Road network signal control strategy generation method based on adversarial learning
Through the adversarial learning method, in the road network signal control, ubiquitous network traffic data and edge computing are used to generate high-quality signal control strategies, which solves the problem of inaccurate signal control in the existing technology, and achieves more accurate and stable traffic management.
Patent Information
- Application Number
- CN202411069181.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-06
- Publication Date
- 2024-12-03
AI Technical Summary
The existing road network signal control strategy is difficult to achieve accurate signal control in complex road network environments, especially the lack of synergy between the center and the edge, which leads to inaccurate control strategies.
Adopting an adversarial learning method, through the adversarial game between traffic prediction and strategy generator, ubiquitous network traffic data are obtained for spatiotemporal feature extraction, combined with real signal schemes to generate optimal control strategies, and use edge computing capabilities to regulate path impedance to avoid data Large quantities and high-dimensional solutions.
It significantly improves the quality and accuracy of the generation strategy, makes it closer to or better than the real signal scheme, has better stability and accuracy, can fully explore the internal laws and correlations of traffic data, and generate high-value characteristic parameters. sequence.
Smart Images

Figure CN119068688A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of road network control strategies, and in particular to a method for generating a road network signal control strategy based on adversarial learning. Background Art
[0002] The signal control strategy of intersections and paths is an important means to ensure the efficiency of intersections and balance the traffic load of paths. In a connected environment, the path selection system is usually an autonomous and dynamic one. The software on each vehicle or mobile terminal device comprehensively selects the optimal route based on the expected travel time of the road network, but it is usually not the best for road network traffic control. However, due to the complexity of the road network scale, the large number of vehicles, and the dynamic randomness of traffic demand, it is impossible to solve the problem of real-time communication of information between both parties and the problem of dimensional disaster in large-scale road networks. Therefore, a single center-side traffic induction cannot be directly applied.
[0003] Traditional traffic control methods usually coordinate single intersections or single lines, and often pay less attention to the dynamic combination of intersections, lacking dynamic spatiotemporal correlation analysis and strategy coordination at the regional level of the road network. Reinforcement learning is a new type of traffic control method, but it focuses more on single intersections, and the mechanism integration of the coordination relationship between adjacent intersections is not deep enough, requiring guidance at the strategy level to form center-edge coordination.
[0004] For example, a "signal control method based on road network division and boundary flow control" disclosed in Chinese patent literature, with the announcement number: CN114038216B, discloses that the city's entire traffic road network is divided into sub-road network clusters using a road network division method based on maximum coordination, and each intersection of the sub-road network is regarded as an intelligent agent, and a multi-agent reinforcement learning method based on a deep neural network is designed to learn and generate the optimal coordinated control strategy for the signal lights in the sub-road network through continuous interaction with the surrounding traffic environment during the execution of actions by the intelligent agents in the sub-road network. At the same time, by calibrating the macro basic graph model of the sub-road network, the flow control is performed on the boundary sections and intersections of the sub-road network, and the traffic operation status and traffic efficiency of the sub-road network are further optimized. However, this solution does not take into account the synergy between the prediction center and the edge, resulting in inaccurate control strategies. Summary of the invention
[0005] In order to solve the problem that the road network signal control strategy in the prior art is not accurate enough, the present invention provides a road network signal control strategy generation method based on adversarial learning, which generates a road network signal strategy through adversarial game between traffic prediction and strategy generator.
[0006] In order to achieve the above object, the present invention provides the following technical solutions: A method for generating a road network signal control strategy based on adversarial learning, comprising: obtaining ubiquitous network traffic data to extract spatiotemporal features to obtain a feature parameter sequence; Obtaining a real signal scheme for decoding and extraction to obtain a feature of the real signal scheme, and combining the feature parameter sequence to generate a strategy to obtain a first strategy; Determine the first signal scheme characteristics of the first strategy and determine the equalization loss in combination with the characteristic parameter sequence, and generate the discrimination loss by determining the accuracy of the first strategy; The prediction loss is generated based on the characteristic parameter sequence combined with the characteristics of the real signal scheme, and the optimal control strategy is determined by combining the discriminant loss and the equilibrium loss. The path impedance is changed by regulating the signal control strategy, and there is no need to obtain the vehicle-level traffic demand at the center. Compared with the path induction method at the center, it avoids the problems of large data volume, high solution dimension, and strong autonomy of travelers; it can fully explore the inherent laws and correlations of traffic data to generate high-value characteristic parameter sequences; through the continuous confrontation of the strategy generation step and the strategy evaluation step, it can prompt the generator to continuously optimize the control strategy it generates to deceive the discriminator. This mechanism can significantly improve the quality of the generated strategy, making it closer to or better than the real signal scheme.
[0007] As a preferred method, when performing spatiotemporal feature extraction, a spatial graph convolution model is used to standardize parameters using the characteristic distribution of each parameter in the data set to generate a corresponding feature parameter sequence. The spatial graph convolution layer can capture local spatial relationships through convolution operations between nodes, while the spatiotemporal convolution layer further introduces the time dimension to achieve comprehensive modeling of spatiotemporal data. The dimensional influence between different features is eliminated, making it easier to flexibly combine different features.
[0008] As a preferred method, a road network graph model is established, and intersections in ubiquitous networked traffic data are used as edges of the graph model, and road sections are used as nodes of the graph model; determining node features through a graph convolution model can intuitively and accurately reflect the topological structure of the traffic network, including the connection relationship between roads and the layout of intersections. It not only describes the physical connection between roads, but also implies the flow direction of traffic flow and possible interaction points.
[0009] Preferably, the feature parameter sequence includes the spatiotemporal features of traffic parameters, including the use of an attention mechanism to adaptively capture the relevance of spatial nodes. Not only the temporal variation characteristics of traffic parameters are considered, but also spatial features are incorporated, which can more comprehensively and accurately reflect the dynamics and complexity of the traffic system. The relevance between different spatial nodes can be adaptively captured, and the weights between different nodes can be dynamically adjusted, so that more attention is paid to nodes that have a greater impact on the results during feature extraction.
[0010] Preferably, after obtaining the real signal scheme, the real signal scheme is encoded and phase vectorized to establish a phase vector queue; the phase vector queue is decoded to obtain the real signal scheme characteristics. By converting the complex signal waveform into a vector form, the data volume is effectively compressed and important phase characteristics are retained.
[0011] As a preferred method, a spatial graph convolution model and a time series model are used as traffic spatiotemporal prediction models, and the real signal scheme features and the first sequence are used as inputs. The prior knowledge and useful information contained in these features can be fully utilized to provide strong support for the prediction model. The fusion of the real signal features and the first sequence information can further enhance the prediction ability of the model, making the prediction results more accurate and reliable.
[0012] Preferably, when generating the discrimination loss, the first strategy is judged by a source discriminator, the input of the source discriminator is the real signal scheme encoding, and the output is the signal scheme source label. In the case of multiple signal sources coexisting, the signal sources can be clearly divided, reducing the error of misjudgment caused by hidden signal confusion.
[0013] Preferably, when determining the balance loss, the first signal scheme feature and the parameters in the spatiotemporal feature of the traffic parameter are differentiated, and the balance loss is determined according to the difference result. The difference between the first signal scheme feature and the spatiotemporal feature of the traffic parameter can be quantified, thereby accurately measuring the degree of deviation between the signal scheme and the actual traffic condition.
[0014] Preferably, when generating the prediction loss, the difference between the actual signal scheme characteristics and the parameters in the spatiotemporal characteristics of the traffic parameters is calculated, and the prediction loss is determined according to the difference result. The deviation between the prediction result and the actual situation can be quantitatively evaluated, and not only the overall trend of the signal scheme can be taken into account, but also the local details and spatiotemporal changes.
[0015] Preferably, the spatiotemporal prediction model has the same parameters when determining the predicted loss and when determining the equilibrium loss; the output of the spatiotemporal prediction model is the traffic parameters for several steps in the future. This ensures the consistency of the model between prediction and equilibrium, reduces the uncertainty introduced by parameter differences, and can adapt faster and maintain stable performance when data or tasks change.
[0016] The present invention has the following advantages: (1) It makes full use of ubiquitous network data and the edge computing capabilities of the vehicle side, and changes the path impedance by regulating the signal control strategy. There is no need to obtain vehicle-level traffic demand at the center. Compared with the center-side path induction method, it avoids the problems of large data volume, high solution dimension, and strong autonomy of travelers. (2) In the model training process, the adversarial learning method is introduced to explore the knowledge representation in the dynamic game equilibrium of traffic. Compared with the existing deep learning method, it takes into account the road network correlation, signal strategy and spatiotemporal evolution, and has better stability and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] The drawings described below are merely exemplary. A person skilled in the art can derive other implementation drawings based on the provided drawings without inventive effort.
[0018] Figure 1 is a schematic diagram of the model logic in the embodiment.
[0019] Figure 2 Schematic diagram of the method steps in the embodiment. DETAILED DESCRIPTION
[0020] Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in the field without making any creative work shall fall within the scope of protection of the present invention.
[0021] like Figure 1 As shown, in a preferred embodiment, the present invention discloses a method for generating a road network signal control strategy based on adversarial learning, comprising the following steps: first, respectively obtain intersection data and road section data, wherein the intersection data and road section data are corresponding data of the current time node obtained in real time. Perform spatiotemporal feature extraction on the intersection data and road section data to obtain spatiotemporal features. Filter intersection signal scheme features in the spatiotemporal features, and predict losses based on the intersection signal scheme features and the spatiotemporal prediction model. And filter intersection spatiotemporal features in the spatiotemporal features.
[0022] At the same time, the real signal scheme is obtained, and the real signal scheme features are generated through the signal encoder according to the real signal scheme. The real signal scheme features are connected in series with the intersection spatiotemporal features to generate the first strategy. The spatiotemporal prediction model is trained according to the real signal scheme features and the intersection spatiotemporal features to determine the prediction loss.
[0023] The method includes establishing a first spatiotemporal prediction model and a second spatiotemporal prediction model with shared parameters. The first spatiotemporal prediction model predicts loss according to the spatiotemporal characteristics of the intersection and the characteristics of the signal scheme at the intersection to obtain the predicted loss. The second spatiotemporal prediction model generates the balanced loss according to the spatiotemporal characteristics of the intersection and the characteristics of the signal scheme of the first strategy.
[0024] It includes judging the first strategy according to the source discrimination model to determine the discrimination loss, combining the equilibrium loss, prediction loss and discrimination loss to generate the total loss, and evaluating the generated first strategy. The equilibrium loss and discrimination loss are weighted and summed according to the fusion weight.
[0025] The ubiquitous network data and spatiotemporal graph model are used to predict the traffic flow and speed of the road network. At the same time, signal scheme coding is introduced into the input data to explore the impact of traffic control measures on road network traffic.
[0026] Signal strategies are generated by encoding and decoding signal schemes. Compared with the reinforcement learning method using behavioral decision-feedback in similar research and technology, the generated signal strategies do not need to rely on existing manually configured scheme libraries, behavior libraries, etc., and have better flexibility and adaptability.
[0027] Through adversarial learning between the traffic predictor and the strategy generator, and between the source discriminator and the strategy generator, the pre-trained model can generate a signal control strategy that is as realistic as possible and can guide travel path selection, with better learning process stability.
[0028] like Figure 2 As shown in the figure, this scheme discloses a method for generating a road network signal control strategy based on adversarial learning, including the following steps: obtaining ubiquitous networked traffic data for spatiotemporal feature extraction to obtain a feature parameter sequence; firstly, by obtaining the vehicle trajectory, traffic flow, and signal light status of the ubiquitous networked traffic data, and using spatiotemporal feature extraction technology, the inherent laws and correlations of the traffic data can be fully explored, thereby generating a high-value feature parameter sequence. This data-driven approach is more accurate and efficient than traditional rule-based or experience-based methods.
[0029] The real signal scheme is obtained for decoding and extraction to obtain the real signal scheme features, which not only considers the current traffic status but also implies comprehensive factors. The first strategy is generated by combining the feature parameter sequence. By decoding the real signal scheme and extracting its features, and combining the spatiotemporal characteristics of traffic parameters for strategy generation, a first strategy that is closer to the actual traffic conditions can be generated.
[0030] Determine the first signal scheme characteristics of the first strategy and determine the equilibrium loss in combination with the characteristic parameter sequence. The equilibrium loss focuses on the balance of the generated strategy in terms of traffic efficiency, safety, fairness, etc. The discrimination loss is generated by judging the accuracy of the first strategy; the discrimination loss focuses on the accuracy of the strategy. A multi-dimensional loss evaluation is achieved.
[0031] The predicted loss is generated based on the characteristic parameter sequence combined with the characteristics of the real signal solution. At this time, the deviation between the characteristic parameter sequence and the characteristics of the real signal solution is determined. The accurate strategy deviation can also be determined for complex and changeable traffic environments. The optimal control strategy is determined by combining the discrimination loss and the equilibrium loss. Make full use of ubiquitous network data and the edge computing capabilities on the vehicle side, change the path impedance through signal control strategy regulation, and do not need to obtain vehicle-level traffic requirements at the center. Compared with the center-side path induction method, it avoids the problems of large data volume, high solution dimension, and strong autonomy of travelers; it can fully explore the inherent laws and correlations of traffic data to generate high-value characteristic parameter sequences. This mechanism can significantly improve the quality of the generated strategy, making it closer to or better than the real signal solution.
[0032] In other embodiments, the present solution includes the use of a spatial graph convolution model. A spatial graph convolution model such as ST-GCN is particularly suitable for processing problems with graph structured data. This model can simultaneously capture spatial relationships such as connections between nodes and temporal dynamics such as the evolution of time series, thereby effectively extracting rich features from spatiotemporal data. The spatial graph convolution layer can capture local spatial relationships through convolution operations between nodes, while the spatiotemporal convolution layer further introduces the time dimension to achieve comprehensive modeling of spatiotemporal data. Parameter standardization using the characteristic distribution of each parameter in the data set helps to eliminate the dimensional influence between different features, making different features numerically comparable. By standardizing the characteristic distribution of each parameter in the data set, the generated feature parameter sequence can be made more stable in terms of numerical value, thereby improving the efficiency and accuracy of model training. In addition, the standardization process can also reduce the sensitivity of the model to outliers and improve the robustness of the model. Generate a corresponding feature parameter sequence. By generating a corresponding feature parameter sequence, different features can be flexibly combined for subsequent analysis and modeling. Including fusing the features extracted by the spatial graph convolution model with other types of feature optical flow features and trajectory features to further improve the performance of the model. In addition, the features can be further optimized according to the requirements of specific tasks, such as feature selection and dimensionality reduction, to improve the efficiency and accuracy of the model.
[0033] In other embodiments, the present solution includes establishing a road network graph model. Determining node features through a graph convolution model can intuitively and accurately reflect the topological structure of the traffic network. The graph convolution model can capture the spatial relationship between nodes and determine the characteristics of the nodes in combination with traffic data to provide accurate input for traffic flow prediction. The intersections in the ubiquitous networked traffic data are used as the edges of the graph model, and the road sections are used as the nodes of the graph model. This modeling method can intuitively reflect the topological structure of the traffic network, including the connection relationship between roads and the layout of intersections. It can not only describe the physical connection between roads, but also imply the direction of traffic flow and possible interaction points.
[0034] In other embodiments, the feature parameter sequence in the present solution includes spatiotemporal features of traffic parameters, which not only considers the time-varying features of traffic parameters, but also incorporates spatial features. This comprehensive spatiotemporal feature modeling can more accurately reflect the dynamics and complexity of the traffic system. Including the use of an attention mechanism to adaptively capture the correlation of spatial nodes. By introducing the attention mechanism, the scheme can adaptively capture the correlation between different spatial nodes. The attention mechanism can dynamically adjust the weights between different nodes, so that in the process of feature extraction, more attention is paid to nodes that have a greater impact on the results, thereby improving the accuracy and efficiency of feature extraction. It can adaptively capture the correlation between different spatial nodes and dynamically adjust the weights between different nodes, so that in the process of feature extraction, more attention is paid to nodes that have a greater impact on the results.
[0035] The weights generated by the attention mechanism can be visualized as a correlation strength graph between different nodes, which helps to determine the mutual influence relationship between nodes in the traffic network. Since the attention mechanism can adaptively adjust the correlation between nodes, the scheme has better adaptability to different traffic scenarios and data sets. Even at different times, locations or traffic conditions, the model can maintain high performance, improving its generalization ability.
[0036] In other embodiments, the present scheme includes, after obtaining the real signal scheme, performing phase vectorization on the encoding of the real signal scheme, by converting the complex signal waveform into a vector form, thereby effectively compressing the amount of data and retaining important phase characteristics; compared with the amplitude information of the signal, the phase information is less affected by noise and interference during the transmission process, so the signal after phase vectorization encoding has a stronger anti-interference ability. Establish a phase vector queue; each vector in the phase vector queue represents a specific feature of the signal, which makes it easier to extract the key features of the signal during the decoding process. Decode the phase vector queue to obtain the real signal scheme characteristics. Be able to respond quickly and feedback the processing results to meet the real-time requirements.
[0037] In other embodiments, the present solution includes the use of a spatial graph convolution model and a time series model as a traffic spatiotemporal prediction model. The spatial graph convolution model is good at capturing the spatial relationship between different nodes in the traffic network, such as road sections and intersections, and can reflect the propagation and mutual influence of traffic flow. The time series model focuses on the analysis and prediction of time series data, and can capture the trend and periodicity of traffic flow over time. And the real signal scheme features and the first sequence are used as input. The prior knowledge and useful information contained in these features can be fully utilized to provide strong support for the prediction model. The fusion of the real signal features with the first sequence information can further enhance the prediction ability of the model, making the prediction results more accurate and reliable.
[0038] As the initial part or specific sequence of the time series, the first sequence contains the historical data required for prediction, which helps the model understand the historical evolution of traffic flow. The model can dynamically adjust the prediction strategy based on the input real signal characteristics and the first sequence information to adapt to changes in traffic conditions.
[0039] In other embodiments, the present solution includes, when generating the discriminant loss, judging the first strategy by a source discriminator, which is specially trained to distinguish signal scheme codes from different sources, and can more accurately identify the source label of the signal scheme. This clear distinction helps to apply the source information more accurately in subsequent processing. The input of the source discriminator is the real signal scheme code, and the output is the signal scheme source label. In the case of multiple signal sources coexisting, the signal sources can be clearly divided, reducing the error of misjudgment caused by hidden signal confusion.
[0040] Encoding the real signal scheme as the input of the source discriminator means that the encoding process has effectively extracted and compressed the features of the original signal. This encoding method helps the source discriminator process and understand the input data more efficiently. The encoded signal scheme contains rich feature information, and the source discriminator will make full use of these features in the process of judging the source label, thereby improving the accuracy and robustness of the discrimination. With the emergence of new signal sources or changes in the characteristics of existing signal sources, the training data and parameters of the source discriminator can be updated to adapt to these changes and maintain the effectiveness and accuracy of the discrimination mechanism.
[0041] In other embodiments, the present scheme includes taking the difference between the first signal scheme feature and the parameters in the spatiotemporal feature of the traffic parameters to quantify the difference between the two, thereby accurately measuring the degree of deviation between the signal scheme and the actual traffic conditions. The difference result directly reflects the degree of mismatch between the signal scheme and the traffic parameters, making the determination of the equilibrium loss more intuitive. The equilibrium loss is determined based on the difference result. As the objective function of model optimization, the value of the equilibrium loss directly guides the adjustment direction of the model parameters. The equilibrium loss determined by the difference can more accurately guide the model to optimize in the direction of reducing the difference between the signal scheme and the traffic parameters. As the model is continuously optimized, the equilibrium loss gradually decreases, and the signal scheme will become closer and closer to the actual traffic conditions, thereby improving the overall performance and prediction accuracy of the model.
[0042] This scheme not only considers the characteristics of the signal scheme, but also integrates the spatiotemporal characteristics of traffic parameters, so that the determination of the balance loss can comprehensively evaluate the comprehensive impact of the signal scheme in time and space. Since traffic conditions are dynamically changing, this scheme considers spatiotemporal characteristics so that the balance loss can dynamically reflect the adaptability of the signal scheme at different times and different spatial locations. Determining the balance loss by taking differences can make the model more robust in the face of interference factors such as noise and outliers. Because the difference operation itself has a certain anti-interference ability, it can reduce the impact of these interference factors on the calculation of the balance loss. This scheme is applicable to a variety of traffic scenarios and data sets because it determines the balance loss based on the differences between signal schemes and traffic parameters, and does not rely on specific traffic models or assumptions.
[0043] In other embodiments, the present solution includes taking the difference between the parameters in the real signal scheme characteristics and the spatiotemporal characteristics of the traffic parameters, and directly taking the difference between the real signal scheme characteristics and the predicted signal scheme characteristics (or the spatiotemporal characteristics of the traffic parameters), so as to quantitatively evaluate the deviation between the prediction result and the actual situation. This deviation is a direct indicator to measure the accuracy of the prediction. The prediction loss obtained by taking the difference not only considers the overall trend of the signal scheme, but also covers the local details and spatiotemporal changes, so it can more comprehensively reflect the performance of the prediction model. The prediction loss is determined according to the difference result.
[0044] As an important feedback signal in the model training process, the prediction loss, its value and change trend, directly guide the adjustment direction of the model parameters. The prediction loss determined by the difference can clearly point out the deficiencies of the model. This scheme takes into account the characteristics of the real signal scheme and the spatiotemporal characteristics of the traffic parameters when generating the prediction loss, and fully captures the complexity of the traffic system. This comprehensive consideration helps the model better understand the dynamic changes of traffic flow. The prediction loss that integrates spatiotemporal characteristics can prompt the model to take into account both time and space dimensions when predicting, thereby enhancing its ability to predict complex traffic conditions.
[0045] In other embodiments, the present solution includes that the spatiotemporal prediction model has the same parameters when determining the prediction loss and when determining the equilibrium loss; this parameter sharing mechanism ensures the consistency of the model between prediction and equilibrium, and reduces the uncertainty and error that may be introduced by parameter differences. Since the model parameters remain unchanged, when the data or task changes, the model can adapt more quickly and maintain stable performance because there is no need to retrain or adjust the model parameters. The output of the spatiotemporal prediction model is the traffic parameters for several steps in the future. This means that the model has strong long-term prediction capabilities. This ability is crucial for fields such as urban traffic planning and traffic flow management because they need to understand the changing trends of traffic conditions over a long period of time in the future. Since the model parameters remain consistent during the prediction and equilibrium process, the model can better capture the spatiotemporal correlations between traffic parameters.
[0046] This includes adjusting the weight distribution of prediction loss or introducing additional regularization terms to balance model complexity and prediction performance. Considering both prediction loss and equilibrium loss can provide a more comprehensive perspective on model performance evaluation.
[0047] In another specific embodiment, the present solution discloses a method for generating a road network signal control strategy based on adversarial learning. Based on historical traffic data, the knowledge data of traffic demand change trends and travelers' path selection behaviors are analyzed, and a road network signal strategy based on path induction is generated through adversarial game between traffic prediction and strategy generator.
[0048] It includes networked traffic parameters based on the road network, and uses the spatiotemporal attention mechanism and graph model to extract the spatiotemporal characteristics of the road network. A traffic predictor is established based on the spatiotemporal characteristics of the road network and the historical signal scheme, and the traffic flow and speed changes of the road network are predicted according to the traffic predictor. The networked traffic parameters include queue length, number of stops, road section speed, road section length, and number of lanes.
[0049] A strategy generator is established to form a recommended signal scheme; based on the spatiotemporal characteristics of the road network, the changes in the road network under the recommended signal scheme are predicted through the traffic forecast period.
[0050] During the model training process, the prediction loss is calculated based on the future actual traffic and speed distribution, and the equilibrium loss is determined with the goal of balancing the road network traffic and speed distribution. The accuracy of the traffic predictor and the effectiveness of the strategy generator are improved through adversarial learning.
[0051] In another embodiment, the scheme includes establishing a spatiotemporal feature extraction model Ext, a spatiotemporal prediction model pre, a signal scheme encoder SigEn, a signal control strategy generator SigGen, a source discriminator SrcDis, and a signal feature conversion conv.
[0052] In another specific embodiment, the scheme includes the following steps.
[0053] Extract spatiotemporal features, including but not limited to through the spatial graph convolution model GCN. This includes obtaining ubiquitous networked traffic data based on fixed detectors placed at intersections or connected floating vehicles. The ubiquitous networked traffic data includes but is not limited to traffic parameters including total flow, number of flow lanes, average flow of flow lanes, green light speed at intersections, road section speed, queue length at the end of traffic lights, delay time, and distance between adjacent intersections. Obtain a ubiquitous networked traffic data set, and perform parameter standardization on the characteristic distribution of each parameter in the ubiquitous networked traffic data set. The corresponding characteristic parameter sequence in the nth section is x n ∈R p*c , where p is the length of the time series and c is the input parameter dimension.
[0054] The initial road network graph model is established as G = (V, E, A), where the road segments are nodes of the graph model, and V is defined as N v The set of road segments, the embedding of the node is expressed as: v n =σ(w FFN1 *x n +b FFN1 ); Among them, w FFN1 With b FFN1 are the weights and Bayesian parameters of the feedforward layer, σ is the activation function. The intersection is the edge of the graph model, and E is defined as N E The shortest distance between the i-th section and the j-th section is d ij , the travel time is τ i,t and τ j,t , define the element w in the dynamic adjacency matrix A i,j : Among them, d σ is the spatial scaling scale, ε is the distance threshold, and coef(·) is the correlation coefficient function of the time series. At this time, the generated Laplace matrix is Where I is the identity matrix, D is the diagonal matrix of A, and λ max is the maximum eigenvalue of the L matrix. Finally, the node features are calculated through the graph convolution model GCN: Among them, w k is the weight parameter of the k-order neighbor of the graph convolution, b g is the Bayesian parameter of graph convolution, T k(·) is the Chebyshev polynomial, and SA is the spatial node association adaptively captured using the attention mechanism: SA=softmax(V SA ·σ((v n ·W SA1 )·W SA2 ·(W SA3 ·v n )+b SA )); Among them, W SA1 , W SA2 , W SA3 is the weight parameter, b SA is the Bayesian parameter, and softmax is the normalization function.
[0055] The time feature analysis is extracted through a time series analysis model, including but not limited to RNN, LSTM, GRU, etc. In this example, GRU is used as an example: fst n =GRU(fs n ).
[0056] A signal control strategy generator is established. Based on the signal scheme, according to the phase start symbol, the cycle length, absolute phase difference, etc. are used for vectorization, which is recorded as For each phase, the phase is vectorized according to the release flow to the inlet, the direction flow to the outlet, the green light duration, the yellow light duration, the red light duration, etc., and recorded as Among them, gt mov1 The green light duration of the phase in the released direction, the last bit is encoded as the phase end symbol. Establish the phase vector sequence M is the phase length after the input length is padded.
[0057] The Transformer model is used to encode and decode the signal scheme. The core of its attention mechanism consists of three matrices: Query Key and Value The feature dimensions are d and d respectively. k With d v , PE is the position code. The attention mechanism realizes the distribution of attention through dot product operation: In the formula, QK T To analyze the similarity of features at different locations, is the scaling factor used for normalization. The encoder (·) in the model consists of a multi-head attention layer, a feedforward network layer, and a regularization layer, and the decoder (·) consists of two layers of multi-head attention layers, a feedforward network layer, and a regularization layer. The signal scheme features extracted by the encoder are combined with the spatiotemporal features of the traffic parameters fst n The series is denoted as fsc n :
[0058] Signal strategy generation via decoder: fsc′ n,m =Decoder(fsc n ,φ n , Mask m ); Among them, Fc(·) is the fully connected layer, Mask m Padding mask for sequence output To perform traffic spatiotemporal prediction, in this embodiment, the spatial graph convolution model GCN and the time series model GRU are used as traffic spatiotemporal prediction models. During the prediction process, the input feature sequence is determined based on the real signal scheme and the generated signal strategy.
[0059] Among them, conv(·) is a sequence feature. The model predicts the traffic parameters of the next few steps: the traffic flow of the road section and the speed change of the road section, which are recorded as y n (t+r) and y′ n (t+r).
[0060] Establish a source discriminator, which includes two convolutional layers and two fully connected layers, and encodes φ with a signal scheme n ,φ′ n As input, the output signal scheme source label sl n , and is provided with a real scheme label. The real scheme label includes: the real signal scheme is 0, and the generated signal strategy is 1.
[0061] Conducting adversarial learning of travel control, in order to train the prediction ability of feature extractor and prediction model, forms the prediction loss function L according to the deviation between the predicted traffic parameters and the supervision target A : L A =RMSE(y n , Yn ); Among them, Y n is the real traffic parameter.
[0062] In order to improve the traffic induction balance ability of the generated signal scheme, the balance loss function L is determined according to the deviation between the predicted traffic parameters and the balanced traffic parameter target. B : L B =RMSE(y′ n , Y′ n ); Among them, Y′ n is the target traffic parameter after equilibrium.
[0063] In order to improve the rationality and effectiveness of the generated signal scheme, the discrimination loss L is determined according to the accuracy of the model in distinguishing whether the signal scheme is real or generated. C : The final total loss function is defined as: L=L A +α B L B -α C L c ; Among them, α B With α C They are fusion weights respectively. In the specific model training process, they include the spatiotemporal prediction model Pre, the signal scheme encoder SigEn, the signal control strategy generator SigGen, the source discriminator SrcDis, and the signal feature conversion conv. Ext , Lr Pre , lr SigEn , lr conv , lr SigGen , lr SrcDis are the training learning rates of different modules, LT B With LT c are the convergence judgment thresholds in local training respectively.
[0064] These modifications or improvements made on the basis of the present invention without departing from the spirit of the present invention all fall within the scope of protection required by the present invention.
Claims
1. A method for generating a road network signal control strategy based on adversarial learning, characterized in that: include: Obtain ubiquitous networked traffic data to extract spatiotemporal features and obtain feature parameter sequences; Obtaining a real signal scheme for decoding and extraction to obtain a feature of the real signal scheme, and combining the feature parameter sequence to generate a strategy to obtain a first strategy; Determine the first signal scheme characteristics of the first strategy and determine the equalization loss in combination with the characteristic parameter sequence, and generate the discrimination loss by determining the accuracy of the first strategy; The prediction loss is generated according to the characteristic parameter sequence combined with the real signal scheme characteristics, and the optimal control strategy is determined by combining the discriminant loss and the equilibrium loss.
2. According to the method for generating a road network signal control strategy based on adversarial learning in claim 1, it is characterized in that: When extracting spatiotemporal features, a spatial graph convolution model is used to standardize parameters using the characteristic distribution of each parameter in the data set to generate a corresponding feature parameter sequence.
3. The method for generating a road network signal control strategy based on adversarial learning according to claim 2, characterized in that: It includes establishing a road network graph model, taking intersections in ubiquitous connected traffic data as edges of the graph model, and taking road sections as nodes of the graph model; and determining node features through a graph convolution model.
4. A method for generating a road network signal control strategy based on adversarial learning according to claim 1, 2 or 3, characterized in that: The feature parameter sequence includes the spatiotemporal characteristics of traffic parameters, including the use of an attention mechanism to adaptively capture spatial node correlations.
5. A method for generating a road network signal control strategy based on adversarial learning according to claim 1, 2 or 3, characterized in that: After obtaining the real signal scheme, the real signal scheme is encoded and phase vectorized to establish a phase vector queue; the phase vector queue is decoded to obtain the real signal scheme characteristics.
6. A method for generating a road network signal control strategy based on adversarial learning according to claim 1 or 2, characterized in that: It includes adopting spatial graph convolution model and time series model as traffic spatiotemporal prediction model, and taking the real signal scheme features and the first sequence as input.
7. The method for generating a road network signal control strategy based on adversarial learning according to claim 5, characterized in that: When generating the discriminative loss, the first strategy is judged by the source discriminator, the input of the source discriminator is the true signal scheme encoding, and the output is the signal scheme source label.
8. The method for generating a road network signal control strategy based on adversarial learning according to claim 4, characterized in that: When determining the balance loss, it includes taking the difference between the first signal scheme characteristic and the parameters in the spatiotemporal characteristic of the traffic parameter, and determining the balance loss according to the difference result.
9. A method for generating a road network signal control strategy based on adversarial learning according to claim 8, characterized in that: When generating the prediction loss, it includes taking the difference between the parameters in the real signal scheme characteristics and the spatiotemporal characteristics of the traffic parameters, and determining the prediction loss based on the difference result.
10. A method for generating a road network signal control strategy based on adversarial learning according to claim 9, characterized in that: The spatiotemporal prediction model has the same parameters when determining the predicted loss and when determining the equilibrium loss; the output of the spatiotemporal prediction model is the traffic parameters of several steps in the future.
