A spatiotemporal causally-driven traffic congestion governance task optimization method and system

Abstract

The application relates to the field of intelligent traffic control, and provides a traffic congestion treatment task optimization method and system driven by space-time causality, which comprises the following steps: acquiring multi-source time sequence data of a city area traffic road network, and constructing a city road network traffic information database; obtaining a road section congestion index of each road section in a first period according to the city road network traffic information database, and defining a road section with a road section congestion index greater than a threshold value as a congestion road section; determining whether there is a causality pointing relationship between each congestion road section in a first time window according to a road section congestion index sequence of each congestion road section in a plurality of prediction steps, and quantifying corresponding causality directions and causality strengths, so as to construct a space-time causality directed graph set; generating a plurality of traffic treatment subtasks, and basing on the space-time causality directed graph set; and generating a scheduling scheme of each traffic treatment subtask according to the task dependency graph model through a reinforcement learning algorithm.

Description

Technical Field

[0001] This invention relates to the field of intelligent traffic control, and in particular to a spatiotemporal causal-driven optimization method and system for traffic congestion management. Background Technology

[0002] With the acceleration of urbanization, traffic congestion has become a prominent problem affecting urban operational efficiency and residents' travel experience. In particular, when recurring traffic congestion points are not effectively managed and occasional traffic congestion points lack proper control, traffic congestion often spreads rapidly from localized areas to the entire city's hotspots, resulting in widespread congestion.

[0003] Currently, traditional traffic congestion management methods mainly focus on the instantaneous state of a single node or local road segment, lacking spatiotemporal causal mining of the dynamic propagation of traffic congestion at the macro-level road network. Some methods attempt to capture the correlation between road network nodes using graph neural networks or correlation analysis. However, correlation does not equal causation, and simple relationship mining cannot accurately determine the source and direction of congestion propagation, easily leading to misjudgment and ineffective scheduling. Research on spatiotemporal causal mining currently mainly adopts simple prior knowledge methods, defining causal relationships based on the chronological order of traffic congestion or abnormal states, or using Bayesian network data-driven methods to establish causal relationships. These methods cannot fully explore the potential causal relationships between city-level spatiotemporal characteristic variables with large-scale nonlinear features. In recent years, influenced by the multi-source nature of road network traffic flow time series, Granger causality tests have become a very effective method for identifying potential causal relationships. However, Granger causality models require the assumption that two causal variables satisfy a linear relationship, which is insufficient in analyzing the causal relationships of non-stationary and nonlinear large-scale traffic spatiotemporal characteristic variables. Furthermore, most existing resource scheduling schemes for traffic congestion management treat task processing as a single, coarse-grained computational task, neglecting the driving influence of spatiotemporal causal relationships in the logical dependencies between tasks and execution constraints. This leads to a disconnect between the scheduling strategy and the inherent causal mechanism between congestion events when using distributed computing resources such as roadside units and the cloud for parallel processing, making it difficult to minimize task execution latency. For example, if a serious traffic accident occurs on the West Second Ring Road in Beijing, causing severe traffic congestion on the distant East Third Ring Road, prioritizing the scheduling of traffic tasks on the West Second Ring Road is necessary, considering spatiotemporal causal relationships.

[0004] Therefore, how to extract directional congestion causal chains from massive traffic data and dynamically organize efficient collaborative task computation of different levels of computing nodes in the spatiotemporal dimension is a key technical challenge for achieving efficient traffic congestion management and proactive traffic congestion intervention. Summary of the Invention

[0005] This invention provides a spatiotemporal causal-driven optimization method and system for traffic congestion management, which can effectively solve the above-mentioned problems.

[0006] This invention is implemented as follows:

[0007] In a first aspect, the present invention provides an optimization method for traffic congestion management tasks driven by spatiotemporal causality, the method comprising:

[0008] Acquire multi-source time-series data of urban area traffic network and construct an urban road network traffic information database;

[0009] Based on the urban road network traffic information database, the road segment congestion index of each road segment in the first cycle is obtained, and the road segment with a congestion index greater than the threshold is defined as a congested road segment.

[0010] Based on the road congestion index sequence of each congested road segment within the first time window, it is determined whether there is a causal relationship between each pair of congested road segments under multiple prediction steps, and the corresponding causal direction and causal strength are quantified, thereby constructing a spatiotemporal causal relationship directed graph.

[0011] Multiple traffic management sub-tasks are generated, each of which is associated with a congested road segment. A task dependency graph model is constructed based on the spatiotemporal causal directed graph, including:

[0012] The logical dependency edges between the traffic management sub-tasks are obtained by mapping the causal pointing relationship in the spatiotemporal causal directed graph. The scheduling weight of the traffic management sub-task includes causal scheduling weight and local scheduling weight. The causal scheduling weight is obtained based on the node features of the corresponding congested road segment in the spatiotemporal causal directed graph. The local scheduling weight is obtained based on the road segment congestion index of the corresponding congested road segment.

[0013] Based on the task dependency graph model, a scheduling scheme for each of the traffic management sub-tasks is generated using a reinforcement learning algorithm.

[0014] Secondly, the present invention provides a spatiotemporal causal-driven traffic congestion management task optimization system, the system comprising:

[0015] The data acquisition module is used to acquire multi-source time-series data of urban area traffic network and build an urban road network traffic information database.

[0016] The index acquisition module is used to obtain the road congestion index of each road segment in the first cycle based on the urban road network traffic information database, and to define the road segment with the road congestion index greater than the threshold as a congested road segment.

[0017] The causal determination module is used to determine whether there is a causal relationship between each pair of congested road segments under multiple prediction steps based on the road segment congestion index sequence of each congested road segment within the first time window, and to quantify the corresponding causal direction and causal intensity, thereby constructing a directed graph of spatiotemporal causal relationships.

[0018] The model building module is used to generate multiple traffic management sub-tasks, each of which is associated with a congested road segment. Based on the spatiotemporal causal directed graph, a task dependency graph model is constructed, comprising:

[0019] The logical dependency edges between the traffic management sub-tasks are obtained by mapping the causal pointing relationship in the spatiotemporal causal directed graph. The scheduling weight of the traffic management sub-task includes causal scheduling weight and local scheduling weight. The causal scheduling weight is obtained based on the node features of the corresponding congested road segment in the spatiotemporal causal directed graph. The local scheduling weight is obtained based on the road segment congestion index of the corresponding congested road segment.

[0020] The scheme generation module is used to generate scheduling schemes for each of the traffic management sub-tasks based on the task dependency graph model and through reinforcement learning algorithms.

[0021] Thirdly, the present invention provides an electronic device, comprising:

[0022] Memory, the memory storing execution instructions; and

[0023] A processor that executes execution instructions stored in the memory, causing the processor to perform the method described in the first aspect.

[0024] Fourthly, the present invention provides a readable storage medium storing executable instructions, which, when executed by a processor, are used to implement the method described in the first aspect.

[0025] Compared with the prior art, the beneficial effects of the present invention are:

[0026] This invention provides a spatiotemporal causal-driven optimization method and system for traffic congestion management tasks, achieving refined optimization and scheduling of traffic congestion management tasks. First, congested road segments are identified based on road segment congestion indices, and a causal quantification algorithm is used to determine the causal relationship between congested road segments within a prediction step size, constructing a directed graph of spatiotemporal causal relationships. This overcomes the problem that traditional correlation analysis cannot accurately determine the direction of congestion propagation. Second, causal relationships are mapped to logical dependency edges between traffic management sub-tasks, and scheduling weights based on causal strength are introduced, realizing a quantitative expression of task execution constraints and solving the problem of inaccurate scheduling caused by neglecting the causal mechanism of congestion propagation in existing technologies. Finally, a scheduling scheme is generated based on the task dependency graph model using a reinforcement learning algorithm, achieving intelligent task scheduling under complex causal constraints and improving the timeliness and effectiveness of traffic congestion management. Attached Figure Description

[0027] Figure 1 This is a flowchart of S100, a spatiotemporal causal-driven traffic congestion management task optimization method provided in an embodiment of the present invention.

[0028] Figure 2 This is a schematic diagram of the structure of the spatiotemporal causal-driven traffic congestion management task optimization system 1000 provided in an embodiment of the present invention. Detailed Implementation

[0029] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be noted that, for ease of description, only the parts relevant to the present invention are shown in the accompanying drawings.

[0030] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined with each other. The technical solution of this invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0031] Unless otherwise stated, the exemplary embodiments / exemplifications shown are to be understood as providing exemplary features of various details that provide ways in which the technical concept of the invention can be implemented in practice. Therefore, unless otherwise stated, the features of the various embodiments / exemplifications may be additionally combined, separated, interchanged and / or rearranged without departing from the technical concept of the invention.

[0032] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The singular forms “a,” “the,” and “the” used in the embodiments of the invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0033] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0034] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0035] The terms "first" and "second" used herein are merely to distinguish similar objects and do not represent a specific ordering of the objects. Understandably, the specific order or sequence of "first" and "second" can be interchanged where permitted. It should be understood that the objects distinguished by "first" and "second" can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein.

[0036] Example 1

[0037] Please refer to Figure 1 This invention provides a spatiotemporal causal-driven optimization method for traffic congestion management, S100.

[0038] Specifically, method S100 includes:

[0039] S102, acquire multi-source time-series data of urban area traffic network and construct urban road network traffic information database;

[0040] S104. Based on the urban road network traffic information database, obtain the road segment congestion index of each road segment in the first cycle, and define the road segment with a congestion index greater than the threshold as a congested road segment.

[0041] S106, Based on the road congestion index sequence of each congested road segment within the first time window, determine whether there is a causal relationship between each pair of congested road segments under multiple prediction steps, and quantify the corresponding causal direction and causal intensity, thereby constructing a spatiotemporal causal relationship directed graph.

[0042] S108, generate multiple traffic management sub-tasks, each of which is associated with a congested road segment, and construct a task dependency graph model based on the spatiotemporal causal directed graph set, wherein the task dependency graph model based on the spatiotemporal causal directed graph set includes:

[0043] The logical dependency edges between the traffic management sub-tasks are obtained by mapping the causal pointing relationship in the spatiotemporal causal directed graph. The scheduling weight of the traffic management sub-task includes causal scheduling weight and local scheduling weight. The causal scheduling weight is obtained based on the node features of the corresponding congested road segment in the spatiotemporal causal directed graph. The local scheduling weight is obtained based on the road segment congestion index of the corresponding congested road segment.

[0044] S110, Based on the task dependency graph model, a scheduling scheme for each of the traffic management sub-tasks is generated using a reinforcement learning algorithm.

[0045] In step S102, multi-source time-series data of the urban road network are acquired and preprocessed to establish an urban road network traffic information database. The urban road network traffic information database is an important foundation for the implementation of the spatiotemporal causal driving mechanism.

[0046] In some implementations, the multi-source time-series data of the urban road network includes: static road network topology data, environmental status data, traffic incident data, and dynamic traffic flow data. Specifically, the static road network topology data includes unique road segment codes, number of lanes, road length, road grade, traffic capacity, and speed limit information extracted from the urban geographic information system; the environmental status data includes visibility and road surface condition information; the traffic incident data includes the time, location, type, and estimated duration of accidents extracted from the traffic control platform; and the dynamic traffic flow data includes the average traffic volume of each road segment during the sampling period.

[0047] Preprocessing includes: cleaning multi-source time-series data and removing outliers; filling missing values ​​with linear interpolation or weighted average of adjacent road segment data based on the autocorrelation of the time series; and mapping data from different sources to the same spatiotemporal reference to construct a standardized spatiotemporal traffic data cube.

[0048] In some implementations, multi-source time-series data of urban area traffic networks are acquired, and an urban road network traffic information database is constructed, including:

[0049] The urban road network traffic information database adopts a relational database architecture based on in-memory computing. The multi-source time-series data is loaded into memory space, and the road network topology data is stored in a structured manner using the T-tree spatial indexing algorithm.

[0050] This database construction method provides efficient and reliable data support for spatiotemporal causal driving mechanisms, ensuring the accuracy and real-time performance of causal mining and dynamic scheduling. Specifically, it adopts a decentralized, fully distributed architecture with an in-memory relational database at its core, using road segments as the basic storage unit to store topological connections and inherent attributes. It achieves millisecond-level fast querying and aggregation of any road segment in any historical state and associated time through efficient spatiotemporal indexing (optimized by the T-tree algorithm). In-memory computing avoids data lag and supports causal quantification algorithms (such as transfer entropy) to dynamically calculate the causal orientation and intensity between congested road segments in real time, achieving precise location of congestion sources. Simultaneously, it addresses the heterogeneity of traffic data, promotes the construction of directed graphs of spatiotemporal causal relationships, and improves the quantification accuracy of task dependency graphs. Based on a "microkernel" module, it implements a general and scalable storage processing mechanism, providing a unified standardized interface to the application layer, supporting edge-cloud collaboration, prioritizing the scheduling of high-degree causal source tasks, and minimizing governance latency through reinforcement learning (Seq2Seq + PPO), thereby improving the real-time performance and resource utilization of congestion handling. This architecture avoids single-point performance bottlenecks in urban transportation hubs through linear scaling design of real-time traffic databases at the city and regional levels; each regional database can be configured with multi-level redundancy, supporting vertical deployment from the center to the edge and horizontal cross deployment of adjacent regions, further improving the reliability and scalability of the system.

[0051] In step S104, the congestion index of each road segment under the periodic window is calculated based on the urban road network traffic information database. Specifically, the calculation fully considers the inherent characteristics of each road (such as the number of lanes, road length, and speed limit) to select congested road segments. Simultaneously, it introduces saturation (the ratio of traffic flow to capacity) and road grade weights to propose a normalized congestion index that integrates traffic flow and speed, used to more scientifically and objectively quantify the road network traffic status. This index comprehensively reflects the actual traffic flow intensity and speed reduction, and adapts to the management needs of different road grades through weighting coefficients, thereby providing accurate congestion status input for subsequent spatiotemporal causal relationship mining.

[0052] In some implementations, the congestion index of each road segment in the first cycle is calculated based on the urban road network traffic information database, including:

[0053] The first cycle of the lower section The congestion index of the aforementioned road section The calculation formula is: ,in , These represent the start and end times of the first cycle, respectively. Indicates road segment The average speed of vehicles passing through during the first cycle , Representing road segments Maximum speed and minimum speed Indicates road segment Traffic flow during the first cycle Indicates road segment Theoretically maximum capacity, and Representing road segments The corresponding velocity coefficient and flow rate coefficient sum to 1. For example, the first cycle is 10 minutes.

[0054] Accordingly, in some implementations, the multi-source time-series data includes at least the data used in the above calculation formula.

[0055] Specifically, the average speed of passing vehicles is collected from GPS trajectory data of floating cars. Floating cars typically refer to taxis, ride-hailing vehicles, buses, and private vehicles equipped with specific navigation SDKs. Once the proportion of floating cars reaches a certain threshold (e.g., 3%-5% of traffic flow), their average speed can represent the overall average speed of that road segment. The trajectory data is preprocessed to remove outliers exceeding preset physical limits and data from stationary vehicles not in motion. Maximum and minimum speeds are calculated based on the preprocessed historical traffic dataset, typically using the highest average speed of the road segment under historical smooth traffic conditions. And the minimum average speed during severe congestion as The theoretical maximum capacity is the maximum theoretical traffic flow that a road segment can handle in the first cycle under ideal conditions (no interference, free flow). It is usually calculated based on road grade, number of lanes, and design capacity standards. Speed ​​coefficient and flow coefficient These are used to reflect the level of traffic experience and road load, respectively. Their values ​​are influenced by the urban area's road management level. When... When the road congestion index is >0.5, the road segment congestion index is... Vehicle speed is the primary factor; conversely, road traffic volume is the primary factor. For example, The value is usually 0.7. The value is set to 0.3. This setting is based on the fact that speed is a fundamental indicator of road congestion, while flow rate acts as an amplification factor to more comprehensively reflect the road's operational status.

[0056] In some implementations, historical congestion index data within a sliding time window is used to periodically (e.g., monthly) calculate and update the congestion determination threshold for each road segment. For example, the congestion index of each road segment is calculated in the first... aA certain percentile is used as a fixed congestion threshold; within the current period, if the congestion index of a road segment exceeds the updated threshold, the road segment is determined to be in a state of traffic congestion. For example, to focus on relatively severe congestion events in historical data, a The threshold is set to 70, which is the 70th percentile of the historical congestion index. Within the first time window, if the congestion index of a road segment exceeds this updated threshold, the road segment is determined to be in a state of traffic congestion. This effectively filters out "congested road segments" that need to be addressed, avoiding the misjudgment of normal traffic fluctuations as congestion events due to an excessively low threshold.

[0057] In step S106, in some embodiments, based on the road segment congestion index sequence of each congested road segment within the first time window, it is determined whether there is a causal relationship between each pair of congested road segments under multiple prediction step sizes, and the corresponding causal direction and causal strength are quantified, including:

[0058] Based on the road congestion index sequence of the first congested road segment and the road congestion index sequence of the second congested road segment within the first time window, the spatiotemporal causal relationship characteristics between the congested road segment pairs are calculated using a causal quantization algorithm under multiple prediction step sizes.

[0059] Based on the spatiotemporal causal relationship characteristics between the congested road segment pairs within the prediction range, it is determined whether there is a causal relationship between the congested road segment pairs, and the corresponding causal direction and causal strength are quantified. The prediction range includes multiple prediction step sizes; for example, the maximum time delay is set to 50 minutes, and the time interval (i.e., the prediction step size) is... If the time interval is 10 minutes, then the maximum step size is 5.

[0060] In some implementations, based on the road segment congestion index sequence of each congested road segment within a first time window, it is determined whether there is a causal relationship between each pair of congested road segments under multiple prediction step sizes, and the corresponding causal direction and causal strength are quantified, including:

[0061] Based on the road congestion index sequence of the first congested road segment within the first time window. Second congested road segment congestion index sequence To obtain each of the prediction step sizes Down, right Transitive entropy and right Transitive entropy ;

[0062] like right Transitive entropy Greater than right Transitive entropy ,and Greater than the significance threshold of causation Then determine the prediction step size. The congested road sections are causally related, and the first congested road section is the cause of the second congested road section.

[0063] The difference in information transmission volume is calculated using the L1 norm (i.e., absolute value) to intuitively reflect the absolute difference in transmission volume between the two directions, serving as the basis for determining the causal direction.

[0064] The first-time window uses a sliding window mechanism, which is adjusted by the window width. and moving step size Data was extracted from the urban road network traffic information database, with a total length of The original time series is divided into continuous subsequences Each sliding window contains Traffic information data, intervals between adjacent windows Data (usually) For example, the movement step of a sliding window. It lasts for 10 minutes.

[0065] Based on the time series of road congestion indexes for congested road segments, i.e., the time series of congested road segment variables, this study employs the theory of transfer entropy to analyze the spatiotemporal causal relationship of urban traffic congestion. It considers the delay in information propagation between variables and, in conjunction with the prediction range, the transfer entropy of the discrete process. The calculation formula is:

[0066] .

[0067] in and Represent two congested road segment variables. and They represent Variables of congested road sections at all times and The value of , Represents variables of congested road sections From the Time's up All Each possible value Represents variables of congested road sections From the Time's up All Each value can be selected. For joint probability, and are the conditional probabilities of having / not having additional information in a given historical state, respectively.

[0068] For example, and All values ​​are set to 1. The reason for this distinction is that in actual calculations, there may be inconsistencies in the sampling frequency of the congested road segment variables. The number of data points can be used to clarify the actual number of data points involved in the calculation.

[0069] For each sliding window, the joint probability density and conditional probability density are calculated using the kernel density estimation method to determine the propagation entropy. The joint probability density / marginal joint probability density is estimated using a Gaussian kernel function. and joint probability density It can be represented as:

[0070] .

[0071] in This represents the window width corresponding to the kernel function.

[0072] Using causality coefficient The causal relationship between variables in congested road sections is measured to determine the directionality and strength of the causal relationship. When the transfer entropy satisfies... Greater than ,and Greater than the significance threshold of causation Then determine the causal variable. yes The cause of the causal relationship between congested roads is... That is, the first congested road segment is the cause of the second congested road segment; conversely, the causal variable... yes The cause of the causal relationship between congested roads is... Based on this rule, we can explore the causal relationships between congested road segments in the urban road network and construct a spatiotemporal causal directed graph of congestion status.

[0073] In some implementations, the road congestion index of the first congested road segment within the first time window is used. The congestion index of the second congested road segment Within the predicted range, right Transitive entropy and right Transitive entropy ,include:

[0074] The first time window adopts an adaptive adjustment time window mechanism based on the road segment congestion index. The length of the first time window is dynamically adjusted according to the fluctuation of the road segment congestion index within a preset retrospective period, and the length of the first time window is limited to between a preset minimum window and a preset maximum window.

[0075] Specifically, the first difference of the congestion index series is introduced to measure the degree of volatility, and the congestion state volatility factor is used. The absolute value of the first-order difference is defined as: .in This indicates the length of the short-term backtracking window. For example, it corresponds to a data interval of 10 minutes. The value range is 30-60 minutes.

[0076] based on Dynamically adjust the length of the sliding window : .in This represents the preset minimum window size, whose value corresponds to the minimum sample size required to calculate the transfer entropy. This indicates the maximum preset window size. This represents the adjustment factor, used for control. The strength of the impact on window size. For example, It lasts for 60 minutes. 5 times That is, 300 minutes. The value range is 0.5-2.

[0077] This design ensures the sliding window length Always in and It decreases adaptively as congestion fluctuations increase, in order to more accurately capture dynamic causal relationships.

[0078] Based on the dynamic characteristics of traffic congestion, an adaptive sliding window method for calculating transfer entropy is adopted to prevent the loss of key causal information and improve the adaptability of transfer entropy in urban traffic network environments. When congestion conditions fluctuate drastically, the sliding window length... The sliding window length is shortened to more precisely capture the causal propagation patterns of congestion events, making scheduling schemes more prone to real-time response; when the congestion index tends to stabilize, the sliding window length is adjusted. The time limit can be appropriately extended to obtain more stable statistical estimates, making the scheduling scheme more predictive.

[0079] The process of constructing a directed graph atlas of spatiotemporal causality is as follows:

[0080] 1. Multi-step mapping: Within the first time window, for each prediction step... The causal relationship coefficient between pairs of congested road segments is calculated using the aforementioned transfer entropy formula. This serves as the causal strength between nodes in the corresponding spatiotemporal causal directed graph slice;

[0081] 2. Graph Tile Construction: Using congested road segments as nodes and directions with significant causal strength as edges, each prediction step is plotted. The calculation results below are respectively constructed as A spatiotemporal causal directed graph slice;

[0082] 3. Atlas Integration: Arrange the spatiotemporal causal directed graph slices in chronological order to form a directed atlas that reflects the dynamic evolution of congestion propagation.

[0083] In step S108, the spatiotemporal causal characteristics of the city-level traffic congestion management task are expressed in a fine-grained quantitative manner. First, based on the functional logic of the task, the city-level task is decomposed into logically independent fine-grained sub-tasks by road segment and their dependencies are analyzed. Then, a causal mapping mechanism is established to map the causal propagation links in the directed graph of spatiotemporal causal relationships to logical dependency edges between sub-tasks, realizing the mapping of spatiotemporal causal relationships to fine-grained traffic congestion management sub-tasks. At the same time, the task dependency weights of each logical dependency edge are dynamically calculated based on the node characteristics of each congested road segment node in the directed graph. Finally, the task dependency topology containing the logical dependency edges and their task dependency weights is injected into the task dependency graph model to quantify the execution constraints of the congestion management task and realize the mapping of the spatiotemporal causal-driven fine-grained traffic congestion management task flow.

[0084] The constructed task dependency graph model covers all active congestion management subtasks in the road network. Based on the existence of causal relationships between tasks, they are divided into two categories: tasks forming causal chains connected by logically dependent edges, and independent tasks with no dependencies on other tasks. This model design enables the scheduling system to simultaneously handle the chain reactions caused by congestion propagation and unrelated local congestion events, achieving globally optimal scheduling of congestion management tasks.

[0085] In some implementations, the method for obtaining the task dependency weight of logically dependent edges includes:

[0086] For each prediction step Normalize the causal strength between nodes: .in, This represents the set of all road segments in the road network. After normalization, all causal strengths are within the interval [0, 1], which facilitates comparison.

[0087] Considering that the causal impact of congestion propagation may decay over time, a decay function is used to weight and aggregate the causal strength under different prediction step sizes. The aggregated weights can be used as the dependency strength between subtasks in the task graph to quantify the execution constraints of congestion management tasks. The calculation formula is as follows: .in, This indicates the prediction step size. This represents the maximum prediction step size; in the example above, the value is 6. This represents the attenuation coefficient. For example, The value is 0.5.

[0088] In some implementations, the task dependency graph model is presented in a visual manner to show the topological dependencies of urban congestion management tasks on road segments.

[0089] In some implementations, the node characteristics include the out-degree and causal strength of the congested road segment in the spatiotemporal causal directed graph.

[0090] The out-degree is used to characterize the causal impact range of a specific congested road segment on the rest of the road network. The larger the out-degree of a node in the directed graph of spatiotemporal causality, the higher the contribution of its associated traffic management sub-tasks to the overall congestion relief, and its causal scheduling weight is positively adjusted accordingly. The more a road segment is the "culprit" (with a wide range of influence and a large sum of causal intensity), the higher its corresponding causal scheduling weight.

[0091] In some implementations, the scheduling weight of the traffic management sub-task includes causal scheduling weight and local scheduling weight. Specifically, the scheduling weight is a weighted fusion of the causal scheduling weight and the local scheduling weight, or it is determined comprehensively through preset rules / parameters. For example, scheduling weight The following formula is used for quantification: Among them Represents a node The congestion index of the road section; Represents nodes in a cause-effect graph The degree of departure; Represents a node The sum of causal strength transmitted outwards. Weighting coefficients. and This can be achieved through the Analytic Hierarchy Process (AHP), expert scoring, or historical data regression optimization to balance the urgency of local congestion with the global causal propagation impact. For example, in expressway sections... The value is 0.4. The value is 0.3. The value is 0.3.

[0092] In some implementations, the task-dependent graph is embedded using a graph attention network (GAT). Specifically, the initial feature vector of each node is composed of task-local attributes (including congestion index, task type, estimated computational requirements, and tolerance latency limit) and graph topological and causal location features (including node statistical features (e.g., out-degree, in-degree) and edge attribute features (e.g., task dependency weights of connected edges)). Through a multi-layer message passing mechanism, nodes aggregate the feature information of their neighboring nodes to obtain a hidden layer representation rich in contextual information. Subsequently, a global pooling operation is used to generate a graph-level embedding vector, which serves as the core component of the state space of the Markov decision process. This graph embedding is concatenated with the dynamic features of the current environment (e.g., current time, node load, and ready task list) to form a complete state vector, which is then input into the subsequent reinforcement learning policy network (e.g., a sequence-to-sequence neural network).

[0093] Specifically, GAT consists of two stacked graph attention layers: The first layer employs a multi-head attention mechanism, setting up four attention heads. Each head independently calculates the importance weights of neighboring nodes to the central node, and the outputs of each head are concatenated to obtain node features with a dimension of 64. This layer allows the model to focus on neighboring nodes from different perspectives (such as strong causal neighbors, weak causal neighbors, topological nearest neighbors, etc.), fully capturing the differentiated influences between tasks. The second layer uses single-head attention, fusing the 64-dimensional features output from the first layer into a final 32-dimensional node embedding. This layer aggregates neighbor information, enabling the node representation to contain the contextual information of its causal chain. Layer normalization and residual connections are applied after each graph attention layer to stabilize the training process. After obtaining the embedding vectors of all nodes, an attention pooling mechanism is used to generate graph-level embeddings: first, a learnable weight score is calculated for each node, and then all node features are weighted and summed to obtain a global graph embedding vector with a dimension of 32. This graph embedding aggregates the structural information of the entire task dependency graph, causal propagation relationships, and the urgency of each task, providing a global perspective for subsequent scheduling decisions.

[0094] This graph-level embedding vector implicitly contains inter-task dependencies, causal propagation importance (learned automatically from the initial topological features and the attention mechanism's automatic learning of neighbor weights), and some task-local attribute information. It is concatenated with the dynamic features of the current environment (including the system's current time, the real-time load of each candidate computing node, and the encoding of the ready task list) to form a complete state vector. This state vector serves as input to the subsequent reinforcement learning policy network.

[0095] Based on this state vector, the policy network calculates a scheduling weight (such as a priority score or execution probability) for each ready task and selects the next scheduled task and its execution node (such as a roadside unit or cloud center) accordingly. In this way, the task dependencies, causal propagation structure, and urgency information contained in the global graph embedding are indirectly used to guide scheduling decisions, thereby achieving end-to-end awareness and optimal scheduling decisions regarding dependency order, causal global impact, and resource heterogeneity.

[0096] This design fully leverages the representational capabilities of graph neural networks to automatically learn the influence of causal chains, thereby improving the model's accuracy in capturing complex causal propagation chains while significantly reducing the complexity of feature engineering.

[0097] In some implementations, logical dependencies are automatically generated from physical causal relationships using a C2D-Mapping algorithm. For example, physically... If it is causal, then the subtask execution sequence generated by the algorithm is as follows: It must take precedence over (or have a higher priority) implement.

[0098] In some implementations, based on the premise that the dynamic information transmission of the local road network emerges as a stable and invariant causal relationship (TCR) at a macroscopic level, each roadside unit continuously monitors the congestion status of the covered road segments and uses a lightweight model combined with the road segment congestion index to determine whether a trigger threshold has been reached. Once triggered, it generates a traffic congestion management sub-task for the corresponding road segment based on its own computing service requirements. This task includes local attributes such as road segment location, task type, estimated computational requirements, and maximum tolerable latency, which serve as the basis for subsequent modeling.

[0099] For example, local attributes are determined in the following ways:

[0100] Predefined (table lookup method): Different governance tasks (such as image recognition tasks, traffic light timing algorithm calculation, and path guidance strategy generation) are assigned fixed type labels during system initialization;

[0101] Empirical estimation: The computational requirements are estimated based on the historical average of the algorithm or linearly based on the amount of data to be processed (such as video resolution and road network node size).

[0102] Business logic (hard constraints): The maximum tolerable delay is determined by traffic safety business rules. For example, traffic light control must respond within 0.5 seconds, while road network macro-guidance can be completed within 10 seconds.

[0103] Specifically, based on the generated sub-tasks, a comprehensive analysis is conducted on the spatiotemporal information of real-time congestion management sub-tasks and the spatiotemporal causal relationships between their respective road segments, constructing a spatiotemporal graph model expressing task dependencies. Weights are assigned to the edges in the graph based on factors such as causal dependencies, execution order, and data transmission latency among sub-tasks, quantifying the strength of spatiotemporal causal constraints and forming a multi-dimensional spatiotemporal dependency task graph. Considering the changing characteristics of management tasks over time and space, a time-series prediction method combined with historical data is used to model the trend of task changes, characterizing the dynamic evolution of spatiotemporal dependency features. Spatiotemporal dependency features are manifested in: the execution order of sub-tasks in the time dimension, the distribution location of sub-tasks in the spatial dimension, and the combined influence of current time and spatial location on execution efficiency in the spatiotemporal interaction dimension. Through the interactive feedback between the model and the traffic environment, edge weights are continuously adjusted and optimized to reflect real-time changes in the system state. Based on the quantitative expression of spatiotemporal dependency features, a "computable pattern" for different congestion management sub-tasks is obtained, transforming abstract management tasks into concrete, computable data forms.

[0104] In some implementations, a task dependency graph model is constructed based on the spatiotemporal causal directed graph atlas, including:

[0105] The scheduling weight also includes a resource scheduling weight, which is obtained based on the estimated computational requirements and maximum tolerable latency of the traffic management sub-task.

[0106] Based on the task dependency graph model, a scheduling scheme for each of the traffic management sub-tasks is generated using a reinforcement learning algorithm, including:

[0107] Based on the task dependency graph model, a scheduling scheme for each traffic management subtask on each candidate computing node is generated using a reinforcement learning algorithm, wherein the scheduling strategy for each traffic management subtask on each candidate computing node is adjusted according to the scheduling weight.

[0108] In some implementations, the scheduling weight further includes a resource scheduling weight, specifically: the scheduling weight is a weighted fusion of causal scheduling weight, local scheduling weight, and resource scheduling weight, or is determined comprehensively through preset rules / parameters. For example, scheduling weight The following formula is used for quantification: .in Indicates resource scheduling weight. Fusion coefficient. It can be adjusted according to the scenario, with a typical value range of [0.3, 0.7].

[0109] In some implementations, the resource scheduling weights are obtained based on the estimated computational requirements and maximum tolerable latency of the traffic management sub-tasks, including:

[0110] Based on the estimated computational requirements With the aforementioned upper limit of tolerance delay The product relationship is used to determine the resource consumption intensity of each traffic management sub-task. ;

[0111] Based on the resource consumption intensity The real-time computing power margin of the corresponding candidate computing nodes The resource scheduling weight is calculated from the ratio of the two values. .

[0112] For example, the following formula can be used for quantification: ; .in Indicates candidate computing nodes The theoretical total computing power (after deducting reserved computing power). This represents the current real-time resource utilization rate (CPU / GPU / memory usage ratio, range [0, 1], collected in real-time by monitoring probes). Further, resource scheduling weights... Softmax normalization is used for probability assignment.

[0113] Furthermore, for city-level traffic congestion management tasks, the computational load estimation is affected by factors such as the number of congestion sources and the operational complexity of the congestion management algorithm. The number of congestion sources directly impacts task parallelism, collaborative complexity, and overall computational resource requirements. The traffic guidance planning, data collection, and processing strategies for each congested road segment all influence the computational load of the entire traffic congestion management optimization system. Therefore, a computational load estimation model is defined. .in The function representing the amount of work. This represents the number of congestion sources on a road segment, i.e., the number of upstream source nodes for that road segment in the directed graph of spatiotemporal causality. Indicates the level of management for congested road sections. This function represents the time complexity calculation function for congestion management algorithms.

[0114] For example, .in Classified as mild according to the congestion index (e.g., ... The three levels are: ∈[0.6, 0.7), medium ([0.7, 0.85), and heavy (≥0.85). The function represents the governance level, with light / medium / heavy corresponding to 1.0, 1.5, and 2.0 respectively. The complexity level factor is determined by the time complexity: O ( n The value is 1; O ( n log n The value is 2;O ( n 2 The value is 4, for example, using a variant of the Dijkstra optimization algorithm. The value is 2. This represents the benchmark coefficient, with a typical value range of

[10] . 5 , 10 7 ], This represents the influence index of the number of sources, with a typical value range of [1, 2].

[0115] In some implementations, a scheduling scheme for each of the traffic management sub-tasks is generated using a reinforcement learning algorithm based on the task dependency graph model, including:

[0116] Obtain the real-time computing power margin of each candidate computing node. ;

[0117] Based on the aforementioned real-time computing power margin For scheduling weight Make corrections and generate scheduling weights. And according to the scheduling weight Adjust the migration or unloading strategy of each of the traffic management sub-tasks between each of the candidate computing nodes.

[0118] In some implementations, the candidate computing nodes include roadside edge computing units and a central cloud platform that are interconnected via a network.

[0119] The two complement each other in terms of resource capabilities and governance roles, including:

[0120] Geographic location: Roadside edge computing units are deployed close to intersections or road sections to achieve near-field perception and real-time response; the central cloud platform is located in a remote data center to aggregate data from across the city.

[0121] Computing power: Edge nodes have limited computing power and are suitable for performing lightweight, local decision-making tasks; the cloud has super computing power to support complex model training, global optimization and long-term analysis.

[0122] Communication latency: Edge execution can achieve latency in the range of microseconds to milliseconds, meeting real-time requirements; cloud communication is in the range of tens to hundreds of milliseconds, suitable for macro-level scheduling that is not extremely low latency.

[0123] Governance perspective: The edge only has a local road network status (single or adjacent intersections); the cloud has a global perspective and can carry out cross-regional collaboration and block the spread of congestion.

[0124] Based on this difference, the reinforcement learning scheduling strategy dynamically balances the load, latency, and computing power of each candidate computing node under the constraints of the task dependency graph. It keeps real-time sensitive tasks and tasks with local causal impacts at the edge, while moving cross-regional collaborative and computationally intensive tasks to the cloud, thereby minimizing governance latency and maximizing resource utilization.

[0125] Specifically, task scheduling to the central cloud platform is mainly based on a joint determination of the following three types of triggering logic:

[0126] Load overflow: When the real-time computing resources (such as CPU and memory usage) of the edge node exceed the safety threshold and cannot handle new tasks, the cloud takes over to avoid local overload and task queuing.

[0127] Global coordination: When the task dependency graph shows that congestion has evolved into a chain propagation across regions, and it is necessary to call the city's historical and real-time data for cross-regional traffic coordination, the cloud can perform macro-level scheduling based on the global view to block the spread of congestion.

[0128] Computational complexity: If the task involves complex deep learning inference, large-scale matrix operations, or global optimization solutions that exceed the computing power of edge nodes, it will be executed in the cloud to ensure the feasibility and timeliness of task computation.

[0129] In some implementations, if the roadside edge computing unit associated with the traffic management subtask meets the computing power carrying capacity threshold, then the task is executed by the roadside edge node.

[0130] If the estimated computational cost of the traffic management sub-task exceeds the remaining capacity of the roadside edge computing unit, or if its causal propagation link involves cross-regional global collaboration requirements, then it will be scheduled to be executed on the central cloud platform.

[0131] The scheduling scheme supports horizontal collaborative migration of tasks between adjacent roadside edge computing units and vertical task offloading to the central cloud platform.

[0132] In step S110, the scheduling weights and task dependency graph model are input as environmental states into the reinforcement learning algorithm. Under the constraint of the execution time of logical dependency edges, the reinforcement learning algorithm generates dynamic scheduling decisions containing the execution priorities of each traffic management sub-task in each scheduling cycle based on the scheduling weights.

[0133] In the spatiotemporal causal-driven congestion management task scheduling scheme, dependencies and scheduling weights, as constraints at different levels, work together in the final decision, forming a mechanism of "hard constraints for ordering and soft weights for optimization": dependencies, as hard constraints, define the insurmountable boundaries of the execution order, ensuring respect for the physical causal laws of congestion propagation (cause and effect); scheduling weights, as soft weights, optimize resource allocation efficiency within the boundaries, achieving optimal allocation of computing resources (prioritizing large tasks and matching computing power); together, they ensure the achievement of the scheduling goal of minimizing global latency.

[0134] In some implementations, within the boundaries defined by dependencies, the policy network calculates scheduling weights for each ready task based on graph embedding encoded by GAT, which is used to optimize resource allocation efficiency and achieve optimal configuration of computing resources. Here, GAT acts as a bridge connecting physical laws and decision optimization, transforming hard constraints into learnable feature representations and providing a global perspective for the calculation of soft weights.

[0135] In some implementations, reinforcement learning strategies employ an event-driven scheduling mechanism that triggers state updates and a new round of decisions upon completion of each task.

[0136] Specifically, when a task is detected to be completed, the task is removed from the task dependency graph, releasing the computing resources it occupies.

[0137] Based on causal dependencies, add subsequent ready tasks to the ready list;

[0138] Based on the updated task dependency graph (including the latest causal dependencies of newly added tasks), the current list of ready tasks, the real-time load of each node, and system time and other state information, the reinforcement learning policy network outputs the execution location (roadside unit or cloud) of the next batch of tasks to be scheduled, so as to minimize the overall governance latency.

[0139] In some implementations, a scheduling scheme for each of the traffic management sub-tasks is generated using a reinforcement learning algorithm based on the task dependency graph model, including:

[0140] A Markov decision model is constructed with the goal of minimizing the total execution delay of the task. The task dependency graph model is used as the input of the environment state, and a reinforcement learning algorithm is used to output the scheduling scheme of each traffic management sub-task.

[0141] The scheduling problem is formalized as a Markov decision process with the optimization objective of minimizing the total task execution delay. A task dependency graph model is used as the core input to the environment state, and a reinforcement learning algorithm is employed to learn the optimal scheduling strategy. This process is analyzed in detail below:

[0142] A Markov decision process consists of four core elements: state space, action space, state transition probabilities, and reward function. In this problem:

[0143] State space: the state at each decision point. It needs to comprehensively reflect the current scheduling environment. This includes snapshot information of the task dependency graph model, such as the set of completed tasks, the set of currently ready tasks (i.e., tasks whose preceding dependencies have been satisfied but have not yet started execution), the attributes of each task (such as computational cost, maximum tolerable latency, position and strength in the causal graph), the current system time, and the real-time load and network status of each candidate computing node (roadside unit and cloud). The structural information of the task dependency graph can be encoded into fixed-dimensional feature vectors using graph embedding techniques (such as GAT), enabling the model to understand the logical constraints and causal propagation relationships between tasks.

[0144] Action Space: At each decision point, the model needs to select one or more ready tasks for scheduling. If only the execution order is considered without considering node allocation, then the action space... To select a task from the set of ready tasks; if the execution location is also determined, the action must specify both the task and the execution node (e.g., ...). =(task) i , node j The size of the action space changes dynamically with the number of ready tasks.

[0145] State transition: When the model selects an action After scheduling a task to a designated node for execution, the environment updates its state: the task is removed from the ready set; if it completes, its successor may become ready; the processing latency and transmission latency of the task are advanced over time; and the load and network status of each node are updated accordingly. The state transition is deterministic (task execution time and transmission latency can be predicted by the model or fed back from actual measurements), which conforms to the Markov property.

[0146] Reward function: Reward This directly reflects the optimization objective, which is to minimize the total execution latency. A common design is a negative latency increment for each step, i.e. =-(current task completion time - previous task completion time), making the cumulative reward negatively correlated with the total latency. Penalties, such as violations of task tolerance latency or resource overload, can be introduced to guide the strategy to meet constraints.

[0147] After training convergence, the policy network can iteratively output scheduling decisions in a real-time environment based on the current task dependency graph state: at each decision point, it observes the current state and samples actions according to the policy distribution (or directly selects the action with the highest probability) until all tasks are scheduled. The final generated scheduling scheme is a series of task allocations to execution nodes and their timing, which can dynamically adapt to the evolution of road network congestion and resource changes, achieving efficient parallel processing of city-level traffic congestion management tasks.

[0148] In some implementations, a scheduling scheme for each of the traffic management sub-tasks is generated using a reinforcement learning algorithm based on the task dependency graph model, including:

[0149] The strategy is solved using a reinforcement learning algorithm in the Markov decision model, and the output is a scheduling action that assigns each traffic management sub-task to the corresponding candidate computing node.

[0150] The reinforcement learning algorithm converts the execution order into task processing queues on each candidate computing node based on the estimated computational requirements and maximum tolerable latency of each traffic management sub-task, as well as the real-time resource status of each candidate computing node. The maximum tolerable latency is influenced by a combination of task processing latency and transmission latency. Task processing latency depends on the computational scale of the task and the real-time available computing power of the execution node, while transmission latency is related to the data size, communication bandwidth, and physical distance. If a task is scheduled for local execution on a roadside unit, the transmission latency is negligible; if scheduled to the central cloud, the complete round-trip transmission overhead must be considered. Under task dependency constraints, the reinforcement learning scheduling strategy must ensure that the actual total latency of the selected execution path does not exceed the tolerable threshold of the task, thereby minimizing global governance latency and dynamically adapting computing resources while meeting real-time requirements.

[0151] In one implementation, the system maintains a network status table that records the average transmission latency experience value from each region to the cloud center, and dynamically corrects it based on real-time network quality (such as RTT detection) as the basis for scheduling decisions.

[0152] In some implementations, a sequence-to-sequence neural network is used to parameterize the scheduling strategy:

[0153] Based on the definition of the state space in reinforcement learning, the scheduling problem is transformed into a prediction problem based on a sequence-to-sequence model: the input sequence is the embedding vector of each congestion management subtask after being encoded by GAT, and arranged in topological order according to the task dependencies; the output sequence is the corresponding task decision sequence, that is, the scheduling plan of all tasks, whose probability can be decomposed by chain rule.

[0154] This sequence-to-sequence model is implemented using the Transformer architecture, where both the encoder and decoder contain multi-layer multi-head self-attention mechanisms.

[0155] Specifically, the attention calculation uses the standard scaled dot product attention mechanism, and its core calculation formula is as follows: .in, , , These represent the query, key, and value matrices, respectively. This represents the key / query dimension for each attention head. =16.

[0156] Multi-head attention is achieved by independently computing eight parallel attention heads, concatenating the results, and then performing a linear transformation to obtain the final output. .in, , , , , All of these are learnable parameter matrices.

[0157] In the encoder, each layer employs self-attention. The input consists of the GAT-encoded embedding vectors of each congestion management subtask, arranged in topological order according to task dependencies, with additional positional encoding (typically using sine / cosine functions or learnable positional embeddings). The output is a sequence of hidden state vectors corresponding to each task. .

[0158] The decoder uses an autoregressive approach at each step. The input includes: the current state vector The previous time step's scheduling decision is embedded (e.g., the encoding of the selected task and node); the context vector is dynamically extracted from the encoder output through a cross-attention mechanism, enabling the decoder to focus on tasks related to the current decision (e.g., upstream tasks on the dependency chain). Internally, the decoder processes the generated sequence through masked self-attention to prevent future information leakage, and fuses the task features from the encoder output through cross-attention to finally output the scheduling probability of the currently ready task.

[0159] Typical parameter configurations are shown in Table 1:

[0160] Table 1 Typical parameter configurations

[0161]

[0162] To ensure the legitimacy of scheduling actions, an action masking mechanism is introduced into the attention calculation of the decoder: a negative infinity mask is applied to the positions corresponding to scheduled tasks and tasks that violate causal dependencies, making their probabilities close to 0 after Softmax. The decoder outputs the scheduling probabilities of all (task, node) combinations in the current ready task set and selects actions accordingly. ,in Describe the selected task. ∈{0,1} represents the execution node (0 for roadside unit, 1 for cloud). This process is repeated until all tasks are scheduled.

[0163] The policy network was trained using the Proximal Policy Optimization (PPO) algorithm, with the following settings: the clipping parameter `clipε` was set to 0.2; the entropy coefficient was linearly decayed from 0.01 to 0.001 to balance exploration and exploitation; and 10% ε-greedy exploration was added in the early stages of training to promote initial exploration. The reward function was designed as follows: .in, This indicates the increment of the scheduling duration. This represents the timeout penalty indicator function, which takes the value 1 if the task times out, and 0 otherwise. This represents the proportion of tasks completed out of all tasks. During training, a value network (Critic) is maintained to estimate state values ​​and updated using mean squared error. An advantage function (such as Generalized Advantage Estimation, GAE) is calculated based on the sampled trajectories and used for pruning target updates in the policy network.

[0164] Convergence criterion: Training is stopped when the change in the average total task latency on the validation set is less than 0.5% over 500 consecutive episodes.

[0165] After training, the trained policy network is deployed in the real-time scheduling system. At each decision point (i.e., when the task completion event is triggered), the system obtains the current task dependency graph model and dynamic environmental features, and constructs a state vector. The input strategy network is used to generate a parallel scheduling scheme for each subtask between the roadside edge computing unit and the central cloud platform. The network outputs scheduling decisions for currently ready tasks, which the system then distributes to the corresponding roadside units or the cloud for execution. After execution, feedback information (such as actual execution latency and governance effectiveness) is collected for continuous model optimization and updates, achieving adaptive scheduling.

[0166] The method provided by this invention first uses a causal quantization algorithm to mine the time series of road traffic flow in the urban road network traffic information database, identifies the spatiotemporal causal relationships between different congested road segments, and constructs a directed graph of spatiotemporal causal relationships of road network congestion status; then, the causal graph is quantified into multiple fine-grained governance sub-tasks; finally, a reinforcement learning algorithm is used for policy training to generate a scheduling scheme for fine-grained tasks, thereby minimizing the execution time of governance tasks and effectively improving the overall efficiency and real-time performance of urban traffic congestion management.

[0167] Example 2

[0168] This invention provides a spatiotemporal causal-driven traffic congestion management task optimization system 1000.

[0169] The traffic congestion management task optimization system 1000 may include corresponding modules for executing one or more steps in the flowchart of the above-described spatiotemporally causally driven traffic congestion management task optimization method. Therefore, each or several steps in the flowchart may be executed by a corresponding module, and the traffic congestion management task optimization system 1000 may include one or more of these modules. A module may be one or more hardware modules specifically configured to execute a corresponding step, or implemented by a processor configured to execute a corresponding step, or stored in a readable storage medium for processor implementation, or implemented through some combination thereof.

[0170] Specifically, such as Figure 2 As shown, the traffic congestion management task optimization system 1000 includes:

[0171] The data acquisition module 1002 is used to acquire multi-source time-series data of urban area traffic network and construct an urban road network traffic information database.

[0172] The index acquisition module 1004 is used to obtain the road congestion index of each road segment in the first cycle based on the urban road network traffic information database, and to define the road segment with the road congestion index greater than the threshold as a congested road segment.

[0173] The causal determination module 1006 is used to determine whether there is a causal relationship between each pair of congested road segments under multiple prediction steps based on the road segment congestion index sequence of each congested road segment within the first time window, and to quantify the corresponding causal direction and causal intensity, thereby constructing a spatiotemporal causal relationship directed graph.

[0174] The model building module 1008 is used to generate multiple traffic management sub-tasks, each of which is associated with a congested road segment, and to construct a task dependency graph model based on the spatiotemporal causal directed graph set. The task dependency graph model constructed based on the spatiotemporal causal directed graph set includes:

[0175] The logical dependency edges between the traffic management sub-tasks are obtained by mapping the causal pointing relationship in the spatiotemporal causal directed graph. The scheduling weight of the traffic management sub-task includes causal scheduling weight and local scheduling weight. The causal scheduling weight is obtained based on the node features of the corresponding congested road segment in the spatiotemporal causal directed graph. The local scheduling weight is obtained based on the road segment congestion index of the corresponding congested road segment.

[0176] The scheme generation module 1010 is used to generate scheduling schemes for each of the traffic management sub-tasks based on the task dependency graph model and through a reinforcement learning algorithm.

[0177] This invention also provides an electronic device, including: a memory storing execution instructions; and a processor or other hardware module executing the execution instructions stored in the memory, causing the processor or other hardware module to execute the above-described spatiotemporal causal-driven traffic congestion management task optimization method.

[0178] This invention also provides a readable storage medium storing execution instructions, which, when executed by a processor, are used to implement the aforementioned spatiotemporal causal-driven traffic congestion management task optimization method.

[0179] The traffic congestion management task optimization system 1000 of the present invention, implemented using a processor-based hardware approach, employs a hardware architecture that can be implemented using a bus architecture. The bus architecture can include any number of interconnect buses and bridges, depending on the specific application and overall design constraints of the hardware. Bus 1100 connects various circuits including one or more processors 1200, memory 1300, and / or hardware modules. Bus 1100 can also connect various other circuits 1400 such as peripheral devices, voltage regulators, power management circuits, external antennas, etc.

[0180] Bus 1100 can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Component Architecture (EISA) bus, etc. Bus 1100 can be divided into address bus, data bus, control bus, etc. For ease of representation, only one connection line is used in this diagram, but this does not indicate that there is only one bus or one type of bus.

[0181] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain. The processor performs the various methods and processes described above. For example, the method embodiments of the invention can be implemented as software programs tangibly contained in a machine-readable medium, such as memory. In some embodiments, part or all of the software program may be loaded and / or installed via memory and / or a communication interface. When the software program is loaded into memory and executed by the processor, one or more steps of the methods described above may be performed. Alternatively, in other embodiments, the processor may be configured to perform one of the methods described above by any other suitable means (e.g., by means of firmware).

[0182] The logic and / or steps represented in the flowchart or otherwise described herein may be specifically implemented in any readable storage medium for use by, or in conjunction with, an instruction execution system, apparatus or device (such as a computer-based system, a processor-based system or other system that can fetch and execute instructions from, an instruction execution system, apparatus or device).

[0183] For the purposes of this specification, a "readable storage medium" can be any means capable of containing, storing, communicating, propagating, or transmitting a program for use in or in conjunction with an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of readable storage media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable read-only memory (CDROM). Furthermore, a readable storage medium can even be paper or other suitable media on which a program can be printed, since a program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in memory.

[0184] It should be understood that various parts of the present invention can be implemented in hardware, software, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0185] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0186] Furthermore, the functional units in the various embodiments of this invention can be integrated into a single processing module, or each unit can exist physically separately, or two or more units can be integrated into a single module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a readable storage medium. The storage medium can be a read-only memory, a disk, or an optical disk, etc.

[0187] Those skilled in the art should understand that the above embodiments are merely for illustrating the present invention and are not intended to limit the scope of the invention. Those skilled in the art can make other changes or modifications based on the above invention, and these changes or modifications still fall within the scope of the present invention.

Claims

1. A spatiotemporal causal-driven optimization method for traffic congestion management, characterized in that, The method includes: Acquire multi-source time-series data of urban area traffic network and construct an urban road network traffic information database, including: The urban road network traffic information database adopts a relational database architecture based on in-memory computing, loads the multi-source time-series data into memory space, and uses the T-tree spatial indexing algorithm to perform structured storage of the road network topology data; Based on the urban road network traffic information database, the road segment congestion index for each road segment in the first cycle is obtained, and road segments with a congestion index greater than a threshold are defined as congested road segments, including: The first cycle of the lower section The congestion index of the aforementioned road section The calculation formula is: ,in , These represent the start and end times of the first cycle, respectively. Indicates road segment The average speed of vehicles passing through during the first cycle , Representing road segments Maximum speed and minimum speed Indicates road segment Traffic flow during the first cycle Indicates road segment Theoretically maximum capacity, and Indicates road segment The corresponding weighting coefficients; Based on the road congestion index sequence of each congested road segment within the first time window, it is determined whether there is a causal relationship between each pair of congested road segments under multiple prediction steps, and the corresponding causal direction and causal strength are quantified, thereby constructing a directed graph of spatiotemporal causal relationships. Multiple traffic management sub-tasks are generated, each of which is associated with a congested road segment. A task dependency graph model is constructed based on the spatiotemporal causal directed graph, including: The logical dependency edges between the traffic management sub-tasks are obtained by mapping the causal pointing relationships in the spatiotemporal causal directed graph. The scheduling weights of the traffic management sub-tasks include causal scheduling weights, local scheduling weights, and resource scheduling weights. The causal scheduling weights are obtained based on the node characteristics of the corresponding congested road segment in the spatiotemporal causal directed graph. The local scheduling weights are obtained based on the road segment congestion index of the corresponding congested road segment. The resource scheduling weights are obtained based on the estimated computational requirements and maximum tolerable latency of the traffic management sub-tasks. Based on the task dependency graph model, a scheduling scheme for each of the traffic management sub-tasks is generated using a reinforcement learning algorithm, including: A Markov decision model is constructed with the goal of minimizing the total execution latency of the task. The task dependency graph model is used as the input of the environment state. A reinforcement learning algorithm is used to generate a scheduling scheme for each traffic management sub-task on each candidate computing node. The scheduling strategy of each traffic management sub-task on each candidate computing node is adjusted according to the scheduling weight.

2. The method as described in claim 1, characterized in that, Based on the road segment congestion index sequence of each congested road segment within the first time window, determine whether there is a causal relationship between each pair of congested road segments under multiple prediction steps, and quantify the corresponding causal direction and causal strength, including: Based on the road congestion index sequence of the first congested road segment within the first time window. Second congested road segment congestion index sequence For each of the aforementioned prediction step sizes, right Transitive entropy and right Transitive entropy ; like right Transitive entropy Greater than right Transitive entropy ,and Greater than the significance threshold of causation Then determine the prediction step size. The congested road sections are causally related, and the first congested road section is the cause of the second congested road section.

3. The method as described in claim 2, characterized in that, Based on the road congestion index of the first congested road segment within the first time window. The congestion index of the second congested road segment Within the predicted range, right Transitive entropy and right Transitive entropy ,include: The first time window adopts an adaptive adjustment time window mechanism based on the road segment congestion index. The length of the first time window is dynamically adjusted according to the fluctuation of the road segment congestion index within a preset retrospective period, and the length of the first time window is limited to between a preset minimum window and a preset maximum window.

4. A spatiotemporal causal-driven traffic congestion management task optimization system, characterized in that, The system includes: The data acquisition module is used to acquire multi-source time-series data of the urban area's road network and construct an urban road network traffic information database, including: The urban road network traffic information database adopts a relational database architecture based on in-memory computing, loads the multi-source time-series data into memory space, and uses the T-tree spatial indexing algorithm to perform structured storage of the road network topology data; The index acquisition module is used to obtain the road congestion index of each road segment in the first cycle based on the urban road network traffic information database, and to define road segments with a congestion index greater than a threshold as congested road segments, including: The first cycle of the lower section The congestion index of the aforementioned road section The calculation formula is: ,in , These represent the start and end times of the first cycle, respectively. Indicates road segment The average speed of vehicles passing through during the first cycle , Representing road segments Maximum speed and minimum speed Indicates road segment Traffic flow during the first cycle Indicates road segment Theoretically maximum capacity, and Indicates road segment The corresponding weighting coefficients; The causal determination module is used to determine whether there is a causal relationship between each pair of congested road segments under multiple prediction steps based on the road segment congestion index sequence of each congested road segment within the first time window, and to quantify the corresponding causal direction and causal intensity, thereby constructing a directed graph of spatiotemporal causal relationships. The model building module is used to generate multiple traffic management sub-tasks, each of which is associated with a congested road segment. Based on the spatiotemporal causal directed graph, a task dependency graph model is constructed, comprising: The logical dependency edges between the traffic management sub-tasks are obtained by mapping the causal pointing relationships in the spatiotemporal causal directed graph. The scheduling weights of the traffic management sub-tasks include causal scheduling weights, local scheduling weights, and resource scheduling weights. The causal scheduling weights are obtained based on the node characteristics of the corresponding congested road segment in the spatiotemporal causal directed graph. The local scheduling weights are obtained based on the road segment congestion index of the corresponding congested road segment. The resource scheduling weights are obtained based on the estimated computational requirements and maximum tolerable latency of the traffic management sub-tasks. The scheme generation module is used to generate scheduling schemes for each of the traffic management sub-tasks based on the task dependency graph model using a reinforcement learning algorithm, including: A Markov decision model is constructed with the goal of minimizing the total execution latency of the task. The task dependency graph model is used as the input of the environment state. A reinforcement learning algorithm is used to generate a scheduling scheme for each traffic management sub-task on each candidate computing node. The scheduling strategy of each traffic management sub-task on each candidate computing node is adjusted according to the scheduling weight.

5. An electronic device, characterized in that, include: The memory stores execution instructions; as well as A processor that executes execution instructions stored in the memory, causing the processor to perform the method according to any one of claims 1-3.

6. A readable storage medium, characterized in that, The readable storage medium stores execution instructions, which, when executed by a processor, are used to implement the method described in any one of claims 1-3.

Images