A resilient traffic scheduling method

Abstract

The present application relates to a kind of toughness traffic scheduling method, comprising: S1: obtaining real-time traffic data and historical traffic flow data;The real-time traffic data includes vehicle GPS track, passenger order request, road sensor data and traffic event report;S2: the real-time traffic data and historical traffic flow data are input into uncertainty perception model, and the traffic state of each road section in target period is predicted, and the future traffic condition of each road section is obtained;S3: according to the future traffic condition of each road section, the optimal scheduling strategy is generated;S4: according to the optimal scheduling strategy, traffic scheduling is carried out.The toughness traffic scheduling method of the present application has the advantages of improving robustness and strain capacity and reducing risk blind.

Description

Technical Field

[0001] This invention relates to the field of traffic scheduling, and in particular to a resilient traffic scheduling method. Background Technology

[0002] Real-time traffic dispatching systems dynamically allocate tasks and plan routes for vehicles in fleets such as buses and ride-hailing services based on real-time traffic conditions and passenger orders. With the development of smart cities and shared mobility, traffic dispatching based on real-time passenger orders and traffic conditions for dynamic bus connections, demand-response bus (DRT), and ride-hailing carpooling has become a key technology for solving the "last mile" problem and meeting personalized travel needs in cities.

[0003] Currently, a data-driven artificial intelligence approach based on "perception-prediction-optimization" is commonly used to improve scheduling efficiency, cost, and user experience. This approach typically integrates multi-source data such as GPS trajectories, order requests, and traffic information, and utilizes deep learning models, such as spatiotemporal graph neural networks, to estimate and predict short-term traffic speeds or travel demand, thus achieving perception and prediction. Subsequently, based on the predicted values, the scheduling problem is modeled as a dynamic vehicle routing problem, and solved using heuristic algorithms or reinforcement learning to output vehicle-order allocation and route planning schemes, achieving optimization and decision-making.

[0004] However, while existing scheduling methods and systems have improved scheduling performance from different perspectives by introducing multi-model adaptation, intelligent resource allocation, or spatiotemporal deep learning, they all heavily rely on and assume that future "point predictions" are accurate or approximately accurate, treating prediction results in different scenarios as deterministic inputs. In reality, short-term prediction errors are unavoidable when dealing with urban transportation systems that are highly complex and inherently uncertain. For existing scheduling methods that rely on accurate "point predictions," the deviation of actual traffic conditions from the predictions will cause a sharp deterioration in system performance, leading to a surge in waiting times, severe delays, and continuous failures in subsequent order matching, resulting in poor robustness that undermines the stability of the scheduling plan. Secondly, whether relying on a single prediction model or a multi-model selection strategy, current scheduling methods cannot quantify and distinguish between a "short average time but large fluctuation (high risk)" scheme and a "slightly longer average time but very stable" scheme, causing the scheduling system to continuously select high-risk schemes, resulting in risk blindness. Moreover, when new or complex emergencies such as traffic accidents or temporary traffic control occur that are not fully reflected in historical data, the prediction models trained based on historical patterns may completely fail, and their ability to cope with emergencies is severely inadequate. Summary of the Invention

[0005] Based on this, the purpose of this invention is to improve the responsiveness and robustness of traffic scheduling methods to emergencies, while reducing the blind spot of wind direction, and to provide a resilient traffic scheduling method.

[0006] A resilient traffic scheduling method includes the following steps:

[0007] S1: Acquire real-time traffic data and historical traffic flow data; the real-time traffic data includes vehicle GPS trajectories, passenger order requests, road sensor data, and traffic incident reports; S2: Input the real-time traffic data and historical traffic flow data into the uncertainty perception model to predict the traffic status of each road segment during the target time period, and obtain the future traffic conditions of each road segment; S3: Generate the optimal scheduling strategy based on the future traffic conditions of each road segment; specifically including: S31a: Generate at least two future traffic state scenarios based on the future traffic conditions of each road segment; each future traffic state scenario is a set of possible road segment speed assignments. S31b: Generate at least one scheduling action based on the real-time traffic data; S32: Perform path planning for each scheduling action based on each future traffic state scenario to obtain each scheduling strategy; S33: Decision evaluation is performed on each of the aforementioned scheduling strategies to obtain the resilience cost; S34: Calculate the overall fitness based on the decision evaluation results of each scheduling strategy; S35: Eliminate scheduling strategies whose fitness does not reach the preset selection threshold to obtain the winning strategies; S36: Perform crossover and mutation on each of the winning strategies to generate a child scheduling strategy; S37: Repeat steps S33 to S36 to iterate until the number of iterations meets the preset number; S38: Filter the scheduling strategies of each child generation obtained in the last iteration to obtain the optimal scheduling strategy.

[0008] S4: Perform traffic scheduling according to the optimal scheduling strategy.

[0009] Compared to existing technologies, this invention introduces a quantitative penalty term for uncertainty and extreme situations in scheduling decisions by employing a resilience function. This allows the system to proactively weigh the expected performance against potential risks during decision-making, prioritizing stable solutions. This avoids the systemic high risk introduced by traditional methods that pursue local theoretical optima, thus enhancing the system's long-term robustness. Subsequently, through continuous iterative optimization of scheduling actions and strategies, the system can still evaluate the most adaptable decisions through multi-scenario simulations when facing entirely new emergencies not covered by historical data. Furthermore, it possesses self-learning and adjustment capabilities, improving the system's overall fault tolerance and intelligence level.

[0010] Furthermore, in step S33, the decision evaluation adopts a resilience cost function, the expression of which is: , in, For path The resilience cost, For travel time, The expected travel time for the path. This is a risk item; This is the worst-case scenario. and These are adjustable weighting coefficients, representing the degree of aversion to volatility risk and extreme situations, respectively; The worst-case term is defined by setting a travel time condition under a preset high confidence level. Quantiles were estimated using Monte Carlo simulation.

[0011] Furthermore, the calculation formula for the screening in step S38 is as follows: , in, This indicates that the optimal solution is sought among various scheduling actions. For scheduling strategy x in various future traffic scenarios The resilience cost in the middle, For scheduling strategy x in all future traffic scenarios The mathematical expectation of resilience cost, The worst-case resilience cost across all scenarios; This is the avoidance coefficient.

[0012] Furthermore, the expression for the overall fitness is: , Where M represents the number of future transportation scenarios. Indicating various future transportation scenarios, is a coefficient.

[0013] Furthermore, step S2 specifically includes: S21: Perform feature encoding on the real-time traffic data and historical traffic flow data respectively to obtain multi-source features with the same embedding representation; the multi-source features include spatiotemporal embedding vectors, trajectory pattern features, order distribution features, and event embedding vectors; S22: Adaptively fuse the multi-source features to obtain the comprehensive feature vector of each road segment; S23: Based on the comprehensive feature vector, calculate the future traffic conditions of each road segment during the target time period using a probabilistic parameterization structure.

[0014] Further, in step S21, the historical traffic flow data is spatiotemporally feature-encoded to obtain the spatiotemporal embedding vector, the vehicle GPS trajectory is trajectory feature-encoded to obtain the trajectory pattern feature, the passenger order request is order feature-encoded to obtain the order distribution feature, and the traffic event is order feature-encoded to obtain the order distribution feature. The spatiotemporal feature encoding employs a spatiotemporal graph neural network to encode the historical traffic flow data along with road network topology structures such as the adjacency matrix. The trajectory feature encoding uses a sequence encoder to extract trajectory pattern features from the vehicle's GPS trajectory; The order feature encoding uses an attention mechanism to aggregate the dynamic features of passenger order requests within the region; The event feature encoding uses a pre-trained language model to extract the event embedding vector from the traffic incident report.

[0015] Furthermore, in step S22, the adaptive fusion adopts a gated fusion mechanism, the expression of which is: , in, It is the sigmoid activation function. For learnable parameters, For element-wise multiplication, Let e ​​be the spatiotemporal embedding vector of road segment node e. The trajectory pattern features of road segment e, The order distribution characteristics of road segment e are as follows: Let e ​​be the event embedding vector for road segment e.

[0016] Furthermore, in step S23, the probabilistic parameterization structure adopts a hybrid density network structure, the expression of which is: , in, The number of mixed components. The weight of the k-th mixture component. ; Let be the mean of the k-th Gaussian components. Let be the standard deviation of the k-th Gaussian component. , and It is learned from the comprehensive feature vector through a neural network.

[0017] Based on the same inventive concept, the present invention also includes an electronic device comprising a processor; a memory for storing a computer program executed by the processor; wherein, when the processor executes the computer program, it implements the resilient traffic scheduling method described above.

[0018] To better understand and implement this invention, the following detailed description is provided in conjunction with the accompanying drawings. Attached Figure Description

[0019] Figure 1 This is an overall flowchart of the resilient traffic scheduling method of the present invention; Figure 2 Here is a structural diagram of the uncertainty perception model; Figure 3 Here is a detailed flowchart of step S2; Figure 4 This is a detailed flowchart of step S3. Detailed Implementation

[0020] This invention addresses the problems of poor adaptability and robustness, as well as risk blindness, in existing traffic scheduling methods by designing a resilient traffic scheduling method. This method changes the basis of scheduling decisions from predicting deterministic points to quantifying probabilistic uncertainties, and further constructs a risk perception assessment model and performs robust optimization for multiple scenarios. It proactively perceives and quantifies the inherent uncertainty of traffic conditions, thereby achieving reliable and highly resilient traffic scheduling with strong anti-interference capabilities.

[0021] Specifically, please refer to Figure 1 , Figure 1 This is an overall flowchart of the resilient traffic scheduling method of the present invention. The present invention provides a multi-task planning method, including: S1: Acquire real-time traffic data and historical traffic flow data. The real-time traffic data includes vehicle GPS trajectories, passenger order requests, road sensor data, and traffic incident reports.

[0022] S2: Input the real-time traffic data and historical traffic flow data into the uncertainty perception model to predict the traffic status of each road segment during the target time period, thereby obtaining the future traffic conditions of each road segment. The uncertainty perception model mathematically represents the degree of uncertainty in traffic conditions such as travel time and speed. After processing the traffic data and historical traffic flow data as input, it outputs a probability distribution representing the traffic conditions over a future period, generating a traffic state probability distribution map covering the entire road network. This associates each road segment or region with a specific probability distribution or its parameters, constituting a fundamental shift from a "deterministic point" to an "uncertain probability field," rather than a single deterministic value, which helps improve the resilience of subsequent decisions.

[0023] Please see Figure 2 and Figure 3 , Figure 2 This is a structural diagram of the uncertainty perception model. Figure 3 This is a detailed flowchart of step S2. Each structure of the uncertainty-aware model processes the input data, including: In the multi-source input coding layer M1, step S21 is performed: feature encoding is performed on the real-time traffic data and historical traffic flow data respectively to obtain multi-source features with the same embedding representation. The multi-source features include spatiotemporal embedding vectors, trajectory pattern features, order distribution features, and event embedding vectors; the feature encoding includes spatiotemporal feature encoding, trajectory feature encoding, order feature encoding, and time feature encoding.

[0024] The spatiotemporal feature encoding employs a spatiotemporal graph neural network to encode the historical traffic flow data along with the road network topology, such as the adjacency matrix. The spatiotemporal graph neural network captures spatial dependencies through graph convolutional layers and obtains the spatiotemporal embedding vector of each node by using gated recurrent units or temporal convolutional layers to study the evolution of the delivery time.

[0025] The trajectory feature encoding employs a sequence encoder to extract trajectory pattern features from the vehicle's GPS trajectory. These trajectory pattern features include road segment capacity and speed fluctuation patterns, among others. The sequence encoder utilizes either an LSTM or Transformer model.

[0026] The order feature encoding uses an attention mechanism to aggregate the dynamic features of passenger order requests within a region to obtain order distribution features; these dynamic features include order density and origin-destination flow.

[0027] The event feature encoding uses a pre-trained language model to extract the event embedding vector of the traffic incident report, which is an unstructured practice report including traffic accidents and temporary traffic control.

[0028] In the multimodal fusion layer M2, step S22 is performed: the multi-source features are adaptively fused to obtain the comprehensive feature vector of each road segment e. The adaptive fusion, performed using a gating fusion mechanism, is calculated using the following formula: , in, It is the sigmoid activation function. For learnable parameters, For element-wise multiplication, Let e ​​be the spatiotemporal embedding vector of road segment node e. The trajectory pattern features of road segment e, The order distribution characteristics of road segment e are as follows: Let e ​​be the event embedding vector for road segment e.

[0029] The probability output layer M3 executes step S23: based on the comprehensive feature vector, it calculates the future traffic conditions of each road segment during the target time period using a probabilistic parameterization structure. The probabilistic parameterization structure employs a hybrid density network structure, with parameters distributed as a mixture of Gaussian distributions as input. Its prediction result is a multimodal distribution that accurately characterizes traffic conditions, representing both smooth and congested traffic. Its expression is: , in, The number of mixed components. The weight of the k-th mixture component. ; The mean of the kth Gaussian components is the desired velocity. Let be the standard deviation of the k-th Gaussian component, which characterizes the degree of uncertainty under that component; and , and It is learned from the comprehensive feature vector through a neural network.

[0030] During training, the uncertainty-aware model is optimized using a negative log-likelihood loss function, the expression of which is: , in, For road section e Actual traffic conditions For each section e The comprehensive feature vector.

[0031] Please see Figure 4 , Figure 4 Here is a detailed flowchart of step S3. S3: Generate the optimal scheduling strategy based on the future traffic conditions of each road segment. Specifically, the genetic algorithm is used to generate the optimal scheduling strategy, and the specific steps include: S31a: Generate at least two future traffic state scenarios based on the future traffic conditions of each road segment. Each of these future traffic state scenarios constitutes a scenario set, denoted as... Each of the aforementioned future traffic state scenarios Each scenario assigns a set of possible road segment speeds. The future traffic scenarios are generated using Monte Carlo or Latin hypercube sampling techniques.

[0032] S31b: Generate at least one dispatching action based on the real-time traffic data. The dispatching action includes assigning the passenger order to one of the ride-hailing vehicles or buses, etc.

[0033] S32: Perform path planning for each scheduling action based on each future traffic state scenario to obtain a scheduling strategy. Each scheduling action obtains a corresponding scheduling strategy based on each future traffic state scenario, that is, each scheduling action obtains a scheduling strategy equal to the number of future traffic state scenarios.

[0034] S33: Decision evaluation is performed on each of the aforementioned scheduling strategies to obtain the resilience cost. The decision evaluation includes risk perception and matching assessment of the road segments. The decision evaluation uses a resilience cost function, the expression of which is: , in, For path The resilience cost, For travel time, To calculate the expected travel time of the path based on the preset expected speed, This is a risk item used to penalize uncertainty; This is the worst-case scenario option, used to guard against extreme risks; and These are adjustable weighting coefficients, representing the degree of aversion to volatility risk and extreme situations, respectively. The core of the risk term is quantifying the volatility of travel time; in this embodiment, standard deviation is used, and its expression is: , , in, Covariance describes the spatiotemporal correlation between road segments.

[0035] The worst-case term is defined by setting a travel time condition under a preset high confidence level. Quantiles, and estimated through Monte Carlo simulation; the Monte Carlo simulation estimate is derived from... N independent samples are drawn from the joint distribution, and the travel time corresponding to each independent sample is calculated. Get the set After sorting them in ascending order, take the first... Each value is used as the worst-case scenario term. In the process of evaluating candidate paths or matching order solutions, the resilience cost of the decision evaluation internalizes the uncertain future road condition information of the multimodal distribution of each road segment into a direct decision cost. In addition to traditional expected costs, it systematically incorporates quantitative penalties for volatility and extreme risks, thereby addressing the technical shortcomings of risk blindness.

[0036] S34: Calculate the overall fitness based on the decision evaluation results of each scheduling strategy. The expression for the overall fitness is: , in, M For the number of future transportation scenarios, Indicating various future transportation scenarios, is a coefficient.

[0037] S35: Eliminate scheduling strategies whose fitness does not reach the preset selection threshold to obtain the winning strategies.

[0038] S36: Perform crossover and mutation on each of the winning strategies to generate a child scheduling strategy.

[0039] S37: Repeat steps S33 to S36 for iteration until the preset number of iterations is met.

[0040] S38: The optimal scheduling strategy is obtained by filtering the child scheduling strategies obtained in the last iteration. The expression used for filtering in this step is: , in, This indicates that the optimal solution is sought among various scheduling actions. For scheduling strategy In various future transportation scenarios The resilience cost in the middle, For scheduling strategy In all future transportation scenarios The mathematical expectation of resilience cost, The worst-case resilience cost across all scenarios; This is the avoidance coefficient; the larger the value, the stronger the avoidance of the worst-case scenario.

[0041] S4: Perform traffic scheduling according to the optimal scheduling strategy.

[0042] Furthermore, during the traffic scheduling process based on the optimal scheduling strategy, the deviation between the actual vehicle trajectory and the prediction is detected in real time. When the deviation exceeds a preset threshold, the new data is fed back to the uncertainty quantification perception model for online updating.

[0043] A dynamic shuttle service between business districts and subway hubs during the evening rush hour in a large city is used as an example. The uncertainty quantification perception model predicts major commuter corridors based on real-time traffic data acquired during the evening rush hour. L The travel time for route 1 follows a normal distribution with an expected value of 24 minutes and a standard deviation of 6 minutes. Alternative corridors... L The travel time of 2 follows a normal distribution with an expected value of 26 minutes and a standard deviation of 2 minutes. The above data indicate that the main commuter corridor is on average faster but unreliable, while the alternative corridor is slightly slower but very stable.

[0044] When a batch of new orders is received, the scheduling plan is performed in step S3 of this invention based on the predicted future traffic conditions. Specifically, the worst-case scenario is preset to the 95th percentile time, and then... , The main commuter corridor was calculated. L Expected travel time for path 1 It lasts for 24 minutes. It lasts for 6 minutes. item Its resilience cost is 34 minutes. It takes 37.2 minutes; similarly, the alternative corridors L Expected travel time for path 2 It lasted 26 minutes. It lasts for 2 minutes. item Its resilience cost is 29.3 minutes. The timeframe was 35.79 minutes. During traffic scheduling, most vehicles were directed to alternative corridors with lower resilience costs and matched with order combinations that allowed for greater flexibility in the time window; a small number of vehicles that had to use the main commuter corridor due to their origin were matched with orders that had more flexible time windows.

[0045] In actual operation, unexpected congestion occurred on the main commuter corridor, with travel time reaching 40 minutes. However, due to risk avoidance decisions made in advance, most vehicles traveled on the stable alternative corridors, thereby stabilizing travel time within 30 minutes and maintaining an overall on-time rate of over 90%. A small number of vehicles traveling on the main commuter corridor did not cause complaints due to the ample time window for passengers. This successfully withstood the impact of uncertainty and demonstrated the resilience of this method in traffic scheduling.

[0046] In another embodiment, the probabilistic parameterization structure can also adopt a parameterization distribution such as a normal distribution, which can directly characterize the degree of uncertainty and provide a simpler representation of future traffic conditions.

[0047] In another embodiment, the probabilistic parameterization structure can also adopt a quantile prediction form, such as using quantile regression technology to directly output the prediction interval of future velocity. ,in, and These are the lower and upper quantiles for a given confidence level, and the width of this interval is an intuitive measure of uncertainty.

[0048] Compared to existing technologies, this invention introduces a quantitative penalty term for uncertainty and extreme situations in scheduling decisions by employing a resilience function. This allows the system to proactively weigh the expected performance against potential risks during decision-making, prioritizing stable solutions. This avoids the systemic high risk introduced by traditional methods that pursue local theoretical optima, thus enhancing the system's long-term robustness. Subsequently, through continuous iterative optimization of scheduling actions and strategies, the system can still evaluate the most adaptable decisions through multi-scenario simulations when facing entirely new emergencies not covered by historical data. Furthermore, it possesses self-learning and adjustment capabilities, improving the system's overall fault tolerance and intelligence level.

[0049] The terminology used in the embodiments of this application is for the purpose of describing particular embodiments only and is not intended to limit the embodiments of this application. The singular forms “a,” “the,” and “the” used in the embodiments and claims of this application are also intended to include the plural forms, unless the context clearly indicates otherwise. It should also be understood that, unless otherwise stated, “a plurality” means two or more; the terms “first,” “second,” “third,” etc., are used only to distinguish and not to describe a particular order or sequence, nor should they be construed as indicating or implying relative importance. The term “and / or” as used herein refers to and includes any or all possible combinations of one or more associated listed items. When the above description relates to drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. In the description of this application, those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0050] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and the present invention also intends to include these modifications and variations.

Claims

1. A resilient traffic scheduling method, characterized in that, Includes the following steps: S1: Acquire real-time traffic data and historical traffic flow data; the real-time traffic data includes vehicle GPS trajectories, passenger order requests, road sensor data, and traffic incident reports; S2: Input the real-time traffic data and historical traffic flow data into the uncertainty perception model to predict the traffic status of each road segment during the target time period, and obtain the future traffic conditions of each road segment; S3: Generate the optimal scheduling strategy based on the future traffic conditions of each road segment; specifically including: S31a: Generate at least two future traffic state scenarios based on the future traffic conditions of each road segment; each future traffic state scenario is a set of road segment speed assignments. S31b: Generate at least one scheduling action based on the real-time traffic data; S32: Perform path planning for each scheduling action based on each future traffic state scenario to obtain each scheduling strategy; S33: Decision evaluation is performed on each of the aforementioned scheduling strategies to obtain the resilience cost; S34: Calculate the overall fitness based on the decision evaluation results of each scheduling strategy; S35: Eliminate scheduling strategies whose fitness does not reach the preset selection threshold to obtain the winning strategies. S36: Perform crossover and mutation on each of the winning strategies to generate a child scheduling strategy; S37: Repeat steps S33 to S36 to iterate until the number of iterations meets the preset number; S38: Filter the scheduling strategies of each child generation obtained in the last iteration to obtain the optimal scheduling strategy; S4: Perform traffic scheduling according to the optimal scheduling strategy.

2. The resilient traffic scheduling method according to claim 1, characterized in that, In step S33, the decision evaluation adopts a resilience cost function, the expression of which is: ， in, For path The resilience cost, For travel time, The expected travel time for the path. This is a risk item; This is the worst-case scenario. and These are adjustable weighting coefficients, representing the degree of aversion to volatility risk and extreme situations, respectively; The worst-case term is defined by setting a travel time condition under a preset high confidence level. Quantiles were estimated using Monte Carlo simulation.

3. The resilient traffic scheduling method according to claim 2, characterized in that, The calculation formula for filtering in step S38 is: ， in, This indicates that the optimal solution is sought among various scheduling actions. For scheduling strategy In various future transportation scenarios The resilience cost in the middle, For scheduling strategy x in all future traffic scenarios The mathematical expectation of resilience cost, The worst-case resilience cost across all scenarios; This is the avoidance coefficient.

4. The resilient traffic scheduling method according to claim 3, characterized in that, The expression for the overall fitness is: ， Where M represents the number of future transportation scenarios. Indicating various future transportation scenarios, is a coefficient.

5. The resilient traffic scheduling method according to claim 1, characterized in that, Step S2 specifically includes: S21: Perform feature encoding on the real-time traffic data and historical traffic flow data respectively to obtain multi-source features with the same embedding representation; the multi-source features include spatiotemporal embedding vectors, trajectory pattern features, order distribution features, and event embedding vectors; S22: Adaptively fuse the multi-source features to obtain the comprehensive feature vector of each road segment; S23: Based on the comprehensive feature vector, calculate the future traffic conditions of each road segment during the target time period using a probabilistic parameterization structure.

6. The resilient traffic scheduling method according to claim 5, characterized in that, In step S21, the historical traffic flow data is spatiotemporally feature-encoded to obtain the spatiotemporal embedding vector, the vehicle GPS trajectory is trajectory feature-encoded to obtain the trajectory pattern feature, the passenger order request is order feature-encoded to obtain the order distribution feature, and the traffic event is order feature-encoded to obtain the order distribution feature. The spatiotemporal feature encoding employs a spatiotemporal graph neural network to encode the historical traffic flow data along with road network topology structures such as the adjacency matrix. The trajectory feature encoding uses a sequence encoder to extract trajectory pattern features from the vehicle's GPS trajectory; The order feature encoding uses an attention mechanism to aggregate the dynamic features of passenger order requests within the region; The event feature encoding uses a pre-trained language model to extract the event embedding vector from the traffic incident report.

7. The resilient traffic scheduling method according to claim 6, characterized in that, In step S22, the adaptive fusion adopts a gated fusion mechanism, the expression of which is: ， in, It is the sigmoid activation function. For learnable parameters, For element-wise multiplication, Let e ​​be the spatiotemporal embedding vector of road segment node e. The trajectory pattern features of road segment e, The order distribution characteristics of road segment e are as follows: Let e ​​be the event embedding vector for road segment e.

8. The resilient traffic scheduling method according to claim 7, characterized in that, In step S23, the probabilistic parameterization structure adopts a hybrid density network structure, the expression of which is: ， in, The number of mixed components. The weight of the k-th mixture component. ; Let be the mean of the k-th Gaussian components. Let be the standard deviation of the k-th Gaussian component. , and It is learned from the comprehensive feature vector through a neural network.

9. The resilient traffic scheduling method according to claim 8, characterized in that, During the training of the uncertainty-aware model, a negative log-likelihood loss function is used for optimization, the expression of which is: ， in, The actual traffic conditions for road segment e. This is the comprehensive feature vector of each road segment e.

10. An electronic device, characterized in that, processor; Memory for storing computer programs executed by the processor; The processor executes the computer program to implement the resilient traffic scheduling method according to any one of claims 1-9.

Images