Edge computing power adaptive allocation method and system in vehicle-road-cloud environment and medium

Abstract

This invention discloses an adaptive edge computing power allocation method, system, and medium in a vehicle-road-cloud environment, relating to the field of computing power resource scheduling. The method includes: collecting multi-dimensional state data in the vehicle-road-cloud environment; predicting the communication connection relationship between vehicle terminals and edge computing nodes based on movement trajectory data within a preset time window to generate a spatiotemporal correlation graph; analyzing task computing power requirement characteristics of tasks to be processed by vehicle terminals based on task queue data; inputting the spatiotemporal correlation graph and task computing power requirement characteristics into a computing power requirement prediction unit to obtain an estimated computing power requirement for edge computing nodes; generating a computing power resource allocation strategy for edge computing nodes based on the estimated computing power requirement and load time-series data; and issuing computing power adjustment instructions to edge computing nodes. This application solves the technical problem of inaccurate computing power allocation in existing edge computing power allocation methods, achieving the technical effect of improving the accuracy of computing power allocation.

Description

Technical Field

[0001] This application relates to the field of computing power resource scheduling, and in particular to the adaptive allocation method, system and medium for edge computing power in a vehicle-road-cloud environment. Background Technology

[0002] Edge computing power allocation in a vehicle-road-cloud environment is a crucial link supporting the collaborative operation of intelligent connected vehicles, roadside sensing devices, and the central cloud. Its allocation effect directly impacts the real-time performance of vehicle task processing and the utilization efficiency of roadside resources. Furthermore, it is a vital foundation for promoting the large-scale application of vehicle-road-cloud integrated technology, possessing extremely high technical research and engineering application value. Currently, the mainstream methods for solving the edge computing power allocation problem in a vehicle-road-cloud environment are mainly divided into two categories: static allocation and simple dynamic allocation. These include pre-setting fixed computing power quotas for each edge node, allocating computing power based on the nearest edge node connected to the vehicle, and making simple adjustments based on the historical average load of edge nodes. Existing methods do not take into account the dynamic nature and multi-factor correlation of the vehicle-road-cloud environment, failing to consider changes in edge node connectivity caused by vehicle movement and real-time fluctuations in roadside task demands. This leads to problems such as mismatch between computing power allocation and real-time task requirements, and edge nodes being prone to overload or resource idleness.

[0003] Currently, in related technologies, edge computing power allocation in the vehicle-road-cloud environment suffers from a technical problem: it cannot adapt to the dynamic changes of the environment, resulting in inaccurate computing power allocation. Summary of the Invention

[0004] This application provides an adaptive edge computing power allocation method, system, and medium in a vehicle-road-cloud environment. It collects three types of multi-dimensional state data from the vehicle-road-cloud environment: vehicle movement trajectories, edge node loads, and roadside equipment task queues. Within a preset time window, it predicts the communication connection between vehicles and edge nodes based on trajectory data and constructs a spatiotemporal correlation graph. It then parses the computing power demand characteristics of tasks to be processed from the task queue data. The spatiotemporal correlation graph and computing power demand characteristics are input into a computing power demand prediction unit to obtain an estimated value of edge node computing power demand. Combined with load time-series data, a computing power resource allocation strategy is generated. Based on the strategy, computing power adjustment instructions are issued to edge nodes, enabling adaptive allocation of edge computing power in a vehicle-road-cloud environment. These technical means solve the technical problem of inaccurate computing power allocation in existing vehicle-road-cloud environments due to the inability to adapt to dynamic environmental changes. This achieves the technical effect of adapting to dynamic changes in the vehicle-road-cloud environment and improving the accuracy of computing power allocation.

[0005] This application provides an adaptive edge computing power allocation method in a vehicle-road-cloud environment, comprising: collecting multi-dimensional state data in the vehicle-road-cloud environment, including vehicle terminal movement trajectory data, edge computing node load time series data, and roadside perception device task queue data, wherein the vehicle-road-cloud environment includes multiple intelligent connected vehicles, multiple edge computing nodes, multiple roadside units, and a central cloud platform; within a preset time window, predicting the communication connection relationship between the vehicle terminal and the edge computing node based on the movement trajectory data to generate a spatiotemporal correlation graph; based on the task queue data, parsing to obtain the task computing power requirement characteristics of the tasks to be processed by the vehicle terminal; inputting the spatiotemporal correlation graph and the task computing power requirement characteristics into a computing power requirement prediction unit to obtain an estimated value of the computing power requirement of the edge computing node; generating a computing power resource allocation strategy for the edge computing node based on the estimated computing power requirement and the load time series data; and issuing computing power adjustment instructions to the edge computing node based on the computing power resource allocation strategy.

[0006] In possible implementations, multi-dimensional state data in the vehicle-road-cloud environment is collected, including vehicle terminal movement trajectory data, edge computing node load time series data, and roadside sensing device task queue data. The following processing is performed: receiving vehicle status messages broadcast by intelligent connected vehicles and parsing them to obtain vehicle terminal movement trajectory data; reading edge computing node load time series data through a resource monitoring agent; extracting descriptive metadata of tasks to be processed from the task scheduling queue and generating task queue data for roadside sensing devices.

[0007] In a possible implementation, within a preset time window, based on the movement trajectory data, the communication connection relationship between the vehicle terminal and the edge computing nodes is predicted, a spatiotemporal correlation graph is generated, and the following processing is performed: the movement trajectory data is used to predict the driving path to obtain the estimated time when the vehicle terminal enters and leaves the communication coverage area of ​​each edge computing node; based on the estimated time, the estimated connection duration between the vehicle terminal and each edge computing node is calculated; based on the signal reception strength indication value and channel occupancy rate of the link between the vehicle terminal and the edge computing node, the task transmission rate is determined; the spatiotemporal correlation graph is constructed with the vehicle terminal and the edge computing node as graph nodes, and the estimated connection duration and the task transmission rate as edge weight attributes.

[0008] In a possible implementation, based on the task queue data, the task computing power requirement characteristics of the tasks to be processed by the vehicle terminal are parsed and obtained, and the following processing is performed: the task queue data is deserialized and parsed to extract the input data volume, algorithm complexity level, and preset priority identifier of each task to be processed; according to the input data volume and the algorithm complexity level, a preset computing power requirement mapping table is queried to obtain the estimated number of cores and the estimated memory usage required for a single task; the estimated number of cores, the estimated memory usage, and the priority identifier are feature-encoded to generate the task computing power requirement characteristics.

[0009] In a possible implementation, the spatiotemporal correlation graph and the task computing power requirement features are input to a computing power requirement prediction unit to obtain the estimated computing power requirement of the edge computing nodes. The following processing is then performed: The computing power requirement prediction unit incorporates a temporal convolutional network based on a graph attention mechanism. The spatiotemporal correlation graph is used as the spatial feature input of the temporal convolutional network, and the task computing power requirement features are used as the initial embedding features of the vehicle terminal nodes. The attention weight coefficients of each edge computing node to the vehicle terminal nodes it connects to are calculated through the graph attention mechanism. Based on the attention weight coefficients, the task computing power requirement features of the vehicle terminal nodes are weighted and aggregated to generate aggregated requirement features of the edge computing nodes. The aggregated requirement features are input to the temporal convolutional layer of the temporal convolutional network to extract temporal dependency features and output the estimated computing power requirement of each edge computing node for multiple time slices within the preset time window.

[0010] In a possible implementation, based on the estimated computing power demand and the load time-series data, a computing power resource allocation strategy for edge computing nodes is generated, and the following processes are performed: extracting the current available computing power resources of the edge computing nodes from the load time-series data; comparing the estimated computing power demand and the current available computing power resources for each edge computing node to obtain the computing power difference; obtaining the dynamic resource topology between edge computing nodes; constructing a computing power resource allocation optimization model with the optimization objectives of minimizing cross-node task migration frequency and maximizing local computing power resource utilization, using the computing power difference as a decision variable constraint and the dynamic resource topology as a resource availability constraint; and solving the computing power resource allocation optimization model to generate a computing power resource allocation strategy.

[0011] This application also provides an adaptive edge computing power allocation system in a vehicle-road-cloud environment, comprising: a multi-dimensional state data acquisition module for acquiring multi-dimensional state data in the vehicle-road-cloud environment, including vehicle terminal movement trajectory data, edge computing node load time-series data, and roadside sensing device task queue data, wherein the vehicle-road-cloud environment includes multiple intelligent connected vehicles, multiple edge computing nodes, multiple roadside units, and a central cloud platform; and a spatiotemporal correlation graph generation module for predicting the communication connection relationship between vehicle terminals and edge computing nodes based on the movement trajectory data within a preset time window, and generating a spatiotemporal correlation graph. The system comprises: a spatiotemporal correlation graph; a task computing power requirement analysis module, used to analyze and obtain the task computing power requirement characteristics of the tasks to be processed by the vehicle terminal based on the task queue data; a computing power requirement prediction module, used to input the spatiotemporal correlation graph and the task computing power requirement characteristics into the computing power requirement prediction unit to obtain the estimated computing power requirement of the edge computing node; a computing power resource configuration strategy generation module, used to generate a computing power resource configuration strategy for the edge computing node based on the estimated computing power requirement and the load time series data; and a computing power adjustment instruction issuance module, used to issue computing power adjustment instructions to the edge computing node based on the computing power resource configuration strategy.

[0012] This application also provides a computer-readable storage medium, including: a computer program stored thereon, which, when executed by a processor, implements an edge computing power adaptive allocation method in a vehicle-road-cloud environment.

[0013] The proposed edge computing power adaptive allocation method, system, and medium in a vehicle-road-cloud environment first collects multi-dimensional state data from the vehicle-road-cloud environment, including vehicle terminal movement trajectory data, edge computing node load time series data, and roadside sensing device task queue data. The vehicle-road-cloud environment includes multiple intelligent connected vehicles, multiple edge computing nodes, multiple roadside units, and a central cloud platform. Then, within a preset time window, based on the movement trajectory data, the communication connection relationship between the vehicle terminal and the edge computing node is predicted, generating a spatiotemporal correlation graph. Next, based on the task queue data, the task computing power requirement characteristics of the tasks to be processed by the vehicle terminal are analyzed. The spatiotemporal correlation graph and the task computing power requirement characteristics are then input into a computing power requirement prediction unit to obtain an estimated value of the computing power requirement of the edge computing node. Based on the estimated computing power requirement and the load time series data, a computing power resource allocation strategy for the edge computing node is generated. Finally, based on the computing power resource allocation strategy, a computing power adjustment command is issued to the edge computing node. Through the above process, the method, system and medium proposed in this application achieve the technical effect of adapting to the dynamic changes of the vehicle-road-cloud environment and improving the accuracy of computing power allocation. Attached Figure Description

[0014] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. Flowcharts are used in this application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from these processes.

[0015] Figure 1 This is a flowchart illustrating the adaptive allocation method for edge computing power in a vehicle-road-cloud environment provided in this application embodiment.

[0016] Figure 2 This is a schematic diagram of the edge computing power adaptive allocation system in a vehicle-road-cloud environment provided in this application embodiment.

[0017] Figure labeling: 10 Multidimensional state data acquisition module, 20 Spatiotemporal correlation graph generation module, 30 Task computing power requirement analysis module, 40 Computing power requirement prediction module, 50 Computing power resource configuration strategy generation module, 60 Computing power adjustment instruction issuance module. Detailed Implementation

[0018] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0019] This application provides an adaptive edge computing power allocation method in a vehicle-road-cloud environment, such as... Figure 1 As shown, the method includes:

[0020] Step S100: Collect multi-dimensional state data in the vehicle-road-cloud environment, including vehicle terminal movement trajectory data, edge computing node load time series data, and roadside sensing device task queue data. The vehicle-road-cloud environment includes multiple intelligent connected vehicles, multiple edge computing nodes, multiple roadside units, and a central cloud platform.

[0021] Specifically, the vehicle-road-cloud environment is a comprehensive and collaborative environment integrating intelligent connected vehicles, edge computing nodes, roadside units, and a central cloud platform. Intelligent connected vehicles are terminals with network communication and data reporting capabilities; edge computing nodes are devices deployed at the edge of the road network, such as near intersections or road segments, possessing local computing power; roadside units are devices installed on the roadside for sensing road conditions and transmitting data, such as cameras and radar; and the central cloud platform is used for global scheduling and data aggregation. Multidimensional state data refers to three core types of data related to vehicles, edge nodes, and roadside devices. Vehicle terminal movement trajectory data includes time-series data such as vehicle position, speed, and direction; edge computing node load time-series data includes data such as CPU utilization, memory usage, and task processing latency at different times; and roadside sensing device task queue data includes queuing data for pending sensing and data processing tasks. The acquisition frequency is set from 100 milliseconds to 1 second according to the scenario requirements. Data transmission uses the PC5 or Uu interface dedicated to vehicle networking. At the same time, the acquired data is subjected to preliminary noise reduction processing. All acquired data is uniformly encapsulated in JSON format and aggregated through the gateway of the edge node.

[0022] In one possible implementation, multi-dimensional state data from the vehicle-road-cloud environment is collected, including vehicle terminal movement trajectory data, edge computing node load time-series data, and roadside sensing device task queue data. Step S100 further includes step S110, receiving vehicle status messages broadcast by intelligent connected vehicles and parsing them to obtain the vehicle terminal's movement trajectory data. Specifically, a vehicle-to-everything (V2X) communication protocol is used to receive the messages, core fields are extracted using a message parsing tool, and the extracted data is standardized to ensure data consistency.

[0023] For example, the PC5 interface is used to receive vehicle status messages broadcast by the vehicle. These messages use the CAN bus message format, with a message ID of 0x100 and a length of 8 bytes. The message is parsed using a CANoe message parsing tool. Bytes 1-2 contain longitude data, bytes 3-4 contain latitude data, byte 5 contains speed data, and byte 6 contains the direction of travel. After parsing, the data is converted to a standardized format; for example, the parsed longitude is converted to 116.39750 degrees, latitude to 39.90882 degrees, speed to 60.0 km / h, and direction of travel to 90 degrees. A timestamp is added to each data entry, forming the vehicle terminal's movement trajectory data.

[0024] Step S120: Read the load time-series data of the edge computing nodes through the resource monitoring agent. Specifically, a lightweight resource monitoring agent is deployed on the edge nodes. The agent uses a timed reading mechanism to read the core load parameters of the edge nodes, sorts the data by timestamp to form time-series data, and uploads it to the data aggregation node.

[0025] For example, using Prometheus as a resource monitoring agent, deployed at the operating system level on edge nodes, the monitoring agent's read cycle is set to 100 milliseconds. The load data read includes CPU utilization, memory usage, task processing latency, and network bandwidth utilization. Each read data is associated with a timestamp and stored in chronological order as a CSV file, forming load time-series data.

[0026] Step S130: Extract the descriptive metadata of the tasks to be processed from the task scheduling queue to generate task queue data for the roadside sensing device. Specifically, the task queue data is a queued list of tasks to be processed in the roadside sensing device, containing descriptive information for each task, such as task type, input data volume, and processing requirements. The descriptive metadata is data used to describe the basic information of the task and is the core component of the task queue data, used to parse the specific requirements of the task. A task scheduling queue monitoring tool is used to monitor the task scheduling queue of the roadside sensing device in real time, extract the descriptive metadata of each task to be processed, deduplicate and classify the metadata, and encapsulate it into a standardized format to form task queue data.

[0027] For example, the task scheduling queue for roadside sensing devices such as high-definition cameras is stored in a Redis queue. The metadata for each task is stored in JSON format, containing six fields: task ID, task type, input data volume, algorithm complexity, priority, and submission time. A Redis client tool monitors this queue, retrieving the metadata of tasks to be processed every 200 milliseconds. The retrieved metadata is deduplicated to avoid processing the same task repeatedly, and categorized by task type, such as image recognition, road condition detection, and data forwarding. The deduplicated and categorized metadata is then encapsulated into a JSON array to generate the task queue data for the roadside sensing devices. Simultaneously, the total number of tasks and the number of tasks of each type are recorded and added to the task queue data.

[0028] Step S200: Within a preset time window, based on the movement trajectory data, predict the communication connection relationship between the vehicle terminal and the edge computing node, and generate a spatiotemporal correlation graph.

[0029] Specifically, the communication coverage of edge nodes is limited, and since vehicles are mobile, their communication connections with edge computing nodes change with location. A preset time window, set according to road network complexity, ranges from 5 to 30 seconds for prediction. Based on historical vehicle movement trajectories, a path prediction algorithm is used to estimate vehicle travel paths. Connection timing parameters are calculated by combining edge node coverage, and transmission rates are calculated by combining communication link parameters. Finally, a spatiotemporal relational graph is generated using a graph construction tool. The spatiotemporal relational graph is a graph structure centered on nodes and edges. Nodes are divided into vehicle terminal nodes and edge computing nodes, and edges represent the communication connections between them. Edge attributes contain two key pieces of information: the estimated connection duration (the time the vehicle stays within the edge node's coverage area) and the task transmission rate (the speed at which task data is transmitted between the vehicle and the edge node). The graph structure dynamically changes over time, reflecting the connection relationships at different points in time, providing spatial and temporal dimensions for computing power allocation.

[0030] In one possible implementation, within a preset time window, based on the movement trajectory data, the communication connection relationship between the vehicle terminal and the edge computing nodes is predicted to generate a spatiotemporal correlation graph. Step S200 further includes step S210, which predicts the driving path based on the movement trajectory data to obtain the estimated times when the vehicle terminal enters and leaves the communication coverage area of ​​each edge computing node. Specifically, a path prediction algorithm that integrates historical trajectory and real-time status is used. Historical vehicle movement trajectory data is input, and a future driving path coordinate sequence is output. Combined with the coordinate information of the edge node coverage area, the estimated times when the vehicle enters and leaves the coverage area are calculated.

[0031] For example, a fusion algorithm combining LSTM and attention mechanisms is used for driving path prediction. This algorithm includes an input layer, an LSTM layer, an attention layer, and an output layer. The LSTM layer contains 64 neurons, and the attention layer contains one attention head. The input layer takes the vehicle's trajectory data from the past 30 seconds as input, and the output layer outputs the coordinate sequence of the driving path for the next 10 seconds. The communication coverage range of each edge node is pre-entered into the system, using a circular coverage model. Each edge node corresponds to a set of coordinates; for example, edge node 1 has coordinates of 116.39700 degrees and 39.90800 degrees, with a coverage radius of 800 meters. Using a coordinate distance calculation algorithm, such as the Haversine formula, the distance between the vehicle's estimated path coordinate sequence and the coordinates of each edge node is calculated. When the distance is first less than or equal to 800 meters, the corresponding time is the estimated time when the vehicle enters the coverage range of that edge node; when the distance is first greater than 800 meters, the corresponding time is the estimated time when the vehicle leaves the coverage range of that edge node. If the vehicle's predicted path never enters the coverage area of ​​a certain edge node, then that edge node has no corresponding entry and exit times and is not included in the connection relationship calculation.

[0032] Step S220: Based on the estimated time, calculate the estimated connection duration between the vehicle terminal and each edge computing node. Specifically, a time difference calculation method is used, subtracting the estimated entry time from the estimated departure time to obtain the estimated connection duration. At the same time, the calculation results are checked for reasonableness, and outliers, such as negative values ​​or values ​​exceeding a preset time window, are removed. Finally, a list of estimated connection durations between the vehicle and each edge node is formed.

[0033] Step S230: Determine the task transmission rate based on the received signal strength index (RSSI) and channel occupancy rate of the link between the vehicle terminal and the edge computing node. Specifically, the roadside sensing device collects the received signal strength index (RSSI) and channel occupancy rate of the link, normalizes the collected data, and substitutes it into a preset transmission rate calculation formula to calculate the task transmission rate.

[0034] For example, roadside radar equipment is used to collect communication link parameters between vehicles and edge nodes. The collection period is set to 100 milliseconds, and the RSSI collection range is -120dBm to -30dBm. The larger the value, the stronger the signal. The channel occupancy rate collection range is 0-1, where 0 represents no occupancy and 1 represents full occupancy. The RSSI is normalized using the formula: RSSI normalized value = (collected RSSI - minimum RSSI) / (maximum RSSI - minimum RSSI). For example, if the collected RSSI is -60dBm, the minimum RSSI is -120dBm, and the maximum RSSI is -30dBm, then the RSSI normalized value = [-60 - (-120)] / [-30 - (-120)] = 60 / 90 ≈ 0.67. The formula for calculating the task transmission rate is: Task transmission rate = Link base transmission rate × RSSI normalized value × (1 - Channel occupancy rate), where the link base transmission rate is set according to the communication interface; the PC5 interface is set to 100Mbps, and the Uu interface is set to 50Mbps. If the collected channel occupancy rate is 0.2, then the task transmission rate = 100 × 0.67 × (1 - 0.2) = 53.6Mbps.

[0035] Step S240: Using the vehicle terminal and edge computing nodes as graph nodes, and the estimated connection duration and task transmission rate as edge weight attributes, construct the spatiotemporal association graph. Specifically, an open-source graph construction tool is used to define the node types and edge attributes of the graph. The identifiers of the vehicle and edge nodes are used as node IDs, and the estimated connection duration and task transmission rate are used as two weight parameters for the edges to construct an undirected graph. The connection between the vehicle and the edge nodes is bidirectional. A timestamp is added to the graph to reflect its spatiotemporal characteristics.

[0036] For example, using the NetworkX open-source graph construction tool, the node types of the graph are first defined. Vehicle nodes have IDs in the format car_XXX, such as car_001 and car_002, while edge nodes have IDs in the format edge_XXX, such as edge_001 and edge_002. Attribute identifiers are added to each node: vehicle nodes have attributes like type=vehicle and current speed=XX km / h, while edge nodes have attributes like type=edge node and coverage radius=XX meters. Edges between vehicles and edge nodes are undirected, and their weight attributes include two parameters: connection duration=XX seconds and transmission rate=XX Mbps. After construction, a timestamp is added to the spatiotemporal graph to identify the corresponding prediction time window. The graph is also stored in GraphML format, which contains a list of nodes, edges, node attributes, and edge attributes, allowing for direct parsing by deep learning models.

[0037] Step S300: Based on the task queue data, analyze and obtain the task computing power requirement characteristics of the tasks to be processed by the vehicle terminal.

[0038] Specifically, task computing power requirement characteristics are a set of core parameters used to describe the computing power resource requirements of a task on edge nodes. These mainly include core count estimates, memory usage estimates, and priority indicators. The core count estimate is the number of CPU cores required to complete the task, the memory usage estimate is the amount of memory required, and the priority is determined based on the urgency of the task, such as high priority, medium priority, and low priority. The computing power allocation of edge computing nodes needs to match specific task requirements. Different tasks, such as image recognition, data transmission, and traffic analysis, have different computing power requirements. A deserialization parsing method is used to extract basic task information, which is then combined with a pre-set computing power requirement mapping table to calculate relevant computing power estimates. Finally, feature encoding integrates this information into a unified computing power requirement feature, clarifying how much computing power the task needs and its priority, providing a task-level basis for computing power requirement prediction.

[0039] In one possible implementation, based on the task queue data, the task computing power requirement characteristics of the tasks to be processed by the vehicle terminal are parsed to obtain the data. Step S300 further includes step S310, which involves deserializing and parsing the task queue data to extract the input data volume, algorithm complexity level, and preset priority identifier for each task to be processed. Specifically, the task queue data generated in step S130 is stored in a serialized format, such as JSON or XML. According to the serialized format of the task queue data, it is parsed using the corresponding deserialization algorithm to extract three core pieces of information for each task to be processed. Among them, the input data volume refers to the size of the data that the task needs to process, the algorithm complexity level refers to the complexity of the task processing algorithm, and the preset priority identifier refers to the urgency of the task.

[0040] For example, task queue data is stored in JSON serialization format and deserialized using Python's `json` module. The `json.loads()` function is called to convert the JSON string into a Python dictionary. Each task's dictionary contains four core fields: "task_id", "input_data", "algorithm_complexity", and "priority". The `task_id` field corresponds to the task ID, the `input_data` field corresponds to the input data volume, the `algorithm_complexity` field corresponds to the algorithm complexity level, and the `priority` field corresponds to the priority indicator. During parsing, the contents of the three target fields are extracted, the input data volume is standardized and converted to MB units, and the algorithm complexity level and priority indicator are validated. If any values ​​deviate from the preset values, the task is identified as abnormal and temporarily excluded from the computational power requirement feature extraction scope.

[0041] Step S320: Based on the input data volume and the algorithm complexity level, query a pre-set computing power requirement mapping table to obtain the estimated number of cores and memory usage required for a single task. Specifically, a computing power requirement mapping table is constructed in advance through experiments and tests. A MySQL database is used to store the correspondence between the input data volume, algorithm complexity level, and estimated number of cores and memory usage. The input data volume and algorithm complexity level extracted in step S310 are used as query conditions. A multi-condition matching query is performed on the mapping table. Based on the range of the input data volume and the algorithm complexity level, the corresponding computing power estimate is matched. If there is no complete match, an approximate value is calculated using interpolation to obtain the estimated number of cores and memory usage required for each task to be processed.

[0042] For example, the pre-defined computing power requirement mapping table is as follows (partial data): 1. Input data volume 0-5MB, low algorithm complexity: estimated number of cores 0.5, estimated memory usage 128MB; 2. Input data volume 5-10MB, low algorithm complexity: estimated number of cores 1, estimated memory usage 256MB; 3. Input data volume 0-5MB, medium algorithm complexity: estimated number of cores 1, estimated memory usage 256MB; 4. Input data volume 5-10MB, medium algorithm complexity: estimated number of cores 1.5, estimated memory usage 384MB; 5. Input data volume 0-5MB, high algorithm complexity: estimated number of cores 1.5, estimated memory usage 384MB; 6. Input data volume 5-10MB, high algorithm complexity: estimated number of cores 2, estimated memory usage 512MB. If the input data volume of a task to be processed is 8MB and the algorithm complexity level is high, query the mapping table and match the entries with an input data volume of 5-10MB and high algorithm complexity, obtaining an estimated core count of 2 cores and an estimated memory usage of 512MB.

[0043] Step S330 involves performing feature encoding on the core count estimate, memory usage estimate, and priority identifier to generate the task computing power requirement feature. Specifically, the core count estimate and memory usage estimate obtained in step S320 are processed using numerical standardization, and the priority identifier extracted in step S310 is processed using one-hot encoding. The processed values ​​are then integrated into a feature vector, which serves as the task computing power requirement feature.

[0044] For example: First, the core count and memory usage estimates are standardized using a min-max normalization algorithm. The normalization range for the core count estimate is set to [0,1] (the preset core count estimate range is 0.5-4 cores), and the normalization range for the memory usage estimate is set to [0,1] (the preset memory usage estimate range is 128MB-2048MB). For example, if the core count estimate is 2 cores, after normalization it becomes (2-0.5) / (4-0.5) = 1.5 / 3.5 ≈ 0.43. If the memory usage estimate is 512MB, after normalization it becomes (512-128) / (2048-128) = 384 / 1920 = 0.2. Priority identifiers are then one-hot encoded: high priority is encoded as [1,0,0], medium priority as [0,1,0], and low priority as [0,0,1]. The standardized core count estimate and memory usage estimate are integrated with the one-hot encoded priority vector to generate a task computing power requirement feature vector. For example, the feature vector for this task is [0.43, 0.2, 1, 0, 0], where the first two elements are the standardized computing power estimate and the last three elements are the priority encoding. If there are multiple tasks to be processed, a corresponding feature vector is generated for each task, forming a task computing power requirement feature set.

[0045] Step S400: Input the spatiotemporal correlation graph and the task computing power demand characteristics into the computing power demand prediction unit to obtain the estimated value of the computing power demand of the edge computing node.

[0046] Specifically, the computing power demand prediction unit is a functional module dedicated to predicting the computing power demand of edge nodes. It incorporates a deep learning model that integrates a graph attention mechanism and a temporal convolutional network. This model uses a spatiotemporal graph as spatial feature input and task computing power demand features as initial node embedding features. The graph attention mechanism calculates attention weights and aggregates demand features, then extracts temporal features through a temporal convolutional layer, ultimately outputting the estimated computing power demand for each time slice. The estimated computing power demand is the predicted amount of computing resources required by an edge node in different time slices within a preset future time window, expressed in terms of the number of CPU cores and memory usage.

[0047] In one possible implementation, the spatiotemporal correlation graph and the task computing power requirement features are input to a computing power requirement prediction unit to obtain an estimated value of the computing power requirement of the edge computing nodes. Step S400 further includes step S410, whereby the computing power requirement prediction unit incorporates a temporal convolutional network based on a graph attention mechanism. The spatiotemporal correlation graph is used as the spatial feature input to the temporal convolutional network, and the task computing power requirement features are used as the initial embedding features of the vehicle terminal nodes. The attention weight coefficients of each edge computing node to the vehicle terminal nodes it connects to are calculated through the graph attention mechanism. Specifically, the computing power requirement prediction unit incorporates a temporal convolutional network based on a graph attention mechanism. This network is a deep learning model that integrates graph attention mechanisms and temporal convolutional networks, which can capture both spatial correlations and temporal dependencies. The spatiotemporal correlation graph generated in step S240 is used as the input of the network spatial features, and the task computing power requirement features generated in step S330 are used as the initial embedding features of the vehicle nodes in the spatiotemporal correlation graph. Through the graph attention mechanism, the attention weight coefficient of each edge node to each vehicle node it connects to is calculated. This coefficient is used to reflect the degree of influence of the vehicle node task on the computing power requirement of the edge node.

[0048] For example, a graph attention layer is constructed, containing two attention heads. A multi-head attention mechanism is used to improve feature extraction. The inputs are the adjacency matrix of the spatiotemporal graph and the initial embedding features of the nodes. The adjacency matrix of the spatiotemporal graph has a dimension of N×N, where N is the total number of nodes. For example, with 10 vehicle nodes and 5 edge nodes, the dimension is 15×15. In the adjacency matrix, an element of 1 indicates that two nodes are connected, and 0 indicates that they are not connected. The initial embedding features of vehicle nodes are the feature vectors generated in step S330, and the initial embedding features of edge nodes are set to their historical load mean vectors. Attention scores are calculated using an attention calculation function, which employs linear transformation and activation functions. First, a linear transformation is performed on the embedding features of each node. Then, the transformed features of edge nodes and vehicle nodes are concatenated, and the attention score is calculated using the LeakyReLU activation function. The Softmax function is used to normalize the attention scores of all vehicle nodes connected to the same edge node, resulting in attention weight coefficients. The sum of the weight coefficients is 1. The more stable the connection and the higher the task priority, the larger the weight coefficient.

[0049] Step S420: Based on the attention weight coefficient, the task computing power requirement features of the vehicle terminal node are weighted and aggregated to generate the aggregated requirement features of the edge computing node. Specifically, based on the attention weight coefficient calculated in step S410, the task computing power requirement features of all vehicle nodes connected to each edge node are weighted and summed. That is, for each edge node, all connected vehicle nodes are traversed, the task computing power requirement feature of each vehicle node is multiplied by the corresponding attention weight coefficient, and then all product results are summed to obtain the aggregated requirement feature of the edge node, which reflects the overall task computing power requirement that the edge node needs to undertake.

[0050] For example, edge node edge_001 connects two vehicle nodes car_001 and car_002. The task computing power requirement of car_001 is [0.43, 0.2, 1, 0, 0], and the attention weight coefficient is 0.8; the task computing power requirement of car_002 is [0.29, 0.1, 0, 1, 0], and the attention weight coefficient is 0.2. The weighted aggregation is calculated by weighting and summing each feature dimension separately. That is, the first element of the aggregated requirement feature = 0.43 × 0.8 + 0.29 × 0.2 = 0.402, the second element = 0.2 × 0.8 + 0.1 × 0.2 = 0.18, the third element = 1 × 0.8 + 0 × 0.2 = 0.8, the fourth element = 0 × 0.8 + 1 × 0.2 = 0.2, and the fifth element = 0 × 0.8 + 0 × 0.2 = 0. The final aggregated demand feature of edge_001 is [0.402, 0.18, 0.8, 0.2, 0]. This feature integrates the task computing power requirements of all the vehicle nodes it connects to. The first two elements reflect the overall computing power demand intensity, and the last three elements reflect the overall task priority distribution.

[0051] Step S430: The aggregated demand features are input into the temporal convolutional layer of the temporal convolutional network to extract temporal dependency features and output the estimated computing power demand of each edge computing node for multiple time slices within the preset time window. Specifically, the aggregated demand features of the edge nodes generated in step S420 are input into the temporal convolutional layer of the temporal convolutional network. The temporal convolutional layer consists of multiple convolutional blocks, each containing a convolutional layer, a batch normalization layer, and an activation function. Temporal features are extracted through convolution operations, and then the features are mapped to the estimated computing power demand through a fully connected layer. The dimension of the estimated value is consistent with the number of time slices.

[0052] For example, a temporal convolutional network contains three convolutional blocks and one fully connected layer. Each convolutional block consists of a convolutional layer, a batch normalization layer, and a ReLU activation function. The convolutional layer uses 32 kernels of size 3 with a stride of 1 and the padding is set to "same" to ensure consistent input and output dimensions. The batch normalization layer has a momentum of 0.9 and an epsilon of 1e-5 to accelerate model training and avoid overfitting. The input aggregation feature is a temporal sequence, such as a preset time window of 10 seconds divided into five 2-second time slices, each corresponding to a specific aggregation feature. This input sequence is fed into the temporal convolutional layer. The first convolutional block performs convolution operations on the input sequence to extract local temporal features. The batch normalization layer normalizes the convolutional results, and the ReLU activation function introduces non-linear features. After processing by three convolutional blocks, the output temporal feature dimension is 5×32. This feature is flattened to obtain a 160-dimensional feature vector, which is then input into a fully connected layer. The fully connected layer contains two hidden layers with 64 and 32 neurons respectively, and the activation function for both is ReLU. Finally, the output layer outputs the estimated computing power requirement, which uses a linear activation function.

[0053] Step S500: Based on the estimated computing power demand and the load time series data, generate a computing power resource allocation strategy for edge computing nodes.

[0054] Specifically, the available computing power resources are extracted from the load time-series data of edge nodes, which is the total computing power of the edge nodes minus the occupied computing power. The difference between the estimated computing power demand of the edge nodes and the current available computing power resources is calculated, with positive numbers indicating insufficient computing power and negative numbers indicating idle computing power. The dynamic resource topology of the edge nodes is obtained, including which edge nodes can mutually schedule computing power, scheduling bandwidth limitations, scheduling latency, etc. An optimization model is constructed with the goal of minimizing the frequency of cross-node task migration and maximizing the utilization rate of local computing power resources. The model is solved through an optimization algorithm to generate a computing power resource allocation strategy. The computing power resource allocation strategy is a specific scheme used to guide the allocation of computing power to edge nodes, including the computing power allocation amount for each edge node, the cross-edge node computing power scheduling scheme, and the task migration strategy.

[0055] In one possible implementation, a computing resource allocation strategy for edge computing nodes is generated based on the estimated computing power demand and the load time-series data. Step S500 further includes step S510, extracting the current available computing power resources of the edge computing nodes from the load time-series data. Specifically, the edge computing node load time-series data generated in step S120 includes various load parameters of the edge nodes at different times. The load time-series data is smoothed using a sliding window averaging method to remove instantaneous fluctuations, and the core load parameters at the current moment are extracted, including the total number of CPU cores, the number of CPU cores already occupied, the total memory, and the memory already occupied. The current available computing power resources are calculated by the difference.

[0056] For example: The load time-series data of the edge node is stored in CSV format, with one data point every 100 milliseconds. The CPU utilization and memory usage data of the most recent 5 data points are extracted, and the smoothed CPU utilization and memory usage are calculated using the sliding window mean method. Given that the edge node has a total of 8 CPU cores and 8GB (8192MB) of memory, the smoothed CPU utilization is 25.3%, and the memory usage is 30.5%. The calculated number of CPU cores currently in use is approximately 2.02 (8 × 25.3%), and the amount of memory currently in use is approximately 2498.56MB (8192 × 30.5%). The current number of available CPU cores is approximately 5.98 (total cores - currently in use), and the current amount of available memory is approximately 5693.44MB (total memory - currently in use).

[0057] Step S520: For each edge computing node, compare the estimated computing power demand with the current available computing power resources to obtain the computing power difference. Specifically, calculate the average estimated computing power demand for each edge node within a preset time window, and subtract the current available computing power resources from the average estimated computing power demand to obtain the computing power difference. A positive number indicates insufficient computing power, requiring additional computing power; a negative number indicates idle computing power, with surplus computing power available for scheduling; and zero indicates matching computing power.

[0058] Step S530: Obtain the dynamic resource topology between edge computing nodes. Specifically, a link detection tool is used to periodically detect the communication link status between edge nodes, collect parameters such as link bandwidth, scheduling latency, and link connectivity, and construct a dynamic resource topology table. The topology table contains four core fields: edge node pairs, link bandwidth, scheduling latency, and connectivity status. At the same time, edge node pairs with disconnected connectivity are removed to ensure the availability of the topology.

[0059] For example, the ping command combined with the iperf tool is used to detect the communication link status between edge nodes, with a detection period of 1 second. Each edge node acts as the initiator, sending detection data packets to all other edge nodes. The receiving end sends back response data packets. The scheduling latency is calculated based on the round-trip time of the data packets. If the round-trip time exceeds 500 milliseconds, the link connectivity is determined to be broken. The iperf tool is used to test the link bandwidth, with a test time of 500 milliseconds. The average value of the test results is taken as the link bandwidth. If the bandwidth is less than 10 Mbps, it is marked as a low-bandwidth link and is not used for computing power scheduling. A dynamic resource topology table is constructed, containing four fields: edge node pairs, link bandwidth, scheduling latency, and connectivity status. An example topology table is shown below (partial data): 1. edge_001-edge_002: Link bandwidth 100Mbps, scheduling latency 50ms, connectivity status: Connected; 2. edge_001-edge_003: Link bandwidth 80Mbps, scheduling latency 60ms, connectivity status: Connected; 3. edge_002-edge_003: Link bandwidth 90Mbps, scheduling latency 55ms, connectivity status: Connected; 4. edge_001-edge_004: Link bandwidth 5Mbps, scheduling latency 800ms, connectivity status: Disconnected. The topology table is updated every second. If the connectivity status changes to disconnected, the edge node pair is temporarily removed from the topology table until the link is restored.

[0060] Step S540: With the optimization objectives of minimizing the frequency of cross-node task migrations and maximizing the utilization rate of local computing resources, and using the computing power difference as a decision variable constraint and the dynamic resource topology as a resource availability constraint, a computing resource allocation optimization model is constructed. Specifically, a multi-objective optimization model framework is adopted. Based on the computing power difference in step S520 and the dynamic resource topology in step S530, two core optimization objectives are determined: first, to reduce the number of cross-node task migrations to avoid increased task latency due to excessive migrations; second, to improve the utilization rate of local computing power on edge nodes to avoid resource waste. Constraints are set, and the two core objectives are integrated using a linear weighted summation method to construct the computing resource allocation optimization model. The weights of the objective function are adjusted according to the actual scenario, and the constraints are used to ensure the feasibility and rationality of the model solution.

[0061] Step S550: Solve the computing resource allocation optimization model to generate a computing resource allocation strategy. Specifically, the particle swarm optimization algorithm is selected to solve the computing resource allocation optimization model. Core parameters are set, and the optimal solution is obtained after iterative convergence. After verification, a standardized allocation strategy is generated.

[0062] For example, the particle swarm size is set to 50, the number of iterations to 100, the inertia weight to 0.7, and the learning factor to 1.5. The solution process is as follows: initialize the particle swarm, calculate the fitness value, update the particle position / velocity, and iterate until convergence to obtain the optimal decision variables. The optimal solution is verified, organized into JSON format, and the local allocation, cross-node scheduling, and task migration schemes for each node are determined to ensure that the constraints are met.

[0063] Step S600: Based on the computing power resource allocation strategy, a computing power adjustment instruction is issued to the edge computing node.

[0064] Specifically, the computing power resource configuration strategy generated in step S550 clarifies the local computing power allocation, cross-node computing power scheduling, and task migration schemes for each edge computing node. Following the process of strategy decomposition, instruction encapsulation, encrypted transmission, and execution feedback, the strategy is decomposed into executable computing power adjustment instructions, which are then sent to each corresponding edge computing node through the communication link of the vehicle-road-cloud environment. This guides the edge nodes to complete the dynamic adjustment of computing power resources, realizes the adaptive allocation of edge computing power, and ensures that the task computing power requirements of all vehicle terminals are met.

[0065] For example: Split the strategy by node ID, generate exclusive instructions in Protobuf format, including instruction ID, adjustment parameters, execution time limit and MD5 checksum, and transmit them encrypted via TLS1.3. Edge nodes verify the instructions and then execute the adjustments. After completion, they report the status. The central cloud monitors the execution status. If it fails, it will be retried, up to 3 times. If the retry fails, the abnormal node is marked and the strategy is adjusted.

[0066] This application's embodiments utilize three types of multi-dimensional state data in the vehicle-road-cloud environment: vehicle movement trajectory, edge node load, and roadside equipment task queue. Within a preset time window, the communication connection between vehicles and edge nodes is predicted based on the trajectory data, and a spatiotemporal correlation graph is constructed. The computing power requirement characteristics of the tasks to be processed are parsed from the task queue data. The spatiotemporal correlation graph and computing power requirement characteristics are input into the computing power requirement prediction unit to obtain the estimated value of edge node computing power requirement. A computing power resource allocation strategy is generated by combining load time series data. Based on the strategy, computing power adjustment instructions are issued to edge nodes to perform adaptive allocation of edge computing power in the vehicle-road-cloud environment. These technical means solve the technical problem of inaccurate computing power allocation in existing vehicle-road-cloud environments, which cannot adapt to the dynamic changes of the environment. This achieves the technical effect of adapting to the dynamic changes of the vehicle-road-cloud environment and improving the accuracy of computing power allocation.

[0067] In the above text, refer to Figure 1 This paper describes in detail an edge computing power adaptive allocation method in a vehicle-road-cloud environment according to embodiments of the present invention. Next, we will refer to... Figure 2 An edge computing power adaptive allocation system in a vehicle-road-cloud environment according to an embodiment of the present invention is described.

[0068] The edge computing power adaptive allocation system in a vehicle-road-cloud environment according to embodiments of the present invention addresses the technical problem of inaccurate computing power allocation in existing edge computing power allocation systems that fail to adapt to dynamic changes in the environment. This results in improved accuracy of computing power allocation. The system aims to adapt to the dynamic changes in the vehicle-road-cloud environment and enhance the accuracy of computing power allocation. The edge computing power adaptive allocation system in a vehicle-road-cloud environment includes: a multi-dimensional state data acquisition module 10, a spatiotemporal correlation graph generation module 20, a task computing power requirement analysis module 30, a computing power requirement prediction module 40, a computing power resource configuration strategy generation module 50, and a computing power adjustment instruction issuance module 60.

[0069] The multi-dimensional state data acquisition module 10 is used to acquire multi-dimensional state data in the vehicle-road-cloud environment, including vehicle terminal movement trajectory data, edge computing node load time-series data, and roadside sensing device task queue data. The vehicle-road-cloud environment includes multiple intelligent connected vehicles, multiple edge computing nodes, multiple roadside units, and a central cloud platform. The spatiotemporal correlation graph generation module 20 is used to predict the communication connection relationship between the vehicle terminal and the edge computing node based on the movement trajectory data within a preset time window, and generate a spatiotemporal correlation graph. The task computing power requirement analysis module 30 is used to... Based on the task queue data, the task computing power requirement characteristics of the tasks to be processed by the vehicle terminal are obtained by parsing; the computing power requirement prediction module 40 is used to input the spatiotemporal correlation graph and the task computing power requirement characteristics into the computing power requirement prediction unit to obtain the estimated computing power requirement of the edge computing node; the computing power resource configuration strategy generation module 50 is used to generate the computing power resource configuration strategy of the edge computing node based on the estimated computing power requirement and the load time series data; the computing power adjustment instruction issuance module 60 is used to issue computing power adjustment instructions to the edge computing node based on the computing power resource configuration strategy.

[0070] The detailed description of the specific configuration of the multi-dimensional state data acquisition module 10 is explained as follows: As mentioned above, it collects multi-dimensional state data in the vehicle-road-cloud environment, including the movement trajectory data of the vehicle terminal, the load time series data of the edge computing node, and the task queue data of the roadside sensing device. The multi-dimensional state data acquisition module 10 may further include: a movement trajectory data acquisition unit for receiving vehicle status messages broadcast by intelligent connected vehicles and parsing them to obtain the movement trajectory data of the vehicle terminal; a load time series data acquisition unit for reading the load time series data of the edge computing node through a resource monitoring agent; and a task queue data acquisition unit for extracting the descriptive metadata of the tasks to be processed from the task scheduling queue and generating the task queue data of the roadside sensing device.

[0071] The detailed description of the specific configuration of the spatiotemporal correlation graph generation module 20 is as follows: As mentioned above, within a preset time window, based on the movement trajectory data, the communication connection relationship between the vehicle terminal and the edge computing nodes is predicted to generate a spatiotemporal correlation graph. The spatiotemporal correlation graph generation module 20 may further include: an estimated time acquisition unit for predicting the driving path of the movement trajectory data to obtain the estimated time when the vehicle terminal enters and leaves the communication coverage area of ​​each edge computing node; an estimated connection duration calculation unit for calculating the estimated connection duration between the vehicle terminal and each edge computing node based on the estimated time; a task transmission rate determination unit for determining the task transmission rate based on the signal reception strength indication value and channel occupancy rate of the link between the vehicle terminal and the edge computing node; and a spatiotemporal correlation graph construction unit for constructing the spatiotemporal correlation graph with the vehicle terminal and the edge computing node as graph nodes and the estimated connection duration and the task transmission rate as edge weight attributes.

[0072] The specific configuration of the task computing power requirement parsing module 30 is described in detail below: As mentioned above, based on the task queue data, the task computing power requirement characteristics of the tasks to be processed by the vehicle terminal are parsed and obtained. The task computing power requirement parsing module 30 may further include: a deserialization parsing unit for deserializing and parsing the task queue data to extract the input data volume, algorithm complexity level, and preset priority identifier of each task to be processed; a computing power requirement mapping unit for querying a preset computing power requirement mapping table according to the input data volume and the algorithm complexity level to obtain the estimated number of cores and the estimated memory usage of a single task; and a feature encoding unit for feature encoding the estimated number of cores, the estimated memory usage, and the priority identifier to generate the task computing power requirement characteristics.

[0073] The detailed description of the specific configuration of the computing power demand prediction module 40 is explained as follows: As mentioned above, the spatiotemporal correlation graph and the task computing power demand features are input to the computing power demand prediction unit to obtain the estimated computing power demand of the edge computing nodes. The computing power demand prediction module 40 may further include: an attention weight coefficient calculation unit, which is used to build a time convolutional network based on graph attention mechanism in the computing power demand prediction unit, input the spatiotemporal correlation graph as the spatial feature of the time convolutional network, and use the task computing power demand features as the initial embedding features of the vehicle terminal nodes, and calculate the attention weight coefficient of each edge computing node to the vehicle terminal nodes it connects to through the graph attention mechanism; a weighted aggregation unit, which is used to perform weighted aggregation of the task computing power demand features of the vehicle terminal nodes according to the attention weight coefficients to generate aggregated demand features of the edge computing nodes; and a temporal dependency feature extraction unit, which is used to input the aggregated demand features to the temporal convolutional layer of the time convolutional network, extract temporal dependency features, and output the estimated computing power demand of each edge computing node for multiple time slices within the preset time window.

[0074] The detailed description of the specific configuration of the computing power resource allocation strategy generation module 50 is explained as follows: As mentioned above, based on the estimated computing power demand and the load time series data, a computing power resource allocation strategy for edge computing nodes is generated. The computing power resource allocation strategy generation module 50 may further include: a current available computing power resource extraction unit for extracting the current available computing power resource of edge computing nodes from the load time series data; a computing power difference calculation unit for comparing the estimated computing power demand and the current available computing power resource for each edge computing node to obtain the computing power difference; a dynamic resource topology acquisition unit for acquiring the dynamic resource topology between edge computing nodes; a computing power resource allocation optimization model construction unit for constructing a computing power resource allocation optimization model with the optimization objectives of minimizing cross-node task migration frequency and maximizing local computing power resource utilization, using the computing power difference as a decision variable constraint and the dynamic resource topology as a resource availability constraint; and a computing power resource allocation strategy generation unit for solving the computing power resource allocation optimization model to generate a computing power resource allocation strategy.

[0075] The edge computing power adaptive allocation system in the vehicle-road-cloud environment provided in the embodiments of the present invention can execute the edge computing power adaptive allocation method in the vehicle-road-cloud environment provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the method.

[0076] Although this application makes various references to certain modules in the system according to the embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy distinction between each other and are not used to limit the scope of protection of this invention.

[0077] Based on the foregoing embodiments, this application also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor of an electronic device, it can implement the edge computing power adaptive allocation method in the vehicle-road-cloud environment as described in any of the foregoing embodiments.

[0078] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. An adaptive edge computing power allocation method in a vehicle-road-cloud environment, characterized in that, The method includes: Collect multi-dimensional state data in the vehicle-road-cloud environment, including vehicle terminal movement trajectory data, edge computing node load time series data, and roadside sensing device task queue data. The vehicle-road-cloud environment includes multiple intelligent connected vehicles, multiple edge computing nodes, multiple roadside units, and a central cloud platform. Within a preset time window, based on the movement trajectory data, the communication connection relationship between the vehicle terminal and the edge computing node is predicted, and a spatiotemporal correlation graph is generated. Based on the task queue data, the task computing power requirement characteristics of the tasks to be processed by the vehicle terminal are obtained by parsing. The spatiotemporal correlation graph and the task computing power requirement characteristics are input into the computing power requirement prediction unit to obtain the estimated computing power requirement of the edge computing node. Based on the estimated computing power demand and the load time series data, a computing power resource allocation strategy for edge computing nodes is generated. Based on the aforementioned computing resource allocation strategy, computing power adjustment instructions are issued to edge computing nodes.

2. The edge computing power adaptive allocation method in a vehicle-road-cloud environment as described in claim 1, characterized in that, Collect multi-dimensional state data in the vehicle-road-cloud environment, including vehicle terminal movement trajectory data, edge computing node load time-series data, and roadside sensing device task queue data, including: Receive vehicle status messages broadcast by intelligent connected vehicles and parse them to obtain the vehicle terminal's movement trajectory data; The resource monitoring agent reads the load time-series data of the edge computing nodes. Extract the descriptive metadata of the tasks to be processed from the task scheduling queue to generate the task queue data for the roadside sensing devices.

3. The edge computing power adaptive allocation method in a vehicle-road-cloud environment as described in claim 1, characterized in that, Within a preset time window, based on the movement trajectory data, the communication connection relationship between the vehicle terminal and the edge computing node is predicted, and a spatiotemporal correlation graph is generated, including: The vehicle trajectory data is used to predict the driving path and obtain the estimated time when the vehicle terminal enters and leaves the communication coverage area of ​​each edge computing node. Based on the estimated time, calculate the estimated connection duration between the vehicle terminal and each edge computing node; The task transmission rate is determined based on the signal reception strength indicator and channel occupancy rate of the link between the vehicle terminal and the edge computing node. Using vehicle terminals and edge computing nodes as graph nodes, and the estimated connection duration and the task transmission rate as edge weight attributes, the spatiotemporal association graph is constructed.

4. The edge computing power adaptive allocation method in a vehicle-road-cloud environment as described in claim 1, characterized in that, Based on the task queue data, the task computing power requirement characteristics of the tasks to be processed by the vehicle terminal are parsed and obtained, including: The task queue data is deserialized and parsed to extract the input data volume, algorithm complexity level, and preset priority identifier of each task to be processed. Based on the input data volume and the algorithm complexity level, query the preset computing power requirement mapping table to obtain the estimated number of cores and memory usage required for a single task; The core count estimate, memory usage estimate, and priority identifier are used for feature encoding to generate the task computing power requirement feature.

5. The edge computing power adaptive allocation method in a vehicle-road-cloud environment as described in claim 1, characterized in that, The spatiotemporal correlation graph and the task computing power requirement characteristics are input into the computing power requirement prediction unit to obtain the estimated computing power requirement of the edge computing node, including: The computing power demand prediction unit has a built-in temporal convolutional network based on graph attention mechanism. The spatiotemporal correlation graph is used as the spatial feature input of the temporal convolutional network, and the task computing power demand feature is used as the initial embedding feature of the vehicle terminal node. The attention weight coefficient of each edge computing node to the vehicle terminal node it is connected to is calculated through the graph attention mechanism. Based on the attention weight coefficient, the task computing power requirement features of the vehicle terminal node are weighted and aggregated to generate the aggregated requirement features of the edge computing node. The aggregated demand features are input into the temporal convolutional layer of the temporal convolutional network to extract temporal dependency features and output the estimated computing power demand of each edge computing node for multiple time slices within the preset time window.

6. The edge computing power adaptive allocation method in a vehicle-road-cloud environment as described in claim 1, characterized in that, Based on the estimated computing power demand and the load time series data, a computing power resource allocation strategy for edge computing nodes is generated, including: Extract the current available computing power resources of the edge computing nodes from the load time series data; For each edge computing node, the estimated computing power demand is compared with the current available computing power resources to obtain the computing power difference; Obtain the dynamic resource topology between edge computing nodes; With the optimization objectives of minimizing the frequency of cross-node task migration and maximizing the utilization rate of local computing resources, and with the computing power difference as the decision variable constraint and the dynamic resource topology as the resource availability constraint, a computing resource configuration optimization model is constructed. The computing resource allocation optimization model is solved to generate a computing resource allocation strategy.

7. An edge computing power adaptive allocation system in a vehicle-road-cloud environment, characterized in that, The system is used to implement the edge computing power adaptive allocation method in the vehicle-road-cloud environment as described in any one of claims 1-6, and the system includes: The multi-dimensional state data acquisition module is used to collect multi-dimensional state data in the vehicle-road-cloud environment, including vehicle terminal movement trajectory data, edge computing node load time series data, and roadside sensing device task queue data. The vehicle-road-cloud environment includes multiple intelligent connected vehicles, multiple edge computing nodes, multiple roadside units, and a central cloud platform. The spatiotemporal correlation graph generation module is used to predict the communication connection relationship between the vehicle terminal and the edge computing node based on the movement trajectory data within a preset time window, and generate a spatiotemporal correlation graph. The task computing power requirement analysis module is used to analyze and obtain the task computing power requirement characteristics of the tasks to be processed by the vehicle terminal based on the task queue data. The computing power demand prediction module is used to input the spatiotemporal correlation graph and the task computing power demand characteristics into the computing power demand prediction unit to obtain the estimated value of computing power demand of edge computing nodes. The computing resource allocation strategy generation module is used to generate a computing resource allocation strategy for edge computing nodes based on the estimated computing demand and the load time series data. The computing power adjustment instruction issuing module is used to issue computing power adjustment instructions to edge computing nodes based on the computing power resource configuration strategy.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the edge computing power adaptive allocation method in the vehicle-road-cloud environment as described in any one of claims 1-6.

Images