Resource scheduling method for power internet of things edge node based on artificial intelligence

Abstract

The application provides a resource scheduling method of an artificial intelligence-based power Internet of Things edge node, can determine a first target edge node in each edge node in the power Internet of Things; when the computing resource utilization rate of the first target edge node exceeds a preset resource rate threshold value, and a task offloading event occurs in a second target edge node within a preset time period, the running parameters of the first target edge node are acquired; the second target edge node is an adjacent node of the first target edge node; the running parameters of the first target edge node are processed by using a prediction model to obtain a resource demand index of the first target edge node; the resource demand index of the first target edge node, a load balancing offset and a time-sensitive task weight are input into a source scheduling strategy executor to obtain a resource scheduling decision matrix; and a resource scheduling operation is performed according to the resource scheduling decision matrix. The dynamic and changeable load demand in the edge computing scene can be met.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a resource scheduling method for edge nodes of the power Internet of Things based on artificial intelligence. Background Technology

[0002] With the rapid development of power Internet of Things (IoT) technology, edge computing nodes in smart grids undertake critical tasks such as real-time data acquisition, equipment status monitoring, and localized decision-making. Traditional power system resource scheduling methods mainly rely on fixed threshold rules or simple polling mechanisms, which are difficult to cope with the dynamic and ever-changing load demands in edge computing scenarios. Summary of the Invention

[0003] The technical problem this application aims to solve is to provide a resource scheduling method and system for edge nodes of the power Internet of Things based on artificial intelligence, capable of proactively predicting resource scheduling. The specific solution is as follows:

[0004] A resource scheduling method for edge nodes of the power Internet of Things based on artificial intelligence includes:

[0005] In response to a resource scheduling instruction, obtain the node identifier in the resource scheduling instruction;

[0006] The first target edge node is determined among the various edge nodes in the power Internet of Things based on the node identifier;

[0007] If the computing resource utilization rate of the first target edge node exceeds a preset resource rate threshold, and a task unloading event occurs at the second target edge node within a preset time period, then the running parameters of the first target edge node are obtained; the second target edge node is an adjacent node of the first target edge node.

[0008] The operating parameters of the first target edge node are processed using a pre-trained prediction model to obtain the resource demand index of the first target edge node.

[0009] The resource demand index, load balancing offset, and time-sensitive task weight of the first target edge node are input into the source scheduling policy executor to obtain the resource scheduling decision matrix.

[0010] Resource scheduling operations are performed on the first target edge node according to the resource scheduling decision matrix.

[0011] Optionally, in the above method, obtaining the operating parameters of the first target edge node includes:

[0012] The edge sensing module obtains at least one of the following operating parameters for the first target edge node: power quality disturbance coefficient, temperature distribution image, channel electromagnetic interference entropy, and dynamic load mutation rate; the power quality disturbance coefficient is a composite quantitative index of the transient disturbance intensity and steady-state distortion degree of the power system; the channel electromagnetic interference entropy characterizes the signal transmission quality; and the dynamic load mutation rate characterizes the load change rate.

[0013] Optionally, in the above method, the prediction model includes convolutional layers, gated recurrent units, and an output layer. The step of processing the operating parameters of the first target edge node using the pre-trained prediction model to obtain the resource demand index of the first target edge node includes:

[0014] The temperature distribution image is input into the convolutional layer of the prediction model to extract the temperature distribution features from the temperature distribution image through the convolutional layer.

[0015] The power quality disturbance coefficient, the channel electromagnetic interference entropy, and the dynamic load mutation rate are input into the gated loop unit of the prediction model, so that the time-series variation characteristics of the power quality disturbance coefficient, the channel electromagnetic interference entropy, and the dynamic load mutation rate can be obtained through the gated loop unit.

[0016] The temperature distribution characteristics and the temporal variation characteristics are input into the output layer of the prediction model, so that the output layer outputs the resource demand index of the first target edge node based on the temperature distribution characteristics and the temporal variation characteristics.

[0017] Optionally, the training process of the prediction model in the above method includes:

[0018] Obtain a first training dataset and an initial prediction model to be trained; the first training dataset includes multiple first training samples and a resource demand index label for each first training sample; each first training sample includes the historical power quality disturbance coefficient, historical temperature distribution image, historical channel electromagnetic interference entropy, and historical dynamic load mutation rate of the edge node of the power Internet of Things.

[0019] The initial prediction model is trained using the first training dataset;

[0020] If the initial prediction model meets the training completion condition, the initial prediction model that meets the training completion condition is taken as the trained prediction model.

[0021] Optionally, in the above method, the step of inputting the resource demand index, load balancing offset, and time-sensitive task weight of the first target edge node into the source scheduling policy executor to obtain the resource scheduling decision matrix includes:

[0022] The load balancing offset of the first target edge node is calculated based on the resource utilization of the first target edge node and the resource utilization of the second target edge node.

[0023] The time-sensitive task weight of the first target edge node is determined based on the priority of each task assigned to the first target edge node;

[0024] The resource demand index, load balancing offset, and time-sensitive task weight of the first target edge node are input into the source scheduling policy executor to obtain the resource scheduling decision matrix.

[0025] Optionally, the construction process of the resource scheduling policy executor in the above method includes:

[0026] The learning environment information of the first target edge node and the second training dataset are obtained; the learning environment information includes a state space, an action space, and a reward function; the state space includes multiple state information of the first target edge node; the action space includes multiple resource scheduling operations; the second training dataset includes multiple second training samples; each second training sample includes historical state data of the edge node of the power Internet of Things, decision result annotations, and performance change data after executing historical resource scheduling operations.

[0027] Construct an initial decision tree based on the learning environment information;

[0028] The initial decision tree is trained using the second training dataset to obtain a resource scheduling policy executor.

[0029] Optionally, in the above method, performing resource scheduling operations on the first target edge node according to the resource scheduling decision matrix includes:

[0030] The computing resources required for each task allocated to the first target edge node are determined based on the resource scheduling decision matrix.

[0031] The computing resources required for each task are allocated to each task.

[0032] Optionally, after performing resource scheduling operations on the first target edge node according to the resource scheduling decision matrix, the above method further includes:

[0033] Obtain resource scheduling effect feedback data of the first target edge node;

[0034] The model parameters of the prediction model are updated using the resource scheduling effect feedback data.

[0035] Optionally, in the above method, obtaining the resource scheduling effect feedback data of the first target edge node includes:

[0036] Determine the resource utilization deviation between the actual resource utilization rate of the first target edge node and the preset target resource utilization rate;

[0037] The task unloading success rate of the first target edge node is determined based on the actual number of tasks successfully unloaded and the total number of tasks.

[0038] The resource utilization deviation and the task unloading success rate are used as feedback data on the resource scheduling effect of the first target edge node.

[0039] Optionally, in the above method, updating the model parameters of the prediction model using the resource scheduling effect feedback data includes:

[0040] If the resource utilization deviation is greater than a preset deviation threshold and the task unloading success rate is lower than a preset success rate threshold, the model parameters of the prediction model are updated using the resource scheduling effect feedback data.

[0041] Based on the above, this application provides a resource scheduling method for edge nodes of a power Internet of Things (IoT) based on artificial intelligence. The method involves: responding to a resource scheduling instruction, obtaining a node identifier from the instruction; determining a first target edge node among the edge nodes in the IoT based on the node identifier; if the computational resource utilization rate of the first target edge node exceeds a preset resource rate threshold, and a task unloading event occurs at a second target edge node within a preset time period, then obtaining the operating parameters of the first target edge node; the second target edge node being an adjacent node of the first target edge node; processing the operating parameters of the first target edge node using a pre-trained prediction model to obtain its resource demand index; inputting the resource demand index, load balancing offset, and time-sensitive task weight of the first target edge node into a source scheduling strategy executor to obtain a resource scheduling decision matrix; and performing resource scheduling operations on the first target edge node based on the resource scheduling decision matrix. Applying the method provided in this application allows for proactive prediction-based resource scheduling, which can meet the dynamically changing load demands in edge computing scenarios. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0043] Figure 1 A flowchart illustrating a resource scheduling method for edge nodes of the power Internet of Things based on artificial intelligence, provided in this application;

[0044] Figure 2 A flowchart illustrating the process of obtaining the resource demand index of a first target edge node, provided in this application;

[0045] Figure 3 A flowchart of a process for obtaining a resource scheduling decision matrix is ​​provided in this application;

[0046] Figure 4 This application provides a schematic diagram of the structure of a resource scheduling system for an edge node of the power Internet of Things based on artificial intelligence. Detailed Implementation

[0047] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0048] In this application, the terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0049] This invention provides a resource scheduling method for edge nodes of the power Internet of Things based on artificial intelligence, applied to electronic devices. The flowchart of the method is shown below. Figure 1 As shown, it specifically includes:

[0050] S101: In response to a resource scheduling instruction, obtain the node identifier in the resource scheduling instruction.

[0051] In this embodiment, the resource scheduling instruction can be triggered when the computing resource utilization rate of the edge node continuously exceeds a first threshold (e.g., ≥75%) for a set duration (e.g., 30 seconds), or when the power quality disturbance coefficient is detected to exceed a safety threshold (e.g., ≥0.1). Alternatively, it can be a resource scheduling instruction from the power grid dispatch center, cloud management platform, or from adjacent nodes.

[0052] Optionally, the node identifier may include a geolocation code, a device type code, and a power grid level identifier. This can be obtained by parsing specific fields in the command message or decoding embedded information in the digital signature.

[0053] S102: Determine the first target edge node among the various edge nodes in the power Internet of Things based on the node identifier.

[0054] In this embodiment, the node identifier is a triplet data structure containing a geographic location code, a device type code, and a power grid level identifier. The geographic location code adopts a composite coding rule of latitude and longitude grid coding and altitude under the WGS-84 coordinate system. The device type code follows a preset logical node classification coding system. The power grid level identifier is a bitmask structure containing a voltage level identifier, a topology level code, and a management domain identifier.

[0055] Optionally, the process of determining the first target edge node based on the node identifier may include:

[0056] A hierarchical search is implemented using the topology graph of the power Internet of Things (IoT). First, spherical distance calculations are performed in the spatial index database based on geographic location codes to filter out a set of candidate nodes within a preset radius (preferably 500 meters). Then, based on equipment type codes, equipment capability matching is performed in the candidate set, and a Bloom filter is used to exclude incompatible equipment types. Based on the bitmask parsing results of the power grid level identifier, it is verified whether the authorization level of the current dispatch command covers the management domain identifier of the target node, triggering a dynamic token verification mechanism for cross-level dispatch requests. When multiple candidate nodes exist, the node that satisfies ΔQ < 0.05 (power quality disturbance coefficient deviation) and has the smallest resource utilization fluctuation variance in the last 30 seconds is selected as the first target edge node. This determination process ensures the real-time status of node states through a distributed consensus algorithm and generates a target node descriptor containing node feature vectors and credibility weights.

[0057] S103: If the computing resource utilization rate of the first target edge node exceeds the preset resource rate threshold, and the second target edge node experiences a task unloading event within a preset time period, then the running parameters of the first target edge node are obtained; the second target edge node is an adjacent node of the first target edge node.

[0058] In this embodiment, the preset time period adopts a dynamic time window adjustment mechanism, and its calculation formula is: T_win=α×(1+e^(-β×N_event)), where α is the base time unit, for example, 5s, β is the event decay coefficient, and N_event is the task offload event count of the second target edge node in the most recent 60 seconds. The detection of the task offload event is implemented through a distributed event bus, specifically including: listening to AVRO format event messages with the "TaskOffload" identifier in the Kafka message queue published by adjacent nodes, and triggering event counting when the resource release amount in the message exceeds the preset event threshold (recommended ≥15% of node computing power).

[0059] In one embodiment provided in this application, based on the above-described scheme, optionally, obtaining the operating parameters of the first target edge node in S103 includes:

[0060] The edge sensing module obtains at least one of the following operating parameters for the first target edge node: power quality disturbance coefficient, temperature distribution image, channel electromagnetic interference entropy, and dynamic load mutation rate; the power quality disturbance coefficient is a composite quantitative index of the transient disturbance intensity and steady-state distortion degree of the power system; the channel electromagnetic interference entropy characterizes the signal transmission quality; and the dynamic load mutation rate characterizes the load change rate.

[0061] S104: Use a pre-trained prediction model to process the operating parameters of the first target edge node to obtain the resource demand index of the first target edge node.

[0062] In one embodiment provided in this application, based on the above scheme, optionally, the prediction model includes a convolutional layer, a gated recurrent unit, and an output layer. The process in S104, which uses a pre-trained prediction model to process the operating parameters of the first target edge node to obtain the resource demand index of the first target edge node, is as follows: Figure 2 As shown, it includes:

[0063] S201: Input the temperature distribution image into the convolutional layer of the prediction model to extract the temperature distribution features in the temperature distribution image through the convolutional layer.

[0064] In this embodiment, the convolutional layer adopts a multi-scale feature extraction architecture, which includes three sets of convolutional kernels arranged in parallel:

[0065] First convolutional group: 3×3 kernel size, 32 channels, ReLU activation, specifically designed to extract hotspot distribution features on the device surface;

[0066] Second convolutional group: 5×5 kernel size, 16 channels, SELU activation, capturing long-range temperature gradient features;

[0067] The third convolutional group: 1×1 kernel size, 8 channels, used for feature dimensionality reduction and channel attention mechanism fusion;

[0068] S202: Input the power quality disturbance coefficient, the channel electromagnetic interference entropy, and the dynamic load mutation rate into the gated loop unit of the prediction model, so as to obtain the time-series variation characteristics of the power quality disturbance coefficient, the channel electromagnetic interference entropy, and the dynamic load mutation rate through the gated loop unit.

[0069] In this embodiment, the gated recurrent unit is configured with a two-layer BiGRU structure with a hidden layer dimension of 64. A causal convolution preprocessing module is introduced during time series processing to extract multi-scale time-frequency features of power quality disturbance coefficients.

[0070] S203: Input the temperature distribution characteristics and the temporal variation characteristics into the output layer of the prediction model, so that the output layer outputs the resource demand index of the first target edge node based on the temperature distribution characteristics and the temporal variation characteristics.

[0071] In this embodiment, a learnable weight allocation mechanism is used when the output layer performs feature fusion, and its expression is:

[0072] RDI=σ(W_c⊙F_temp+W_gru⊙F_time+b)

[0073] Where σ is the Sigmoid activation function, ⊙ represents the Hadamard product, W_c and W_gru are trainable weight matrices, F_temp is the temperature distribution feature vector, and F_time is the time-series feature vector.

[0074] S105: Input the resource demand index, load balancing offset, and time-sensitive task weight of the first target edge node into the source scheduling policy executor to obtain the resource scheduling decision matrix.

[0075] In one embodiment provided in this application, based on the above scheme, optionally, the process of inputting the resource demand index, load balancing offset, and time-sensitive task weight of the first target edge node into the source scheduling policy executor to obtain the resource scheduling decision matrix is ​​as follows: Figure 3 As shown, it includes:

[0076] S301: Calculate the load balancing offset of the first target edge node based on the resource utilization of the first target edge node and the resource utilization of the second target edge node.

[0077] In this embodiment, the load difference between nodes can be calculated by dynamically evaluating the real-time resource utilization of the first target edge node and the adjacent second target edge node through a weighted difference algorithm.

[0078] The evaluation can be calculated using node utilization differences and load fluctuations. Node utilization difference involves proportionally comparing the current resource utilization of the first target node with the average utilization of its neighboring nodes; the difference is dynamically adjusted based on preset reinforcement learning weight coefficients. Load fluctuations are assessed by calculating the standard deviation of resource utilization for both nodes over the past ten minutes, measuring the impact of load stability on the system. When multiple neighboring nodes exist, an inversely proportional weighting strategy based on communication latency is used, prioritizing the load status of low-latency nodes.

[0079] S302: Determine the time-sensitive task weight of the first target edge node based on the priority of each task assigned to the first target edge node.

[0080] In this embodiment, weight levels are assigned based on task type, with fault diagnosis tasks receiving the highest weight, followed by real-time control tasks, and batch calculation tasks receiving the lowest weight. A time-sensitive enhancement factor is applied to tasks with deadline constraints, which decays exponentially over time to ensure that tasks nearing their deadlines receive higher resource allocation. The corrected weight values ​​of each task are proportionally integrated with their resource requirements to ultimately generate a time-sensitive weight index that reflects the urgency of all tasks globally.

[0081] S303: Input the resource demand index, load balancing offset, and time-sensitive task weight of the first target edge node into the source scheduling policy executor to obtain the resource scheduling decision matrix.

[0082] In this embodiment, an optimization model is established that includes node computing power limits, inter-task dependencies, and communication bandwidth constraints. The Pareto optimality algorithm is used to balance three main objectives: maximizing resource utilization, improving task completion rate, and reducing energy consumption. A two-dimensional decision matrix is ​​generated to describe the mapping relationship between tasks and resources. The row vectors of the matrix represent specific task allocation schemes, and the column vectors specify the allocation ratios of CPU, GPU, and storage resources, strictly adhering to the maximum resource carrying capacity threshold for each task. When the state update delay of an adjacent node exceeds a set threshold, the historical best decision cache is automatically activated, and conservative resource pre-allocation is implemented based on local resource characteristics to ensure the scheduling continuity of the system under abnormal operating conditions.

[0083] S106: Perform resource scheduling operation on the first target edge node according to the resource scheduling decision matrix.

[0084] In one embodiment provided in this application, based on the above-described scheme, optionally, the step of performing resource scheduling operation on the first target edge node according to the resource scheduling decision matrix includes:

[0085] The computing resources required for each task allocated to the first target edge node are determined based on the resource scheduling decision matrix.

[0086] The computing resources required for each task are allocated to each task.

[0087] In this embodiment, based on the correspondence between task identifiers and resource codes in the resource scheduling decision matrix, the types and quantitative indicators of computing resources required for each task are analyzed. The types of computing resources include, but are not limited to: CPU core utilization, GPU memory allocation, storage space occupancy, and communication bandwidth reservation ratio. For time-sensitive tasks, a reserved resource incremental overlay mechanism is adopted, adding a dynamic buffer margin (preferably 15%-25% of the base value) to the basic resource allocation value. For batch computing tasks, a resource time-sharing reuse strategy is implemented, dynamically releasing inactive resources according to the task execution stage.

[0088] In this embodiment, a hierarchical resource injection operation is performed based on the parsing results, including: allocating exclusive computing units to high-priority tasks to ensure that their CPU / GPU resource usage is not interfered with by other tasks; establishing a dynamic resource pool lending mechanism for real-time control tasks with fluctuating requirements, allowing temporary access to idle resources of adjacent nodes within a preset safety threshold; and verifying the effectiveness of resource allocation through the virtualization layer monitoring module to ensure that the resource quotas of each task meet hardware-level isolation constraints.

[0089] Optionally, the following metrics are continuously monitored during task execution: the deviation rate between actual resource utilization and allocated values, and the compliance of task progress lag with timing constraints; when the deviation rate exceeds the fault tolerance threshold (typically 10%) or the timing compliance is lower than the preset standard (preferably 95%), GPU acceleration resource reallocation is enabled for computationally intensive lagging tasks; for communication-constrained tasks, the bandwidth allocation weight is dynamically increased and the data compression algorithm is optimized.

[0090] In one embodiment provided in this application, based on the above-described scheme, optionally, after performing resource scheduling operations on the first target edge node according to the resource scheduling decision matrix, the method further includes:

[0091] Obtain resource scheduling effect feedback data of the first target edge node;

[0092] The model parameters of the prediction model are updated using the resource scheduling effect feedback data.

[0093] In one embodiment provided in this application, based on the above-described scheme, optionally, obtaining the resource scheduling effect feedback data of the first target edge node includes:

[0094] Determine the resource utilization deviation between the actual resource utilization rate of the first target edge node and the preset target resource utilization rate;

[0095] The task unloading success rate of the first target edge node is determined based on the actual number of tasks successfully unloaded and the total number of tasks.

[0096] The resource utilization deviation and the task unloading success rate are used as feedback data on the resource scheduling effect of the first target edge node.

[0097] In this embodiment, a double-buffered incremental learning strategy is adopted when the update condition is triggered:

[0098] Create a mirror copy of the prediction model and inject feedback data to form an incremental training set;

[0099] Applying the Elastic Weight Consolidation (EWC) algorithm to constrain the shift of important parameters, the loss function is added as follows:

[0100] L_total=L_new+λΣF_i(θ_i-θ_old_i)^2

[0101] Where F_i is the diagonal element of the Fisher information matrix; the prediction error rate of the updated model in shadow mode is verified by A / B testing, and a hot switch is performed when the error improvement rate is ≥15%.

[0102] In one embodiment provided in this application, based on the above-described scheme, optionally, updating the model parameters of the prediction model using the resource scheduling effect feedback data includes:

[0103] If the resource utilization deviation is greater than a preset deviation threshold and the task unloading success rate is lower than a preset success rate threshold, the model parameters of the prediction model are updated using the resource scheduling effect feedback data.

[0104] By applying the method provided in the embodiments of this application, resource scheduling can be performed through proactive prediction, which can meet the dynamic and ever-changing load requirements in edge computing scenarios.

[0105] In one embodiment provided in this application, based on the above-described scheme, optionally, the training process of the prediction model includes:

[0106] Obtain a first training dataset and an initial prediction model to be trained; the first training dataset includes multiple first training samples and a resource demand index label for each first training sample; each first training sample includes the historical power quality disturbance coefficient, historical temperature distribution image, historical channel electromagnetic interference entropy, and historical dynamic load mutation rate of the edge node of the power Internet of Things.

[0107] The initial prediction model is trained using the first training dataset;

[0108] If the initial prediction model meets the training completion condition, the initial prediction model that meets the training completion condition is taken as the trained prediction model.

[0109] In this embodiment, the training completion condition includes a dual convergence criterion:

[0110] 1. The loss function (using Huber loss) shows a decrease of ≤1e-4 over 5 consecutive epochs;

[0111] 2. The prediction error rate (MAPE) of the resource demand index on the validation set is ≤8.5%;

[0112] During training, an adaptive moment estimation optimizer (AdamW) was used in conjunction with cosine annealing for learning rate scheduling. The initial learning rate was set to 3e-4, and the minimum learning rate was decayed to 1e-6. To prevent overfitting, DropPath regularization was applied after the GRU layer, with a dropout probability set to 0.2.

[0113] In one embodiment provided in this application, based on the above-described scheme, optionally, the construction process of the resource scheduling policy executor includes:

[0114] The learning environment information of the first target edge node and the second training dataset are obtained; the learning environment information includes a state space, an action space, and a reward function; the state space includes multiple state information of the first target edge node; the action space includes multiple resource scheduling operations; the second training dataset includes multiple second training samples; each second training sample includes historical state data of the edge node of the power Internet of Things, decision result annotations, and performance change data after executing historical resource scheduling operations.

[0115] Construct an initial decision tree based on the learning environment information;

[0116] The initial decision tree is trained using the second training dataset to obtain a resource scheduling policy executor.

[0117] In this embodiment, the reward function is designed as a multi-objective optimization function:

[0118] R=ω_1×(1-|U_t-U_t^target|)+ω_2×T_success+ω_3×E_saved

[0119] Where: U_t is the actual resource utilization rate, U_t^target is the target utilization rate (preferably 82%); T_success is the time-sensitive task completion rate; E_saved=Σ(P_base-P_alloc) / P_max, representing the energy efficiency improvement; ω_1,ω_2,ω_3 are adjustable weight coefficients (default values ​​0.5, 0.3, 0.2).

[0120] In this embodiment, the initial decision tree adopts the Gradient Boosting Decision Tree (GBDT) architecture, and resource scheduling domain knowledge constraints are introduced during feature splitting, including:

[0121] Cross-level resource allocation is prohibited when the weight of a time-sensitive task is greater than 0.7;

[0122] When the power quality disturbance factor is greater than 0.15, local redundant resource allocation will be enforced first.

[0123] During training, SHAP value analysis is used to rank feature importance, and the maximum depth of the decision tree (ranging from 8 to 15 layers) is dynamically adjusted to balance model complexity and generalization ability.

[0124] See Figure 4 This is a schematic diagram of the structure of a resource scheduling method system for edge nodes of the power Internet of Things based on artificial intelligence, provided in an embodiment of this application. The system includes:

[0125] The first acquisition unit 401 is used to acquire the node identifier in the resource scheduling instruction in response to the resource scheduling instruction;

[0126] Determining unit 402 is used to determine a first target edge node among the edge nodes in the power Internet of Things based on the node identifier;

[0127] The second acquisition unit 403 is used to acquire the running parameters of the first target edge node when the computing resource utilization rate of the first target edge node exceeds a preset resource rate threshold and a task unloading event occurs in the second target edge node within a preset time period; the second target edge node is an adjacent node of the first target edge node.

[0128] The processing unit 404 is used to process the operating parameters of the first target edge node using a pre-trained prediction model to obtain the resource demand index of the first target edge node.

[0129] The execution unit 405 is used to input the resource demand index, load balancing offset and time-sensitive task weight of the first target edge node into the source scheduling policy executor to obtain the resource scheduling decision matrix.

[0130] The scheduling unit 406 is used to perform resource scheduling operations on the first target edge node according to the resource scheduling decision matrix.

[0131] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0132] Finally, it should be noted that in this paper, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0133] For ease of description, the above system is described by dividing it into various functional units. Of course, in implementing this application, the functions of each unit can be implemented in one or more software and / or hardware.

[0134] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of this application.

[0135] The above provides a detailed description of a resource scheduling method for edge nodes of the power Internet of Things based on artificial intelligence. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A resource scheduling method for edge nodes of the power Internet of Things based on artificial intelligence, characterized in that, include: In response to a resource scheduling instruction, obtain the node identifier in the resource scheduling instruction; The first target edge node is determined among the various edge nodes in the power Internet of Things based on the node identifier; If the computing resource utilization rate of the first target edge node exceeds a preset resource rate threshold, and a task unloading event occurs at the second target edge node within a preset time period, then the running parameters of the first target edge node are obtained; the second target edge node is an adjacent node of the first target edge node. The operating parameters of the first target edge node are processed using a pre-trained prediction model to obtain the resource demand index of the first target edge node. The resource demand index, load balancing offset, and time-sensitive task weight of the first target edge node are input into the source scheduling policy executor to obtain the resource scheduling decision matrix. Perform resource scheduling operations on the first target edge node according to the resource scheduling decision matrix; The step of inputting the resource demand index, load balancing offset, and time-sensitive task weight of the first target edge node into the source scheduling policy executor to obtain the resource scheduling decision matrix includes: The load balancing offset of the first target edge node is calculated based on the resource utilization of the first target edge node and the resource utilization of the second target edge node. The time-sensitive task weight of the first target edge node is determined based on the priority of each task assigned to the first target edge node; The resource demand index, load balancing offset, and time-sensitive task weight of the first target edge node are input into the source scheduling policy executor to obtain the resource scheduling decision matrix.

2. The method according to claim 1, characterized in that, The step of obtaining the operating parameters of the first target edge node includes: The edge sensing module obtains at least one of the following operating parameters for the first target edge node: power quality disturbance coefficient, temperature distribution image, channel electromagnetic interference entropy, and dynamic load mutation rate; the power quality disturbance coefficient is a composite quantitative index of the transient disturbance intensity and steady-state distortion degree of the power system; the channel electromagnetic interference entropy characterizes the signal transmission quality; and the dynamic load mutation rate characterizes the load change rate.

3. The method according to claim 2, characterized in that, The prediction model includes convolutional layers, gated recurrent units, and an output layer. The step of processing the operating parameters of the first target edge node using the pre-trained prediction model to obtain the resource demand index of the first target edge node includes: The temperature distribution image is input into the convolutional layer of the prediction model to extract the temperature distribution features from the temperature distribution image through the convolutional layer. The power quality disturbance coefficient, the channel electromagnetic interference entropy, and the dynamic load mutation rate are input into the gated loop unit of the prediction model, so that the time-series variation characteristics of the power quality disturbance coefficient, the channel electromagnetic interference entropy, and the dynamic load mutation rate can be obtained through the gated loop unit. The temperature distribution characteristics and the temporal variation characteristics are input into the output layer of the prediction model, so that the output layer outputs the resource demand index of the first target edge node based on the temperature distribution characteristics and the temporal variation characteristics.

4. The method according to claim 2 or 3, characterized in that, The training process of the prediction model includes: Obtain a first training dataset and an initial prediction model to be trained; the first training dataset includes multiple first training samples and a resource demand index label for each first training sample; each first training sample includes the historical power quality disturbance coefficient, historical temperature distribution image, historical channel electromagnetic interference entropy, and historical dynamic load mutation rate of the edge node of the power Internet of Things. The initial prediction model is trained using the first training dataset; If the initial prediction model meets the training completion condition, the initial prediction model that meets the training completion condition is taken as the trained prediction model.

5. The method according to claim 1, characterized in that, The construction process of the resource scheduling policy executor includes: The learning environment information of the first target edge node and the second training dataset are obtained; the learning environment information includes a state space, an action space, and a reward function; the state space includes multiple state information of the first target edge node; the action space includes multiple resource scheduling operations; the second training dataset includes multiple second training samples; each second training sample includes historical state data of the edge node of the power Internet of Things, decision result annotations, and performance change data after executing historical resource scheduling operations. Construct an initial decision tree based on the learning environment information; The initial decision tree is trained using the second training dataset to obtain a resource scheduling policy executor.

6. The method according to claim 1, characterized in that, The step of performing resource scheduling operations on the first target edge node according to the resource scheduling decision matrix includes: The computing resources required for each task allocated to the first target edge node are determined based on the resource scheduling decision matrix. The computing resources required for each task are allocated to each task.

7. The method according to claim 1, characterized in that, After performing resource scheduling operations on the first target edge node according to the resource scheduling decision matrix, the method further includes: Obtain resource scheduling effect feedback data of the first target edge node; The model parameters of the prediction model are updated using the resource scheduling effect feedback data.

8. The method according to claim 7, characterized in that, The step of obtaining resource scheduling effect feedback data for the first target edge node includes: Determine the resource utilization deviation between the actual resource utilization rate of the first target edge node and the preset target resource utilization rate; The task unloading success rate of the first target edge node is determined based on the actual number of tasks successfully unloaded and the total number of tasks. The resource utilization deviation and the task unloading success rate are used as feedback data on the resource scheduling effect of the first target edge node.

9. The method according to claim 8, characterized in that, The step of updating the model parameters of the prediction model using the resource scheduling effect feedback data includes: If the resource utilization deviation is greater than a preset deviation threshold and the task unloading success rate is lower than a preset success rate threshold, the model parameters of the prediction model are updated using the resource scheduling effect feedback data.

Images