Traffic signal lamp time intelligent adjustment system and method

Abstract

The present application relates to the technical field of traffic, in particular to a traffic signal lamp time intelligent adjustment system and method; the method comprises collecting traffic flow related data, taking the intersection as the node of the graph, constructing the physical distance graph, the flow related graph and the upstream and downstream relationship graph with static edges, dynamic edges and semantic edges, respectively performing graph convolution operation on the three kinds of graph structures, extracting spatial features, calculating space-time attention weight, fusing space-time features, fusing the feature representation of the three kinds of graph structures, constructing fusion features, predicting the traffic flow of a certain intersection according to the fusion features of the intersection, and adjusting the time of the traffic signal lamp of the intersection according to the predicted traffic flow of the intersection and the predicted traffic flow of other intersections adjacent to the intersection; the method solves the problems of unreasonable fixed time of traffic signal lamp and poor coordination of multiple intersections.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of traffic technology, in particular to a traffic signal lamp time intelligent adjustment system and method. BACKGROUND

[0002] After in-depth analysis, it is not difficult to find that the main reasons are: 1. Traffic facilities cannot meet the growing demand and the unreasonable structure of the road network; 2. The vehicles passing through the peak hours are mainly fixed on the commuting route. Traffic signal lights rely on fixed timing schemes or limited sensor data, and traffic facilities cannot be fully utilized. The unreasonable timing of traffic lights makes the congestion worse. Therefore, improving traffic management and improving traffic efficiency have become a problem to be solved.

[0003] However, current traffic signal control relies on fixed timing schemes or limited sensor data, which has the following defects:

[0004] 1. Insufficient data sensing: the coverage density of traditional sensors (such as magnetic coils) is low, and it is difficult to fully capture real-time traffic dynamics, especially in complex intersections or extreme weather conditions, with high risk of failure.

[0005] 2. Low prediction accuracy: static models cannot adapt to the nonlinear changes of traffic flow (such as tidal phenomenon and sudden congestion), resulting in signal timing lag or imbalance.

[0006] 3. Poor coordination between multiple intersections: existing systems mainly focus on single intersection optimization, and regional green wave coordination relies on manual experience, which is difficult to dynamically respond to traffic fluctuations.

[0007] 4. High dependence on hardware: traditional solutions require large-scale deployment of dedicated detection equipment, which is costly and difficult to maintain. SUMMARY

[0008] In view of the above problems of the prior art, the present application aims to provide a traffic signal lamp time intelligent adjustment system and method to solve the problems of unreasonable fixed time of traffic signal lamp and poor coordination between multiple intersections.

[0009] In order to solve the above problems, the present application adopts the following technical solutions:

[0010] On the one hand, the present application provides a traffic signal lamp time intelligent adjustment system, comprising a data acquisition module, a graph structure construction module, a feature fusion module, a traffic flow prediction module and a traffic signal lamp time adjustment module.

[0011] The data collection module is configured to collect traffic flow related data, including historical average traffic flow, speed and occupancy, historical vehicle travel paths obtained according to license plate numbers and the proportion of vehicles with travel paths to traffic flow, road segment length, number of lanes, slope and curvature of the road segment, weather conditions, date, period, holiday type and special events.

[0012] The graph structure construction module is configured to take intersections as nodes of a graph, and construct a physical distance graph, a traffic related graph and an upstream and downstream relationship graph using static edges, dynamic edges and semantic edges.

[0013] The feature fusion module is configured to perform graph convolution operations on the three kinds of graph structures respectively, extract spatial features, calculate spatio-temporal attention weights, fuse spatio-temporal features, fuse feature representations of the three kinds of graph structures, and construct fused features.

[0014] The traffic flow prediction module is configured to predict traffic flow at an intersection according to the fused features of the intersection.

[0015] The traffic signal time adjustment module is configured to adjust the time of a traffic signal at an intersection according to the predicted traffic flow at the intersection and the predicted traffic flow at other intersections adjacent to the intersection.

[0016] As an implementable manner, the construction of the physical distance graph includes: taking the physical connection relationship between road network intersections as static edges, and constructing the physical distance graph by weighting and summing the road segment length, the number of lanes, the slope and the curvature as the edge weight of the static edge.

[0017] As an implementable manner, the construction of the traffic related graph includes: calculating the Spearman correlation coefficient between intersections as the dynamic edge between intersections according to the historical average traffic flow, speed and occupancy of the intersections, and constructing the traffic related graph.

[0018] As an implementable manner, the construction of the upstream and downstream relationship graph includes: using the historical vehicle travel paths obtained according to license plate numbers and the proportion of vehicles with travel paths to traffic flow, and weather conditions, date, period, holiday type and special events, and using a causal discovery algorithm to calculate the causal transmission relationship of traffic flow as a semantic edge, and constructing the upstream and downstream relationship graph.

[0019] As an implementable manner, the feature fusion module adopts a graph attention network (GAT) for a physical distance graph to differentially aggregate neighbor node information, uses a graph wave neural network (GWNN) to process an undirected weighted graph for a traffic correlation graph, and adopts a gated graph neural network (GGNN) to model a directed causal propagation process for an upstream and downstream relationship graph. After a spatial feature vector output by each branch is extracted for a local timing pattern by a time convolution network (TCN), the spatial feature vector is sent to a spatio-temporal cross-attention network. The spatio-temporal cross-attention network first calculates cross-graph attention weights between the three kinds of graph features, and then combines spatial position encoding and timestamp embedding to align and weight fuse the multi-graph features in the spatio-temporal dimension, and outputs fused features.

[0020] As an implementable manner, the traffic prediction module adopts an encoder-decoder architecture, wherein the encoder is constituted by stacked spatio-temporal attention layers for capturing dynamic evolution laws of fused features in the current and past multiple time periods, and the decoder part introduces causal convolution and dilated convolution structures to perform multi-step prediction on traffic flow in the future 15 to 60 minutes.

[0021] As an implementable manner, the traffic signal light time adjustment module is configured to receive future traffic prediction results of a current intersection and K adjacent upstream and downstream intersections, construct a regional collaborative optimization objective function, and model the objective function as a mixed integer linear programming (MILP) problem for minimizing regional average delay time and maximum queue length, and constraint conditions include minimum green light time, cycle length range, and phase sequence logic.

[0022] In another aspect, the application provides a traffic signal light time intelligent adjustment method, comprising:

[0023] Collecting traffic flow related data, wherein the traffic flow related data includes historical average traffic flow, speed and occupancy, a path of a vehicle obtained according to a license plate number, a proportion of vehicles with the path, a length of a road section, a number of lanes, a slope and a curvature of the road section, weather conditions, a date, a period, a holiday type and a special event;

[0024] Taking an intersection as a node of a graph, and constructing a physical distance graph, a traffic correlation graph and an upstream and downstream relationship graph by using static edges, dynamic edges and semantic edges;

[0025] Performing graph convolution operations on the three kinds of graph structures respectively, extracting spatial features, calculating spatio-temporal attention weights, fusing spatio-temporal features, and constructing fused features by fusing feature representations of the three kinds of graph structures;

[0026] Performing traffic flow prediction of an intersection according to fused features of the intersection;

[0027] The time of the traffic signal light of the intersection is adjusted according to the predicted traffic flow of the intersection and the predicted traffic flow of other intersections adjacent to the intersection.

[0028] As an implementable manner, the constructing the physical distance graph comprises: constructing the physical distance graph based on the physical connection relationship of the road network intersection and the intersection as a static edge, and weighting and summing the road segment length, the number of lanes, the slope and the curvature as the edge weight of the static edge.

[0029] As an implementable manner, the constructing the flow correlation graph comprises: calculating the Spearman correlation coefficient of the intersection and the intersection according to the historical average traffic flow, speed and occupancy of the intersection as the dynamic edge of the intersection and the intersection, and constructing the flow correlation graph.

[0030] As an implementable manner, the constructing the upstream and downstream relationship graph comprises: obtaining the historical vehicle driving path according to the license plate number, and calculating the causal transmission relationship of the traffic flow as the semantic edge by using a causal discovery algorithm, and constructing the upstream and downstream relationship graph by using the proportion of the vehicle with the driving path and the traffic flow, and the weather condition, the date, the period, the holiday type and the special event.

[0031] As an implementable manner, the graph attention network GAT is used for the physical distance graph to differentially aggregate neighbor node information; the graph wavelet neural network GWNN is used for the flow correlation graph to process the undirected weighted graph; the gated graph neural network GGNN is used for the upstream and downstream relationship graph to model the directed causal propagation process; the spatial feature vectors output by each branch are sent into the space-time cross-attention network after local time sequence patterns are extracted by the time convolution network TCN; the space-time cross-attention network firstly calculates the cross-graph attention weight between the three kinds of graph features, and then combines the spatial position encoding and the time stamp embedding to align and weight fuse the multi-graph features in the space-time dimension, and outputs the fused features.

[0032] As an implementable manner, in the constructing the fused features, an encoder-decoder architecture is used, wherein the encoder is composed of stacked space-time attention layers, which are used to capture the dynamic evolution law of the fused features in the current and past multiple time periods, and the decoder part introduces causal convolution and dilated convolution structures to perform multi-step prediction of the traffic flow in the future 15 to 60 minutes.

[0033] As an implementable manner, in the adjusting the time of the traffic signal light of the intersection, the future flow prediction results of the current intersection and its adjacent K upstream and downstream intersections are received, a regional collaborative optimization objective function is constructed, the objective function is to minimize the regional average delay time and the maximum queue length, and is modeled as a mixed integer linear programming MILP problem, and the constraint conditions include the minimum green light time, the cycle length range and the phase sequence logic.

[0034] The beneficial effects of the present application are that: the present application collects traffic flow related data, takes intersections as nodes of a graph, constructs a physical distance graph, a flow related graph and an upstream and downstream relationship graph with static edges, dynamic edges and semantic edges, respectively performs graph convolution operations on the three kinds of graph structures, extracts spatial features, calculates spatio-temporal attention weights, fuses spatio-temporal features, fuses feature representations of the three kinds of graph structures, constructs fusion features, predicts the traffic flow of a certain intersection according to the fusion features of the intersection, adjusts the time of the traffic signal light of the intersection according to the predicted traffic flow of the intersection and the predicted traffic flow of other intersections adjacent to the intersection, and solves the problems of unreasonable fixed time of traffic signal light and poor coordination of multiple intersections. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 It is a schematic diagram of a traffic signal light time intelligent adjustment system.

[0036] Figure 2 It is a schematic diagram of a traffic signal light time intelligent adjustment method flow. DETAILED DESCRIPTION

[0037] The present application will be further described in detail below in combination with specific embodiments.

[0038] It should be noted that these embodiments are only used to illustrate the present application, and are not limiting the present application, and simple improvements of the present method under the concept of the present application all belong to the scope of the present application.

[0039] Embodiment 1

[0040] Reference Figure 1 It is a traffic signal light time intelligent adjustment system, which comprises a data acquisition module 100, a graph structure construction module 200, a feature fusion module 300, a flow prediction module 400 and a traffic signal light time adjustment module 500.

[0041] The data acquisition module 100 is used for acquiring traffic flow related data; the traffic flow related data comprises historical average traffic flow, speed and occupancy, the path of the historical vehicle obtained according to the license plate number and the proportion of the vehicle with the driving path to the traffic flow, the length of the road section, the number of lanes, the slope and the curvature of the road section, the weather condition, the date, the period, the holiday type and the special event.

[0042] The data acquisition module realizes the perception of the traffic state through the deployment of multiple types of sensors and external information systems at each intersection. For example:

[0043] Magnetic coil detector: deployed in front of each entrance lane stop line, reports traffic flow (vehicles / 5min), average speed (km / h) and lane occupancy rate (%) every 5 minutes.

[0044] Video detector: combined with YOLOv7 target detection model and DeepSORT multi-target tracking algorithm, for license plate recognition, to obtain the possible driving trajectory of the vehicle at the intersection through the license plate, and to calculate the proportion of vehicles on the path to the total traffic flow (e.g., the E→G→F path accounts for 78.6%).

[0045] Weather bureau real-time weather interface: e.g., obtain rainfall level (no rain / light rain / moderate rain / heavy rain) and visibility (m) every 10 minutes. Calendar service interface: automatically determine the current date as "workday", "weekend", "legal holiday", "date", etc.

[0046] Event notification interface: whether there are special events such as traffic accidents, road construction, large-scale activities, etc.

[0047] Based on the intersection, a historical vehicle driving path database is constructed to record effective and complete paths (no interruption from start to end). For incomplete paths (e.g., only appearing at the starting or ending intersection), a probability path completion algorithm based on road network topology and traffic distribution is used to recover and improve path coverage. All data are normalized by unified spatiotemporal index (UTC timestamp + GIS coordinates) and filtered by outlier value filtering and missing value interpolation modules to ensure data quality, forming a high-quality structured dataset as the input basis for subsequent graph structure dynamic modeling. The graph structure construction module 200 is used to construct a physical distance graph, a traffic-related graph, and an upstream and downstream relationship graph based on the intersection as the node of the graph and the static edge, dynamic edge, and semantic edge.

[0048] The construction of the physical distance graph includes: based on the physical connection relationship between the road network intersection and the intersection as the static edge, the length of the road segment, the number of lanes, the slope, and the curvature are weighted and summed as the edge weight of the static edge to construct the physical distance graph.

[0049] The intersections / crossings in the urban traffic network are mapped to a set of graph nodes V = {v1, v2,..., vN}, and the actual road connection relationship is used to establish bidirectional static edges to form the basic topology of the physical distance graph , is the static edge set. For example, there is a bidirectional six-lane main road connecting between intersection A and intersection B, then add edges (vA, vB) and (vB, vA). The weight of each edge is calculated by weighting the geometric properties and traffic capacity of the corresponding road segment, and the formula is as follows:

[0050] wherein, is the length of the road segment (m), is the number of lanes, ranging from 2 to 6, is the absolute value of the slope, ranging from 0 to 0.15, is the curvature (1 / m), 、 、 、 are learnable parameters. Taking the A→B link as an example, assuming its length is 900 m, the number of lanes is 5, the slope is 0.07, and the curvature is 0.003, then: The calculation result is 51.64. This weight is mainly dominated by the reciprocal of the curvature term, reflecting that the turning of this link is relatively sharp, and the traffic efficiency is relatively low, so the spatial propagation weight in the graph neural network needs to be reduced. After the physical distance graph is constructed, its topological structure remains unchanged during system operation, and is used to reflect the inherent spatial proximity and physical accessibility of the road network. The construction of the flow-related graph includes: calculating the Spearman correlation coefficient between intersections and intersections as the dynamic edge between intersections and intersections according to the historical average traffic flow, speed and occupancy of the intersection, and constructing the flow-related graph.

[0051] Based on the time series of traffic flow, speed and occupancy of each intersection in the historical period, the Spearman rank correlation coefficient between any two intersections is calculated using a sliding time window (such as the last 24 hours) as the initial weight of the dynamic edge, and an undirected flow-related graph is constructed. To enhance timeliness, the system updates the correlation matrix every 5 or 15 minutes, and performs significance test (p < 0.05, Bonferroni corrected) on the correlation coefficient, and only retains the significant correlation edges. The correlation strength is normalized by Sigmoid and used as the adjacency weight of graph convolution operation, which is used to capture the implicit association between non-directly connected intersections that have statistical synergistic trends.

[0052] For example, during the evening peak (17:00-19:00), the system detects that the traffic flow , speed , occupancy between intersection C and intersection D, then: Significance test (two-tailed t-test, degrees of freedom df=286) is performed on this correlation, and the p-value=0.0001 < 0.05 (still significant after Bonferroni correction), which is determined as a strong positive correlation, and the edge (vC, vD) is retained, and the weight is set to 0.704. For negative correlation cases (such as one intersection congestion and the other smooth), negative weights are retained to represent the reverse linkage trend. All edges that pass the significance test constitute an undirected weighted graph , is the set of associated edges, The weight set is used as the adjacency input in the graph convolutional network. This graph is dynamically updated every 5 or 15 minutes, capturing implicit associations between intersections that are not directly connected but exhibit statistically coordinated trends, such as synchronous fluctuations caused by shared commuter corridors or the same event. Constructing the upstream and downstream relationship graph involves using historical vehicle routes obtained based on license plate numbers, the proportion of vehicles with routes and traffic flow, weather conditions, dates, cycles, holiday types, and special events. A causal discovery algorithm is used to calculate the causal transmission relationships of traffic flow as semantic edges to construct the upstream and downstream relationship graph.

[0053] The upstream and downstream relationship graph uses intersections as nodes. Initial semantic edges are identified using either the PC algorithm or the LiNGAM causal discovery algorithm, combined with the direction and proportion of traffic flow propagation in historical vehicle paths, to determine intersection pairs with causal influence. Causal strength is quantified using transfer entropy.

[0054] ;

[0055] Where X and Y represent intersection nodes, It quantifies the intensity of the causal influence of the traffic state at the upstream intersection X on the traffic state at the downstream intersection Y; For the upstream intersection X Traffic conditions at any given time, including traffic flow, percentage of vehicles on the route, and environmental characteristics. For the downstream intersection Y "Traffic status characteristics" at any given time Let Y be the "traffic state characteristics" of the downstream intersection at time t. For conditional proportions, For the known "downstream historical state" "and upstream historical status" "At that time, the current state of the downstream is y t The probability of occurrence (i.e., the accuracy of predicting the current downstream traffic flow by combining upstream information), For only the "downstream historical state" At that time, the current state of the downstream is y. t The probability of occurrence (i.e., "prediction accuracy without considering upstream information"), the meaning of the ratio: if upstream X has a significant causal effect on downstream Y (e.g., traffic peak in X will lead to subsequent traffic growth in Y), then the numerator will be significantly greater than the denominator (or less than, depending on the direction of the effect), and the ratio will deviate from 1; if they are unrelated, the ratio will be close to 1.

[0056] The initial directed and weighted causal graph is generated by conditional causal inference with weather, holiday, and other contextual variables. During system operation, an incremental causal inference mechanism is introduced: every 5 minutes, real-time multivariate time series (flow, speed, occupancy) are collected, and the data stream within the sliding window (the last 60 minutes, granularity of 5 minutes) is used to reevaluate the significance and weight changes of existing causal edges through dynamic Bayesian networks or online Granger causal detection methods. If the causal strength of an edge decreases by more than a preset threshold (e.g., 30%) for three consecutive evaluations, the deletion mechanism is triggered. If a new strong causal relationship (p < 0.01, corrected by FDR multiple testing) is detected, a semantic edge is dynamically added.

[0057] For example, taking the morning peak as an example, the system analyzes the vehicle path from intersection E to intersection F, a total of 2,100 vehicles, of which 1,650 (accounting for 78.6%) pass through the E→G→F path, indicating that there is a potential causal effect from E to G and G to F. The causal strength is quantified using transfer entropy (Transfer Entropy), with intersection G as the receiving end Y and intersection E as the sending end X, and the calculation result is , which is significantly higher than the random permutation test threshold (e.g., 0.16 bit), indicating that there is a causal relationship. Further conditional causal inference is introduced with contextual variables: in sunny conditions, ; in moderate rain conditions (rainfall intensity > 10 mm / h), ; during holidays, the value drops to 0.22, indicating that changes in travel patterns have weakened the original path dependence. The system performs incremental causal inference every 5 minutes: using the last 1 hour of high-frequency data (5-minute granularity) as input to the dynamic Bayesian network model, reevaluate the confidence of all existing causal edges. If the causal strength of an edge decreases by more than a preset threshold (e.g., 30%) for three consecutive evaluations, the deletion mechanism is triggered. If a new strong causal relationship is detected (p < 0.01, corrected by FDR multiple testing), a semantic edge is dynamically added.

[0058] For example, during the dispersal of a large sports event, the system automatically identifies a strong causal flow from the sports stadium exit H to the subway station I , p = 0.0007), and dynamically adds the directed edge (vH, vI), ensuring that subsequent predictions can timely reflect the path of sudden flow propagation. TEX→Y is finally quantified by the weighted sum of the above probability terms (weight is joint probability): "the 'prediction gain' of the historical traffic state of the upstream intersection X (including path, vehicle flow proportion, environmental factors) on the current traffic state of the downstream intersection Y".

[0059] If TEX→Y > 0: it indicates that X has a causal effect on Y (X is the upstream intersection of Y), the larger the value, the stronger the impact;

[0060] Combined with PC algorithm or LiNGAM causal discovery algorithm, the combination of TEX→Y>0 can be screened from all intersection pairs as the semantic edge of the upstream and downstream relationship graph, and the weight of the edge is the value of TEX→Y.

[0061] The feature fusion module 300 is used for performing graph convolution operations on the three kinds of graph structures respectively, extracting spatial features, calculating spatio-temporal attention weights, fusing spatio-temporal features, fusing feature representations of the three kinds of graph structures, and constructing fused features.

[0062] The graph attention network GAT is used for the physical distance graph to differentially aggregate neighbor node information; the graph wavelet neural network GWNN is used for the traffic correlation graph to process the undirected weighted graph; the gated graph neural network GGNN is used for the upstream and downstream relationship graph to model the directed causal propagation process; the spatial feature vectors output by the branches are sent to the spatio-temporal cross-attention network after local temporal patterns are extracted by the time convolution network TCN; the spatio-temporal cross-attention network first calculates the cross-graph attention weights between the three kinds of graph features, and then combines the spatial position encoding and the timestamp embedding to align and weight fuse the multi-graph features in the spatio-temporal dimension, and outputs the fused features.

[0063] Specifically, special graph neural network branches are designed for the three kinds of graph structures:

[0064] The graph attention network (GAT) is used for the physical distance graph branch, 8 attention heads are set, each head outputs 16-dimensional features, and finally 128-dimensional spatial features are spliced, neighbor node information is differentially aggregated through the multi-head mechanism; the graph wavelet neural network (GWNN) is used for the traffic correlation graph branch, which is used for undirected weighted graphs, separates low-frequency smooth features and high-frequency mutation patterns, and captures regional collaborative fluctuations; the gated graph neural network (GGNN) is used for the upstream and downstream relationship graph branch, which models the delay propagation process of vehicle flow from upstream to downstream through the GRU unit. Temporal convolution network (TCN) is used to extract temporal dynamic features, TCN includes 3 layers of causal convolution, and the expansion factors are 1, 2 and 4 in turn, and the receptive field covers the past 60 minutes of data (12 time steps). The output spatio-temporal feature The spatio-temporal cross-attention module is inputted after: time encoding: hours (1-24), days (1-7), and minutes (0-59) are respectively embedded into 8-dimensional vectors, and then spliced and mapped into 128-dimensional time features through MLP; spatial encoding: the latitude and longitude of the intersection are discretized into grid indexes to generate a learnable position encoding vector.

[0065]

[0066]

[0067] The key (K) and the value (V) are respectively from the TCN output of different graph branches, and finally a unified fusion feature representation is generated .

[0068] The traffic flow prediction module 400 is used for predicting the traffic flow of an intersection according to the fusion features of the intersection.

[0069] An encoder-decoder architecture is adopted, wherein the encoder is composed of stacked spatio-temporal attention layers, which are used to capture the dynamic evolution law of the fusion features in the current and past time periods, and the decoder part introduces a causal convolution and dilated convolution structure, and is used to perform multi-step prediction on the traffic flow in the future 15 to 60 minutes.

[0070] The encoder is composed of 3 layers of stacked spatio-temporal attention layers, which are used to capture the time evolution law of the fusion features in the past 60 minutes. The decoder part introduces a causal convolution and dilated convolution structure, and supports multi-step autoregressive prediction on the traffic flow in the future 15, 30, 45 and 60 minutes. In order to improve the prediction robustness, an uncertainty estimation mechanism is integrated into the module: 100 independent prediction paths are generated in the inference stage through the Monte Carlo DropPath method (dropout rate 0.1), the mean and variance are calculated, and the prediction result with confidence interval is output. At the same time, when the external event label is activated (such as “heavy rain red alert” or “major traffic accident”), the system automatically calls a pre-trained abnormal scene adjustment subnetwork, which is specially trained on historical extreme scene data and can more accurately capture abnormal traffic evolution patterns.

[0071] Encoder structure and operation mechanism:

[0072] The input of each spatio-temporal attention block is the node state at the previous moment and the fusion feature vector f i ∈R 128 at the current moment (from the feature fusion module). The graph convolution sublayer performs graph convolution operations based on three updated graph structures:

[0073] The graph attention network (GAT) is used on the physical distance graph to calculate the importance weight of the neighbor nodes and realize differentiated information aggregation;

[0074] The graph wavelet neural network (GWNN) is used on the traffic correlation graph to separate the low-frequency trend and high-frequency fluctuation components;

[0075] The gated graph neural network (GGNN) is used on the dynamic causal graph to model the delay propagation process of the vehicle flow from upstream to downstream.

[0076] The spatial feature vectors output by each graph convolution sub-layer are concatenated and fed into the temporal convolution sub-layer. The temporal convolution sub-layer adopts a dilated causal convolution structure, and the dilation factor is set to 1, 2, and 4 in different layers, respectively, so that the receptive field expands layer by layer. Taking the first layer as an example, the convolution kernel size is 3 and the dilation factor is 1, so it covers 3 consecutive time steps; the dilation factor of the second layer is 2, covering 5 historical points with an interval of 1 time step; the dilation factor of the third layer is 4, which can capture long-term dependencies of the past 12 time steps (i.e., the past 60 minutes). This structure effectively avoids the gradient vanishing problem of traditional RNN and can model the evolution of traffic patterns over a long time span.

[0077] Residual connections and layer normalization mechanisms are introduced within each spatio-temporal attention block to improve training stability. After three layers of stacking, the encoder outputs a hidden state sequence H = [h1, h2,..., hT], where T = 12 represents a 60-minute history window (5 minutes per time step), and ht∈R 128 is the context representation for the t-th time step.

[0078] The autoregressive prediction process of the decoder:

[0079] The decoder adopts an autoregressive structure and supports multi-step rolling prediction of traffic flow for the next 15, 30, 45, and 60 minutes. Initially, the hidden state h12 of the last time step of the encoder is used as the initial input to the decoder. For the k-th step prediction (k = 1, 2, 3, 4, corresponding to the next 15-60 minutes):

[0080] The prediction result of the previous step is concatenated with the fusion feature fi(t+k) of the current time step;

[0081] The concatenated features are input into the spatio-temporal attention block in the decoder to generate the context vector for the current step.

[0082] The traffic flow prediction value y^t k for the current step is output by the fully connected layer.

[0083] To enhance the model's ability to adapt to abnormal traffic patterns, the system integrates an external event gating mechanism. When a special event marker such as a "heavy rain red alert" or a "major traffic accident" is detected, the system automatically invokes a pre-trained abnormal scenario adjustment sub-network. This sub-network is trained on historical extreme scenario data (such as typhoon days and large-scale accident road closures) and dynamically adjusts the encoder output weights through an attention gating mechanism. For example, under a heavy rain red alert, the system automatically reduces the physical connection weight of road segments affected by flooding while enhancing the causal strength of detour paths, thereby improving prediction accuracy under extreme conditions.

[0084] The traffic signal time adjustment module 500 is configured to adjust the time of the traffic signal at the intersection according to the predicted traffic flow at the intersection and the predicted traffic flow at other intersections adjacent to the intersection.

[0085] The future traffic prediction results of the current intersection and its adjacent K upstream and downstream intersections are received, a regional collaborative optimization objective function is constructed, the objective function is to minimize the regional average delay time and the maximum queue length, and is modeled as a mixed integer linear programming MILP problem, the constraint conditions include the minimum green light time, the cycle length range, and the phase sequence logic.

[0086] The traffic signal time adjustment module receives the future traffic prediction results of the current intersection and its adjacent K=4 upstream and downstream intersections, and constructs a regional collaborative optimization objective function: wherein, is the regional average delay time, is the maximum queue length, may be 0.6, may be 0.4, and is a weight coefficient. The problem is modeled as a mixed integer linear programming (MILP) problem, and the decision variables include the green light time of each phase, the cycle length (constrained between 90-180 seconds), and the phase sequence logic. The solution adopts a branch and bound algorithm combined with a heuristic initial value generation strategy based on traffic ratio, and completes a single optimization calculation in an average of 1.9 seconds. The optimization results are issued to each intersection signal controller in the form of a phase timing table, supporting automatic switching between regular timing and emergency green wave coordination mode, and realizing dynamic and collaborative regional signal control. See Figure 2 A traffic signal time intelligent adjustment method, comprising:

[0087] S100, collecting traffic flow related data; the traffic flow related data includes historical average traffic flow, speed and occupancy, the path of the vehicle obtained according to the license plate number and the proportion of the vehicle with the driving path to the traffic flow, the length of the road section, the number of lanes, the slope and the curvature of the road section, the weather condition, the date, the period, the holiday type and the special event;

[0088] S200, taking the intersection as the node of the graph, and constructing the physical distance graph, the traffic related graph and the upstream and downstream relationship graph with static edges, dynamic edges and semantic edges;

[0089] S300, performing graph convolution operation on the three kinds of graph structures respectively, extracting spatial features, calculating spatio-temporal attention weights, fusing spatio-temporal features, fusing feature representations of the three kinds of graph structures, and constructing fused features;

[0090] S400, predicting the traffic flow of a certain intersection according to the fused features of the intersection;

[0091] S500, adjusting the time of the traffic signal light of the intersection according to the predicted traffic flow of the intersection and the predicted traffic flow of other intersections adjacent to the intersection.

[0092] Wherein, the physical distance graph is constructed by taking the physical connection relationship of the road network intersection as a static edge, and weighting and summing the road segment length, the number of lanes, the slope and the curvature as the edge weight of the static edge.

[0093] Wherein, the flow correlation graph is constructed by calculating the Spearman correlation coefficient of the intersection according to the historical average traffic flow, speed and occupancy of the intersection as the dynamic edge of the intersection.

[0094] The upstream and downstream relationship graph is constructed by using the historical vehicle driving path obtained according to the license plate number, and the proportion of vehicles with driving path and traffic flow, and weather conditions, date, period, holiday type and special event, and using a causal discovery algorithm to calculate the causal transmission relationship of traffic flow as a semantic edge.

[0095] Wherein, the graph attention network GAT is used for the physical distance graph to differentiate and aggregate neighbor node information; the graph wavelet neural network GWNN is used for the flow correlation graph to process the undirected weighted graph; the gated graph neural network GGNN is used for the upstream and downstream relationship graph to model the directed causal propagation process; the spatial feature vectors output by each branch are sent to the spatio-temporal cross-attention network after extracting local time sequence patterns by the time convolution network TCN; the spatio-temporal cross-attention network first calculates the cross-graph attention weight between the three kinds of graph features, and then combines the spatial position encoding and time stamp embedding to align and weight fuse the multi-graph features in the space-time dimension, and outputs the fused features.

[0096] Wherein, in the construction of the fused features, an encoder-decoder architecture is used, wherein the encoder is composed of stacked spatio-temporal attention layers for capturing the dynamic evolution law of the fused features in the current and past multiple time periods, and the decoder part introduces causal convolution and dilated convolution structure for multi-step prediction of traffic flow in the future 15 to 60 minutes.

[0097] Wherein, in the adjustment of the time of the traffic signal light of the intersection, the future traffic flow prediction results of the current intersection and its adjacent K upstream and downstream intersections are received, a regional collaborative optimization objective function is constructed, the objective function is to minimize the regional average delay time and the maximum queue length, and is modeled as a mixed integer linear programming MILP problem, and the constraint conditions include the minimum green light time, the cycle length range and the phase sequence logic.

[0098] Based on the method, taking "future 15-60 minute traffic flow prediction" as the core scene, the mainstream indicators MAE (mean absolute error, unit: vehicle / 5min) and MAPE (mean absolute percentage error) in the field of traffic prediction are used to compare the performance differences of single graph model (physical distance graph, traffic correlation graph, upstream and downstream relationship graph) and three graph fusion model, see Table 1.

[0099] Table 1 Performance difference table of single graph model and three graph fusion model

[0100]

[0101] The three graph fusion model avoids the one-sidedness of the single graph model (such as the physical distance graph ignoring traffic coordination and the traffic correlation graph lacking causal logic) by integrating three types of key information (physical topology, traffic statistical correlation, and causal transmission), so the prediction error of each time length is significantly lower than that of the single graph model, and the short-term prediction (15 minutes) is improved most obviously, which meets the demand of real-time signal adjustment.

[0102] Taking "real-time adjustment of causal edges of upstream and downstream relationship graph" as the core, the performance of static causal graph (without incremental update) and dynamic causal update (incremental adjustment every 5 minutes) is compared, and the adaptive ability of abnormal scenes (heavy rain, event dispersal) is verified, see Table 2.

[0103] Table 2 Adaptability verification table

[0104]

[0105] Incremental causal inference every 5 minutes (dynamic Bayesian network), deletion of weak causal edges, and addition of strong causal edges, dynamic update can quickly adapt to changes in traffic paths (such as the addition of H→I causal edge in event dispersal), so the accuracy of causal identification and prediction precision are improved more significantly in abnormal scenes, which meets the real-time traffic dynamics.

[0106] Taking "K=4 adjacent intersection areas" as the optimization object, the core indicators (delay time, queue length, and traffic efficiency) of traditional fixed timing, single intersection independent optimization, and regional collaborative optimization of the present invention are compared, and the results are shown in Table 3.

[0107] Table 3

[0108]

[0109] The MILP model minimizes the average delay, maximum queue length, etc. as constraints, such as green light ≥20 seconds, cycle 90-180 seconds), regional collaborative optimization breaks through the "local optimum" limitation of single intersection, and through the coordination of green light duration and cycle of multiple intersections, it realizes the maximization of regional overall efficiency, so the delay and queue length are significantly reduced.

[0110] Finally, it is to be understood that all the terms and expressions used herein including technical jargon are those customarily used by those skilled in the art. It should also be noted that, as used in this specification and the appended claims, the singular forms "a," "an" and "the" include plural referents unless the context clearly dictates otherwise.

Claims

1. A traffic signal light time intelligent adjustment system, characterized in that, The application relates to a traffic flow prediction method and device based on multi-graph fusion. The data collection module is used for collecting traffic flow related data; the traffic flow related data comprises historical average traffic flow, speed and occupancy, a historical vehicle driving path obtained according to a license plate number and a proportion of vehicles with the driving path to traffic flow, a road section length, a lane number, a slope and a curvature of a road section, weather conditions, a date, a period, a holiday type and an event; The graph structure construction module is used for constructing a physical distance graph, a flow related graph and an upstream and downstream relationship graph by taking an intersection as a node of a graph, and taking a static edge, a dynamic edge and a semantic edge; The feature fusion module is used for performing graph convolution operations on three kinds of graph structures respectively, extracting spatial features, calculating space-time attention weights, fusing space-time features, fusing feature representations of the three kinds of graph structures and constructing fused features; The traffic flow prediction module is used for predicting traffic flow of an intersection according to fused features of the intersection; The traffic signal time adjustment module is used for adjusting a traffic signal time of an intersection according to predicted traffic flow of the intersection and predicted traffic flow of other intersections adjacent to the intersection; The physical distance graph is constructed by taking a physical connection relationship between road network intersections as a static edge, and taking a weighted sum of a road section length, a lane number, a slope and a curvature as an edge weight of the static edge; The flow related graph is constructed by calculating a Spearman correlation coefficient between intersections as a dynamic edge between the intersections according to historical average traffic flow, speed and occupancy of the intersections; The upstream and downstream relationship graph is constructed by using a historical vehicle driving path obtained according to a license plate number and a proportion of vehicles with the driving path to traffic flow, and weather conditions, a date, a period, a holiday type and an event, and using a causal discovery algorithm to calculate a causal transmission relationship of traffic flow as a semantic edge, and constructing the upstream and downstream relationship graph; The feature fusion module adopts a graph attention network GAT to differentially aggregate neighbor node information for the physical distance graph, uses a graph wavelet neural network GWNN to process a non-directional weighted graph for the flow related graph, and adopts a gated graph neural network GGNN to model a directional causal propagation process for the upstream and downstream relationship graph; after spatial feature vectors output by the branches are subjected to local time sequence mode extraction by a time convolution network TCN, the spatial feature vectors are sent to a space-time cross attention network; the space-time cross attention network firstly calculates cross-graph attention weights between three kinds of graph features, and then combines spatial position coding and time stamp embedding to align and weight fuse multi-graph features in a space-time dimension, and outputs fused features; The traffic signal time adjustment module is used for receiving future traffic flow prediction results of a current intersection and K upstream and downstream intersections of the current intersection, constructing a regional collaborative optimization objective function, modeling a mixed integer linear programming MILP problem by taking minimization of regional average delay time and maximum queue length as the objective function, and taking minimum green light time, a cycle length range and phase sequence logic as constraint conditions.

2. The traffic signal timing intelligent adjustment system of claim 1, wherein, The traffic prediction module adopts an encoder-decoder architecture, wherein the encoder is composed of stacked spatio-temporal attention layers for capturing the dynamic evolution law of the fused features of the current and past multiple time periods, and the decoder part introduces causal convolution and dilated convolution structures for multi-step prediction of traffic flow in the future 15 to 60 minutes.

3. A traffic signal time intelligent adjustment method, characterized by, Comprise: Collect traffic flow related data; the traffic flow related data includes historical average traffic flow, speed and occupancy, the path of the historical vehicle obtained according to the license plate number and the proportion of the vehicle with the driving path and the traffic flow, the length of the road section, the number of lanes, the slope and the curvature of the road section, the weather condition, the date, the period, the holiday type and the event; Take the intersection as the node of the graph, and construct the physical distance graph, the traffic related graph and the upstream and downstream relationship graph with static edges, dynamic edges and semantic edges; Respectively, the graph convolution operation is performed on the three kinds of graph structures, the spatial features are extracted, the spatio-temporal attention weight is calculated, the spatio-temporal features are fused, the feature representation of the three kinds of graph structures is fused, and the fused features are constructed; According to the fused features of a certain intersection, the traffic flow of the intersection is predicted; According to the predicted traffic flow of the intersection and the predicted traffic flow of other intersections adjacent to the intersection, the time of the traffic signal of the intersection is adjusted; The construction of the physical distance graph comprises: based on the physical connection relationship between the road network intersections as the static edge, the weighted sum of the length of the road section, the number of lanes, the slope and the curvature as the edge weight of the static edge, the physical distance graph is constructed; The construction of the traffic related graph comprises: calculating the Spearman correlation coefficient between the intersections as the dynamic edge between the intersections according to the historical average traffic flow, speed and occupancy of the intersections, and constructing the traffic related graph; The construction of the upstream and downstream relationship graph comprises: using the path of the historical vehicle obtained according to the license plate number and the proportion of the vehicle with the driving path and the traffic flow, and the weather condition, the date, the period, the holiday type and the event, using the causal discovery algorithm to calculate the causal transmission relationship of the traffic flow as the semantic edge, and constructing the upstream and downstream relationship graph; Wherein, for the physical distance graph, the graph attention network GAT is used to aggregate the neighbor node information in a differentiated manner; for the traffic related graph, the graph wavelet neural network GWNN is used to process the undirected weighted graph; for the upstream and downstream relationship graph, the gated graph neural network GGNN is used to model the directed causal propagation process; the spatial feature vectors output by each branch are sent to the spatio-temporal cross attention network after extracting the local time sequence mode by the time convolution network TCN; the spatio-temporal cross attention network first calculates the cross-graph attention weight between the three kinds of graph features, and then combines the spatial position coding and time stamp embedding to align and weight fuse the multi-graph features in the spatio-temporal dimension, and outputs the fused features; Receive the future traffic prediction results of the current intersection and its K upstream and downstream intersections, construct a regional collaborative optimization objective function, the objective function is to minimize the regional average delay time and the maximum queue length, model it as a mixed integer linear programming MILP problem, and the constraint conditions include the minimum green light time, the period length range, the phase sequence logic.

Images