Edge computing network-oriented energy efficiency service deployment and delivery method and system

Abstract

The invention provides an edge computing network-oriented energy efficiency service deployment and delivery method and system, and relates to the technical field of edge computing, and the method comprises the steps: extracting multi-dimensional service features, carrying out the coding fusion of the extracted multi-dimensional service features, and generating a global context feature representation; in combination with a system load dynamic prediction result, adaptively adjusting an upper layer decision time scale based on the global context feature representation; under the adjusted upper-layer time scale, executing a service deployment decision and a base station dormancy / activation switching decision; under the fine-grained lower-layer time scale, executing a service delivery decision and a resource allocation decision based on the real-time service request; and coordinating the decision process through a double-time-scale hierarchical learning framework. According to the method provided by the invention, the long-term network overhead can be effectively reduced, the dynamic service request can be effectively responded, the processing capability of the time coupling relationship between deployment and delivery weeks is improved, and the system cost is reduced.

Description

Technical Field

[0001] The present invention relates to the field of edge computing technology, and in particular to a method and system for deploying and delivering energy efficiency services for edge computing networks. Background Art

[0002] Mobile Edge Computing (MEC) significantly reduces service latency and improves user experience by moving computing and storage capabilities to edge nodes. However, existing energy-efficient service deployments still face three major challenges: First, multi-dimensional dependencies are difficult to capture. The widespread geographical distribution of edge nodes and the complexity of service semantics make traditional sequence models (such as RNN / LSTM) and local feature extraction methods (such as CNN) unable to effectively model global contextual dependencies.

[0003] Second, the deployment and delivery cycles are not decoupled. Existing methods ignore the inherent timescale differences between service deployment (long cycle) and service delivery (short cycle). Frequently switching deployment strategies can easily lead to high overhead, while static delivery strategies are difficult to respond to dynamic requests.

[0004] Third, the decision-making mechanism lacks adaptability. Current solutions rarely dynamically adjust the decision timescale based on service semantics and system load, making them inadequately adaptable in dynamic network environments.

[0005] Representative solutions such as random optimization, heuristic algorithms, and single-layer DRL have not solved the above problems and have many disadvantages such as high system cost, latency sensitivity, and low service response rate. Summary of the Invention

[0006] In order to solve at least some of the technical problems in the related art, the present invention provides an energy efficiency service deployment and delivery method and system for edge computing networks.

[0007] In order to achieve the above-mentioned purpose, the technical solution adopted by the present invention includes: according to the first aspect of the present invention, a method for energy-efficient service deployment and delivery for edge computing networks is provided, including: step S1: extracting multidimensional service features, wherein the multidimensional service features include geographic location information of the service, service semantic features and service popularity features; step S2: encoding and fusing the extracted multidimensional service features through a self-attention encoder to generate a global context feature representation; step S3: adaptively adjusting the upper-level decision time scale based on the global context feature representation in combination with the dynamic prediction results of the system load; step S4: executing service deployment decisions and base station sleep / activation switching decisions at the adjusted upper-level time scale; step S5: executing service delivery decisions and resource allocation decisions based on real-time service requests at a fine-grained lower-level time scale; step S6: coordinating the decision-making processes of steps S4 and S5 through a dual-time-scale hierarchical learning framework.

[0008] Optionally, the multi-dimensional service feature extraction specifically includes: Step S1-1: expressing the geographical location information of the service as: ,in, Indicates the The two-dimensional geographic coordinates of the edge servers, Indicates the The longitude of the edge server, Indicates the The latitude of the edge server, Indicates the total number of edge servers.

[0009] Step S1-2: Express the semantic features of the service as: Where, , represents the time variable, represents a discrete range of time, where is the total number of time periods considered by the system, Indicates the edge servers at a point in time Different service attributes under Step S1-3: Generate service popularity features according to modified Zipf distribution: Where, For the edge servers at a point in time Next service The flow ratio, For service In service Ranked in descending order of popularity, is the platform coefficient, is the skewness parameter, and Indicates the type of service, For the edge servers at a point in time Next service The traffic intensity, and meet the ,In the feature encoding stage, service popularity is expressed as: Where, is the traffic distribution set for all services, For the time point Next Traffic distribution of edge servers; Step S1-4: Concatenate the above two-dimensional geographic coordinates, semantic features, and service popularity features to obtain the input tensor: .

[0010] Optionally, the feature encoding process based on the self-attention mechanism includes: Step S2-1: Input tensor Generate query matrix by linear transformation , key matrix Sum Matrix ; Step S2-2: Calculate attention weighted features: Where, is the bond matrix The transposed matrix, dot product Used to measure the similarity between each pair of query and key, A factor that scales the dot product result to prevent the dot product value from being too large in high-dimensional data; Step S2-3: Output context feature vector through feedforward neural network : Where, 、 are the weight matrices of the two layers respectively, 、 are the corresponding bias vectors respectively.

[0011] Optionally, the adaptive adjustment of the upper layer decision time scale includes: Step S3-1: Constructing a system load sequence: , where For the The actual observed load of the system at the moment, is the actual observed load of the system at the current moment For the future The predicted load at the moment, For the future Forecasted load at the moment; Step S3-2: Fusion of temporal coding vectors Generate a load signature representation: Where, Indicates that the load data Mapped to a high-dimensional vector space through the embedding layer, is the embedding vector associated with the time position; Step S3-3: Extract load dynamic features through Transformer encoder : Where, Indicates the operation of pooling the input data. To represent the query vector in the load prediction task, is the representation of the key vector in the load prediction task, is the representation of the value vector in the load prediction task; Step S3-4: Concatenate context feature vectors Dynamic characteristics of load , input multi-layer perceptron to generate adaptive upper layer update interval : Where, and are the weight matrix and bias vector of the time scale decision module respectively.

[0012] Optionally, in step S4, the upper layer decision cycle is an adaptive interval, and the output action is a base station switching state.

[0013] Optionally, in step S5, the lower-layer decision cycle is a fixed fine-grained time slot, and the output action is a service delivery path and a resource allocation ratio.

[0014] Optionally, the upper network in step S6 adopts the Dueling architecture: Step S6-1: Separately model the state value function and the action advantage function: Where, represents the state value function, Indicates time State of the moment , represents the set of parameters of the state-value function, represents the advantage function of the action in the state, Indicates time The action selected by the agent at that moment , represents the parameter set of the action advantage function, is the parameter set of the upper network, is the advantage function; Step S6-2: Update network parameters by minimizing the timing difference error: Among them, the target Q value The calculation formula is: Where, is the loss function, is the expected value, For intelligent agents of function, Indicates time The state of the agent at each moment, Indicates time The action of the agent at each moment, For time The instant reward obtained by the agent at each moment, is the discount factor and , Indicates time The state of the agent at each moment, Indicates time The action of the agent at each moment, Represents the parameters of the target network.

[0015] Optionally, the lower layer decision in step S6 adopts a multi-agent deep deterministic policy gradient algorithm, where the policy network parameters and evaluate network parameters Update according to the following rules: 、 , where and are the learning rates of the policy network and the evaluation network, is the gradient of the policy network, is the gradient of the evaluation network.

[0016] Optionally, step S6 further includes a soft update mechanism, and the target network parameters are updated according to the following rules: 、 、 , where are the parameters of the target policy network, The parameters of the target evaluation network are, are the parameters of the target Critic network, They respectively represent the update rate of the upper target network, the target parameter update rate of the lower Actor network, and the target parameter update rate of the lower Critic network.

[0017] According to the second aspect of the present invention, an edge computing system is also provided for executing the energy efficiency service deployment and delivery method for edge computing networks described in any technical solution in the first aspect of the present invention, the edge computing system including a feature extraction module, a self-attention encoder module, a time scale control module, an upper-layer decision module, a lower-layer decision module and a hierarchical coordinator; the feature extraction module is configured to perform the extraction of multi-dimensional service features; the self-attention encoder module is configured to perform the encoding fusion of multi-dimensional service features; the time scale control module is configured to perform the adjustment of the adaptive upper-layer decision time scale; the upper-layer decision module is configured to perform service deployment decisions and base station sleep / activation switching decisions; the lower-layer decision module is configured to perform service delivery decisions and resource allocation decisions; the hierarchical coordinator is configured to perform dual-time-scale decision coordination.

[0018] Beneficial effects: 1. Through the above technical solution, first, the method of the present invention adopts the self-attention mechanism to replace the traditional RNN / CNN, and integrates geographic location, service semantics and popularity features in parallel, significantly improving the global dependency modeling capability and feature extraction efficiency, and meeting the real-time decision-making needs of edge computing.

[0019] Second, the method of the present invention introduces a dual time scale mechanism for the upper and lower layers. The upper layer dynamically adjusts the deployment strategy response cycle to reduce frequent switching overhead; the lower layer fine-grainedly controls service delivery to improve the response speed and service quality to sudden requests.

[0020] Third, the method of the present invention avoids resource mismatch, improves computing and storage utilization, accelerates strategy convergence, and effectively reduces the long-term operating cost of the system through a two-layer collaboration and hierarchical learning mechanism.

[0021] 2. Other beneficial effects or advantages of the present invention will be described in detail in the specific implementation manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without paying any creative labor.

[0023] Figure 1 1 is a flow chart of a method for deploying and delivering energy efficiency services for edge computing networks provided by an exemplary embodiment of the present invention; Figure 2 is a schematic diagram of deployment and delivery of dual-time-scale energy-efficient services in a mobile edge computing (MEC) network provided by an exemplary embodiment of the present invention; Figure 3 1 is a schematic diagram of context-aware adaptive time-scale control based on deep feature fusion provided by an exemplary embodiment of the present invention; Figure 4 is a schematic diagram of a dual-time-scale hierarchical learning process for energy efficiency service deployment and delivery provided by an exemplary embodiment of the present invention; Figure 5 a is a schematic diagram of rewards in the upper network learning process for energy efficiency service deployment under a single time scale provided by an exemplary embodiment of the present invention; Figure 5 b is a schematic diagram of changes in loss of an upper layer network provided by an exemplary embodiment of the present invention; Figure 6 a to 6d are network training results of hierarchical learning at dual time scales provided by an exemplary embodiment of the present invention, wherein, Figure 6 a is the upper level reward, Figure 6 b is the upper layer loss, Figure 6 c is the lower-level reward, Figure 6 d is the lower layer loss; Figure 7 a to Figure 7 d is a schematic diagram of the total system cost under different training rounds, the cumulative distribution function (CDF) of the total system cost, the total system cost under different computing capabilities, and the total system cost under different storage capacities, respectively, provided by an embodiment; Figure 8 a to Figure 8 d are respectively the service delay under different training rounds, the cumulative distribution function (CDF) of the service delay, the service delay under different computing capabilities, and the service delay under different storage capacities provided by an exemplary embodiment; Figure 9 a to 9d are respectively the service request response rates under different training rounds, the cumulative distribution function (CDF) of the service request response rate, the service request response rates under different computing capabilities, and the service request response rates under different storage capacities provided by an exemplary embodiment; Figure 10 a to Figure 10 d are respectively the computing resource utilization under different training rounds, the cumulative distribution function (CDF) of the computing resource utilization, the computing resource utilization under different computing capabilities, and the computing resource utilization under different storage capacities provided by an exemplary embodiment; Figure 11 a to Figure 11d are storage resource utilization under different training rounds, cumulative distribution function (CDF) of storage resource utilization, storage resource utilization under different computing capabilities, and storage resource utilization under different storage capacities, respectively, provided by an exemplary embodiment. DETAILED DESCRIPTION

[0024] The following first further describes the existing related technologies and the technical problems they have.

[0025] The rapid development of the Internet of Things, big data, and 5G / 6G technologies has led to a surge in terminal devices and emerging services, creating unprecedented network traffic pressure. To meet the demand for low latency and diverse services, Mobile Edge Computing (MEC) moves computing and storage capabilities to the edge of the network, enabling rapid responses and personalized services tailored to the user's location. MEC not only reduces the load on cloud centers and shortens data transmission paths, but also significantly improves service latency sensitivity and computing efficiency. Leveraging its localized coverage capabilities, MEC supports context-aware service deployment and is widely used in scenarios such as smart cities, the Industrial Internet of Things, and autonomous driving.

[0026] To cope with the surge in traffic, operators continue to expand cellular network capacity, with ultra-dense networks (UDN) becoming a key trend in 5G and subsequent architectures. While densely deployed small cells and edge servers improve service efficiency and responsiveness, they also significantly increase system energy consumption, further exacerbating the ICT industry's carbon emissions burden. Research shows that cellular infrastructure and data servers already account for the majority of ICT energy consumption, making energy efficiency an increasingly critical issue. Therefore, optimizing energy usage in MEC systems while ensuring service quality has become a critical issue that needs to be addressed.

[0027] Current energy efficiency optimization methods in MEC mainly include random optimization, heuristic algorithms and deep learning strategies. Although they have reduced costs and improved response efficiency to a certain extent, they still face several key challenges.

[0028] First, the distributed and geographically heterogeneous nature of edge nodes leads to complex, multi-dimensional dependencies in service deployment. Traditional methods such as RNNs and CNNs struggle to capture global context and long-range dependencies, making them inadequate for large-scale, real-time decision-making. Furthermore, service deployment and delivery differ fundamentally in their timescales. Deployment involves long-term planning, and frequent handoffs incur high overhead; delivery requires immediate response, and static strategies can easily lead to sluggish responses. Existing methods generally fail to decouple and optimize these two aspects, making it difficult to balance service quality and system energy efficiency. Furthermore, most strategies employ fixed decision cycles and lack the ability to dynamically adjust timescales based on service semantics or system load. This lack of adaptability makes them difficult to cope with highly dynamic network environments. Therefore, there is an urgent need for an energy-efficient service deployment method that integrates multi-dimensional contextual information, decouples deployment and delivery timescales, and possesses adaptive scheduling capabilities to ensure efficient operation of MEC systems in dynamic environments.

[0029] The technical solution of the present invention is described in detail below with reference to the accompanying drawings.

[0030] like Figures 1 to 11 As shown in Figure d, the present invention provides an energy-efficiency service deployment and delivery method for edge computing networks, comprising the following steps: Step S1: Extracting multidimensional service features, including geographic location information, service semantic features, and service popularity features; Step S2: Encoding and fusing the extracted multidimensional service features using a self-attention encoder to generate a global context feature representation. In this embodiment, multidimensional service feature extraction and self-attention encoding effectively capture the global dependencies of services, improve feature fusion efficiency, and thus improve computational efficiency. Step S3: Adaptively adjusting the upper-layer decision time scale based on the global context feature representation in conjunction with dynamic system load prediction results. In this embodiment, dynamically adjusting the decision interval based on service semantics and load predictions avoids redundant switching and thereby reduces network fluctuations. Step S4: Executing service deployment decisions and base station sleep / activation switching decisions within the adjusted upper-layer time scale. In this embodiment, upper-layer service deployment and base station switching can proactively optimize long-term energy efficiency, reduce deployment costs, and thus reduce system costs. Step S5: At the fine-grained lower-level time scale, service delivery and resource allocation decisions are made based on real-time service requests. In this implementation, lower-level service delivery and resource allocation enable dynamic response to requests, reduce latency, and ensure service quality. Step S6: The decision-making processes of steps S4 and S5 are coordinated through a dual-time-scale hierarchical learning framework. This dual-time-scale hierarchical collaboration allows for explicit decoupling of deployment and delivery cycles, alleviating the temporal coupling issue.

[0031] Through the above technical solutions, in general, the method of the present invention can systematically solve the defects of the background technology through a three-level progressive mechanism of multi-dimensional feature fusion → adaptive time control → dual-scale hierarchical decision.

[0032] An exemplary embodiment of the present invention is described in detail below with reference to the accompanying drawings.

[0033] 1. Describing System Models and Problem Modeling 1. System Model Assume that the mobile user set is ,in, Represents the set of all users. The base station (BS) and MEC server are deployed in coordination and are uniformly represented as , Indicates a base station set. The service type is Number, among which, A collection of all service types. In the time slot The service request initiated is recorded as: The three parameters above respectively mean the data size of the service request (unit: bit), the required computational intensity (unit: number of CPU cycles per bit), and the delay requirement (unit: millisecond). The system runs in discrete time slots, and the time set is .user In the time slot The position is expressed as: .set up Indicates time slot A subset of users who initiate service requests. All service requests will be routed to nearby edge servers or cloud centers for processing. The cloud center is denoted as , which has sufficient computing resources and is considered to be able to host all types of services. In contrast, edge servers have limited resources and only deploy some local related services, and can only process corresponding requests when the target service has been deployed.

[0034] Figure 2 This diagram illustrates energy-efficient service deployment and delivery in a dual-timescale MEC network. The network consists of multiple interconnected base stations and edge servers, some of which are active, while those with lower loads are dormant to conserve energy. Edge servers with deployed services process corresponding service requests.

[0035] 2. Energy Consumption Model: The energy consumption of a base station (BS) equipped with an edge server primarily consists of fixed energy consumption and load-dependent energy consumption. The former is determined by components such as the microwave module and rectifier, while the latter depends on the operating status of devices such as the transceiver and digital signal processing unit. Therefore, the total energy consumption of all base stations and edge servers in a time slot can be expressed as: , where Indicates the Base stations in time slots The service load, For the The maximum load capacity of a base station. Indicates base station Resource utilization. Define a binary variable Indicates base station The running state, when When , it indicates that the base station is in the activated state; when , it indicates that the base station is in sleep state. Indicates the fixed power consumption of the base station in the activated state, Indicates the maximum power consumption of the base station under full load, Indicates the fixed power consumption of the base station in sleep mode.

[0036] 3. Service deployment model: Considering the wide distribution of edge nodes, limited resources and regional characteristics of services, strategic deployment is required to improve service availability and effectively respond to local user requests. Service deployment decisions are made through a binary variable The table shows that if the service In the time slot Deployed on edge servers On, then ,otherwise . The system is in time slot The total deployment cost in is: Among them, the deployment cost and Services Required storage resources Indicator function The definition is as follows: The above formula shows that if a service was not deployed in the previous time slot and needs to be deployed in the current time slot, corresponding deployment overhead will be incurred.

[0037] 4. QoS Model: Quality of Service (QoS) reflects the service performance under the current service deployment and base station switching strategy. In latency-sensitive edge computing scenarios, QoS focuses on the user-perceived service response delay, including the wireless transmission delay between the user and the base station and the transmission delay between base stations. Specifically, the spectrum efficiency between the base station and the user is expressed as: ,in, is the transmit power, is the noise power, is the channel gain, whose value is determined by the time slot Internal base station With users Therefore, the wireless data transmission rate is calculated as: ,in, represents the sub-bandwidth resources allocated by OFDM technology. With base station The wireless transmission delay between is: .

[0038] When the local MEC server cannot process a user request (for example, the target service is not deployed or resources are insufficient), the request can be responded to by other available edge servers in the MEC network. When processing these service requests, the data may need to be transmitted over multiple hops. Base stations usually communicate through wired connections, and the transmission delay is: ,in, Indicates that in the Dijkstra shortest path, adjacent base stations and Links between If the request is ultimately processed by the cloud center (such as during peak load), set ,in, represents the large delay introduced by the wide area network (WAN). The computational delay of the service request on the edge server is: ,in, Indicates that it is assigned to the user The proportion of computing resources, Represents an edge server For cloud processing, the computing latency is ,in represents the computing rate of the cloud center. Therefore, the system The QoS degradation cost can be expressed as: .

[0039] 2. Problem Modeling In energy-efficient service deployment, idle base stations (BSs) and edge servers can be put into sleep mode to reduce energy consumption. However, frequent state switching may cause network fluctuations and hardware performance degradation. Therefore, the total cost of base station switching in a time slot can be expressed as: in, and Represents base stations In most practical scenarios, due to the additional energy consumption, signaling overhead and warm-up delay in the reactivation process, there are usually .

[0040] To characterize this cost asymmetry, we set ,in Therefore, the system is in the time slot The total cost of is composed of the following four parts: energy consumption cost, service deployment cost, quality of service (QoS) degradation cost and base station switching cost. Its total cost function is defined as: in, , , , is a preset weight factor that reflects the system's preference for each cost. Considering the difference in optimization cycles between service deployment and service delivery, this paper constructs a dual-time-scale optimization framework to coordinate the decision-making process between the upper and lower layers.

[0041] Among them, service delivery decision is represented by a binary variable Indicates that if the user Service requests in time slots Base station If the response ,otherwise Based on this, the overall cost minimization problem of energy-efficient service deployment and delivery in MEC networks can be expressed as: The upper layer operates with a large time step, i.e. , and the lower layer is further subdivided into smaller granularity time steps within each upper layer time step, recorded as The relationship between the upper and lower time scales is defined as ,in, Indicates the number of adaptive lower layer time slots contained in each upper layer time interval. In order to adapt to the dynamic system state, It is not a static preset value, but is dynamically adjusted based on service semantics and system-level load. It is important to note that base station switching, service deployment, and service delivery are all binary decision variables, while resource allocation is a continuous decision variable.

[0042] Constraints ( 、 、 ) describes the dependencies between decision variables. Constraints ( ) means: To provide services, the base station must be in the active state; the constraint ( ) means: only the edge servers that have deployed the corresponding services can respond to user requests; the constraint ( ) means that resources can only be allocated after a service request is received.

[0043] In addition, the constraints ( ) stipulates that each service request can only be served by one node at most; Constraint ( ) is the storage resource capacity constraint of the edge server, where Representation node Storage capacity; Constraints ( ) is the computing resource capacity constraint of the edge node; constraint ( ) is the delay constraint of the service request.

[0044] The above optimization problem faces the following key challenges: First, the dual-time-scale energy-efficient service deployment and delivery problem is modeled as a mixed-integer nonlinear programming problem. Please see the proof of Theorem 1 below. This problem is NP-hard, which means it is difficult to solve in polynomial time. Second, due to the lack of complete future information and the strict latency requirements of MEC networks for service delivery, offline optimization methods are not feasible. Third, online methods need to deal with highly dynamic network states and the complex time coupling relationships introduced by different optimization cycles. While ensuring real-time response to service requests, it is also necessary to minimize the long-term network cost of the system. In particular, the decision time scale needs to be dynamically adjusted according to the service context and system-level load status.

[0045] NP-Hardness of the Energy Efficiency Service Deployment Problem The dual-time-scale energy-efficient service deployment and delivery problem involves the joint optimization of multiple decision-making dimensions, including base station switching, service deployment, service delivery, and resource allocation, and is subject to dynamic constraints such as resource capacity and latency requirements. To analyze the computational complexity of this problem, we perform a polynomial time reduction on the classic NP-complete problem, the 2D BinPacking problem, and prove that The following theorem formally verifies the NP-hardness of the proposed optimization model.

[0046] Theorem 1. Dual-time-scale energy efficiency service deployment and delivery problem (Problem ) is NP-hard.

[0047] Proof: The problem is proved by performing a polynomial time reduction from the known NP-hard problem, the 2D Bin Packing Problem. The NP-hardness of the two-dimensional packing problem is defined as follows: given a set of two-dimensional rectangular items, each item With width and height ; and a set of containers of equal size, each container The width is , the height is The goal is to put all items into as few containers as possible, satisfying the following two constraints: (1) no two items in the same container can overlap; (2) the size of each item cannot exceed the size of the assigned container. The two-dimensional bin packing problem can be formally modeled as the following optimization problem: in, Is a binary decision variable, indicating the container Whether it is used. Is a binary decision variable, representing the item Whether it is placed in the container middle.

[0048] Mapping relationship of decision variables: Each edge server Corresponding to a container in the two-dimensional bin packing problem , each service request Corresponding to an item Service Request The required computing resources and storage resources are mapped to items Width and height , and the server Available resources (i.e. ) corresponds to the size of the container and In terms of decision variables, base station switching decision Direct mapping to use variables for containers in the two-dimensional bin packing problem Service delivery decisions Mapped to , indicating a request Whether the server Processing. Resource allocation variables Mapped to , reflects the capacity actually allocated by the server for the request. Service deployment variables is omitted to simplify the construction, but this does not affect the complexity class of the problem. This mapping effectively preserves the combinatorial structure of the original problem while maintaining its computational complexity, thereby completing a polynomial time reduction from the two-dimensional bin packing problem to the dual-time-scale energy efficiency service deployment problem.

[0049] The mapping relationship between the objective function and the constraints: In terms of the objective function mapping, the problem The system cost minimization objective in can be simplified to minimizing the number of active edge servers, i.e. , which is similar to the two-dimensional packing problem The objective function in In terms of constraint mapping, the resource constraints on width and height in the two-dimensional packing problem (constraints and ) corresponds to The capacity constraints of computing resources and storage resources in and ). In addition, the constraint and Can be merged into , that is, only when the edge server Only when in active state can it respond to service requests , this logical constraint is equivalent to the two-dimensional packing problem . Therefore, the equivalence between the two can be summarized as follows: Therefore, since every instance of the 2D Bin Packing Problem can be reduced to a The example of the problem, based on the Karp reduction rule, proves the proposed optimization problem is NP-hard.

[0050] 4. Context-aware adaptive timescale control based on deep feature fusion To address dynamic service demands and system load fluctuations, this paper proposes a context-aware control framework based on deep feature fusion. This framework utilizes a Transformer encoder to extract multidimensional contextual features such as geographic location, semantic associations, and service popularity. It also uses a self-attention mechanism to model global dependencies, improving feature extraction efficiency. Furthermore, a dual-stream encoding strategy is employed to jointly model service semantics and system load. The fused features are then fed into a lightweight controller, enabling adaptive decision-making timescale adjustment to efficiently cope with dynamic network environments.

[0051] A. Feature Encoding In an MEC environment, edge servers are typically deployed in specific geographic areas, and their utilization is affected by the server's geographic location and the distribution of users within its coverage area. To capture this spatial correlation, the geographic information of the edge server is represented as: Different types of services have different characteristics. For example, video streaming services usually require a large storage space, while autonomous driving services require high computing resources. To capture this heterogeneity of services, service semantics is introduced, which is expressed as: ;in, , including the index of all services related to the edge server, as well as related attributes of the request, such as request size, resource requirements, and latency sensitivity. In addition, service deployment and delivery decisions are also affected by user distribution and user preferences. Service preferences are usually manifested as some services having higher popularity. Therefore, deploying popular services with regional relevance in the MEC environment can significantly improve the response efficiency of local requests. Service popularity follows a modified Zipf distribution, specifically expressed as: ; and are the platform coefficient and skewness parameter, respectively. Representation Service In service Ranked in descending order of popularity. In the feature encoding stage, service popularity is expressed as: ;in, Finally, the three types of embedded features, geographic information, service semantics, and service popularity, are concatenated into a unified tensor. , this tensor will be fed into the self-attention module of the Encoder as input to further extract multi-dimensional context features.

[0052] B. Context-aware service feature extraction The multi-dimensional feature extraction mechanism based on the Transformer model consists of multiple layers of stacked encoders. Each encoder performs the following operations in sequence: a self-attention module, a feedforward neural network module, and appends a residual connection and a normalization layer after each module. Input features First, it is projected onto the query via the learnable matrix ( ),key( ) and value ( ) three spaces, and then sent to the self-attention module to obtain the attention weighted feature vector : ;in, The input features are transformed by different linear matrices , , Get the dot product Used to measure the similarity between each pair of query and key. Scaling helps stabilize the gradient when processing long sequences. The output of the self-attention module in the Encoder It is input into its feedforward neural network module, which consists of two fully connected layers: the first fully connected layer uses ReLU as the activation function, and the second fully connected layer uses a linear activation function. The processing of this feedforward module can be expressed as: in, , Represent the weight matrices of the two layers respectively, , is the corresponding bias vector. The final output vector is The contextual characteristics of the service are represented, and this representation will be used in the subsequent adaptive time scale control module.

[0053] C. Load-aware feature fusion time scale control To further improve the temporal adaptability of the model, a network load prediction module was introduced into the original Transformer-based architecture. Specifically, a unified time series was defined, and the historical load and future predicted load values ​​were fused and expressed as follows: ;in, Indicates the actual observed load of the system at the current moment (such as CPU utilization, request volume, etc.), Indicates the future The load sequence is first transformed through the embedding layer, mapping each scalar load value into a high-dimensional vector representation. In order to characterize the temporal position relationship of each load value in the sequence, a time encoding vector is introduced. , the overall expression is: ;in, Represents an embedding vector related to the time position. Unlike the fixed position encoding method used in the standard Transformer, an adaptive time encoding mechanism is introduced here to more accurately capture the dynamic characteristics of system load changes over time. By introducing future load predictions, the scheduler can predict the fluctuation trend of system load and actively adjust the time scale of upper-level decisions accordingly. This forward-looking design enhances the time sensitivity of the model, enabling it to shorten the decision interval when the system is highly loaded and extend the update cycle when the system is stable, thereby improving overall efficiency without sacrificing responsiveness.

[0054] Load sequence after embedding It is then input into the Transformer encoder to capture the temporal dependencies and dynamic patterns between historical and predicted load values. Within the encoder, the self-attention mechanism enables each time step in the sequence to interact with all other time steps, effectively modeling long-range dependencies and trend changes. This mechanism calculates the similarity between any two time steps and generates a weighted representation to capture the importance of each load observation in the context. The output vector of the attention module is then aggregated through average pooling to generate the final load context representation vector: This aggregation operation enables the system to highlight and capture key fluctuation points or inflection points, which are crucial for the dynamic adjustment of decision time scales. The resulting load context vector The service-level feature representation obtained by the Transformer encoder is The fused feature vectors are then fed into the time scale control module to generate an adaptive upper layer update interval. : ;in, and are the weight matrix and bias vector of the time scale decision module respectively. , represents a variable upper-layer decision timescale, whose value is dynamically adjusted based on service semantics and system load trends. This adaptive mechanism allows the system to increase the update frequency when load fluctuates drastically and reduce it when the system is stable, thereby saving computation and switching overhead. Figure 3 This paper demonstrates a context-aware timescale control process based on deep feature fusion. The control module itself is implemented as a lightweight multilayer perceptron and trained with system-level cost feedback signals to learn context-aware scheduling policies. Because the lower layers operate at fine-grained timescales (e.g., milliseconds), this high-resolution architecture enables upper-layer updates to be flexibly triggered at any decision point. Ultimately, this framework achieves extremely high adaptability and responsiveness under dynamic system conditions.

[0055] 5. Energy Efficiency Service Deployment and Delivery Method Based on Dual-Time-Scale Hierarchical Learning In order to solve the problem of essential difference in optimization cycle between service deployment and service delivery, and adapt to the random dynamics of service requests, this paper proposes a dual-time-scale energy-efficient service deployment and delivery algorithm based on hierarchical learning.

[0056] A. Transformation of the Two-Level Optimization Problem Inspired by the difference in optimization cycles between service deployment and service delivery, the overall optimization problem is decomposed into a two-layer hierarchical structure. Figure 4 As shown in the figure, the upper-layer strategy performs service deployment and base station switching on a larger timescale to reduce the long-term operating costs of the MEC network; the lower-layer strategy performs service delivery and resource allocation on a smaller timescale to ensure the system's real-time responsiveness. To effectively solve this two-layer optimization problem, we model it as a partially observable multi-agent Markov decision process (POM-MDP) and propose a cost-aware distributed reinforcement learning algorithm (Cost-Aware Distributed Reinforcement Learning) that operates collaboratively on both timescales.

[0057] 1. State space: The joint state space is defined as ,in and Represent the state vectors of the upper layer and the lower layer respectively. The state of the upper layer used for service deployment is defined as , which contains information such as predicted traffic information, extracted multi-dimensional service semantic features, base station switching status of the previous time step, service deployment status of the previous time step, and the current adaptive upper layer update interval. The state of the lower layer for service delivery is defined as , which includes the current base station switching status, the current service deployment status, real-time user service requests, and user location information. It is important to note that when the upper layer makes a decision, the service request has not yet arrived. Therefore, the upper layer network needs to learn proactive service deployment and base station switching strategies based on predicted traffic and service context features.

[0058] 2. Observation space: The observation space is defined as ,in and Represent the observation information of the upper and lower intelligent agents respectively. The observation space of the upper network is consistent with its state space, which includes information such as predicted traffic, multi-dimensional service characteristics, historical status and update cycle. The observation space of the lower network is defined as , which includes edge servers The base station switching state, edge server Service deployment status, edge server Received local service requests and user location information. It's important to note that in this two-layer learning architecture, the lower network layer adopts a distributed structure. Each agent (such as an edge server) makes service delivery and resource allocation decisions based solely on its local observations, without relying on a central controller or frequent information exchange. This design is highly consistent with the distributed architecture of MEC systems, enabling flexible and efficient control at the network edge.

[0059] 3. Action Space: In the proposed dual-time-scale hierarchical learning framework, each agent takes action according to its observation information and strategy. Select an action to perform from the action space. The action space is defined as , where the upper level action Represents base station switching and service deployment decisions, defined as . Lower level action For service delivery and resource allocation decisions, it is defined as Because the upper and lower layer strategies are executed alternately at different time scales, the upper layer strategy is invoked periodically to make forward-looking deployment decisions, while the lower layer strategy is continuously optimized based on the real-time network status to achieve coordinated control of energy-efficient service deployment and real-time service delivery.

[0060] 4. Reward function: To reduce the long-term cost of the system during service deployment and delivery, this study defines a reward function under dual time scales: Specifically, the upper layer reward function is defined as , indicating that at each adaptive upper layer time interval The average system cost within is the starting time of the lower layer corresponding to the upper layer time step. The lower layer reward function is defined as ,in Represents the number of all lower-level time steps in a complete episode, Indicates at time These two reward functions at different timescales jointly drive a hierarchical learning process: the upper layer focuses on long-term service deployment efficiency, while the lower layer ensures the real-time responsiveness of service delivery. By introducing a learning mechanism based on expected rewards, the framework coordinates decisions between the upper and lower timescales, effectively alleviating the temporal coupling between service deployment and delivery cycles.

[0061] B. Hierarchical Deep Reinforcement Learning Method at Dual Time Scales: Service Deployment and Delivery for Energy Efficiency In the proposed dual-time-scale hierarchical learning framework, the upper-layer network adopts an improved Dueling Network architecture, which separates the state value function from the action advantage function, thereby more effectively evaluating the relative merits of each action. The upper-layer learning focuses on formulating forward-looking service deployment and base station switching strategies at larger time scales to achieve long-term system cost reduction. The lower-layer network adopts a multi-agent deep deterministic policy gradient algorithm to handle service delivery and resource allocation tasks at small time scales. This method is applicable to continuous action spaces and can solve coordination problems between multiple agents, thereby improving the system's response efficiency and resource utilization. By combining the long-term planning capabilities of the upper layer with the real-time response capabilities of the lower layer, the proposed framework achieves collaborative optimization across different time scales, while improving long-term energy efficiency while ensuring short-term service quality.

[0062] 1. Upper-layer learning: The upper-layer network adopts the Dueling architecture to separate the state value function and the action advantage function into models, which can be expressed as follows: in, A value function representing a state. represents the advantage function of the action in the state. is the parameter set of the upper network. This decomposition method can significantly improve the stability and discrimination ability of Q-value estimation when the rewards of different actions are close, which helps to more efficiently learn long-term deployment strategies.

[0063] The goal of the upper layer learning is to minimize the long-term operating cost of the system by optimizing the service deployment strategy and base station switching strategy. To this end, the action value function It is iteratively updated by minimizing the temporal difference error, which is defined as the mean squared error between the predicted Q value and the target Q value: Among them, the target Q value The calculation formula is: ;in, is a discount factor that balances immediate rewards with future rewards. Represents the parameters of the target network, which is a delayed copy of the main network and is mainly used to provide relatively stable target values ​​during training, alleviate the oscillation caused by frequent updates of the main network, and thus improve the stability of model training. Main network parameters The update is done by gradient descent: ;in, is the learning rate, which controls the step size of each parameter update. By iteratively minimizing the above loss function, the upper-layer network can effectively learn forward-looking service deployment and base station switching strategies, thereby achieving long-term system cost reduction and energy efficiency improvement in MEC networks.

[0064] 2. Lower-layer learning: Lower-layer learning focuses on enabling real-time service delivery and resource allocation on a smaller timescale. This requires the system to rapidly respond to dynamically changing service requests. Considering the distributed nature of edge servers and the need for independent yet collaborative decision-making, this problem is modeled as a multi-agent reinforcement learning (MARL) task. To effectively address complex environments involving continuous action spaces and multi-agent interactions, this paper employs the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm. In this approach, each agent corresponds to an edge server and independently optimizes its policy based on local observations. During execution, each agent adopts a decentralized execution strategy, while a centralized training strategy is employed during the training phase. This facilitates collaborative learning and scalability in dynamic MEC environments. This mechanism balances responsiveness with collaborative efficiency, making it an ideal choice for meeting the real-time and distributed characteristics of edge computing.

[0065] Specifically, each agent Equipped with two neural networks: a policy network , used to observe the status of the local Output deterministic actions to perform service delivery and resource allocation. An evaluation network , used based on global state and joint actions Evaluate the value of the current action. The policy network is trained by maximizing the expected return estimated by the evaluation network, and its optimization objective is: ; This goal encourages the policy network to continuously improve its policy based on the evaluation of the evaluation network. The training goal of the evaluation network is to minimize the mean square error between the predicted Q value and the target Q value: Among them, the target Q value is defined as: ;in, is the immediate reward received at the time step. is a discount factor used to balance immediate and future rewards. , are the parameters of the target policy network and the target evaluation network respectively. Finally, the parameter update rules of the policy network and the evaluation network are as follows: in, and are the learning rates of the policy network and the evaluation network, respectively.

[0066] 3. Parameter synchronization: To ensure the stability of the training process and improve the convergence speed, this paper adopts a soft update mechanism between the main network and the target network to synchronize parameters. The soft update rules of the target network are as follows: in, These represent the update rates of the upper target network, the lower actor network's target parameter update rates, and the lower critic network's target parameter update rates, respectively. These update rate parameters control the rate at which the target network approaches the main network's parameters with each step. A smaller update rate can smoother changes in the target network, effectively preventing drastic fluctuations in target values ​​and improving the stability of the overall training process.

[0067] As described below, Algorithm 1 shows the pseudocode flow of the proposed DT-EESD (Dual-Timescale Energy-EfficientService Deployment and Delivery) algorithm. This dual-timescale hierarchical learning framework jointly optimizes at two levels: the upper layer performs high-level service deployment and base station switching decisions; the lower layer performs fine-grained service delivery and resource allocation control. To improve the flexibility and responsiveness of the strategy, the framework introduces a context-aware adaptive timescale control mechanism that dynamically adjusts the decision update frequency based on service semantic characteristics and system load dynamics. By combining a hierarchical asynchronous learning mechanism with an expected return-based optimization strategy, the framework effectively alleviates the temporal coupling problem between service deployment and delivery, achieving the coordinated optimization of energy efficiency and service quality.

[0068] Algorithm 1: Context-aware dual-time-scale hierarchical learning approach for energy-efficient service deployment and delivery 6. Performance Evaluation This application evaluates the performance of the proposed DT-EESD framework through a large number of experiments, compares it with a variety of advanced algorithms and classic baseline methods, and comprehensively analyzes its performance in terms of learning convergence, total system cost, service latency, and resource utilization.

[0069] A. Experimental Setup The experiment was implemented using the open-source framework PyTorch. The simulation environment consisted of eight fixed base stations (each equipped with an edge server) and 100 randomly moving users. The edge servers had 25Gbit storage and 10GHz computing power. The service request data volume ranged from [0.5, 1]Mbit, the computational intensity from [100, 500]CPU cycles / bit, and the popularity followed an MZipf distribution. In the DT-EESD framework, the upper network had two hidden layers, a learning rate of 1e-4, and a batch size of 32; the lower-layer actor network had three hidden layers, with both the actor and critic learning rates of 1e-4, a batch size of 256, and the Adam optimizer. The TransformerEncoder had an input dimension of 176, a hidden layer dimension of 176, four attention heads, and a dropout rate of 0.1. Fully connected layer configuration: first layer 232→512, second layer 512→256; action advantage function output layer 256→80, state value function layer 56→1.

[0070] B. Comparison Algorithms Dual-layer Learning without multidimensional features (DL-WM) does not include the extracted multidimensional context features in the state representation, but still adopts a dual-time-scale learning framework. To ensure fairness, it is expanded to the same dual-layer structure as DT-EESD.

[0071] Single-layer Learning (SL) is a method that simultaneously makes service deployment and delivery decisions in a single-layer learning framework. The entire learning process only involves the reward and loss feedback of the lower layer and does not consider the optimization structure across time scales.

[0072] Cycle-reward-based Dual-layer Learning (CRDL) addresses the issue of differing optimization cycles between the upper and lower layers in dual-layer learning. CRDL introduces a periodic reward mechanism and conducts dual-layer joint learning on this basis to obtain feasible energy-efficient service deployment and delivery strategies.

[0073] Linear Programming with Randomized Rounding (LPRR) formulates the problem as a relaxed linear programming model. It first seeks the optimal solution for continuous variables and then uses randomized rounding to convert this solution into a feasible discrete decision within resource constraints. It's important to note that LPRR is not a learning-based algorithm but rather a heuristic optimization method.

[0074] The above method is used as a comparison baseline to evaluate the performance advantages of DT-EESD from the following aspects: convergence speed of the learning process, total cost of the system, service response latency, and resource utilization efficiency.

[0075] C. Analysis of numerical results 1) Convergence and training loss. The reward and loss curves during the training phase are key indicators for evaluating the learning efficiency and stability of the algorithm. To comprehensively evaluate the convergence performance of this method in terms of service deployment and long-term cost optimization, we first trained the upper-layer EESD algorithm at a single time scale and compared it with DQN and Dueling DQN without Prioritized Experience Replay (DDQN-WP). The results are shown in the figure below. Figure 5 a and Figure 5 As shown in b. At the same time, Figure 6 a to Figure 6 Figure d shows the training performance of the DT-EESD dual-time-scale architecture. Due to the randomness of the network environment, the reward curve fluctuates during training. To verify the robustness of convergence, multiple experiments were conducted to reduce accidental interference. The results show that EESD converges faster and has higher cumulative rewards, demonstrating its ability to effectively extract multidimensional features and accelerate policy learning with the help of advantage functions. DT-EESD exhibits fast and stable convergence at both the upper and lower layers, thanks to the good synergy between the forward-looking deployment strategy of the upper layer and the real-time service response mechanism of the lower layer. Compared with single-layer methods, the dual-time-scale structure effectively alleviates the temporal coupling problem of upper and lower layer decision-making by introducing an expected reward mechanism, thereby improving the stability and adaptability of the overall system.

[0076] 2) Total system cost. This method aims to reduce overall system operating costs through energy-efficient service deployment and delivery strategies. Evaluations are conducted over 800 time slices, using a cumulative distribution function and a sliding average cost per 100 slices to balance network fluctuations and ensure fair results. Figure 7 a to Figure 7 Figure d shows that DT-EESD has the lowest total cost among all compared algorithms. At 10GHz computing power, DT-EESD reduces costs by 31.75%, 12.07%, 18.25%, and 21.98% compared to SL, LPRR, DL-WM, and CRDL, respectively; the overall average cost is reduced by 30.51%, 13.92%, 17.81%, and 22.46%. This advantage is primarily due to three aspects: 1) dual-timescale joint optimization, balancing long-term energy efficiency and short-term responsiveness; 2) context-aware feature extraction, integrating location, semantics, popularity, and traffic predictions to achieve precise deployment and minimize resource waste; and 3) an adaptive timescale control mechanism, dynamically adjusting policies based on load, improving real-time performance and reducing switching overhead, thereby further reducing costs.

[0077] 3) Service delay: Service delay reflects response speed and resource scheduling efficiency, and is a key indicator for measuring system performance. Figure 8 a to Figure 8 d compares the performance of each algorithm in terms of total service latency. DT-EESD achieves the lowest latency across all settings. Specifically, at 6GHz computing power, DT-EESD reduces latency by 28.85%, 42.98%, 30.78%, and 25.46% compared to DL-WM, SL, CRDL, and LPRR, respectively. At 12GHz, DT-EESD maintains its advantage, with latency reductions of 20.26%, 37.83%, 25.79%, and 13.97%. This performance improvement is attributed to DT-EESD's dual-timescale collaborative optimization mechanism: the upper layer implements forward-looking deployment to pre-configure resources and avoid allocation lags; the lower layer rapidly responds to dynamic requests through real-time scheduling. These two mechanisms work together to achieve efficient, low-latency service delivery.

[0078] 4) Service request response rate: The response rate reflects the system's ability to handle user requests, which is particularly critical in a dynamic environment. Figure 9 a to Figure 9 As shown in Figure d, DT-EESD achieved the highest response rate among all algorithms. At 16 GHz computing power, DT-EESD achieved a 96.24% response rate, significantly outperforming DL-WM (93.58%), SL (89.59%), CRDL (90.95%), and LPRR (93.13%). This advantage stems primarily from its layered learning architecture and multi-dimensional feature extraction capabilities. By combining service semantics with traffic prediction, it enables proactive deployment. Its semantic and load-aware strategies enable more accurate service placement in high-demand areas. This collaborative and context-driven approach effectively improves service reachability and overall responsiveness.

[0079] 5) Computing and storage resource utilization: Figure 10 a to Figure 11 Figure d shows the performance of each algorithm in terms of compute and storage resource utilization. Resource utilization is low in the early stages of training, primarily due to policy convergence and insufficient exploration. Overall, DT-EESD maintains the highest resource utilization throughout the training process. Specifically, at 8GHz of computing power, DT-EESD achieves 99.13% compute resource utilization; at 10Gbit of storage capacity, storage resource utilization reaches 97.68%. This demonstrates that DT-EESD achieves efficient coordinated scheduling of compute and storage resources through a dual-time-scale optimization framework. Combining context-aware feature extraction with proactive deployment strategies further improves resource scheduling efficiency, reduces idle rates, and maximizes resource utilization.

[0080] The present invention also provides an edge computing system for executing the energy efficiency service deployment and delivery method for edge computing networks described in any technical solution in the first aspect of the present invention, wherein the edge computing system includes a feature extraction module, a self-attention encoder module, a time scale control module, an upper-layer decision module, a lower-layer decision module and a hierarchical coordinator; the feature extraction module is configured to perform the extraction of multi-dimensional service features; the self-attention encoder module is configured to perform the encoding fusion of multi-dimensional service features; the time scale control module is configured to perform the adjustment of the adaptive upper-layer decision time scale; the upper-layer decision module is configured to perform service deployment decisions and base station sleep / activation switching decisions; the lower-layer decision module is configured to perform service delivery decisions and resource allocation decisions; the hierarchical coordinator is configured to perform dual-time-scale decision coordination.

[0081] The above are only specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions within the technical scope disclosed by the present invention should be covered by the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A method for deploying and delivering energy efficiency services for edge computing networks, characterized in that: include: Step S1: extracting multi-dimensional service features, wherein the multi-dimensional service features include geographical location information of the service, service semantic features, and service popularity features; Step S2: The extracted multi-dimensional service features are encoded and fused through the self-attention encoder to generate a global context feature representation; Step S3: Adaptively adjusting the upper-layer decision time scale based on the global context feature representation in combination with the system load dynamic prediction result; Step S4: executing service deployment decision and base station sleep / activation switching decision under the adjusted upper layer time scale; Step S5: Execute service delivery decisions and resource allocation decisions based on real-time service requests at a fine-grained lower-level time scale; Step S6: Coordinate the decision-making process of steps S4 and S5 through a dual-time-scale hierarchical learning framework.

2. The energy efficiency service deployment and delivery method for edge computing networks according to claim 1 is characterized in that: The multi-dimensional service feature extraction specifically includes: Step S1-1: Express the geographical location information of the service as: in, Indicates the The two-dimensional geographic coordinates of the edge servers, Indicates the The longitude of the edge server, Indicates the The latitude of the edge server, Indicates the total number of edge servers; Step S1-2: Express the semantic features of the service as: Where, , represents the time variable, represents a discrete range of time, where is the total number of time periods considered by the system, Indicates the edge servers at a point in time Different service attributes under Step S1-3: Generate service popularity features according to modified Zipf distribution: Where, For the edge servers at a point in time Next service The flow ratio, For service In service Ranked in descending order of popularity, is the platform coefficient, is the skewness parameter, and Indicates the type of service, For the edge servers at a point in time Next service The traffic intensity, and meet the ,In the feature encoding stage, service popularity is expressed as: Where, is the traffic distribution set for all services, For the time point Next Traffic distribution of edge servers; Step S1-4: Concatenate the above two-dimensional geographic coordinates, semantic features, and service popularity features to obtain the input tensor: .

3. The energy efficiency service deployment and delivery method for edge computing networks according to claim 2, characterized in that: The feature encoding process based on the self-attention mechanism includes: Step S2-1: Input tensor Generate query matrix by linear transformation , key matrix Sum Matrix ; Step S2-2: Calculate attention weighted features: Where, is the bond matrix The transposed matrix, dot product Used to measure the similarity between each pair of query and key, A factor that scales the dot product result to prevent the dot product value from being too large in high-dimensional data; Step S2-3: Output context feature vector through feedforward neural network : Where, 、 are the weight matrices of the two layers respectively, 、 are the corresponding bias vectors respectively.

4. The energy efficiency service deployment and delivery method for edge computing networks according to claim 3, characterized in that: The adaptive adjustment of the upper layer decision time scale includes: Step S3-1: Build system load sequence: Where, For the The actual observed load of the system at the moment, is the actual observed load of the system at the current moment, For the future The predicted load at the moment, For the future Forecasted load at the moment; Step S3-2: Fusion of temporal coding vectors Generate a load signature representation: Where, Indicates that the load data Mapped to a high-dimensional vector space through the embedding layer, is the embedding vector associated with the time position; Step S3-3: Extract load dynamic features through Transformer encoder : Where, Indicates the operation of pooling the input data. To represent the query vector in the load prediction task, is the representation of the key vector in the load prediction task, is the representation of the value vector in the load prediction task; Step S3-4: Concatenate context feature vectors Dynamic characteristics of load , input multi-layer perceptron to generate adaptive upper layer update interval : Where, and are the weight matrix and bias vector of the time scale decision module respectively.

5. The energy efficiency service deployment and delivery method for edge computing networks according to claim 1, characterized in that: In step S4, the upper layer decision cycle is an adaptive interval, and the output action is a base station handover state.

6. The energy efficiency service deployment and delivery method for edge computing networks according to claim 1, characterized in that: In step S5, the lower-layer decision cycle is a fixed fine-grained time slot, and the output action is the service delivery path and resource allocation ratio.

7. The method for deploying and delivering energy efficiency services for edge computing networks according to claim 1, wherein: The upper layer network in step S6 adopts Dueling architecture: Step S6-1: Separate the state value function and the action advantage function into models: Where, represents the state value function, Indicates time State of the moment , represents the set of parameters of the state-value function, represents the advantage function of the action in the state, Indicates time The action selected by the agent at that moment , represents the parameter set of the action advantage function, is the parameter set of the upper network, is the advantage function; Step S6-2: Update network parameters by minimizing the timing difference error: Among them, the target Q value The calculation formula is: Where, is the loss function, is the expected value, For intelligent agents of function, Indicates time The state of the agent at each moment, Indicates time The action of the agent at each moment, For time The instant reward obtained by the agent at each moment, is the discount factor and , Indicates time The state of the agent at each moment, Indicates time The action of the agent at each moment, Represents the parameters of the target network.

8. The method for deploying and delivering energy efficiency services for edge computing networks according to claim 1, wherein: The lower layer decision in step S6 adopts the multi-agent deep deterministic policy gradient algorithm, where the policy network parameters are and evaluate network parameters Update according to the following rules: Where, and are the learning rates of the policy network and the evaluation network, is the gradient of the policy network, is the gradient of the evaluation network.

9. The method for deploying and delivering energy efficiency services for edge computing networks according to claim 1, wherein: The step S6 also includes a soft update mechanism, and the target network parameters are updated according to the following rules: Where, are the parameters of the target policy network, The parameters of the target evaluation network are, are the parameters of the target Critic network, They respectively represent the update rate of the upper target network, the target parameter update rate of the lower Actor network, and the target parameter update rate of the lower Critic network.

10. An edge computing system, characterized in that: The energy efficiency service deployment and delivery method for an edge computing network according to any one of claims 1 to 9 is configured to be executed, wherein the edge computing system comprises: a feature extraction module configured to perform extraction of multi-dimensional service features; A self-attention encoder module configured to perform encoding fusion of multi-dimensional service features; a time scale control module configured to perform adaptive upper layer decision time scale adjustment; An upper-layer decision module configured to execute service deployment decisions and base station sleep / activation switching decisions; A lower-level decision module, configured to execute service delivery decisions and resource allocation decisions; A hierarchical coordinator configured to perform dual-timescale decision coordination.

Images