestamp when the c-th congestion event is obtained, and it is assumed that the first observation is made at timestamp 0; Introducing a mask vector To express t c Whether a congestion event occurs at a certain time, its value is 1 / 0 respectively, which is expressed by the following formula: Where: t c Indicates the timestamp of the congestion event; represents the mask vector; Indicates the time interval between adjacent consecutive congestion events; In order to capture the temporal correlation pattern of congestion events at dual time granularity, an adaptively learnable prior intensity function model is designed and integrated into the traditional gated recurrent network (GRU). The formula is: Among them, ["|"] represents the connection symbol, are model learning parameters; represents the cumulative vector of congestion events; After obtaining the results, in order to model the time correlation between historical congestion events and the cumulative number of congestion, we multiply λ(t c ) to update the temporal features and hidden states of the original gated recurrent unit input It is then used as a priori model and combined with the GRU model to form the neural point process gated recurrent unit NPPGRU; NPPGRU is updated as follows: Among them, u represents the update gate, r represents the reset gate, represents the candidate hidden variable, σ represents the activation function, W r ,W u ,U r ,U u ,U and represents the learning weight parameter, b r ,b u and represents the bias term; [⊙] represents the dot product, and [*] represents the matrix product; The mask vector based on the congestion event Directly input the neural point process gated recurrent unit model to assist in the temporal correlation modeling of events and obtain the temporal correlation model.

Images