A power edge iot agent task allocation method

Abstract

The application provides a power edge Internet of Things agent task allocation method, which comprises the following steps: defining a task importance, a task set J={j}, a processor set P={p}, and a task allocation variable u j,p ; the task occupies the execution time of the processor, and the resource cannot exceed the load capacity of the processor; defining an environment history data set as epsilon=[e1, e2,..., e N′ ], state space modeling; and action space modeling. Through a DQN algorithm and a local correction process, the calculation efficiency is improved without reducing the performance.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power communication, and particularly relates to a power edge Internet of Things agent task allocation method. BACKGROUND

[0002] A large number of sensing devices gradually access the power Internet of Things, and the amount of data generated increases day by day, and the edge computing technology can better cope with the huge data processing demand of the power Internet of Things. Therefore, the power Internet of Things agent based on edge computing emerges as the times require. It refers to deploying a series of edge agent nodes with edge intelligent processing or servers with communication access, business logic processing, and data storage functions at the business terminal and cloud center of the power Internet of Things, and completing the processing or preprocessing of the tasks initiated by the business in a nearby manner. In the prior art, the edge computing technology is widely used in the construction of low-latency, lightweight, high-efficiency, and high-reliability power Internet of Things business.

[0003] The power Internet of Things has many applications that require low latency and high reliability, such as real-time data processing and optimization, real-time robot path planning, and real-time system monitoring. In view of the high latency and low robustness characteristics of cloud computing business, the task allocation in edge computing needs to be studied in the prior art, that is, how to place the tasks initiated by the business terminal to multiple edge nodes for completion, and optimize the target under the constraint conditions of business resources, latency, etc. SUMMARY

[0004] In view of the above problems, the present application is proposed in order to provide a power edge Internet of Things agent task allocation method which overcomes the above problems or at least partially solves the above problems.

[0005] According to one aspect of the present application, a power edge Internet of Things agent task allocation method is provided, which comprises:

[0006] Defining the importance of the task, the task set is J={j}, and the importance of the task j is I j =H(J;θ)-H(J\{j};θ\(θ j )), wherein θ j represents the model parameter of the task j, θ represents the parameter vector θ={θ j}, and the decision function H(·) outputs the final decision performance;

[0007] The processor set is P={p}, and the task allocation variable is defined as a binary variable u j,p ; u j,p =1 indicates that the task j is allocated to the processor p for execution, otherwise, u j,p =0; each task can only be executed on one processor,

[0008] The task occupies the execution time of the processor, and the resource cannot exceed the load capacity of the processor; represented as And Wherein, t j represents the execution time of the task j, T represents the total execution time limit, v j represents the resource required for executing the task j, V p represents the total amount of resources of the processor.

[0009] The benefit function is defined as the resource occupation considering the importance of the task, represented as

[0010] The environment history data set is defined as ε=[e1,e2,...,e N′ ], wherein e j represents the corresponding environment variable; the environment adopts a KNN classification algorithm to define the environment history data set, represented as e=kNN(ε,Z); wherein Z represents the perception data.

[0011] The environment is modeled in two dimensions, one dimension representing the object type, and the object type includes power load prediction, state analysis and fault positioning, and the other dimension representing the available processor.

[0012] The element in the environment matrix is a task, represented as Wherein, I j represents the importance of the task, V p represents the processor capacity.

[0013] The state space is modeled, and the state is defined as whether the current task is selected by the system, and the state space is represented as Wherein, s ij =1 represents that the current task is selected by the system, otherwise s ij =0.

[0014] The action space is modeled, and at each time point, the scheduler selects any subset of N×M tasks.

[0015] Optionally, the state space modeling further comprises:

[0016] The state space is defined as Π={π|π∈{1,2,...,M}}, π=j represents that the jth task is executed by the current processor at the current time.

[0017] Optionally, the allocation method further comprises:

[0018] The reward function is modeled, and when the edge Internet of Things agent node reaches the terminal state, the reward function r(t)=∑ j∈J I j , otherwise r(t)=0.

[0019] Optionally, the allocation method further comprises: the optimization process is defined as a DQN process, Q(s, p; 0, J), wherein 0 represents a neural network weight parameter variable, and a feature space X composed of an environment and an initial state is given;

[0020] The task allocation variable u is solved according to the DQN algorithm under the power edge Internet of Things agent environment, u <- F1(J, X) = F1(J, (e, s0)) = Q(s, p; 0, J);

[0021] The DQN process is modified into a cooperative collaborative learning model, and a local modification process F(J, X) = F(J, (C, R)) = w1F1(J, C) + w2F2(J, R) is added after the DQN process; wherein F1(J, C) represents the DQN process, and F2(J, R) represents the local modification process.

[0022] Optionally, the allocation method further comprises:

[0023] The DQN process feature engineering considers two factors:

[0024] The first factor is a historical accurate prediction data set, which refers to a past optimal decision set; if the industrial demand, configuration and environment do not change too much, the past selected task is likely to be selected again in this period;

[0025] The second factor is prediction accuracy, the similarity between the prediction performance and the actual performance of the past given task.

[0026] Optionally, the allocation method further comprises: calculating a local modification decision set; and calculating u <- F2(J, X) = SVM(X; w, J) according to the feature engineering by using an SVM algorithm.

[0027] The power edge Internet of Things agent task allocation method provided by the application comprises the following steps: defining task importance, task set J = {j}; processor set P = {p}, defining task allocation variable u as a binary variable u j,p ; the task occupies the execution time of the processor, and the resource cannot exceed the load capacity of the processor; defining an environment historical data set as e = [e1, e2,..., e N′ ], modeling a state space; and modeling an action space. Through the DQN algorithm and the local modification process, the calculation efficiency is improved without reducing the performance.

[0028] The above description is only a summary of the technical scheme of the application, in order to more clearly understand the technical means of the application, the application can be implemented according to the content of the specification, and in order to make the above and other purposes, characteristics and advantages of the application more obvious and easy to understand, the following specific embodiments of the application are described. BRIEF DESCRIPTION OF DRAWINGS

[0029] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed to be used in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort based on these drawings.

[0030] Figure 1 A flow chart of a power edge Internet of Things agent task allocation method provided by the embodiment of the present application is shown in FIG. 1.

[0031] Figure 2 A power edge Internet of Things agent system architecture diagram provided by the embodiment of the present application is shown in FIG. 2.

[0032] Figure 3 A processing time diagram of a task allocation system with different processor numbers provided by the embodiment of the present application is shown in FIG. 3.

[0033] Figure 4 A processing time diagram of a task allocation system with different data input sizes provided by the embodiment of the present application is shown in FIG. 4.

[0034] Figure 5 A processing time diagram of a task allocation system with different bandwidth limits provided by the embodiment of the present application is shown in FIG. 5. DETAILED DESCRIPTION

[0035] Exemplary embodiments of the present disclosure will be described more fully hereinafter with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it is to be understood that the present disclosure can be embodied in various forms without being limited by the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the present disclosure to those skilled in the art.

[0036] The terms "include" and "have" and any variations thereof in the specification and claims of the present application and the drawings are intended to cover the non-exclusive inclusion, for example, the inclusion of a series of steps or units.

[0037] The technical solutions of the present application will be described in further detail below with reference to the drawings and embodiments.

[0038] The application solves the task allocation problem of power edge Internet of Things agents by introducing task importance and data-driven ideas. First, the task importance is defined, and the benefit function based on the task importance is maximized under the constraints of resources and execution time to establish the power edge Internet of Things agent task allocation optimization problem. The application proposes a data-driven collaborative task allocation method, which improves the computational efficiency without reducing the performance through the DQN algorithm and the local correction process. The simulation proves that the processing time of the method is better than that of the prior art, and the method provides an effective and practical mechanism for reducing the resources required for executing complex operations on edge devices.

[0039] As shown in Figure 1 and Figure 2 , the specific steps of the application scheme include the following.

[0040] Step one: define the task importance. Assuming that the task set is J={j}, the importance of task j is I j =H(J; θ)-H(J\{j}; θ\(θ j )), wherein θ j represents the model parameters of task j, and θ represents the parameter vector θ={θ j}. The decision function H(·) outputs the final decision performance.

[0041] Step two: assuming that the processor set is P={p}. Define the task allocation variable as a binary variable u j,p . u j,p =1 indicates that task j is allocated to processor p for execution, otherwise, u j,p =0. Each task can only be executed on one processor, that is In addition, the execution time and resources occupied by all tasks on the processor cannot exceed the load capacity of the processor.

[0042] That is and , wherein t j represents the execution time of task j, T represents the total execution time limit, v j represents the resources required for executing task j, and V p represents the total amount of resources of the processor.

[0043] The benefit function of the application is defined as the resource occupancy considering the task importance, that is The optimization problem to be solved by the application is to maximize the benefit function under the constraints of resources and execution time, and finally to obtain the task allocation decision variable.

[0044] Step three: define the environment history data set as ε=[e1, e2,..., e N′ ]. Among them, e jwhere represents the corresponding environmental variable.

[0045] The environment adopts KNN classification algorithm to define the environmental history dataset, e = kNN (ε, Z), where Z represents the perception data. The environment modeling of the present application is in two dimensions, one dimension represents the object type (power load forecasting, state analysis, fault location, etc.), and the other dimension represents the available processors. The elements in the environment matrix are regarded as the tasks in the present application, where I j represents the task importance, V p represents the processor capacity.

[0046] Step four: state space modeling. The state in the present application is defined as whether the current task is selected by the system, and the state space is represented as where s ij = 1 represents that the current task is selected by the system, otherwise s ij = 0.

[0047] Step five: action space modeling. At each time point, the scheduler may select any subset of N x M tasks. A huge action space is needed, which leads to the impossibility of execution on the edge agent device. In order to solve this problem, the present application only allows the edge agent device to perform only one action at each step. The state space is defined as Π = {π | π ∈ {1, 2,..., M}}. π = j represents that the jth task is executed by the current processor at the current time.

[0048] Step six: reward function modeling. The purpose of defining the reward function is to maximize the overall task importance when the edge Internet of Things agent node reaches the termination state, and all tasks in the current system have been allocated, the reward function r (t) = ∑ j∈J I j , otherwise r (t) = 0.

[0049] Step seven: the optimization process is defined as the DQN process, Q (s, π; θ, J). Where θ represents the neural network weight parameter variable. Given the feature space X composed of the environment and the initial state,

[0050] The present application solves the task allocation variable u according to the DQN algorithm under the power edge Internet of Things agent environment, u ← F1 (J, X) = F1 (J, (e, s0)) = Q (s, π; θ, J).

[0051] Step eight: the individual DQN process will cause too much difference from the real environment. Therefore, step seven is revised to a cooperative collaborative learning model, and a local correction process F (J, X) = F (J, (C, R)) = ω1F1 (J, C) + ω2F2 (J, R) is added after the DQN process. Where F1 (J, C) represents the DQN process, and F2 (J, R) represents the local correction process.

[0052] Step nine: Local process feature engineering considers two factors: one is the historical accurate prediction dataset, which refers to the past optimal decision set. If the industrial demand, configuration and environment do not change too much, the past selected task is likely to be selected again in this period; the second is the prediction accuracy, that is, the similarity between the prediction performance and the actual performance of the given task in the past. It is also an important factor. Accurate prediction indicates that the final decision will be more reliable.

[0053] Step ten: Calculate the local correction decision set. Using the SVM algorithm, according to the feature engineering, calculate u←F2(J,X)=SVM(X;ω,J).

[0054] Beneficial effect: Through the DQN algorithm and the local correction process, the computing efficiency is improved without reducing the performance.

[0055] The above specific embodiments further illustrate the purpose, technical solutions and beneficial effects of the present application. It should be understood that the above is only a specific embodiment of the present application and is not used to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A method for task allocation of power edge IoT agent, characterized in that, The allocation method comprises: Definition of task importance, task set is J = {j}, the importance of task j is I j = H(J; θ) - H(J \ {j}; θ \ {θ j}) where θ j represents the model parameters of task j, θ represents the parameter vector θ = {θ j}, the decision function H(·) outputs the final decision performance; The set of processors is P = {p} and the task assignment variables are binary variables u j,p ; u j,p = 1 if task j is assigned to processor p for execution, otherwise, u j,p = 0; each task can only be executed on one processor, The task occupies the execution time of the processor, and the resource cannot exceed the load capacity of the processor; represented as And Wherein, t j The execution time of task j, T represents the total execution time limit, v j The resource required for executing task j, V p The total amount of resources of the processor; The benefit function is defined as the resource occupancy considering the importance of the task, represented as The environment history dataset is defined as ε = [e1, e2,..., e N ], wherein e j represents a corresponding environment variable; the environment adopts a KNN classification algorithm to define the environment history dataset, represented as e = kNN(ε, Z); wherein Z represents the perception data; The environment modeling is two-dimensional, one dimension representing object types including power load forecasting, state analysis and fault location, and the other dimension representing available processors; The elements in the environment matrix are tasks, denoted as e = [I j ×V p ] N×M , where I j represents the importance of the task, V p represents the total amount of resources of the processor, N represents that there are N environment variables in the environment historical data set; M is the maximum dimension of the state space; State space modeling, state defined as whether the current task is selected by the system, state space represented as S = [s ij ] N×M , where s ij = 1 indicates that the current task is selected by the system, otherwise s ij = 0; Action space modeling, at each time point, the scheduler selects any subset of N x M tasks. 2.The power edge-IoT agent task allocation method of claim 1, wherein, The state space modeling further comprises: The state space is defined as Π = {π | π ∈ {1, 2,..., M}}, and π = j represents that the jth task is executed by the current processor at the current time. 3.The power edge-IoT agent task allocation method of claim 1, wherein, The allocation method further comprises: The reward function is modeled as r(t) = ∑ j∈J I j when the edge IoT agent node reaches the terminal state, and r(t) = 0 otherwise. 4.The power edge-IoT agent task allocation method of claim 2, wherein, The allocation method further comprises: the optimization process is defined as a DQN process, Q(s, π; θ, J), wherein θ represents neural network weight parameter variables, and a feature space X composed of an environment and an initial state is given; Solving the task allocation variable u under the power edge Internet of Things agent environment according to the DQN algorithm u = F1(J, X) = F1(J, (e, s0)) = Q(s, π; θ, J); F1 represents a local correction process function; e represents an environment history data set, that is, an interaction data set stored in reinforcement learning; The DQN process is modified into a cooperative collaborative learning model, and a local correction process F(J, X) = F(J, (C, R)) = ω1F1(J, C) + ω2F2(J, R) is added after the DQN process; wherein F1(J, C) represents a local correction process, and F2(J, R) represents a DQN process.

5. The power edge IoT agent task allocation method according to claim 4, wherein, The allocation method further comprises: The DQN process feature engineering considers two factors: The first factor is a historical accurate prediction data set, which refers to the past optimal decision set; if the industrial demand, configuration and environment do not change much, the past selected tasks are likely to be selected again in this period; The second factor is prediction accuracy, the similarity between the prediction performance and the actual performance of the past given task.

Images