Electric power service feature sensing resource scheduling method and device for heterogeneous edge cloud system and computer equipment

Abstract

The invention relates to a heterogeneous edge cloud system-oriented power service feature sensing resource scheduling method and device and computer equipment. The method comprises the steps that a target function is constructed with the minimum total energy consumption of an electric power service computing network in an electric power computing task as a target, and constraint conditions of the target function comprise a task unloading decision constraint condition, a task inseparable constraint condition, a transmitting power constraint condition of mobile equipment and a task time delay constraint condition; constructing a first loss function of the student network according to the target function and the constraint condition, and training the student network based on the first loss function and the teacher network to obtain a resource scheduling model; and inputting the power service characteristics into a resource scheduling model to obtain an optimal task unloading decision and optimal transmitting power output by the resource scheduling model. By adopting the method, accurate and efficient resource scheduling in a heterogeneous edge cloud environment can be realized.

Description

Technical Field

[0001] The present application relates to the technical field of power systems, and in particular to a method, apparatus, and computer equipment for scheduling power service feature-aware resources for heterogeneous edge cloud systems. Background Art

[0002] With the intelligent and digital transformation of power systems, the demand for power services is becoming increasingly complex and diverse. With the large-scale integration of smart grids, distributed energy resources, and renewable energy, power systems face greater dynamism and complexity. This is particularly true in areas such as load scheduling, real-time monitoring, intelligent forecasting, and optimized control. Achieving efficient and accurate resource scheduling and management has become a key technical challenge in power systems.

[0003] Early power systems relied on centralized control and management, performing large-scale computing and data processing in remote data centers. However, with the increasing variety of power equipment and more frequent changes in load demand, the limitations of traditional centralized computing models in terms of real-time performance, reliability, and flexibility have gradually become apparent.

[0004] To address these issues, edge computing, a new computing architecture, has been gradually adopted. By migrating computing tasks from centralized data centers to edge cloud servers close to the data source, edge computing can significantly reduce data transmission latency, alleviate the computing burden on central nodes, and improve response speed.

[0005] However, the diverse range of mobile devices and edge cloud servers involved in power systems, most of which come from different manufacturers, present significant heterogeneity, increasing the complexity of resource scheduling. Therefore, how to accurately and efficiently schedule resources in a heterogeneous edge cloud environment has become an urgent problem that needs to be solved. Summary of the Invention

[0006] Based on this, it is necessary to provide a power service feature-aware resource scheduling method, device and computer equipment for heterogeneous edge cloud systems that can achieve accurate and efficient resource scheduling in heterogeneous edge cloud environments to address the above technical problems.

[0007] In the first aspect, the present application provides a power service feature-aware resource scheduling method for a heterogeneous edge cloud system, the method comprising: constructing an objective function with the goal of minimizing the total energy consumption of the power service computing network in the power computing task, and the constraints of the objective function include task offloading decision constraints, task indivisibility constraints, mobile device transmission power constraints and task delay constraints; wherein, the power service computing network includes multiple mobile devices and a heterogeneous edge cloud system; constructing a first loss function of the student network according to the objective function and the constraints, and training the student network based on the first loss function and the teacher network to obtain a resource scheduling model; inputting the power service features into the resource scheduling model to obtain the optimal task offloading decision and optimal transmission power output by the resource scheduling model, and scheduling the resources in the power service computing network based on the optimal task offloading decision and the optimal transmission power.

[0008] In one embodiment, an objective function is constructed with the goal of minimizing the total energy consumption of the power service computing network in the power computing task, including: constructing an objective function with the goal of minimizing the sum of the energy consumed by the power service computing network in the power computing task, the energy consumed by the local execution of the task by each mobile device, and the energy consumed by the execution of the task by each edge cloud server in the heterogeneous edge cloud system.

[0009] In one embodiment, a first loss function of the student network is constructed according to the objective function and the constraint conditions, including: constructing a first unloading decision function and a first resource allocation function with the channel parameters and the first model parameters of the student network as variables; the first unloading decision function is the unloading decision function of the student network, and the first resource allocation function is the resource allocation function of the student network; the first loss function is constructed according to the first unloading decision function, the first resource allocation function, the objective function, the constraint conditions and the balance factor.

[0010] In one embodiment, the method further includes: constructing a first activation layer function based on the first offloading decision function and the task offloading decision constraint; the first activation layer function is the offloading decision activation layer function of the student network; constructing a second activation layer function based on the first resource allocation function and the transmission power constraint; the second activation layer function is the resource allocation activation layer function of the student network; and using the first activation layer function and the second activation layer function as the output layer of the student network.

[0011] In one embodiment, the method also includes: constructing a second offloading decision function and a second resource allocation function with the channel parameters and the second model parameters of the teacher network as variables; the second offloading decision function is the offloading decision function of the teacher network, and the second resource allocation function is the resource allocation function of the teacher network; constructing a second loss function of the teacher network according to the second offloading decision function, the second resource allocation function, the objective function, the constraint conditions and the penalty proportional factor; the structure of the second loss function is the same as that of the first loss function; constructing a third activation layer function based on the second offloading decision function; the structure of the third activation layer function is different from that of the first activation layer function; the third activation layer function is the offloading decision activation layer function of the teacher network; constructing a fourth activation layer function based on the second resource allocation function and the transmission power constraint conditions; the structure of the fourth activation layer function is the same as that of the second activation layer function; the fourth activation layer function is the resource allocation activation layer function of the teacher network; using the third activation layer function and the fourth activation layer function as the output layer of the teacher network.

[0012] In one embodiment, the student network is trained based on the first loss function and the teacher network to obtain a resource scheduling model, including: inputting sample data into the student network and the teacher network respectively to obtain the output of the student network and the output of the teacher network; updating the second model parameters of the teacher network based on the output of the teacher network and the second loss function; obtaining the gradient loss based on the output of the teacher network and the output of the student network; and updating the first model parameters of the student network based on the gradient loss, the output of the student network and the first loss function; updating the balance factor in the first loss function and the penalty proportional factor in the second loss function; and cyclically inputting sample data into the student network and the teacher network respectively to obtain the output of the student network and the output of the teacher network until the training conditions are met to obtain the resource scheduling model.

[0013] In a second aspect, the present application further provides a power service feature-aware resource scheduling device for a heterogeneous edge cloud system, the device comprising:

[0014] An objective function construction module is used to construct an objective function with the goal of minimizing the total energy consumption of the power service computing network in the power computing task, and the constraints of the objective function include task offloading decision constraints, task indivisibility constraints, mobile device transmission power constraints, and task delay constraints; wherein the power service computing network includes multiple mobile devices and a heterogeneous edge cloud system;

[0015] A loss function construction module is used to construct the first loss function of the student network according to the objective function and the constraints;

[0016] A training module, configured to train the student network based on the first loss function and the teacher network to obtain a resource scheduling model;

[0017] The scheduling module is used to input the power service characteristics into the resource scheduling model, obtain the optimal task offloading decision and optimal transmission power output by the resource scheduling model; and schedule the resources in the power service computing network based on the optimal task offloading decision and optimal transmission power.

[0018] In one embodiment, the objective function construction module is specifically used to construct an objective function with the goal of minimizing the sum of the energy consumed by the offloading transmission of each mobile device task in the power service computing network, the energy consumed by each mobile device in locally executing the task, and the energy consumed by each edge cloud server in the heterogeneous edge cloud system in executing the task.

[0019] In one embodiment, the loss function construction module is specifically used to construct a first unloading decision function and a first resource allocation function with the channel parameters and the first model parameters of the student network as variables; the first unloading decision function is the unloading decision function of the student network, and the first resource allocation function is the resource allocation function of the student network; the first loss function is constructed according to the first unloading decision function, the first resource allocation function, the objective function, the constraint conditions and the balance factor.

[0020] In one embodiment, the device also includes an activation layer function construction module, which is used to construct a first activation layer function based on the first offloading decision function and the task offloading decision constraint condition; the first activation layer function is the offloading decision activation layer function of the student network; the second activation layer function is constructed based on the first resource allocation function and the transmission power constraint condition; the second activation layer function is the resource allocation activation layer function of the student network; and the first activation layer function and the second activation layer function are used as the output layer of the student network.

[0021] In one embodiment, the activation layer function construction module is also used to construct a second unloading decision function and a second resource allocation function with the channel parameters and the second model parameters of the teacher network as variables; the second unloading decision function is the unloading decision function of the teacher network, and the second resource allocation function is the resource allocation function of the teacher network; the second loss function of the teacher network is constructed according to the second unloading decision function, the second resource allocation function, the objective function, the constraint conditions and the penalty proportional factor; the structure of the second loss function is the same as that of the first loss function; a third activation layer function is constructed based on the second unloading decision function; the structure of the third activation layer function is different from that of the first activation layer function; the third activation layer function is the unloading decision activation layer function of the teacher network; a fourth activation layer function is constructed based on the second resource allocation function and the transmission power constraint conditions; the structure of the fourth activation layer function is the same as that of the second activation layer function; the fourth activation layer function is the resource allocation activation layer function of the teacher network; the third activation layer function and the fourth activation layer function are used as the output layer of the teacher network.

[0022] In one embodiment, the training module is specifically used to input sample data into the student network and the teacher network respectively to obtain the output of the student network and the output of the teacher network; update the second model parameters of the teacher network based on the output of the teacher network and the second loss function; obtain the gradient loss based on the output of the teacher network and the output of the student network; and update the first model parameters of the student network based on the gradient loss, the output of the student network and the first loss function; update the balance factor in the first loss function and the penalty scale factor in the second loss function; and cyclically input the sample data into the student network and the teacher network respectively to obtain the output of the student network and the output of the teacher network, until the training conditions are met, so as to obtain a resource scheduling model.

[0023] In a third aspect, the present application further provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any one of the methods described in the first aspect when executing the computer program.

[0024] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any one of the methods described in the first aspect.

[0025] In a fifth aspect, the present application further provides a computer program product, comprising a computer program, which, when executed by a processor, implements the steps of any one of the methods described in the first aspect above.

[0026] The aforementioned power service feature-aware resource scheduling method, apparatus, and computer device for heterogeneous edge cloud systems construct an objective function based on minimizing the total energy consumption of a power service computing network in power computing tasks. The objective function's constraints include task offloading decision constraints, task indivisibility constraints, mobile device transmit power constraints, and task latency constraints. The power service computing network includes multiple mobile devices and a heterogeneous edge cloud system. A first loss function for a student network is constructed based on the objective function and constraints. The student network is trained based on the first loss function and the teacher network to obtain a resource scheduling model. The power service features are then input into the resource scheduling model to obtain the optimal task offloading decision and optimal transmit power outputs from the resource scheduling model. Resources in the power service computing network are then scheduled based on the optimal task offloading decision and optimal transmit power. In other words, the resource scheduling model obtained through training in this application realizes the solution of the objective function, wherein the resource scheduling model takes into account the task offloading decision constraints, transmission power constraints and task delay constraints during the training process, and the input data of the resource scheduling model is the power service characteristics. Based on this, the optimal task offloading decision and optimal transmission power obtained are obtained by analyzing the power service characteristics, so that the resource scheduling of this application can adapt to different power service characteristics, thereby realizing accurate and efficient resource scheduling in a heterogeneous edge cloud environment. It can be understood that by obtaining power service characteristics in real time and using the resource scheduling model for resource scheduling, dynamic perception and adaptive resource scheduling can be achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following briefly introduces the drawings required for use in the embodiments of the present application or related technical descriptions. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying any creative work.

[0028] Figure 1 1 is a flow chart of a method for power service feature-aware resource scheduling for a heterogeneous edge cloud system in one embodiment;

[0029] Figure 2 A schematic diagram of the architecture of a power service computing network in one embodiment;

[0030] Figure 3 A schematic diagram of a training principle in one embodiment;

[0031] Figure 4 1 is a flow chart of a training method in one embodiment;

[0032] Figure 5A structural block diagram of a power service feature-aware resource scheduling device for a heterogeneous edge cloud system in one embodiment;

[0033] Figure 6 FIG. 1 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0034] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0035] With the intelligent and digital transformation of power systems, the demand for power services is becoming increasingly complex and diverse. With the large-scale integration of smart grids, distributed energy resources, and renewable energy, power systems face greater dynamism and complexity. This is particularly true in areas such as load scheduling, real-time monitoring, intelligent forecasting, and optimized control. Achieving efficient and accurate resource scheduling and management has become a key technical challenge in power systems.

[0036] Early power systems relied on centralized control and management, performing large-scale computing and data processing in remote data centers. However, with the increasing variety of power equipment and more frequent changes in load demand, the limitations of traditional centralized computing models in terms of real-time performance, reliability, and flexibility have gradually become apparent. For example, in specific power application scenarios such as distributed power generation, smart meter data processing, and power load forecasting, the traditional architecture's insufficient computing power, high communication latency, and heavy data transmission burden have become particularly prominent. As the scale of power systems continues to expand, the limitations of traditional scheduling models have become more apparent, and they can no longer effectively meet the diverse and real-time requirements of power services.

[0037] To address these issues, edge computing, a new computing architecture, is gaining increasing adoption. By migrating computing tasks from centralized data centers to edge cloud servers close to the data source, edge computing can significantly reduce data transmission latency, alleviate the computational burden on central nodes, and improve response speed. In power systems in particular, edge computing not only improves real-time data processing capabilities but also enables more flexible resource allocation and task scheduling.

[0038] Currently, much research focuses on building edge computing frameworks and optimizing basic scheduling strategies. However, most traditional resource scheduling methods rely on static models or simple load balancing strategies, failing to account for the diverse characteristics and real-time changes of power services. Power services (such as generation, distribution, and load scheduling) have varying time sensitivities, computing requirements, and service quality requirements. Traditional scheduling strategies are unable to fully perceive and adapt to these service characteristics.

[0039] In addition, the power system involves a wide variety of mobile devices and edge cloud servers, most of which come from different manufacturers and have significant heterogeneity. Traditional heterogeneous edge cloud scheduling methods typically use heuristic algorithms or convex optimization methods to solve the task offloading and resource allocation problems. For example, decomposition methods based on genetic algorithms, particle swarm optimization, or Lagrangian duality require high linearization and convexity assumptions. In practical applications, heterogeneous edge cloud scheduling involves a variety of dynamic characteristics, such as the time-varying channel state, the heterogeneity of edge server computing power, and the multi-objective trade-off between task latency and energy consumption. These factors make it difficult for traditional heterogeneous edge cloud scheduling methods to quickly generate effective scheduling strategies. Especially in scenarios where multiple devices and multiple servers interact, the combination of task offloading and resource allocation further increases the difficulty of solving the problem. Therefore, traditional heterogeneous edge cloud scheduling methods have difficulty handling the nonlinear and complex characteristics of actual heterogeneous edge cloud environments. In addition, traditional heterogeneous edge cloud scheduling methods require high computational complexity, especially when the number of devices and task scale increase, the solution efficiency drops significantly, and there is a lack of robust handling of goal conflicts in multi-objective optimization problems, which makes it easy to fall into local optimal solutions.

[0040] In summary, in the context of dynamically changing power service demands and heterogeneous computing resources, how to perform accurate and efficient resource scheduling based on power service feature perception has become an urgent problem to be solved. Based on this, it is necessary to propose effective technical means to solve the above problems. The following specific embodiments describe in detail the technical solutions of this application and how the technical solutions of this application solve the above technical problems. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0041] In an exemplary embodiment, Figure 1 As shown, a method for power service feature-aware resource scheduling for a heterogeneous edge cloud system is provided, which is described by applying the method to a computer device, and includes the following steps 101 to 103. Among them:

[0042] Step 101, constructing an objective function with the goal of minimizing the total energy consumption of the power service computing network in the power computing task, and the constraints of the objective function include task offloading decision constraints, task indivisibility constraints, mobile device transmission power constraints and task delay constraints.

[0043] Among them, the architecture of the power service computing network is as follows: Figure 2As shown, the power service computing network includes mobile devices, heterogeneous edge cloud systems and base stations. It can be understood that there are multiple mobile devices and base stations, and the heterogeneous edge cloud system includes multiple edge cloud servers. Each edge cloud server comes from a different manufacturer and is heterogeneous.

[0044] In an optional implementation, an objective function is constructed with the goal of minimizing the sum of the energy consumed by the task offloading and transmission of each mobile device in the power computing task, the energy consumed by each mobile device in executing the task locally, and the energy consumed by each edge cloud server in the heterogeneous edge cloud system in executing the task.

[0045] Specifically, the power service computing network includes mobile devices and Edge cloud servers are configured, and each mobile device has the ability to perform computationally intensive tasks. Each edge cloud server represents a computing node from a different manufacturer. In resource scheduling, time is divided into time slots of fixed length, and each time slot is divided into three stages: the task processing decision stage, the task offloading and calculation stage, and the result feedback stage. In the task processing decision stage, the mobile device dynamically selects the task processing method based on characteristics such as channel state information, the computational amount of the task, and delay constraints. The task can be executed locally by the mobile device, or offloaded to a heterogeneous edge cloud system through a base station and executed by an edge cloud server. The key to this stage is the task processing method to maximize computing efficiency. In the task offloading and calculation stage, the mobile device uses orthogonal frequency division multiple access technology and power control mechanism to simultaneously offload the task to multiple edge cloud servers for execution. In the result feedback stage, after the edge cloud server completes the calculation, it transmits the processing result back to the mobile device through the communication link. In this application, the task is holistic and cannot be further split, that is, the task must be completely executed by the mobile device or completely offloaded to a certain edge cloud server. In order to achieve efficient task scheduling and power control, all base stations in the power service computing network have dynamic network status monitoring capabilities, which can perceive the channel status, task energy consumption and task data size of mobile devices in real time.

[0046] Mobile devices offload tasks to edge cloud servers via wireless links. The communication process is affected by factors such as channel status, transmission power, and bandwidth. Orthogonal frequency division multiple access is used to allocate spectrum resources. Mobile devices offload task data to edge cloud servers via uplinks. During this process, the upload rate r of mobile device n is n It can be expressed as:

[0047]

[0048] Where W represents bandwidth; p n represents the transmission power of mobile device n; h nrepresents the channel gain of mobile device n; σ represents the variance of the complex Gaussian channel white noise.

[0049] Therefore, the task offloading time t required for mobile device n to offload task data to the edge cloud server is n,m It can be expressed as:

[0050]

[0051] Among them, a n,m represents the task offloading decision, a n,m =0 means the task is executed locally on the mobile device, a n,m =1 means the task is executed on the edge cloud server; B n Indicates the amount of task data.

[0052] Assuming that the computing power of the edge cloud server is sufficient to support multi-task parallel processing and is not limited by the computational load of a single task. At the same time, since the amount of computational result data is smaller than the task data, the downlink delay and energy consumption are negligible compared to the uplink. Therefore, the energy consumed by task offloading transmission is E n,m It can be expressed as:

[0053] E n,m =p n t n,m (3)

[0054] when a n,m =0 When the task is executed locally on the mobile device, for mobile device n, the time t consumed by the mobile device to execute the task locally n It can be expressed as:

[0055]

[0056] Among them, C n represents the number of cycles per bit of data processed by the CPU of mobile device n; f n Indicates the CPU speed of mobile device n.

[0057] Therefore, the energy consumed by the mobile device to execute the task locally is n It can be expressed as:

[0058] E n =p nl t n (5) Among them, p nl represents the computing power of mobile device n.

[0059] when a n,m =1 When the task is executed on the edge cloud server, for edge cloud server m, the time t consumed by the edge cloud server to execute the task m It can be expressed as:

[0060]

[0061] Among them, C m represents the number of cycles per bit of data processed by the CPU of edge cloud server m; f m Indicates the CPU speed of edge cloud server m.

[0062] Therefore, the energy consumed by the edge cloud server to perform tasks is m It can be expressed as:

[0063] E m =p m t m (7) Among them, p m represents the computing power of edge cloud server m.

[0064] In other words, the energy consumed by the mobile device task offloading transmission is E n,m , which can be determined based on the transmission power of the mobile device and the task offloading time required for the mobile device to offload the task data to the edge cloud server; the energy consumed by the mobile device to perform the task locally E n , which can be determined based on the computing power of the mobile device and the time consumed by the mobile device to execute the task locally; the energy consumed by the edge cloud server to execute the task E m , which can be determined based on the computing power of the edge cloud server and the time consumed by the edge cloud server to execute the task.

[0065] Based on this, by jointly optimizing the task offloading decision and transmission power, the objective function constructed with the goal of minimizing the total energy consumption of the power service computing network in processing power tasks can be expressed as:

[0066]

[0067] Among them, four constraints ensure the feasibility and efficiency of resource scheduling. The task offloading decision constraint (C1) is used to ensure that each task can only choose one offloading target in each time slot, either completely executed locally or completely offloaded to an edge cloud server. The task indivisibility constraint (C2) is used to stipulate that the task cannot be split between multiple edge cloud servers, that is, each task is either processed locally by the mobile device or by a single edge cloud server. The transmission power constraint (C3) when the mobile device uploads tasks to the heterogeneous edge cloud system is used to ensure that the transmission power of each mobile device does not exceed its hardware and energy limitations, that is, it cannot exceed the maximum transmission power p of the mobile device. maxTo avoid equipment damage or excessive energy consumption due to excessive transmission. The task delay constraint (C4) is used to ensure that each task is completed within the specified time limit. That is, the sum of the task's upload delay, calculation delay, and feedback delay must meet the system's real-time requirements, that is, it cannot exceed the specified time T.

[0068] Step 102: construct a first loss function of the student network according to the objective function and the constraints, and train the student network based on the first loss function and the teacher network to obtain a resource scheduling model.

[0069] The student network can be a deep neural network (DNN) model, so the resource scheduling model is also a deep neural network model.

[0070] The objective function can be solved by a trained model to facilitate dynamic resource scheduling based on the dynamically changing characteristics of power services.

[0071] In one possible implementation, a first offloading decision function and a first resource allocation function are constructed with channel parameters and first model parameters as variables; the first offloading decision function is the offloading decision function of the student network, and the first resource allocation function is the resource allocation function of the student network; and a first loss function is constructed based on the first offloading decision function, the first resource allocation function, the objective function, the constraint conditions, and the balance factor.

[0072] Specifically, the channel parameter can be represented by h, which may include bandwidth and channel gain, etc. The first model parameter can be represented by θ. Therefore, the first offloading decision function and the first resource allocation function can be expressed as:

[0073]

[0074] In order to meet the task offloading decision constraints, the first activation layer function can be constructed based on the first offloading decision function and the task offloading decision constraints, and the first activation layer function is used as the output layer of the student network. The first activation layer function can be expressed as:

[0075]

[0076] Among them, m n =1-f n / C n B n , is the task offloading ratio; the symbolic function sign(·) is expressed as:

[0077]

[0078] In order to meet the transmit power constraint, a second activation layer function can be constructed based on the first resource allocation function and the transmit power constraint, and the second activation layer function is also used as the output layer of the student network. The second activation layer function can be expressed as:

[0079]

[0080] Based on this, the objective function can be transformed into:

[0081]

[0082] At the same time, in order to ensure that the task is completed within the specified time, that is, to meet the task delay constraint, the time constraint term P(θ) is introduced. When the task is completed within the specified time, the loss value is 0. Therefore, the objective function can be transformed into minimizing the first loss function of the student network. The first loss function can be expressed as:

[0083]

[0084]

[0085] in, Represents the expected value of the loss function for multi-channel sample data. The training process of the student network focuses on minimizing the system energy consumption of the mobile device while satisfying the constraints. ω is a balance factor used to balance the gap between different items. It should be noted that the balance factor ω needs to be carefully adjusted as a hyperparameter. If ω is too small, the first loss function will focus on the overall energy consumption of the system and output illegal constraints; if it is too large, it will focus on the constraint rules and ignore the system energy consumption. Therefore, the subgradient method is used to balance the hyperparameters. The update formula of the balance factor ω is formula (16):

[0086]

[0087] Here, α represents the step size.

[0088] However, when the back-propagation algorithm is used to update the first model parameters of the student network, the gradient will disappear and the first model parameters cannot be effectively updated. Therefore, this application designs a teacher network to provide effective gradient information to the student network to help the student network avoid the problem of gradient disappearance of the student network, thereby accelerating the convergence of the student network.

[0089] The teacher network and the student network have the same structure except for the activation layer that handles the task offloading decision. The construction process of the teacher network is as follows:

[0090] A second offloading decision function and a second resource allocation function are constructed using the channel parameters and the second model parameters as variables; a second loss function of the teacher network is constructed based on the second offloading decision function, the second resource allocation function, the objective function, the constraint conditions, and the penalty scaling factor; the structure of the second loss function is the same as that of the first loss function;

[0091] constructing a third activation layer function based on the second unloading decision function; the structure of the third activation layer function is different from that of the first activation layer function;

[0092] constructing a fourth activation layer function based on the second resource allocation function and the transmit power constraint condition; the structure of the fourth activation layer function is the same as that of the second activation layer function;

[0093] The third activation layer function and the fourth activation layer function are used as the output layer of the teacher network.

[0094] Specifically, the channel parameter is represented by h, which includes bandwidth and channel gain, etc. The second model parameter can be represented by θ teac Therefore, the second offloading decision function and the second resource allocation function can be expressed as:

[0095]

[0096] In the teacher network, the task offloading decision is converted into a continuous constraint, that is, the third activation layer function can be expressed as:

[0097]

[0098] The student network serves as the task offloading network model and shares the fully connected neural network layers before the activation layer with the teacher network. The teacher network transforms task offloading decisions into continuous constraints, so that during backpropagation, parameter updates in the teacher network also pass information to the student network.

[0099] The structure of the fourth activation layer function is the same as that of the second activation layer function. Therefore, the fourth activation layer function can be expressed as:

[0100]

[0101] The structure of the second loss function is the same as that of the first loss function. Therefore, the second loss function can be expressed as:

[0102]

[0103]

[0104] Where μ is the penalty scaling factor, which is used to balance the gap between unused items. It should be noted that the penalty scaling factor μ needs to be carefully adjusted as a hyperparameter. If μ is too small, the second loss function will focus on the overall energy consumption of the system and output illegal constraints; if it is too large, it will focus on the constraint rules and ignore the system energy consumption. Therefore, the subgradient method is used to balance the hyperparameters. The update formula of the penalty scaling factor μ is formula (22):

[0105]

[0106] The teacher network and the student network have the same structure except for the activation layer that handles task offloading decisions. This can effectively propagate and share information between the teacher network and the student network, and reduce the information loss caused by the structural mismatch between the teacher network and the student network, so that the student network can directly obtain lossless gradients.

[0107] The above introduces the model structure and loss function of the student network and the model structure and loss function of the teacher network. The following will introduce the specific training process. The training principle is as follows Figure 3 As shown, the training process is as follows Figure 4 As shown, the training steps are as follows:

[0108] Initialize the parameters of the student network and the teacher network; input sample data into the student network and the teacher network respectively to obtain the output of the student network and the output of the teacher network; update the second model parameters of the teacher network based on the output of the teacher network and the second loss function; obtain the gradient loss based on the output of the teacher network and the output of the student network; and update the first model parameters of the student network based on the gradient loss, the output of the student network and the first loss function; update the balance factor in the first loss function and the penalty scale factor in the second loss function; cyclically input the sample data into the student network and the teacher network respectively to obtain the output of the student network and the output of the teacher network, until the training conditions are met, so as to obtain the resource scheduling model.

[0109] The second model parameters of the teacher network are updated based on the output of the teacher network and the second loss function, that is, the second model parameters can be updated using the Adam optimizer method. The first model parameters of the student network are updated based on the gradient loss, the output of the student network and the first loss function, and the first model parameters can also be updated using the Adam optimizer method.

[0110] The training condition can be that the loss value of the student network reaches the preset loss value, or the number of iterations reaches the preset number. Figure 4 In the example, the training condition is to determine whether the number of iterations reaches the preset number.

[0111] The above training process is briefly as follows:

[0112] 1. Initialize parameters. Specifically, initialize the first model parameter θ of the student network and the second model parameter θ of the teacher network teac , balancing factor ω, penalty scaling factor μ, channel parameter h and step size α.

[0113] 2. Generate sample data. Sample data is also called channel sample data. The batch size for model training can be 1000, which means that 1000 channel sample data can be generated.

[0114] 3. Use the Adam optimizer to train the teacher network (i.e., use the Adam optimizer to minimize the loss function of formula (21)) and update the second model parameter θ teac .

[0115] 4. Use the Adam optimizer to train the student network (i.e., use the Adam optimizer to minimize the loss function of formula (15)) and update the first model parameter θ.

[0116] 5. Use the subgradient method (i.e., formula (16) and formula (22)) to update the balance factor ω and the penalty scaling factor μ respectively.

[0117] The training process described above is a joint training process. The teacher network, leveraging its extensive knowledge and experience, provides effective gradient information to the student network. This helps the student network avoid vanishing gradients, particularly during complex task offloading decisions, thereby accelerating model convergence. By sharing the task offloading decision output network, the teacher and student networks can be trained collaboratively within the same framework. The teacher network provides accurate decision guidance, while the student network independently learns and optimizes based on this information. The teacher network not only provides lossless gradient information in each round of training but also provides more accurate parameter update directions for the student network. Since the teacher network's parameters have typically been pre-trained to high accuracy over a long period of time, it provides a reliable guide for the student network's initial training, preventing the student network from getting stuck in local minima or non-convergence due to poor initial parameters or data noise. By introducing a joint training mechanism, the alternating training of the teacher and student networks further enhances the model's global optimization capabilities. By guiding the student network's training, the teacher network effectively avoids oscillations and local minima during the student network's optimization process, enabling the student network to search within a wider range of the solution space and ultimately achieve a better global optimal solution. In addition, the feature joint method can enhance the adaptability of the student network to different tasks and scenarios, improve its performance in actual task offloading decisions, and ensure the stability and efficiency of task scheduling in dynamic environments.

[0118] Step 103: Input the power service characteristics into the resource scheduling model to obtain the optimal task offloading decision and optimal transmission power output by the resource scheduling model, and schedule the resources in the power service computing network based on the optimal task offloading decision and optimal transmission power.

[0119] Among them, power service characteristics include bandwidth requirements, channel status, and task computation complexity.

[0120] Based on the trained resource scheduling model, the power service characteristics perceived by the base station are input into the resource scheduling model in real time. The resource scheduling model can output the optimal task offloading decision and the optimal transmission power in real time, so that the resources in the power service computing network can be scheduled in real time based on the optimal task offloading decision and the optimal transmission power. In turn, accurate and efficient resource scheduling based on the perception of power service characteristics can be achieved in a dynamically changing power service demand and heterogeneous computing resource environment.

[0121] The above-mentioned power service feature-aware resource scheduling method for heterogeneous edge cloud systems constructs an objective function based on minimizing the total energy consumption of the power service computing network in power computing tasks. The objective function's constraints include task offloading decision constraints, task indivisibility constraints, mobile device transmit power constraints, and task latency constraints. The power service computing network includes multiple mobile devices and a heterogeneous edge cloud system. A first loss function is then constructed for the student network based on the objective function and constraints. The student network is trained based on the first loss function and the teacher network to obtain a resource scheduling model. The power service features are then input into the resource scheduling model, which outputs the optimal task offloading decision and optimal transmit power. Resources in the power service computing network are then scheduled based on these optimal task offloading decisions and optimal transmit power. In other words, the resource scheduling model obtained through training in this application realizes the solution of the objective function, wherein the resource scheduling model takes into account the task offloading decision constraints, transmission power constraints and task delay constraints during the training process, and the input data of the resource scheduling model is the power service characteristics. Based on this, the optimal task offloading decision and optimal transmission power obtained are obtained by analyzing the power service characteristics, so that the resource scheduling of this application can adapt to different power service characteristics, thereby realizing accurate and efficient resource scheduling in a heterogeneous edge cloud environment. It can be understood that by obtaining power service characteristics in real time and using the resource scheduling model for resource scheduling, dynamic perception and adaptive resource scheduling can be achieved.

[0122] This application has the following advantages:

[0123] (1) An intelligent task offloading decision-making mechanism is realized through deep learning technology, which can automatically adjust the task processing method according to the characteristics of the power service. The power service computing network can perceive the task data volume, channel status and computing power of mobile devices in real time, thereby dynamically optimizing the task offloading decision. This method significantly improves the system's response speed and computing efficiency, and can adapt to the ever-changing service needs in the power system in real time, ensuring that tasks are processed on time and reducing delays.

[0124] (2) In response to the heterogeneity of computing nodes provided by different vendors in the edge cloud, a joint optimization algorithm based on DNN is proposed to efficiently coordinate and allocate heterogeneous resources. The system flexibly schedules tasks based on the characteristics of each edge cloud server and avoids computing bottlenecks caused by uneven resource distribution. This optimization scheme significantly improves resource utilization efficiency, avoids overloaded nodes and computing bottlenecks, and thus ensures efficient operation of the system.

[0125] (3) Orthogonal frequency division multiple access (OFDMA) technology is used to allocate spectrum resources, and combined with a power control mechanism to optimize the communication link during task offloading. Through this optimization, the system reduces communication overhead and scheduling delays, ensuring that tasks can be completed on time while reducing the network burden during task offloading.

[0126] (4) In response to the dynamic problems in the edge cloud environment, a joint training mechanism is introduced. This method uses deep neural networks to learn multi-dimensional input features such as channel state information, task characteristics, and device resource distribution, and dynamically generates optimal task offloading and resource allocation strategies. At the same time, through the design of activation layers, it ensures that the optimization results strictly meet the constraints, including task indivisibility, power constraints, and delay requirements. In addition, in response to the shortcomings of traditional methods in non-convex optimization problems, a joint optimization loss function is introduced to combine energy consumption optimization and delay constraints in a unified optimization goal. By setting a balance factor, a dynamic trade-off between multiple objectives is achieved. In network training, the subgradient method is combined with an adaptive learning rate strategy to solve the gradient vanishing problem in deep neural network training and improve the efficiency and convergence of parameter updates. By combining deep learning and optimization theory, this method not only overcomes the problem of insufficient adaptability of traditional methods to complex dynamic scenarios, but also significantly improves the efficiency of task scheduling and system performance.

[0127] It should be understood that, although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence according to the instructions of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above can include multiple steps or multiple stages, and these steps or stages are not necessarily executed and completed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0128] Based on the same inventive concept, an embodiment of the present application further provides a device for implementing the aforementioned method for implementing the power service feature-aware resource scheduling for heterogeneous edge cloud systems. The solution to the problem provided by this device is similar to the solution described in the aforementioned method. Therefore, the specific limitations of one or more embodiments of the device for implementing the power service feature-aware resource scheduling for heterogeneous edge cloud systems provided below can be found in the aforementioned limitations of the method for implementing the power service feature-aware resource scheduling for heterogeneous edge cloud systems, and will not be repeated here.

[0129] In an exemplary embodiment, Figure 5 As shown, a power service feature-aware resource scheduling device for a heterogeneous edge cloud system is provided. The power service feature-aware resource scheduling device 500 for a heterogeneous edge cloud system includes: an objective function construction module 501, a loss function construction module 502, a training module 503 and a scheduling module 504, wherein:

[0130] An objective function construction module 501 is configured to construct an objective function with the goal of minimizing the total energy consumption of the power service computing network in the power computing task, wherein the objective function has constraints including task offloading decision constraints, task indivisibility constraints, mobile device transmission power constraints, and task latency constraints; wherein the power service computing network includes multiple mobile devices and a heterogeneous edge cloud system;

[0131] A loss function construction module 502 is used to construct a first loss function of the student network according to the objective function and the constraint conditions;

[0132] A training module 503 is configured to train the student network based on the first loss function and the teacher network to obtain a resource scheduling model;

[0133] The scheduling module 504 is used to input the power service characteristics into the resource scheduling model, obtain the optimal task offloading decision and optimal transmission power output by the resource scheduling model; and schedule the resources in the power service computing network based on the optimal task offloading decision and optimal transmission power.

[0134] In one embodiment, the objective function construction module 501 is specifically used to construct an objective function with the goal of minimizing the sum of the energy consumed by the offloading transmission of each mobile device task in the power computing network, the energy consumed by each mobile device in locally executing the task, and the energy consumed by each edge cloud server in the heterogeneous edge cloud system in executing the task.

[0135] In one embodiment, the loss function construction module 502 is specifically used to construct a first offloading decision function and a first resource allocation function with channel parameters and first model parameters as variables; the first offloading decision function is the offloading decision function of the student network, and the first resource allocation function is the resource allocation function of the student network; the first loss function is constructed according to the first offloading decision function, the first resource allocation function, the objective function, the constraint conditions and the balance factor.

[0136] In one embodiment, the device also includes an activation layer function construction module, which is used to construct a first activation layer function based on the first offloading decision function and the task offloading decision constraint condition; the first activation layer function is the offloading decision activation layer function of the student network; the second activation layer function is constructed based on the first resource allocation function and the transmission power constraint condition; the second activation layer function is the resource allocation activation layer function of the student network; and the first activation layer function and the second activation layer function are used as the output layer of the student network.

[0137] In one embodiment, the activation layer function construction module is also used to construct a second offloading decision function and a second resource allocation function with channel parameters and second model parameters as variables; the second offloading decision function is the offloading decision function of the teacher network, and the second resource allocation function is the resource allocation function of the teacher network; the second loss function of the teacher network is constructed according to the second offloading decision function, the second resource allocation function, the objective function, the constraint conditions and the penalty proportional factor; the structure of the second loss function is the same as that of the first loss function; a third activation layer function is constructed based on the second offloading decision function; the structure of the third activation layer function is different from that of the first activation layer function; the third activation layer function is the offloading decision activation layer function of the teacher network; a fourth activation layer function is constructed based on the second resource allocation function and the transmission power constraint conditions; the structure of the fourth activation layer function is the same as that of the second activation layer function; the fourth activation layer function is the resource allocation activation layer function of the teacher network; the third activation layer function and the fourth activation layer function are used as the output layer of the teacher network.

[0138] In one embodiment, the training module 503 is specifically used to input sample data into the student network and the teacher network respectively to obtain the output of the student network and the output of the teacher network; update the second model parameters of the teacher network based on the output of the teacher network and the second loss function; obtain the gradient loss based on the output of the teacher network and the output of the student network; and update the first model parameters of the student network based on the gradient loss, the output of the student network and the first loss function; update the balance factor in the first loss function and the penalty proportional factor in the second loss function; and cyclically input the sample data into the student network and the teacher network respectively to obtain the output of the student network and the output of the teacher network, until the training conditions are met to obtain a resource scheduling model.

[0139] Each module in the aforementioned heterogeneous edge cloud system-oriented power service feature-aware resource scheduling device can be implemented in whole or in part through software, hardware, or a combination thereof. Each module can be embedded in or independent of a processor in a computer device in hardware form, or can be stored in a computer device memory in software form, so that the processor can call and execute the corresponding operations of each module.

[0140] In an exemplary embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as shown in FIG. Figure 6 As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O) and a communication interface. The processor, memory and input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database of the computer device is used to store sample data and loss values, etc. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a power service feature-aware resource scheduling method for a heterogeneous edge cloud system is implemented.

[0141] Those skilled in the art will understand that Figure 6 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0142] In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor implements the steps of any one of the above method embodiments when executing the computer program.

[0143] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of any one of the above method embodiments are implemented.

[0144] In one embodiment, a computer program product is provided, comprising a computer program, which, when executed by a processor, implements the steps of any one of the method embodiments described above.

[0145] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, database or other media used in the embodiments provided in this application may include at least one of non-volatile memory and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory may include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchains. The processor involved in the various embodiments provided herein may be, but are not limited to, a general-purpose processor, a central processing unit (CPU), a graphics processing unit (GPU), a digital signal processor (DSP), a programmable logic unit (PLC), a data processing logic unit based on quantum computing, an artificial intelligence (AI) processor, and the like.

[0146] The technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0147] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.

Claims

1. A power service feature-aware resource scheduling method for heterogeneous edge cloud systems, characterized in that: The method comprises: An objective function is constructed with the goal of minimizing the total energy consumption of the power service computing network in power computing tasks, and the constraints of the objective function include task offloading decision constraints, task indivisibility constraints, mobile device transmission power constraints, and task delay constraints; wherein the power service computing network includes multiple mobile devices and a heterogeneous edge cloud system; Constructing a first loss function of the student network according to the objective function and the constraint conditions, and training the student network based on the first loss function and the teacher network to obtain a resource scheduling model; The power service characteristics are input into the resource scheduling model to obtain the optimal task offloading decision and the optimal transmission power output by the resource scheduling model, and the resources in the power service computing network are scheduled based on the optimal task offloading decision and the optimal transmission power.

2. The method according to claim 1, characterized in that The first loss function of the student network is constructed according to the objective function and the constraint condition, comprising: A first offloading decision function and a first resource allocation function are constructed using the channel parameter and the first model parameter of the student network as variables; the first offloading decision function is the offloading decision function of the student network, and the first resource allocation function is the resource allocation function of the student network; The first loss function is constructed according to the first offloading decision function, the first resource allocation function, the objective function, the constraint condition and the balance factor.

3. The method according to claim 2, characterized in that The method further comprises: Constructing a first activation layer function based on the first offloading decision function and the task offloading decision constraint condition; the first activation layer function is the offloading decision activation layer function of the student network; Constructing a second activation layer function based on the first resource allocation function and the transmit power constraint condition; the second activation layer function is the resource allocation activation layer function of the student network; The first activation layer function and the second activation layer function are used as the output layer of the student network.

4. The method according to claim 3, characterized in that The method further comprises: Constructing a second offloading decision function and a second resource allocation function with the channel parameter and the second model parameter of the teacher network as variables; the second offloading decision function is the offloading decision function of the teacher network, and the second resource allocation function is the resource allocation function of the teacher network; A second loss function of the teacher network is constructed according to the second offloading decision function, the second resource allocation function, the objective function, the constraint condition, and the penalty scaling factor; the structure of the second loss function is the same as that of the first loss function; Constructing a third activation layer function based on the second offloading decision function; the structure of the third activation layer function is different from the structure of the first activation layer function; the third activation layer function is the offloading decision activation layer function of the teacher network; Constructing a fourth activation layer function based on the second resource allocation function and the transmit power constraint condition; the structure of the fourth activation layer function is the same as the structure of the second activation layer function; the fourth activation layer function is the resource allocation activation layer function of the teacher network; The third activation layer function and the fourth activation layer function are used as the output layer of the teacher network.

5. The method according to claim 4, characterized in that The step of training the student network based on the first loss function and the teacher network to obtain a resource scheduling model includes: Inputting sample data into the student network and the teacher network respectively to obtain the output of the student network and the output of the teacher network; Updating the second model parameters of the teacher network based on the output of the teacher network and the second loss function; Obtaining a gradient loss based on an output of the teacher network and an output of the student network; and updating the first model parameters of the student network based on the gradient loss, the output of the student network, and a first loss function; Updating the balance factor in the first loss function and the penalty scaling factor in the second loss function; The steps of inputting the sample data into the student network and the teacher network respectively to obtain the output of the student network and the output of the teacher network are looped until the training conditions are met to obtain the resource scheduling model.

6. The method according to claim 1, characterized in that The objective function is constructed with the goal of minimizing the total energy consumption of the power service computing network in the power computing task, including: An objective function is constructed with the goal of minimizing the sum of the energy consumed by the offloading transmission of each mobile device task, the energy consumed by each mobile device in locally executing tasks, and the energy consumed by each edge cloud server in the heterogeneous edge cloud system in executing tasks in the power service computing network.

7. A power service feature-aware resource scheduling device for heterogeneous edge cloud systems, characterized in that: The device comprises: An objective function construction module is configured to construct an objective function with the goal of minimizing the total energy consumption of the power service computing network in the power computing task, wherein the constraints of the objective function include a task offloading decision constraint, a task indivisibility constraint, a mobile device transmission power constraint, and a task delay constraint; wherein the power service computing network includes a plurality of the mobile devices and a heterogeneous edge cloud system; A loss function construction module, configured to construct a first loss function of the student network according to the objective function and the constraint conditions; A training module, configured to train the student network based on the first loss function and the teacher network to obtain a resource scheduling model; The scheduling module is used to input the power service characteristics into the resource scheduling model to obtain the optimal task offloading decision and the optimal transmission power output by the resource scheduling model; and schedule the resources in the power service computing network based on the optimal task offloading decision and the optimal transmission power.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images