A vehicle-road cooperation edge computing power scheduling method

Abstract

The application discloses a kind of vehicle-road cooperation edge power scheduling methods.There is including for multiple vehicles, obtain the driving sequence data of vehicle and roadside cooperation data;Based on the mixed structure model constructed in advance, the driving sequence data, roadside cooperation data are processed, determine prediction task vector and predicted channel state;For all predicted prediction task vectors, according to the prediction data in prediction task vector, determine the task processing priority corresponding to the prediction task vector, to utilize task processing priority, determine task decomposition information, based on task decomposition information, the target task corresponding to the prediction task vector is decomposed into tasks;Based on the predicted channel state of the target task after decomposition and the current task execution information of each network node in resource graph, the target task corresponding to the prediction task vector is carried out power resource scheduling.By pre-constructed mixed structure model and resource graph, the power scheduling performance of vehicle-road cooperation is improved.

Description

Technical Field

[0001] This invention relates to the field of resource scheduling technology, and in particular to a method for scheduling edge computing power in vehicle-road cooperative systems. Background Technology

[0002] With the development of intelligent connected vehicles and autonomous driving technology, more and more in-vehicle applications and sensing devices are connected to the vehicle-road cooperative network. A large number of tasks are generated during the operation of the vehicle cooperative network. The processing of these tasks often has high requirements for latency and reliability. Therefore, there is an urgent need for a resource scheduling method that can meet both timeliness and reliability requirements.

[0003] The current main computing paradigms for resource scheduling include vehicle-local computing, edge computing, and remote cloud computing. Each computing paradigm has its own advantages, but when faced with dynamic and complex traffic environments and diverse in-vehicle business needs, a single computing method can hardly fully support the differentiated performance indicators of tasks. It has problems with the timeliness and reliability of resource scheduling and cannot meet the requirements for latency and reliability. Summary of the Invention

[0004] This invention provides a vehicle-road cooperative edge computing power scheduling method to solve the problems of poor timeliness and reliability of resource scheduling.

[0005] According to one aspect of the present invention, a vehicle-road cooperative edge computing power scheduling method is provided, comprising: For multiple vehicles, vehicle driving sequence data and roadside cooperative data are acquired. The driving sequence data includes vehicle driving data corresponding to multiple collection times, and the roadside cooperative data is environmental perception data collected based on roadside perception units fixed on the driving road. Based on a pre-built hybrid structure model, driving sequence data and roadside cooperative data are processed to determine the prediction task vector and prediction channel state. The prediction task vector is used to reflect the prediction data related to the computing power task within the prediction time window, and the prediction channel state is used to reflect the prediction channel resources among each computing power node. For all predicted task vectors, the task processing priority corresponding to the predicted task vector is determined based on the predicted data in the predicted task vector. The task processing priority is used to determine the task decomposition information, and the target task corresponding to the predicted task vector is decomposed based on the task decomposition information. Based on the predicted channel state of the decomposed target task and the current task execution information of each network node in the resource graph, computing resources are scheduled for the target task corresponding to the predicted task vector. The resource graph includes multiple network nodes. Each network node includes at least dynamic attributes, static attributes, and a queue of tasks to be processed. Static attributes reflect the total computing power and total memory capacity of the network node, while dynamic attributes reflect the current computing rate and memory utilization of the network node. The queue of tasks to be processed includes multiple tasks to be executed, which reflects the instantaneous load of the network node.

[0006] Optionally, the hybrid structure model includes a graph neural network and a long short-term memory network. Based on the pre-built hybrid structure model, driving sequence data and roadside coordination data are processed to determine the prediction task vector and prediction channel state. This includes: extracting features from the roadside coordination data based on the graph neural network to obtain spatial interaction features and coordination intention features between vehicles; and processing the driving sequence data based on the long short-term memory network to obtain the vehicle's learning time and spatial joint dynamic characteristics. The first processing unit in the hybrid structure model processes the spatial interaction features, coordination intention features, and joint dynamic characteristics to output the vehicle's prediction task vector and prediction channel state within the prediction time. The prediction task vector includes at least the vehicle's task type, task prediction priority, estimated data volume, and latency requirement information. The prediction channel state includes the prediction channel gain and available bandwidth between the vehicle and the preset computing power node.

[0007] Optionally, based on the predicted data in the predicted task vector, the task processing priority corresponding to the predicted task vector is determined, including: mapping the task type in the predicted data to a first preset space to obtain the important attributes of the task type; and determining the urgency attribute based on the latency requirement information in the predicted data; determining the decision metric attribute based on the important attribute and the urgency attribute; and determining the task processing priority corresponding to the predicted task vector based on the decision metric attribute and the priority threshold range determined based on the resource map.

[0008] Optionally, the priority threshold range corresponding to the resource graph is dynamically determined based on the following method: The global average resource utilization rate of the resource graph is determined based on the total computing power of each network node in the resource graph, the computing rate at the current moment, and the total number of nodes, where the total number of nodes is the total number of network nodes in the resource graph; a first reference value of the resource graph at the current moment is determined based on the global average resource utilization rate and a high threshold determination function; where the high threshold determination function is determined based on a first critical baseline value and the resource utilization rate under critical congestion conditions; a second reference value of the resource graph at the current moment is determined based on the global average resource utilization rate and a low threshold determination function; and the priority threshold range is determined based on the first and second reference values.

[0009] Optionally, task decomposition information is determined using task processing priority, including: when the task processing priority is the first priority, the task decomposition information is no task decomposition; when the task processing priority is the second priority, the task decomposition information is decomposing the target task by combining task execution logic and resource graph; when the task processing priority is the third priority, the task decomposition information is decomposing the target task into subtasks based on the target constraint function; wherein, the target constraint function is a function of constraining computational load.

[0010] Optionally, based on the predicted channel state of the decomposed target task and the current task execution information of each network node in the resource graph, computing resources are scheduled for the target task corresponding to the predicted task vector. This includes: for the target task, if the target task is not decomposed, determining the allocatable resource information of the network node based on the current task execution information of each network node in the resource graph; determining the target node based on the allocatable resource information, the predicted channel state of the target task, and the delay constraint function, so as to execute the target task after unloading the node task of the target node; wherein, the target node is the node in the network node used to process the target task.

[0011] Optionally, based on the predicted channel state of the decomposed target task and the current task execution information of each network node in the resource graph, computing resources are scheduled for the target task corresponding to the predicted task vector. This includes: when the target task has been decomposed, a joint optimization scheduler based on deep learning processes the global state of the resource graph and the characteristics of the queue of tasks to be scheduled, the predicted channel state of the target task and the current task execution information of each network node to determine the network node identifier of each subtask related to the target task and the resource allocation information of the network node.

[0012] Optionally, the method further includes: after the target task is executed, obtaining the actual latency and actual energy consumption of the target task; substituting the actual latency and actual energy consumption into the target evaluation function, outputting the target evaluation attribute, and adjusting the parameters in the joint optimization scheduler based on the target evaluation attribute.

[0013] Optionally, the method further includes: determining network nodes of the resource graph, wherein the network nodes are composed of vehicles, roadside units, edge servers, and cloud center nodes; determining task execution information of the network nodes based on vehicle data, roadside data fed back by roadside units, data processing data corresponding to edge servers, and execution data corresponding to cloud center nodes, so as to schedule computing resources for the target task corresponding to the predicted task vector based on the task execution information.

[0014] Optionally, the method further includes: after determining the resource scheduling information corresponding to the target task, updating the task execution information of each network node in the resource graph, so as to perform computing resource scheduling based on the updated task execution information after receiving the target task again.

[0015] According to another aspect of the present invention, a vehicle-road cooperative edge computing power scheduling device is provided, comprising: The data acquisition module is used to acquire vehicle driving sequence data and roadside cooperative data for multiple vehicles. The driving sequence data includes vehicle driving data corresponding to multiple collection times, and the roadside cooperative data is environmental perception data collected based on roadside perception units fixed on the driving road. The prediction task vector and prediction channel state determination module is used to process driving sequence data and roadside coordination data based on a pre-built hybrid structure model to determine the prediction task vector and prediction channel state. The prediction task vector is used to reflect the prediction data related to the computing power task within the prediction time window, and the prediction channel state is used to reflect the prediction channel resources between each computing power node. The task decomposition processing module is used to determine the task processing priority corresponding to the predicted task vector based on the predicted data in the predicted task vector for all predicted task vectors, so as to determine the task decomposition information using the task processing priority, and to decompose the target task corresponding to the predicted task vector based on the task decomposition information. The computing resource scheduling and processing module is used to schedule computing resources for the target task corresponding to the predicted task vector based on the predicted channel state of the decomposed target task and the current task execution information of each network node in the resource graph. The resource graph includes multiple network nodes, and each network node includes at least dynamic attributes, static attributes, and a queue of tasks to be processed. The static attributes reflect the total computing power and total memory capacity of the network node, the dynamic attributes reflect the current computing rate and memory utilization of the network node, and the queue of tasks to be processed includes multiple tasks to be executed, which reflects the instantaneous load of the network node.

[0016] According to another aspect of the present invention, an electronic device is provided, the electronic device comprising: At least one processor; and A memory that is communicatively connected to at least one processor; wherein, The memory stores a computer program that can be executed by at least one processor, such that the at least one processor is able to execute the vehicle-road cooperative edge computing power scheduling method of any embodiment of the present invention.

[0017] According to another aspect of the present invention, a computer-readable storage medium is provided, which stores computer instructions for causing a processor to execute and implement the vehicle-road cooperative edge computing power scheduling method of any embodiment of the present invention.

[0018] According to another aspect of the present invention, a computer program product is provided, comprising a computer program that, when executed by a processor, implements a vehicle-road cooperative edge computing power scheduling method as described in any of the embodiments of the present invention.

[0019] The technical solution of this invention involves acquiring vehicle driving sequence data and roadside cooperative data for multiple vehicles. The driving sequence data includes vehicle driving data corresponding to multiple acquisition times, while the roadside cooperative data is environmental perception data collected by roadside sensing units fixed on the driving road. The driving sequence data and roadside cooperative data are processed based on a pre-built hybrid structure model to determine prediction task vectors and prediction channel states. The prediction task vectors reflect prediction data related to computing power tasks within the prediction time window, and the prediction channel states reflect prediction channel resources among computing power nodes. For all predicted prediction task vectors, the corresponding prediction task vector is determined based on the prediction data within the prediction task vectors. The task processing priority is used to determine task decomposition information, and the target task corresponding to the predicted task vector is decomposed based on the task decomposition information. Based on the predicted channel state of the decomposed target task and the current task execution information of each network node in the resource graph, the computing resources of the target task corresponding to the predicted task vector are scheduled. The resource graph includes multiple network nodes, and each network node includes at least dynamic attributes, static attributes, and a queue of tasks to be processed. The static attributes are used to reflect the total computing power and total memory capacity of the network node, the dynamic attributes are used to reflect the current computing rate and memory utilization of the network node, and the queue of tasks to be processed includes multiple tasks to be executed, which are used to reflect the instantaneous load of the network node. This solution achieves deep collaboration between vehicle-side time-series data and roadside perception data. It completes accurate dual prediction of computing power and channel through a hybrid structure model, and forms a closed-loop process through priority determination, task decomposition, and resource map linkage scheduling. This enables advance prediction of computing power requirements and channel resources, reasonable allocation of tasks based on priorities, and fine-grained scheduling by combining node static computing power, memory, dynamic load, and utilization. It solves the problems of poor timeliness and reliability of resource scheduling, effectively improves computing power resource utilization, reduces task latency, and enhances the stability and response efficiency of the vehicle-road cooperative system.

[0020] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart of a vehicle-road cooperative edge computing power scheduling method provided in Embodiment 1 of the present invention; Figure 2 A flowchart of a vehicle-road cooperative edge computing power scheduling method applicable to embodiments of the present invention; Figure 3 This is a flowchart of a vehicle-road cooperative edge computing power scheduling method provided in Embodiment 2 of the present invention; Figure 4 This is a schematic diagram of the structure of a vehicle-road cooperative edge computing power scheduling device provided in Embodiment 3 of the present invention; Figure 5 This is a schematic diagram of the structure of an electronic device that implements the vehicle-road cooperative edge computing power scheduling method of the present invention. Detailed Implementation

[0023] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0024] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0025] Example 1 Figure 1This is a flowchart of a vehicle-road cooperative edge computing power scheduling method provided in Embodiment 1 of the present invention. This embodiment is applicable to situations involving vehicle-road cooperative edge computing power scheduling. This method can be executed by a vehicle-road cooperative edge computing power scheduling device, which can be implemented in hardware and / or software. This device can be configured in electronic devices such as computers and servers. Figure 1 As shown, the method includes: S110. For multiple vehicles, acquire vehicle driving sequence data and roadside cooperative data. The driving sequence data includes vehicle driving data corresponding to multiple acquisition times, and the roadside cooperative data is environmental perception data collected based on roadside perception units fixed on the driving road.

[0026] Specifically, driving sequence data can be understood as vehicle-related data acquired by preset data acquisition devices at multiple consecutive acquisition times for multiple vehicles. This data dynamically reflects the vehicle's changing operating status and trajectory over time, including but not limited to vehicle speed, acceleration, steering angle, and GPS (Global Positioning System) location information. For example, driving sequence data can be represented as... ,in, The data in the table represent the time series of the vehicle's speed, acceleration, steering angle, GPS position, and other states at time t. Roadside cooperative data can be specifically understood as environmental perception data collected by roadside sensing units fixedly deployed on the road, used to provide information on the external environment of the entire road area. Driving sequence data and roadside cooperative data constitute the core foundational data for vehicle-road cooperative scenarios, originating from the vehicle and roadside respectively.

[0027] Specifically, by continuously collecting data from multiple vehicles on the road, driving sequence data composed of vehicle driving information at multiple collection times is obtained. At the same time, relying on roadside sensing units deployed at fixed locations on the road, corresponding roadside environmental sensing collaborative data is collected and provided, realizing the synchronous acquisition of vehicle dynamic driving information and roadside full-domain environmental information.

[0028] In this embodiment, the individual vehicle motion characteristics and the overall road environment status can be collected comprehensively and in real time, forming a complementary support between vehicle and road data. This provides multi-source, reliable and spatiotemporally synchronized basic data support for subsequent data processing, task prediction and resource scheduling, effectively improving the integrity of data and the accuracy of environmental perception.

[0029] S120. Based on the pre-built hybrid structure model, the driving sequence data and roadside cooperative data are processed to determine the prediction task vector and prediction channel state. The prediction task vector is used to reflect the prediction data related to the computing power task within the prediction time window, and the prediction channel state is used to reflect the prediction channel resources between each computing power node.

[0030] Specifically, the hybrid structure model can be understood as a pre-built comprehensive model used to fuse and process vehicle driving sequence data and roadside cooperative data. It can jointly learn the temporal and environmental features of multi-source heterogeneous information. Optionally, the hybrid structure model can be jointly constructed by graph neural networks and long short-term memory networks. The predicted task vector can be understood as predicted data representing the scale, type, and requirements of computing power tasks within the prediction time window. It can be obtained by processing driving sequence data and roadside cooperative data by the hybrid structure model, including but not limited to vehicle task type, task prediction priority, estimated data volume, and latency requirements. The predicted channel state can be understood as the predicted channel resource results reflecting the future channel gain and available bandwidth between each computing power node. It can be obtained by processing driving sequence data and roadside cooperative data by the hybrid structure model. The predicted task vector and predicted channel state together provide key support for the prediction of computing power tasks and channel resources in vehicle-road cooperative scenarios.

[0031] Specifically, based on a pre-built hybrid structure model, the vehicle driving sequence data and roadside collaborative data collected from multiple sources are jointly processed. Through the fusion analysis of spatiotemporal sequence and environmental perception information by the hybrid structure model, the predicted task vector, which is used to characterize the relevant information of computing power tasks within the prediction time window, and the predicted channel state, which is used to characterize the communication resource status between each computing power node, are output.

[0032] In this embodiment, the pre-built hybrid structure model can fully integrate vehicle-side dynamic data and roadside environmental data. By using the hybrid structure model, it is possible to accurately predict the future computing power task requirements and channel resource conditions in advance, providing a reliable predictive basis for subsequent task priority division, task decomposition and computing power scheduling, and effectively improving the system's foresight and decision-making rationality.

[0033] Optionally, the hybrid structure model includes a graph neural network and a long short-term memory network. Based on the pre-built hybrid structure model, driving sequence data and roadside coordination data are processed to determine the prediction task vector and prediction channel state. This includes: extracting features from the roadside coordination data based on the graph neural network to obtain spatial interaction features and coordination intention features between vehicles; and processing the driving sequence data based on the long short-term memory network to obtain the vehicle's learning time and spatial joint dynamic characteristics. The first processing unit in the hybrid structure model processes the spatial interaction features, coordination intention features, and joint dynamic characteristics to output the vehicle's prediction task vector and prediction channel state within the prediction time. The prediction task vector includes at least the vehicle's task type, task prediction priority, estimated data volume, and latency requirement information. The prediction channel state includes the prediction channel gain and available bandwidth between the vehicle and the preset computing power node.

[0034] Specifically, spatial interaction features can be understood as characteristics representing the spatial positional relationships and mutual influences between vehicles and between vehicles and roadside facilities, which can be extracted from roadside collaborative data using graph neural networks. Collaborative intention features can be understood as characteristics reflecting the collaborative driving intentions, interactive behaviors, and group driving trends of vehicles under roadside perception, which can be obtained by processing driving sequence data using long short-term memory networks.

[0035] Specifically, the hybrid structure model consists of a graph neural network (GNN) and a long short-term memory (LSTM) network. When processing data, the GNN extracts features from the roadside coordination data to obtain spatial interaction features and coordination intention features between vehicles. At the same time, the LSTM network processes the driving sequence data to obtain the spatiotemporal joint dynamic characteristics of vehicles. The above features are then input into the first processing unit of the hybrid structure model, and finally outputs a predicted task vector containing task type, task prediction priority, estimated data volume and time delay requirements within the prediction time, as well as a predicted channel state containing the prediction channel gain and available bandwidth between vehicles and computing nodes.

[0036] For example, a hybrid model combining graph neural networks and long short-term memory networks is used to process the collected data. This involves processing roadside coordination data. Input to the GNN layer to extract spatial interaction and collaborative intent features between vehicles. ; drive sequence data The time-series data is input into an LSTM layer to learn the joint temporal and spatial dynamics. The hybrid structure model outputs in the future window. The prediction task vector within Each prediction task It can be represented as , For task type, To predict task priorities, To estimate the amount of data, For delay requirements; and for predicting channel state. This includes the predicted channel gain between the vehicle and each computing node. and available bandwidth .

[0037] In this embodiment, a hybrid structural model including graph neural networks and long short-term memory networks is used to process driving sequence data and roadside coordination data to determine the prediction task vector and prediction channel state. This fully integrates the spatiotemporal characteristics of vehicle-road data, takes into account the temporal driving patterns of vehicles and the spatial coordination relationship of the roadside, and achieves accurate multi-dimensional prediction of computing power task requirements and channel resources. This provides a more comprehensive and reliable basis for subsequent task scheduling, and improves prediction accuracy and the rationality of system decision-making.

[0038] S130. For all predicted task vectors, determine the task processing priority corresponding to the predicted task vector based on the predicted data in the predicted task vector, so as to use the task processing priority to determine the task decomposition information, and perform task decomposition on the target task corresponding to the predicted task vector based on the task decomposition information.

[0039] Specifically, task processing priority can be understood as a mechanism to differentiate the urgency and importance of different computing power tasks. It can be determined by comprehensively considering predicted data such as task type, estimated data volume, and latency requirements in the predicted task vector to establish the order of task execution. Task decomposition information can be understood as the basis for task splitting based on priority, including but not limited to subtask division methods, processing order, and resource requirements. It is used to break down the target task into units suitable for execution by computing power nodes. Task processing priority and task decomposition information together provide standards and guidance for subsequent computing power resource scheduling and efficient task execution.

[0040] Specifically, for all generated prediction task vectors, core prediction data such as task type, estimated data volume, and latency requirements are extracted from the prediction task vectors. Using preset priority determination rules (such as weight scoring, threshold comparison, machine learning classification, etc.), a corresponding task processing priority is matched for each prediction task vector, resulting in the task processing priority for each prediction task vector. Then, according to the task processing priority and a preset task decomposition strategy, the task decomposition information corresponding to each prediction task vector is determined. Based on this task decomposition information, the target task corresponding to the prediction task vector is structurally and adaptively decomposed. For example, the preset task decomposition strategy can be set as follows: first task priority is not decomposed; second task priority is decomposed based on execution logic and resource graph; and third task priority is decomposed based on the target constraint function.

[0041] In this embodiment, based on the predicted data in the predicted task vector, the task processing priority corresponding to the predicted task vector is determined, and the task decomposition information is determined using the task processing priority. Then, the target task corresponding to the predicted task vector is decomposed, which realizes the dynamic, accurate and differentiated task decomposition, avoids task congestion and resource waste, helps to solve the problem of huge differences in task service quality requirements in vehicle-road cooperative systems, lays the foundation for subsequent efficient and orderly computing power scheduling and task execution, and effectively improves system processing efficiency and real-time response capability.

[0042] Optionally, task decomposition information is determined using task processing priority, including: when the task processing priority is the first priority, the task decomposition information is no task decomposition; when the task processing priority is the second priority, the task decomposition information is decomposing the target task by combining task execution logic and resource graph; when the task processing priority is the third priority, the task decomposition information is decomposing the target task into subtasks based on the target constraint function; wherein, the target constraint function is a function of constraining computational load.

[0043] The first priority can be understood as representing highly urgent and real-time computing tasks, i.e., the highest priority task. Tasks with the first priority do not need to be decomposed to ensure response speed. The second priority can be understood as representing ordinary important tasks, i.e., the next highest priority task. It can be reasonably decomposed according to task execution logic (i.e., the execution order, dependencies, and processing flow within the task) and resource graph (including the static computing power, memory, dynamic load, and global resource information of the waiting queue for each network node). The third priority can be understood as regular or non-real-time tasks, i.e., the priority of tasks that are neither the highest nor the next highest priority task. It can be decomposed into fine-grained subtasks based on a target constraint function centered on computational constraints. This hierarchical and definitional system makes the task decomposition strategy more targeted, achieving precise matching between task characteristics and resource conditions. The objective constraint function is a computational constraint rule specifically used for the decomposition of ordinary tasks with the third priority. It takes the computational workload of the task as the core constraint condition and uses mathematical methods to limit the maximum computational workload, total computational workload and minimum number of decompositions of subtasks. This ensures that the objective task is broken down into fine-grained computing units with small computational workload, simple dependencies and flexible scheduling, so that it can accurately match the fragmented computing resources in the network. While controlling the decomposition overhead, it maximizes the global resource utilization and system throughput.

[0044] Specifically, a differentiated task decomposition strategy can be pre-defined based on task processing priority levels. This strategy determines the corresponding task decomposition information based on different task processing priorities. The differentiated task decomposition strategy can be set as follows: For the highest priority, the decomposition information without task decomposition is directly adopted to ensure the rapid execution of core tasks; for the second priority, the target task is adaptively and topologically optimized by combining task execution logic and real-time resource graphs to obtain several independently schedulable subtasks; for the third priority, the target task is split into multiple fine-grained, low-dependency subtasks based on a target constraint function constrained by computational load. In other words, the highest-level task is not decomposed and is treated as an indivisible task unit. Not decomposing tasks is primarily to eliminate all additional latency and uncertainty caused by decomposition, communication synchronization, and result aggregation, providing the most robust and fastest execution path for the highest-level task. For the next highest priority tasks, adaptive and topologically optimized decomposition is performed. The system mainly decomposes tasks based on their own logic while simultaneously querying the resource graph in real time. The task is broken down into N independently schedulable subtasks, i.e. Each subtask is executed in parallel on multiple available nodes. The expected latency gains are calculated and compared with the increased communication and scheduling overhead due to task decomposition, thus making a dynamic decision on whether to decompose and the granularity of decomposition. The goal is to find the optimal balance between parallel acceleration and decomposition costs. For ordinary tasks, the system performs maximum fine-grained decomposition, breaking them down into numerous basic computational units with small computational requirements and simple dependencies. The goal of fine-grained decomposition is no longer simply to optimize the latency of individual tasks, but to create the maximum number of flexibly schedulable computational particles for the entire system, filling any available fragmented computing resources in the network, thereby maximizing global resource utilization and total system throughput.

[0045] In this embodiment, differentiated and refined decomposition strategies can be determined based on task processing priorities. This ensures that high-priority tasks are executed quickly, while allowing medium and low-priority tasks to be reasonably split according to resource conditions and constraints. This effectively balances task execution efficiency, resource utilization, and system stability, and improves the flexibility and rationality of overall computing power scheduling.

[0046] S140. Based on the predicted channel state of the decomposed target task and the current task execution information of each network node in the resource map, the computing resources of the target task corresponding to the predicted task vector are scheduled.

[0047] The resource graph includes multiple network nodes. Each network node includes at least dynamic attributes, static attributes, and a queue of tasks to be processed. Static attributes reflect the total computing power and total memory capacity of the network node, while dynamic attributes reflect the current computing rate and memory utilization of the network node. The queue of tasks to be processed includes multiple tasks to be executed, which reflects the instantaneous load of the network node.

[0048] It should be noted that the resource map is a temporal and spatial joint heterogeneous computing power resource map. This is a global resource visualization and management platform composed of multiple network nodes, providing a structured and global system view for subsequent scheduling decisions. The node set V defining this graph contains four types of heterogeneous nodes, namely... ,in, Indicates a vehicle node. Indicates roadside unit, Indicates edge server, This represents a central cloud node; each node contains static attributes reflecting total computing power and total memory capacity, dynamic attributes reflecting real-time computing rate and memory utilization, and a queue of pending tasks reflecting the node's instantaneous load. It comprehensively presents the distribution and real-time operating status of computing resources across the entire network, i.e., every node in the network. Both can be represented by a dynamic attribute vector. Quantize the node status: ;in, and These are static attributes of a node, representing its total computing power and total memory capacity, respectively; while and These are dynamic attributes of a node, representing its current computation and memory utilization, respectively. This indicates the current length of the task queue to be processed at that node, reflecting its instantaneous load. Define the edge set of the graph. This represents a dynamically changing communication link between nodes. Define any edge. The weight is Defined as a vector, used to evaluate the communication quality of the link: ; in, Let be the end-to-end communication delay from node i to node j. This represents the current available bandwidth of the link. This weight value is not simply a real-time measurement, but rather a combination of real-time network probe data and predicted channel conditions. The resulting weighted fusion is used to enhance the model's predictive power and robustness to network jitter. The attributes of all nodes in the graph are periodically updated at a preset high frequency. weight of edges .

[0049] Specifically, after the target task is decomposed, computing resources are scheduled by combining the predicted channel state corresponding to the decomposed task, as well as the static and dynamic attributes of each node in the resource graph containing multiple network nodes and the information of the task queue to be processed. The static attributes in the resource graph reflect the total computing power and total memory capacity of the node, the dynamic attributes reflect the real-time computing rate and memory utilization of the node, and the task queue to be processed reflects the instantaneous load of the node.

[0050] In this embodiment, a spatiotemporal heterogeneous computing power resource map is used to uniformly model the computing power at the cloud, edge, and terminal levels. By dynamically updating node attributes and link weights, the instantaneous topology and resource status of the entire vehicle-road cooperative system are accurately depicted. This structured global view provides a solid foundation for subsequent refined, topology-oriented scheduling decisions, solving the problem that traditional methods are difficult to effectively manage heterogeneous resources. It can combine predicted channel resources with real-time node resource status and load conditions to achieve more accurate and adaptive computing power allocation, ensuring the reliability of task transmission, making full use of node resources, avoiding overload, and significantly improving the overall scheduling efficiency, task execution success rate, and resource utilization of the system.

[0051] Optionally, based on the predicted channel state of the decomposed target task and the current task execution information of each network node in the resource graph, computing resources are scheduled for the target task corresponding to the predicted task vector. This includes: for the target task, if the target task is not decomposed, determining the allocatable resource information of the network node based on the current task execution information of each network node in the resource graph; determining the target node based on the allocatable resource information, the predicted channel state of the target task, and the delay constraint function, so as to execute the target task after unloading the node task of the target node; wherein, the target node is the node in the network node used to process the target task.

[0052] Specifically, allocable resource information can be understood as the remaining computing power, bandwidth, storage, and other available resources for new tasks at each network node in the resource graph after deducting the resources occupied by the currently executing task. This can be calculated from the static total capacity and dynamic real-time utilization rate of each network node in the resource graph, as well as current task execution information such as the pending queue. The latency constraint function can be understood as the constraint condition centered on task latency requirements, used to judge whether a task can be completed within a specified time. It is a key basis for scheduling decisions, namely, a mathematical judgment condition used to quantify the maximum latency limit that must be met during task execution, ensuring that task processing does not time out. The target node can be understood as the optimal network processing node ultimately used to receive and process the target task, determined by comprehensively considering allocable resource information, predicting channel state, and filtering using the latency constraint function. It is the carrier for offloading and executing computing power tasks.

[0053] Specifically, for undecomposed target tasks, the available resource information of each node is calculated based on the current task execution information of each network node in the resource graph. Then, the available resources, the predicted channel state of the task, and the delay constraint function are combined to select a suitable target node for processing the task. After unloading the original node task of the target node, the target task is executed to complete the scheduling of computing resources. This ensures that undecomposed tasks can be processed with the highest priority and without delay, that is, ensures that the highest priority task can be processed with the highest priority and without delay.

[0054] For example, preemptive scheduling based on deterministic rules is performed on the highest-level task. When the highest-level task appears, the scheduler will immediately schedule it in the current graph. In the middle, search and determine the optimal unloading node to minimize the total latency. The total latency calculation model comprehensively considers both the communication and computation stages; therefore, the expression for the latency constraint function is as follows: ; Where i represents the source vehicle and j represents the candidate target node. and These are the data volume of the task and the required computation period, respectively. For the available bandwidth of the link, The computing resources that can be allocated to node j for this task. Once the optimal node is found, the system will unconditionally and forcibly preempt any resources occupied by any low-priority tasks running on it and allocate them all to the highest-priority task to ensure that it receives the highest priority and zero-delay processing.

[0055] For sub-high-level tasks, general tasks, and other subtasks, the system initiates a joint optimization scheduler based on deep reinforcement learning (DRL). This scheduler will optimize the current graph. Its state space consists of the global state and the characteristics of the queue of tasks to be scheduled. Based on this state, the DRL scheduling decision engine then defines its action space. Each task k in the process outputs a set containing the target node to be unloaded, the resource allocation ratio to be calculated, and the bandwidth allocation ratio, i.e. .

[0056] In this embodiment, precise node matching can be performed by combining real-time node resources, channel status, and latency requirements. The highest-level task is not decomposed but is regarded as an indivisible task unit. This eliminates all the additional latency and uncertainty caused by decomposition, communication synchronization, and result aggregation, providing the most robust and fastest execution path for the highest-level task, prioritizing the execution of the highest-priority task, and achieving a low-latency effect.

[0057] Optionally, based on the predicted channel state of the decomposed target task and the current task execution information of each network node in the resource graph, computing resources are scheduled for the target task corresponding to the predicted task vector. This includes: when the target task has been decomposed, a joint optimization scheduler based on deep learning processes the global state of the resource graph and the characteristics of the queue of tasks to be scheduled, the predicted channel state of the target task and the current task execution information of each network node to determine the network node identifier of each subtask related to the target task and the resource allocation information of the network node.

[0058] The deep learning-based joint optimization scheduler refers to an intelligent scheduling unit built upon deep learning for global collaborative decision-making, responsible for uniformly processing multi-source scheduling information. Global state refers to the comprehensive state of the overall resource distribution, load level, and channel conditions of all network nodes presented in the resource graph. The characteristics of the queue of tasks to be scheduled refer to key attribute information such as the number, priority, computing power requirements, latency requirements, and data volume of subtasks to be executed. Current task execution information refers to real-time status information reflecting the instantaneous load, operating status, and remaining capacity of each network node, such as the real-time computing rate, memory utilization, and length of the queue of tasks to be processed. Network node identifiers refer to the unique numbers or identity information used to distinguish and locate each computing node. Network node resource allocation information refers to the specific resource quotas and usage rules allocated by the scheduler to the corresponding node, including computing power, memory, and bandwidth; these elements collectively support the intelligent and efficient computing power scheduling of the decomposed subtasks.

[0059] Specifically, when the target task has been decomposed, a joint optimization scheduler based on deep learning is used to uniformly process the global state of the resource graph, the characteristics of the queue of tasks to be scheduled, the predicted channel state of the target task, and the current execution information of each network node. This automatically determines the network node identifier and resource allocation information of each sub-task, thus completing fine-grained computing resource scheduling.

[0060] In this embodiment, deep learning is used to achieve global state awareness and collaborative optimization. It can intelligently match the characteristics of the decomposed subtasks, the real-time status of nodes, and channel conditions to complete fine-grained computing resource scheduling. It can realize multi-granularity task decomposition and hierarchical scheduling strategies based on quantized priority. Combined with deterministic rule preemption and deep reinforcement learning optimization, it realizes differentiated processing of the highest-priority task and other tasks, ensuring the deterministic low latency of high-priority tasks and the collaborative optimization of the overall system resource utilization. It significantly improves the accuracy of scheduling decisions and global resource utilization, effectively balances node load, reduces task execution latency, and enhances the system's adaptive scheduling capability and overall operating efficiency in multi-task and multi-node scenarios.

[0061] Based on the above embodiments, the method further includes: after the target task is executed, obtaining the actual latency and actual energy consumption of the target task; substituting the actual latency and actual energy consumption into the target evaluation function, outputting the target evaluation attribute, and adjusting the parameters in the joint optimization scheduler based on the target evaluation attribute.

[0062] In this context, the actual latency of the target task refers to the actual time consumed from task issuance to completion. Actual energy consumption refers to the actual energy consumed during task execution and data transmission at network nodes. Actual latency and actual energy consumption are key indicators reflecting task execution performance. The target evaluation attribute refers to the comprehensive evaluation result obtained by substituting actual latency and actual energy consumption into the target evaluation function. This attribute is used to quantitatively evaluate the merits of scheduling strategies and provide direct and reliable feedback for parameter optimization of the deep learning joint optimization scheduler. The target evaluation function can be a reward function, which is determined by the actual latency. and actual energy consumption The structure, expressed as follows: ; in, and To balance the weighting coefficients of latency and energy consumption, As a baseline energy consumption, This is a negative penalty triggered when a task times out. By maximizing long-term cumulative rewards, deep learning jointly optimizes the DRL scheduler, enabling it to learn complex strategies for efficient resource allocation and task offloading in dynamically changing environments.

[0063] Specifically, after the target task is completed, the system collects the actual latency and energy consumption data of the task, substitutes them into a preset target evaluation function to calculate the target evaluation attribute, and then iteratively updates and optimizes the internal parameters of the deep learning joint optimization scheduler based on the evaluation attribute. In this embodiment, after the scheduling decision engine completes a scheduling decision and the task is executed, the system will issue a reward to the scheduling decision engine. Then, the parameters in the joint optimization scheduler are adjusted based on the obtained reward value.

[0064] In this embodiment, a closed-loop mechanism of scheduling-execution-evaluation-feedback-optimization is formed. By continuously correcting the scheduling model through real latency and energy consumption data, the accuracy and rationality of the scheduling strategy are continuously improved, effectively reducing the overall system latency and energy consumption, and enhancing the adaptive capability and long-term operating performance of computing power scheduling.

[0065] Based on the above embodiments, the method further includes: determining network nodes of the resource graph, wherein the network nodes are composed of vehicles, roadside units, edge servers and cloud center nodes; determining task execution information of the network nodes based on vehicle data, roadside data fed back by roadside units, data processing data corresponding to edge servers and execution data corresponding to cloud center nodes, so as to schedule computing resources for the target task corresponding to the predicted task vector based on the task execution information.

[0066] Understandably, the network nodes of the resource graph are various entity units that constitute the vehicle-road cooperative full-domain computing power system. They are composed of vehicles, roadside units, edge servers, and cloud center nodes, and are the core carriers for task processing, data transmission, and computing power supply. Task execution information is node status information obtained by comprehensively analyzing vehicle data, roadside data, edge server processed data, and cloud center execution data. It is used to reflect the current computing power load, resource consumption, task queue, and operation status of each network node, and is a key basis for accurate computing power scheduling and task allocation.

[0067] Specifically, the process first identifies the various network nodes in the resource graph, which are composed of vehicles, roadside units, edge servers, and cloud center nodes. Then, it collects vehicle data, roadside data fed back by roadside units, data processing data from edge servers, and execution data from cloud center nodes. Based on this, it comprehensively determines the task execution information of each network node, and then uses this as a basis to schedule computing resources for the target task corresponding to the predicted task vector.

[0068] In this embodiment, an integrated vehicle-road-edge-cloud resource perception system is constructed, which can comprehensively acquire real-time status information of multi-level network nodes, providing rich, accurate, and multi-dimensional data support for computing power scheduling, realizing hierarchical, collaborative, and efficient task scheduling and resource allocation, and significantly improving the overall resource utilization, task processing capability, and operational stability of the system.

[0069] Based on the above embodiments, the method further includes: after determining the resource scheduling information corresponding to the target task, updating the task execution information of each network node in the resource graph, so as to perform computing resource scheduling based on the updated task execution information after receiving the target task again.

[0070] Specifically, after determining the resource scheduling information corresponding to the target task, the system will update the task execution information of each network node in the resource graph in real time, including the status of node load, resource occupancy, and pending task queue, so that the resource graph is always up-to-date. This will enable more accurate and reasonable scheduling of computing resources based on the updated task execution information when the target task is received again in the future.

[0071] In this embodiment, dynamic synchronization and closed-loop updates of resource status are achieved, ensuring the real-time nature and accuracy of scheduling criteria, effectively avoiding node overload or resource idleness, and improving the rationality of task allocation and the overall operating efficiency of the system.

[0072] In a specific embodiment, such as Figure 2 The diagram illustrates a process flow of a vehicle-road cooperative edge computing power scheduling method. The system first inputs vehicle data and roadside cooperative data, which are then processed by the "Dynamic Demand Prediction and Resource Graph Construction" module to generate predicted task vectors and a dynamic resource graph. Subsequently, the system proceeds to the "Task Priority Determination" stage, dividing tasks into two branches: high-priority tasks and second-highest priority and ordinary tasks. High-priority tasks are directly executed using "rule-based preemptive scheduling," while second-highest priority and ordinary tasks undergo "task decomposition" before entering "DRL optimized scheduling." Finally, the scheduling results from both branches converge in the "Final Scheduling Decision and Resource Allocation Scheme," achieving priority-based differentiated scheduling. This ensures the immediate execution of high-priority tasks while optimizing resource utilization for low- and medium-priority tasks through decomposition and reinforcement learning, thereby improving the overall task scheduling efficiency and resource utilization of the vehicle-road cooperative system.

[0073] The technical solution of this embodiment acquires vehicle driving sequence data and roadside cooperative data for multiple vehicles. The driving sequence data includes vehicle driving data corresponding to multiple acquisition times, and the roadside cooperative data is environmental perception data collected by roadside sensing units fixed on the driving road. The driving sequence data and roadside cooperative data are processed based on a pre-built hybrid structure model to determine prediction task vectors and prediction channel states. The prediction task vectors reflect the prediction data related to the computing power task within the prediction time window, and the prediction channel states reflect the prediction channel resources among various computing power nodes. For all predicted prediction task vectors, the corresponding prediction task vector is determined based on the prediction data within the prediction task vectors. Task processing priority is used to determine task decomposition information, and the target task corresponding to the predicted task vector is decomposed based on the task decomposition information. Based on the predicted channel state of the decomposed target task and the current task execution information of each network node in the resource graph, computing resources are scheduled for the target task corresponding to the predicted task vector. The resource graph includes multiple network nodes, and each network node includes at least dynamic attributes, static attributes, and a queue of tasks to be processed. Static attributes are used to reflect the total computing power and total memory capacity of the network node, dynamic attributes are used to reflect the current computing rate and memory utilization of the network node, and the queue of tasks to be processed includes multiple tasks to be executed, which are used to reflect the instantaneous load of the network node. This solution achieves deep collaboration between vehicle-side time-series data and roadside perception data. It completes accurate dual prediction of computing power and channel through a hybrid structure model, and forms a closed-loop process through priority determination, task decomposition, and resource map linkage scheduling. This enables advance prediction of computing power requirements and channel resources, reasonable allocation of tasks based on priorities, and fine-grained scheduling by combining node static computing power, memory, dynamic load, and utilization. It solves the problems of poor timeliness and reliability of resource scheduling, effectively improves computing power resource utilization, reduces task latency, and enhances the stability and response efficiency of the vehicle-road cooperative system.

[0074] Example 2 Figure 3 This is a flowchart of a vehicle-road cooperative edge computing power scheduling method provided in Embodiment 2 of the present invention. The method in this embodiment is a further optimization of the method in the above embodiments. Optionally, the task type in the prediction data is mapped to a first preset space to obtain the important attributes of the task type; and, based on the latency requirement information in the prediction data, the urgency attribute is determined; based on the important attribute and the urgency attribute, the decision metric attribute is determined; and based on the decision metric attribute and the priority threshold range determined based on the resource graph, the task processing priority corresponding to the predicted task vector is determined. Figure 3 As shown, the method includes: S310. For multiple vehicles, acquire vehicle driving sequence data and roadside cooperative data. The driving sequence data includes vehicle driving data corresponding to multiple acquisition times, and the roadside cooperative data is environmental perception data collected based on roadside perception units fixed on the driving road.

[0075] S320. Based on the pre-built hybrid structure model, the driving sequence data and roadside cooperative data are processed to determine the prediction task vector and prediction channel state. The prediction task vector is used to reflect the prediction data related to the computing power task within the prediction time window, and the prediction channel state is used to reflect the prediction channel resources between each computing power node.

[0076] S330. For all predicted task vectors, map the task types in the predicted data to the first preset space to obtain the important attributes of the task types, and determine the urgency attribute based on the time delay requirement information in the predicted data.

[0077] The important attributes of task types refer to the features quantified after mapping task types to a first preset space. These features characterize the importance, business value, and scope of impact of the computing task itself. It should be noted that the first preset space refers to the mapping space between task types and importance weights. The quantification of importance is a static mapping based on domain knowledge. The system pre-defines a task criticality level table, which includes all known task types. Mapped to a normalized importance weight This weight is used to measure the importance of the task in terms of driving safety, implementation of core driving functions, and user experience optimization. The urgency attribute refers to an indicator determined based on task timeliness requirements; time urgency... The quantization of latency is a dynamic process used to reflect the task's sensitivity to latency. To accurately capture the non-linear characteristic of "the closer to the deadline, the more urgent it becomes," the Sigmoid function is used to quantify latency requirements. Normalization is performed: ; in, It is the standardized urgency score of task k. This serves as the benchmark for the system's maximum tolerable delay. and This is a curve adjustment parameter used to define the degree to which urgency changes with relative latency. This function ensures that tasks with extremely stringent latency requirements have an urgency score that approaches 1 infinitely. The important attributes of task type and urgency together provide core quantitative basis for task priority calculation, decomposition strategy selection, and computing resource scheduling.

[0078] Specifically, for all predicted task vectors, the task type is first extracted from the predicted data and mapped to the first preset space for feature representation, thereby obtaining important attributes that can reflect the core value and impact of the task, namely important weight data. At the same time, the urgency attribute of the task is calculated and determined based on the time delay requirement information in the predicted data, thus completing the quantitative extraction of multi-dimensional features of the task.

[0079] In this embodiment, spatial mapping is used to quantify the importance of tasks, and latency requirements are combined to clarify the urgency of tasks. This provides a quantifiable and standardized basis for subsequent task priority determination, task decomposition, and computing power scheduling, thereby improving the rationality and accuracy of task sorting and resource allocation.

[0080] S340. Determine the decision measurement attributes based on the importance and urgency attributes.

[0081] Among them, the decision metric attribute refers to a comprehensive decision index obtained through quantitative calculation after considering the important attributes of the comprehensive task type and the urgency attributes related to latency. It is used to uniformly measure the priority level of tasks and serves as the core basis for subsequently determining task processing priorities, formulating task decomposition plans, and carrying out computing resource scheduling. This makes scheduling decisions more scientific, balanced, and quantifiable. The quantization expression is: ,in, and It is a system parameter, and satisfies .

[0082] Specifically, by combining the important attributes of the task type with the urgency attribute of the task, a weighted fusion or quantitative calculation is performed to ultimately obtain a decision metric attribute used for uniformly evaluating task priority. The important attributes and the urgency attribute of the task can be weighted and fused according to the quantitative expression of the decision metric attribute to obtain the decision metric attribute.

[0083] In this embodiment, importance and urgency are organically combined to form a single and intuitive decision indicator, making task sorting, decomposition and scheduling more scientific and consistent, simplifying scheduling logic, improving decision-making efficiency, and ensuring that highly important and urgent tasks are given priority.

[0084] S350. Determine the task processing priority corresponding to the predicted task vector based on the decision metric attribute and the priority threshold range determined based on the resource map.

[0085] The priority threshold range is a dynamically defined level interval based on the real-time load, resource availability, and other global states of each network node in the resource graph. It is used to classify and judge the decision metrics of tasks. The task processing priority is the order of task execution determined by comparing the decision metrics with this threshold range. It determines whether a task is decomposed, how it is decomposed, and the priority order of computing power allocation. It is a key basis for achieving orderly and efficient computing power scheduling.

[0086] Specifically, the importance and urgency attributes of the task are first integrated to form a decision metric attribute. Then, the corresponding priority threshold range is determined by combining the network node load and available resources in the resource graph. By comparing the decision metric attribute with the threshold range, the task processing priority corresponding to the predicted task vector is directly determined.

[0087] In this embodiment, the overall value of the task and the dynamic classification of the network resource status can be combined to make priority determination more accurate and adaptive, avoid resource imbalance caused by fixed thresholds, and improve scheduling rationality and overall system efficiency.

[0088] Optionally, the task processing priority corresponding to the predicted task vector is determined based on the decision metric attribute and the priority threshold range determined based on the resource map, including: when the decision metric attribute is greater than the maximum value in the priority threshold range, the task processing priority is determined as the first priority; when the decision metric attribute is within the priority threshold range, the task processing priority is determined as the second priority; when the decision metric attribute is less than the minimum value in the priority threshold range, the task processing priority is determined as the third priority.

[0089] Specifically, a decision metric attribute is first formed by comprehensively considering the importance and urgency of tasks. A dynamic priority threshold range is then determined based on the real-time status of the resource map. The decision metric attribute is then compared with this threshold range. When the decision metric attribute is greater than the maximum value of the threshold range, the task is assigned the highest priority; if it is within the threshold range, it is assigned the second priority; and if it is less than the minimum value, it is assigned the third priority. This achieves priority allocation in a quantitative and standardized manner, and by combining it with real-time resource status, the determination is more closely aligned with the system's carrying capacity. This ensures that high-priority tasks are prioritized, ordinary tasks are rationally allocated, and routine tasks are executed in an orderly manner, significantly improving the rationality, stability, and resource utilization of task scheduling. It also provides a clear and reliable basis for subsequent differentiated task decomposition and intelligent computing power scheduling.

[0090] Optionally, the priority threshold range corresponding to the resource graph is dynamically determined based on the following method: The global average resource utilization rate of the resource graph is determined based on the total computing power of each network node in the resource graph, the computing rate at the current moment, and the total number of nodes, where the total number of nodes is the total number of network nodes in the resource graph; a first reference value of the resource graph at the current moment is determined based on the global average resource utilization rate and a high threshold determination function; where the high threshold determination function is determined based on a first critical baseline value and the resource utilization rate under critical congestion conditions; a second reference value of the resource graph at the current moment is determined based on the global average resource utilization rate and a low threshold determination function; and the priority threshold range is determined based on the first and second reference values.

[0091] The global average resource utilization rate refers to the average utilization of computing resources across the entire network, calculated based on the total computing power, current computing rate, and total number of nodes in the resource graph. The calculation formula is as follows: ; in, This represents the total computing power of the i-th node. Let represent the computation rate of the i-th node at time t. This represents the total number of computing nodes in the network.

[0092] The first reference value is the upper limit of the priority interval, obtained by combining the global average resource utilization rate with a high-threshold determination function. This high-priority threshold boundary is primarily driven by security requirements but is also affected by extreme system loads. It is specifically calculated using the high-threshold determination function, whose expression is: ; in, It is a preset, relatively high baseline value used to ensure that the vast majority of security tasks can be identified. It is the resource utilization rate at which the system enters a critical congestion state. It is a non-negative coefficient. This formula proves that when the system is extremely congested, the threshold for becoming the highest priority will be appropriately raised to protect core security tasks from being interfered with by too many "secondary urgent" tasks.

[0093] The second reference value is the lower limit of the priority interval obtained through the low-threshold determination function, i.e., the low-priority threshold boundary. The low-priority threshold is mainly used to distinguish between ordinary tasks and tasks that require certain performance guarantees. It is more dynamic and is specifically calculated by the low-threshold determination function, the expression of which is: ; in, It is a preset, lower baseline value. The coefficient is non-negative. This formula proves that the higher the system load, the higher the threshold for becoming a "normal" task. This will prompt more tasks to be classified as second-highest priority, thereby using adaptive decomposition to seek opportunities for parallel processing and alleviate system pressure.

[0094] The priority threshold range refers to a quantized interval, composed of a first reference value and a second reference value, used to classify task processing priorities and dynamically and adaptively adjusted according to the overall network resource status. This is used to determine the priority of each task. In the case that the definition satisfies The task is the highest level task, and meets the requirements. The task is a sub-high-level task; the remaining tasks are ordinary tasks.

[0095] Specifically, the global average resource utilization rate is first calculated based on the total computing power, real-time computing rate and total number of nodes of each network node in the resource map. Then, the first reference value is obtained by using a high threshold determination function composed of the first critical baseline value and the resource utilization rate under critical congestion. At the same time, the second reference value is obtained by using a low threshold determination function. Finally, the priority threshold range for the current moment is dynamically generated from these two values.

[0096] In this embodiment, the threshold can be adaptively adjusted in real time according to the resource load of the entire network, so that the task priority determination is highly matched with the system resource status, effectively avoiding node congestion or resource idleness, and improving the rationality, robustness and global resource utilization of computing power scheduling.

[0097] S360. Utilize task processing priority to determine task decomposition information, and decompose the target task corresponding to the predicted task vector based on the task decomposition information.

[0098] S370. Based on the predicted channel state of the decomposed target task and the current task execution information of each network node in the resource map, the computing resources of the target task corresponding to the predicted task vector are scheduled.

[0099] The resource graph includes multiple network nodes. Each network node includes at least dynamic attributes, static attributes, and a queue of tasks to be processed. Static attributes reflect the total computing power and total memory capacity of the network node, while dynamic attributes reflect the current computing rate and memory utilization of the network node. The queue of tasks to be processed includes multiple tasks to be executed, which reflects the instantaneous load of the network node.

[0100] The technical solution of this embodiment acquires driving sequence data and roadside cooperative data from multiple vehicles, processes them through a hybrid structure model to obtain predicted task vectors and predicted channel states, maps task types to a first preset space to obtain important attributes, determines urgency attributes based on latency requirements, calculates decision metric attributes, and determines task processing priorities by combining the priority threshold range dynamically generated from the resource map. Based on the priorities, decomposition information is generated and the target task is decomposed. Finally, based on the predicted channel states of the decomposed tasks and the current task execution information of each network node in the resource map, computing resources are scheduled. This solution integrates vehicle-road cooperative data to achieve accurate prediction of tasks and channels, scientifically divides priorities through multi-attribute quantification and dynamic thresholds, and completes efficient scheduling by combining intelligent decomposition and global resource perception, forming a fully intelligent closed loop from perception, prediction, decision-making to scheduling, significantly improving the task processing efficiency, resource utilization, and operational reliability of the vehicle-road cooperative system.

[0101] Example 3 Figure 4 This is a schematic diagram of the structure of a vehicle-road cooperative edge computing power scheduling device provided in Embodiment 3 of the present invention. Figure 4 As shown, the device includes: The data acquisition module 410 is used to acquire vehicle driving sequence data and roadside cooperative data for multiple vehicles. The driving sequence data includes vehicle driving data corresponding to multiple collection times, and the roadside cooperative data is environmental perception data collected based on roadside perception units fixed on the driving road. The prediction task vector and prediction channel state determination module 420 is used to process driving sequence data and roadside cooperative data based on a pre-built hybrid structure model to determine the prediction task vector and prediction channel state. The prediction task vector is used to reflect the prediction data related to the computing power task within the prediction time window, and the prediction channel state is used to reflect the prediction channel resources between each computing power node. The task decomposition processing module 430 is used to determine the task processing priority corresponding to the predicted task vector based on the predicted data in the predicted task vector for all predicted task vectors, so as to determine the task decomposition information using the task processing priority, and to decompose the target task corresponding to the predicted task vector based on the task decomposition information. The computing resource scheduling and processing module 440 is used to schedule computing resources for the target task corresponding to the predicted task vector based on the predicted channel state of the decomposed target task and the current task execution information of each network node in the resource map. The resource map includes multiple network nodes, and each network node includes at least dynamic attributes, static attributes, and a queue of tasks to be processed. The static attributes are used to reflect the total computing power and total memory capacity of the network node, the dynamic attributes are used to reflect the current computing rate and memory utilization of the network node, and the queue of tasks to be processed includes multiple tasks to be executed, which are used to reflect the instantaneous load of the network node.

[0102] The technical solution of this embodiment involves a data acquisition module acquiring vehicle driving sequence data and roadside coordination data from multiple vehicles. The driving sequence data includes vehicle driving data corresponding to multiple acquisition times, while the roadside coordination data is environmental perception data collected by roadside sensing units fixed on the driving road. A prediction task vector and prediction channel state determination module processes the driving sequence data and roadside coordination data based on a pre-built hybrid structure model to determine the prediction task vector and prediction channel state. The prediction task vector reflects the prediction data related to the computing power task within the prediction time window, and the prediction channel state reflects the prediction channel resources among various computing power nodes. A task decomposition processing module, for all predicted prediction task vectors, determines the prediction task vector based on the prediction data within the prediction task vectors. The task processing priority corresponding to the predicted task vector is determined, and task decomposition information is determined based on the task processing priority. The target task corresponding to the predicted task vector is decomposed based on the task decomposition information. The computing resource scheduling processing module performs computing resource scheduling on the target task corresponding to the predicted task vector based on the predicted channel state of the decomposed target task and the current task execution information of each network node in the resource map. The resource map includes multiple network nodes, and each network node includes at least dynamic attributes, static attributes, and a queue of tasks to be processed. The static attributes are used to reflect the total computing power and total memory capacity of the network node, the dynamic attributes are used to reflect the current computing rate and memory utilization of the network node, and the queue of tasks to be processed includes multiple tasks to be executed, which are used to reflect the instantaneous load of the network node. This solution achieves deep collaboration between vehicle-side time-series data and roadside perception data. It completes accurate dual prediction of computing power and channel through a hybrid structure model, and forms a closed-loop process through priority determination, task decomposition, and resource map linkage scheduling. This enables advance prediction of computing power requirements and channel resources, reasonable allocation of tasks based on priorities, and fine-grained scheduling by combining node static computing power, memory, dynamic load, and utilization. It solves the problems of poor timeliness and reliability of resource scheduling, effectively improves computing power resource utilization, reduces task latency, and enhances the stability and response efficiency of the vehicle-road cooperative system.

[0103] Based on the above embodiments, optionally, the prediction task vector and prediction channel state determination module 420 is specifically used to extract features from roadside cooperative data based on graph neural networks to obtain spatial interaction features and cooperative intent features between vehicles, and to process driving sequence data based on long short-term memory networks to obtain the learning time and spatial joint dynamic characteristics of vehicles; and to process the spatial interaction features, cooperative intent features and joint dynamic characteristics based on the first processing unit in the hybrid structure model to output the prediction task vector and prediction channel state of vehicles within the prediction time; wherein, the prediction task vector includes at least the vehicle's task type, task prediction priority, estimated data volume and latency requirement information, and the prediction channel state includes the prediction channel gain and available bandwidth between the vehicle and the preset computing power node.

[0104] Optionally, the task decomposition processing module 430 is specifically used to map the task type in the prediction data to a first preset space to obtain the important attributes of the task type, and to determine the urgency attribute based on the latency requirement information in the prediction data; to determine the decision metric attribute based on the important attribute and the urgency attribute; and to determine the task processing priority corresponding to the prediction task vector based on the decision metric attribute and the priority threshold range determined based on the resource map.

[0105] Optionally, the task decomposition processing module 430 is specifically used to determine the global average resource utilization rate of the resource graph based on the total computing power of each network node in the resource graph, the computing rate at the current moment, and the total number of nodes, wherein the total number of nodes is the total number of network nodes in the resource graph; determine the first reference value of the resource graph at the current moment based on the global average resource utilization rate and a high threshold determination function, wherein the high threshold determination function is determined based on a first critical baseline value and the resource utilization rate under critical congestion conditions; determine the second reference value of the resource graph at the current moment based on the global average resource utilization rate and a low threshold determination function; and determine the priority threshold range based on the first reference value and the second reference value.

[0106] Optionally, the task decomposition processing module 430 is specifically used to decompose the target task into subtasks when the task processing priority is the first priority; when the task processing priority is the second priority, the task decomposition information is to decompose the target task by combining the task execution logic and the resource graph; and when the task processing priority is the third priority, the task decomposition information is to decompose the target task into subtasks based on the target constraint function; wherein, the target constraint function is a function of the constraint computation amount.

[0107] Optionally, the computing power resource scheduling and processing module 440 is specifically used to determine the allocatable resource information of the network nodes based on the current task execution information of each network node in the resource graph if the target task is not decomposed; and to determine the target node based on the allocatable resource information, the predicted channel state of the target task, and the delay constraint function, so as to execute the target task after unloading the node task of the target node; wherein, the target node is the node in the network node used to process the target task.

[0108] Optionally, the computing power resource scheduling and processing module 440 is specifically used to process the global state of the resource graph and the characteristics of the queue of tasks to be scheduled, the predicted channel state of the target task and the current task execution information of each network node by a joint optimization scheduler based on deep learning when the target task has been decomposed, and to determine the network node identifier of each subtask related to the target task and the resource allocation information of the network node.

[0109] Optionally, the device is also used to obtain the actual latency and actual energy consumption of the target task after the target task is executed; substitute the actual latency and actual energy consumption into the target evaluation function, output the target evaluation attribute, and adjust the parameters in the joint optimization scheduler based on the target evaluation attribute.

[0110] Optionally, the device is also used to determine network nodes in the resource map, wherein the network nodes are composed of vehicles, roadside units, edge servers, and cloud center nodes; based on vehicle data, roadside data fed back by roadside units, data processing data corresponding to edge servers, and execution data corresponding to cloud center nodes, the task execution information of the network nodes is determined, so as to schedule computing resources for the target task corresponding to the predicted task vector based on the task execution information.

[0111] Optionally, the device is also used to update the task execution information of each network node in the resource graph after determining the resource scheduling information corresponding to the target task, so as to perform computing resource scheduling based on the updated task execution information after receiving the target task again.

[0112] The vehicle-road cooperative edge computing power scheduling device provided in this embodiment of the invention can execute the vehicle-road cooperative edge computing power scheduling method provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0113] Example 4 Figure 5This is a schematic diagram of the structure of an electronic device provided in Embodiment 4 of the present invention. The electronic device 10 is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (such as helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0114] like Figure 5 As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded into the RAM 13 from storage unit 18. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0115] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0116] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as vehicle-to-everything (V2X) edge computing power scheduling methods.

[0117] In some embodiments, the vehicle-to-infrastructure (V2I) edge computing power scheduling method can be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program can be loaded and / or installed on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the V2I edge computing power scheduling method described above can be performed. Alternatively, in other embodiments, processor 11 can be configured to execute the V2I edge computing power scheduling method by any other suitable means (e.g., by means of firmware).

[0118] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0119] Computer programs used to implement the vehicle-road cooperative edge computing power scheduling method of the present invention can be written in any combination of one or more programming languages. These computer programs can be provided to the processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the functions / operations specified in the flowcharts and / or block diagrams are implemented. The computer programs can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0120] Example 5 Embodiment 5 of the present invention also provides a computer-readable storage medium storing computer instructions for causing a processor to execute a vehicle-road cooperative edge computing power scheduling method, the method comprising: For multiple vehicles, vehicle driving sequence data and roadside cooperative data are acquired. The driving sequence data includes vehicle driving data corresponding to multiple collection times, and the roadside cooperative data is environmental perception data collected based on roadside perception units fixed on the driving road. Based on a pre-built hybrid structure model, driving sequence data and roadside cooperative data are processed to determine the prediction task vector and prediction channel state. The prediction task vector is used to reflect the prediction data related to the computing power task within the prediction time window, and the prediction channel state is used to reflect the prediction channel resources among each computing power node. For all predicted task vectors, the task processing priority corresponding to the predicted task vector is determined based on the predicted data in the predicted task vector. The task processing priority is used to determine the task decomposition information, and the target task corresponding to the predicted task vector is decomposed based on the task decomposition information. Based on the predicted channel state of the decomposed target task and the current task execution information of each network node in the resource graph, computing resources are scheduled for the target task corresponding to the predicted task vector. The resource graph includes multiple network nodes. Each network node includes at least dynamic attributes, static attributes, and a queue of tasks to be processed. Static attributes reflect the total computing power and total memory capacity of the network node, while dynamic attributes reflect the current computing rate and memory utilization of the network node. The queue of tasks to be processed includes multiple tasks to be executed, which reflects the instantaneous load of the network node.

[0121] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0122] To provide interaction with an object, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the object; and a keyboard and pointing device (e.g., a mouse or trackball) through which the object provides input to the electronic device. Other types of devices can also be used to provide interaction with the object; for example, feedback provided to the object can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the object can be received in any form (including sound input, voice input, or tactile input).

[0123] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., a computer with a graphical user interface or web browser through which an item can interact with the implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0124] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and VPS services, such as high management difficulty and weak business scalability.

[0125] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0126] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for scheduling edge computing power in vehicle-road cooperative systems, characterized in that, include: For multiple vehicles, the driving sequence data and roadside cooperative data of the vehicles are acquired. The driving sequence data includes vehicle driving data corresponding to multiple collection times, and the roadside cooperative data is environmental perception data collected based on roadside perception units fixed on the driving road. The driving sequence data and the roadside coordination data are processed based on a pre-built hybrid structure model to determine the prediction task vector and the prediction channel state. The prediction task vector is used to reflect the prediction data related to the computing power task within the prediction time window, and the prediction channel state is used to reflect the prediction channel resources among the computing power nodes. For all predicted task vectors, the task processing priority corresponding to the predicted task vector is determined based on the predicted data in the predicted task vector. The task processing priority is then used to determine task decomposition information, and the target task corresponding to the predicted task vector is decomposed based on the task decomposition information. Based on the predicted channel state of the target task after decomposition, and the current task execution information of each network node in the resource map, computing resources are scheduled for the target task corresponding to the predicted task vector. The resource graph includes multiple network nodes, each of which includes at least dynamic attributes, static attributes, and a queue of tasks to be processed. The static attributes reflect the total computing power and total memory capacity of the network node, the dynamic attributes reflect the current computing rate and memory utilization of the network node, and the queue of tasks to be processed includes multiple tasks to be executed, which reflect the instantaneous load of the network node.

2. The method according to claim 1, characterized in that, The hybrid structure model includes a graph neural network and a long short-term memory network. The process of using the pre-built hybrid structure model to process the driving sequence data and the roadside coordination data to determine the prediction task vector and the prediction channel state includes: Feature extraction is performed on roadside coordination data based on graph neural networks to obtain spatial interaction features and coordination intention features between vehicles. In addition, the driving sequence data is processed based on long short-term memory networks to obtain the learning time and spatial joint dynamic characteristics of the vehicles. The first processing unit in the hybrid structure model processes the spatial interaction features, the collaborative intent features, and the joint dynamic characteristics to output the predicted task vector and predicted channel state of the vehicle within the prediction time. The predicted task vector includes at least the vehicle's task type, task prediction priority, estimated data volume, and latency requirement information, and the predicted channel state includes the predicted channel gain and available bandwidth between the vehicle and the preset computing power node.

3. The method according to claim 1, characterized in that, The step of determining the task processing priority corresponding to the prediction task vector based on the prediction data in the prediction task vector includes: The task types in the predicted data are mapped to a first preset space to obtain the important attributes of the task types, and the urgency attribute is determined based on the time delay requirement information in the predicted data. Based on the importance attribute and the urgency attribute, determine the decision measurement attribute; The task processing priority corresponding to the predicted task vector is determined based on the decision metric attribute and the priority threshold range determined based on the resource map.

4. The method according to claim 3, characterized in that, The priority threshold range corresponding to the resource map is dynamically determined based on the following method: The global average resource utilization rate of the resource graph is determined based on the total computing power of each network node in the resource graph, the computing rate at the current moment, and the total number of nodes, wherein the total number of nodes is the total number of network nodes in the resource graph. Based on the global average resource utilization rate and the high threshold determination function, a first reference value for the resource map at the current moment is determined; wherein, the high threshold determination function is determined based on a first critical benchmark value and the resource utilization rate under critical congestion conditions; Based on the global average resource utilization rate and the low threshold determination function, the second reference value of the resource map at the current moment is determined; The priority threshold range is determined based on the first reference value and the second reference value.

5. The method according to claim 1, characterized in that, The step of determining task decomposition information using the task processing priority includes: When the task processing priority is the first task priority, the task decomposition information is no task decomposition; When the task processing priority is the second task priority, the task decomposition information is the decomposition of the target task by combining the task execution logic and the resource map; When the task processing priority is the third task priority, the task decomposition information is the sub-task decomposition of the target task based on the target constraint function; The objective constraint function is a function of the constraint computation quantity.

6. The method according to claim 1, characterized in that, The step of scheduling computing resources for the target task corresponding to the predicted task vector based on the predicted channel state of the decomposed target task and the current task execution information of each network node in the resource graph includes: For the target task, if the target task is not decomposed, the allocatable resource information of the network node is determined based on the current task execution information of each network node in the resource graph. Based on the allocable resource information, the predicted channel state of the target task, and the delay constraint function, a target node is determined so that the target task can be executed after the node task of the target node is unloaded. The target node is a node in the network node used to process the target task.

7. The method according to claim 1, characterized in that, The step of scheduling computing resources for the target task corresponding to the predicted task vector based on the predicted channel state of the decomposed target task and the current task execution information of each network node in the resource graph includes: When the target task has been decomposed, the joint optimization scheduler based on deep learning processes the global state of the resource graph, the characteristics of the queue of tasks to be scheduled, the predicted channel state of the target task, and the current task execution information of each network node to determine the network node identifier of each subtask related to the target task and the resource allocation information of the network node.

8. The method according to claim 1, characterized in that, The method further includes: After the target task is completed, obtain the actual latency and actual energy consumption of the target task; Substitute the actual latency and actual energy consumption into the target evaluation function to output the target evaluation attribute, and adjust the parameters in the joint optimization scheduler based on the target evaluation attribute.

9. The method according to claim 1, characterized in that, The method further includes: The network nodes of the resource graph are determined, wherein the network nodes are composed of vehicles, roadside units, edge servers, and cloud center nodes; Based on the vehicle data, the roadside data fed back by the roadside unit, the data processing data corresponding to the edge server, and the execution data corresponding to the cloud center node, the task execution information of the network node is determined, so as to schedule computing resources for the target task corresponding to the predicted task vector based on the task execution information.

10. The method according to claim 1, characterized in that, The method further includes: After determining the resource scheduling information corresponding to the target task, the task execution information of each network node in the resource graph is updated so that computing resources can be scheduled based on the updated task execution information after the target task is received again.

Images