Method for routing and resource scheduling in large-scale heterogeneous local power communication network based on hybrid expert

Abstract

The application discloses a large-scale heterogeneous local power communication network routing and resource scheduling method based on a hybrid expert, and the method comprises the following steps: step S1, establishing a large-scale heterogeneous local communication network routing and resource scheduling optimization problem according to the differentiated QoS requirements of services on data transmission rate, time delay and reliability, jointly optimizing the transmission path of each service and the transmission medium, channel and data modulation mode on the path, and realizing the minimum difference between the required performance of the service and the supply performance of the network; and step S2, solving the routing and resource scheduling optimization problem established in step S1 based on a hybrid expert algorithm, obtaining the transmission path of each service and the transmission medium, channel and data modulation mode scheme on the path, and the minimum value of the performance supply-demand difference.

Description

Technical Field

[0001] This invention relates to the field of communication network technology, and in particular to a routing and resource scheduling method for large-scale heterogeneous local power communication networks based on hybrid experts. Background Technology

[0002] Against the backdrop of energy transition, the high proportion of distributed clean energy integration is leading to profound changes in power grid dispatching and operation modes. The power grid is evolving from traditional single-dispatch to a complex system of coordinated control of power generation, grid, load, and storage, and multi-level collaboration of transmission, distribution, and microgrids. This transformation has not only spurred a series of innovative power system application scenarios, such as distributed renewable energy grid integration and consumption, vehicle-grid interaction, and virtual power plants, but has also greatly enriched the connotation of power services. Today, power services not only cover traditional areas such as distribution automation, but also widely involve new services such as renewable energy output forecasting and peak shaving and valley filling assisted by charging piles. These services exhibit significant diversity and differences in terms of generation cycles and QoS requirements, placing more stringent demands on the performance of power communication networks, such as ultra-high bandwidth, ultra-high speed, ultra-low latency, and ultra-high reliability. Furthermore, with the large-scale integration of new power electronic devices, the number of power grid terminals is increasing dramatically, and "response" is gradually extending to the end of the distribution network. Information collection and information exchange between platforms will experience explosive growth, and the demand for high-speed collaborative interaction between devices is increasing, necessitating ubiquitous and flexible local power communication networks to provide real-time response speeds and wide-coverage access capabilities. However, given the limited communication resources and the ever-increasing demands for service performance, how to provide on-demand services and improve network capacity while ensuring service QoS has become a huge challenge for local power communication networks.

[0003] Currently, heterogeneous data communication architectures integrating power line carrier (PLC) and wireless communication are a promising approach to meeting the ubiquitous connectivity, ultra-high reliability, and high-capacity communication requirements of large-scale heterogeneous power local communication networks. Based on this architecture, numerous scholars have proposed corresponding routing and resource scheduling schemes. However, both optimization-based and heuristic resource allocation and scheduling methods require precise mathematical models and parameters. As network size and optimization variables increase, both methods suffer from solution difficulties and time consumption, thus failing to adequately adapt to the ever-growing application scenarios, network topology, and the increasing complexity and uncertainty of network resources in the power industry.

[0004] In recent years, Deep Reinforcement Learning (DRL) has attracted widespread attention due to its outstanding performance in handling the curse of state space dimension and dynamic environmental changes. It involves an agent continuously exploring and learning from mistakes in the environment, adjusting its behavioral strategy based on immediate rewards to maximize cumulative returns. However, the large number of terminals in local power communication networks leads to excessively large action spaces and computational complexity in existing DRL methods, making convergence difficult and rendering them unsuitable for real-world scenarios. Furthermore, for large-scale multi-service problems, training all services together in the same DRL model may result in negative transfer, degrading method performance.

[0005] Recent advancements in the Expert Hybrid (MoE) framework offer an effective solution by employing a suite of AI models as expert specialists, significantly reducing the need for training models specific to individual business needs. The Expert Hybrid (MoE) framework is a simple yet effective integration architecture that allows for efficient switching between AI models based on their online performance within the integration. Summary of the Invention

[0006] To address the aforementioned issues, this invention discloses a routing and resource scheduling method for large-scale heterogeneous local power communication networks based on hybrid experts. This method aims to meet the differentiated QoS requirements of power services as much as possible in resource-constrained local power communication networks. It can also process diverse power services in parallel, thereby improving the effectiveness of resource allocation.

[0007] The technical solution adopted in this invention is as follows:

[0008] Step S1: Based on the differentiated QoS requirements of services for data transmission rate, latency, and reliability, establish a large-scale heterogeneous local communication network routing and resource scheduling optimization problem, jointly optimize the transmission path of each service and the transmission medium, channel, and data modulation method on the path, so as to minimize the difference between the performance required by the service and the performance provided by the network.

[0009] Step S2: Based on the hybrid expert algorithm, solve the routing and resource scheduling optimization problem established in step S1 to obtain the transmission path of each service, the transmission medium, channel and data modulation scheme on the path, and the minimum performance supply and demand difference.

[0010] The aforementioned method for routing and resource scheduling in large-scale heterogeneous local communication networks based on hybrid experts, in step S1, involves a large-scale heterogeneous local communication network comprising three types of nodes: central nodes, aggregation nodes, and end nodes. These nodes communicate via high-speed power line carrier (High PLC, HPLC) links, wireless radio frequency (RF) links, and low-power links (e.g., LoRa, WiFi, Zigbee). Central nodes communicate with aggregation nodes via dual-mode communication (HPLC link + RF link), aggregation nodes communicate with each other via dual-mode communication or single-mode communication (HPLC link only or RF link only), and end nodes communicate with aggregation nodes via low-power links. The large-scale heterogeneous power line local communication network is modeled as a graph structure. ,in Represents a set of nodes. Represents a set of links. HPLC link set and wireless link set Composition. Assume the set of all available channels in the network is... ,in Represents the HPLC channel set. Let represent the set of RF channels. Assume the set of all available data modulation schemes in the network is . These include Binary Phase Shift Keying (BPSK), Quadrature Phase Shift Keying (QPSK), and 16-ary Quadrature Amplitude Modulation (16QAM).

[0011] make Represents a set of business items, a 5-tuple. Indicates business Differentiated needs, among which and For business The source node and the destination node, , , and Indicates business Size (bits), data transmission rate requirements, end-to-end latency requirements, and end-to-end reliability requirements. Let Indicates from the source node to the destination node The routing service aims to find an end-to-end path, including the transmission medium, channel, and data modulation scheme along that path. The sequence of forwarding nodes along this path can be represented as follows: Let vector ,in 0-1 variables Used to represent business Select node? and between A type of link; if it is, then the value is 1, otherwise it is 0. Let the vector... , where 0-1 variables Used to represent nodes and between Does the link use a channel? If yes, then the value is 1; otherwise, it is 0. Simultaneously, a vector is defined. , where 0-1 variables Represents a node and between Channel on the link Select modulation method Based on the differences between the channel models of HPLC and wireless communication, using... Represents a node and Signal-to-noise ratio of inter-HPLC links, Represents a node and Inter-wireless link signal-to-noise ratio, services At the node and between Type link channel The data transmission rate is

[0012]

[0013] in, for Type link channel The bandwidth. When the service... At the node and Once the transmission link, transmission channel, and data modulation method are determined, the transmission delay... for

[0014]

[0015] in, Indicates business Size (bits) Indicates the data modulation method. Service slave node Transmit to node During the process, the nodes Called a node The parent node, at this time the node The hierarchy is higher than the node High. Business In the parent node Queuing delay at the location is

[0016]

[0017] in Indicates business Reaching the parent node The size of the business waiting to be processed in its cache queue at that time. Indicates the parent node The business processing rate. For compute-intensive businesses, define binary variables. Indicates computationally intensive business Select node? Perform the calculation; if yes, return 1; otherwise, return 0. Assume... For nodes The available computing power (e.g., CPU cycles per second) then the business At the node The computation delay at that location is

[0018]

[0019] in , This refers to a set of nodes with computing resources. Then the business... Path end-to-end delay for

[0020] +

[0021] If the business No computing resources are required. Typically, the error rate distribution in reality approximates a Poisson distribution, therefore business... From node Transmit to node Appear The probability of a bit error is

[0022]

[0023] in Indicating the modulation method The bit error rate is below that. Then the service... At the node and Reliability during inter-transmission for

[0024]

[0025] business Path end-to-end reliability for

[0026]

[0027] The optimization objective of routing and resource scheduling optimization in large-scale heterogeneous local communication networks is to minimize the performance supply-demand gap, i.e., the difference between the performance required by the service and the performance supplied by the network. This gap consists of three parts: data transmission rate, end-to-end latency, and end-to-end reliability. Performance supply and demand differences for

[0028]

[0029] in , and It is an adjustable weighting coefficient, when the business... Increase when more sensitive to latency When business Increase when higher reliability requirements are required , , .

[0030] This paper proposes a routing and resource scheduling optimization problem for large-scale heterogeneous local power communication networks, which jointly optimizes the transmission paths, transmission media, transmission channels, and data modulation methods of all services in the network.

[0031] :

[0032] st

[0033]

[0034]

[0035]

[0036]

[0037]

[0038]

[0039]

[0040]

[0041] C1, C3, and C4 represent business functions. The requirements for link data transmission rate, end-to-end latency, and reliability. C2 represents... The total amount of traffic transmitted on a single link cannot exceed the maximum data transmission rate of that link. C5 indicates that at node The computing resources required by the services being computed must not exceed the maximum computing capacity of that node. C6 indicates business. At the node and Only one link type can be selected for inter-port transmission. C7 indicates the service. At the node and between Only one channel can be selected for transmission via a type-2 link. C8 indicates the service. At the node and between Channel of type link Only one modulation scheme can be selected during uplink transmission. C9 indicates the service. Only one node can be selected for calculation.

[0042] The aforementioned method for routing and resource scheduling in large-scale heterogeneous local communication networks based on hybrid experts establishes a unified Mixture of Experts (MoE) model in step S2 to simultaneously achieve relay node selection and resource allocation and scheduling between node pairs. Taking the source node as an example, the hybrid expert model consists of three sub-modules: an expert network, a gating network, and a tower network. These three networks are specifically:

[0043] S21. Shared Experts Based on Graph Attention Networks (GATs) in Expert Networks: An optimal next-hop selector is constructed using a shared expert model based on GATs. The optimal next hop needs to be selected from first-order neighbor nodes that satisfy the node's characteristics. First, three high-quality node features are selected: load rate, betweenness centrality, and level. Due to the different dimensions of the features, the feature data cannot be directly input into the GAT model for training. Betweenness centrality is a benefit-type feature; a larger value indicates better network performance. Load rate and level are cost-type features; a smaller value indicates better network performance. Therefore, a maximum-minimum normalization method is used to perform dimensionless processing on the data.

[0044] Secondly, train a shared weight matrix for all nodes in the graph. This serves as a mapping between the input and output features of a node. During the computation, the weight matrix is ​​first used to map the features of each high-level neighbor node of the central node, and the results of these feature mappings are concatenated. Then, a feedforward neural network is input, mapping the concatenated vector to a single value. Finally, through... Activation function, obtain nodes right Importance :

[0045]

[0046] in This is for concatenation operations. As input to the GAT model, , The input feature dimension. For nodes to share a weight matrix, Here is the weight matrix of the feedforward neural network. This is the output feature dimension.

[0047] Finally, compute nodes Neighboring nodes The probability of being selected as the optimal next hop :

[0048]

[0049] in It is a node (include The high-level first-order nearest neighbor set of ) Represents a node right The importance of each node is considered. The self-attention layer outputs the probability that a neighboring node is selected as the optimal next hop, and selects the neighbor with the highest probability as the center node of the next self-attention layer. Similarly, each self-attention layer generates the probability that the first-order neighbors of the center node are the optimal next hop, and the neighbor with the highest probability is selected as the center node of the next self-attention layer. The operation terminates when the first-order neighbors of a center node contain the destination node.

[0050] S22. Task-Specific QoS Experts Based on Multi-Arm Slot Machines in Expert Networks: Three experts are set up to learn the mapping relationship between different service QoS requirements and the allocation of inter-node pair transmission media, channels, and data modulation schemes. These three experts use the same multi-arm slot machine structure. Inter-node pair transmission media include HPLC and RF, which are combined with channel selection. While selecting the optimal next hop, the available HPLC / RF channels and data modulation schemes between node pairs are traversed sequentially to construct the swing arm. The set of HPLC and RF channels available between node pairs is as follows: The selected modulation scheme set is as follows Then the rocker arm assembly and each of the rocker arms for

[0051]

[0052] in, Indicates the number of rocker arms. and It is a rocker arm The channel and data modulation scheme selected between node pairs , In each iteration, it is necessary to start from... Choose one of the rocker arms to execute. Assume... Indicates end-to-end path The selected set of rocker arms is used to calculate the data transmission rate, latency, and reliability for each rocker arm, and the optimization objective of the routing and resource scheduling optimization problem in claim 2 is set as the reward.

[0053]

[0054] Using incremental form At the current iteration number Expected return valuation at that time Update

[0055]

[0056] in It is the number of iterations. Expected return valuation at that time It is a rocker arm The return express The number of executions. Furthermore, decay is employed. - Greedy strategy for selecting the rocker arm Exploring Probability The update method is

[0057]

[0058] in, This represents the initial value of the exploration probability. This represents the decay coefficient. Assume the maximum number of iterations is... In each iteration, the rocker arm The selection method is as follows:

[0059]

[0060] when Much larger At that time, it can traverse all rocker arms and find the optimal rocker arm.

[0061] Analysis shows that the rocker arm selected for execution in each iteration is... Closely related; with the number of iterations The increase, The rate of exploration gradually decreases, and the multi-armed slot machine model gradually transitions from an initial exploration-oriented approach of selecting new arms to a more exploitative approach of selecting the known optimal arm, while still attempting new arms with a small probability. Therefore, when Much larger In this way, it is basically possible to traverse all rocker arms and find the optimal rocker arm;

[0062] S23. A gated network is a softmax classifier that outputs the probability distribution of each expert being selected. For computationally intensive services with multiple QoS requirements, where data transmission rate, latency, and reliability are all critical, then each expert is selected.

[0063] S24. The tower network consists of a Fully Connected Network (FCN) layer. The FCN layer, as the final output module of this model, is composed of independent network layer structures. The results from the aggregated experts are then forwarded to the service-specific FCN network. The FCN layer uses the aggregated information to perform service-specific processing and generate the final output. Each service is associated with its own dedicated FCN layer and selects the ReLU function as its activation function.

[0064] The beneficial effects of this invention are:

[0065] 1. This invention addresses the challenges of surging demand for communication applications in local power communication networks, differentiated QoS requirements for services, and limited communication resources. It establishes a mechanism to minimize the performance supply-demand gap in large-scale heterogeneous local power communication networks. By jointly optimizing the transmission paths, transmission media, transmission channels, and data modulation methods of all services in large-scale heterogeneous local power communication networks, it aims to ensure that the resources allocated to services precisely meet the QoS requirements of those services. This allows for providing services to more services under limited resource conditions, thereby increasing the total number of successfully transmitted services across the entire network.

[0066] 2. A hybrid expert model incorporating both an expert network and a gating network was designed. Different expert models learn routing and resource scheduling strategies in a specialized and parallel manner to meet the diverse service requirements for data transmission rate, latency, and reliability, achieving faster convergence. Compared to various baseline methods, the proposed method demonstrates better performance in terms of performance supply and demand differences and successful service transmission rates. Attached Figure Description

[0067] Figure 1 It is a resource allocation and scheduling framework for large-scale heterogeneous local power communication networks based on hybrid experts;

[0068] Figure 2 It is an instantiation of LPCN node distribution;

[0069] Figure 3 This represents the performance supply-demand difference under different business volumes;

[0070] Figure 4 It represents the success rate of control-type service transmissions;

[0071] Figure 5 It represents the success rate of transmission for computationally intensive services.

[0072] Figure 6 It represents the success rate of data acquisition services. Detailed Implementation

[0073] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. It should be noted that the terms "front," "rear," "left," "right," "up," and "down" used in the following description refer to directions in the accompanying drawings, and the terms "inner" and "outer" refer to directions toward or away from the geometric center of a specific component, respectively.

[0074] This embodiment proposes a routing and resource scheduling method for large-scale heterogeneous local power communication networks based on hybrid experts, including the following steps:

[0075] Step S1: Based on the differentiated QoS requirements of services for data transmission rate, latency, and reliability, establish a large-scale heterogeneous local communication network routing and resource scheduling optimization problem, jointly optimize the transmission path of each service and the transmission medium, channel, and data modulation method on the path, so as to minimize the difference between the performance required by the service and the performance provided by the network.

[0076] Step S2: Based on the hybrid expert algorithm, solve the routing and resource scheduling optimization problem established in step S1 to obtain the transmission path of each service, the transmission medium, channel and data modulation scheme on the path, and the minimum performance supply and demand difference.

[0077] Step S3: Evaluate the performance of the proposed method under different service quantities and different service QoS requirements.

[0078] In step S1, the large-scale heterogeneous local communication network includes three types of nodes: central nodes, aggregation nodes, and end nodes. The connections between these three types of nodes include high-speed power line carrier (High PLC, HPLC) links, wireless radio frequency (RF) links, and low-power links (e.g., LoRa, WiFi, Zigbee). Central nodes communicate with aggregation nodes via dual-mode communication (HPLC link + RF link), aggregation nodes communicate with each other via dual-mode communication or single-mode communication (HPLC link only or RF link only), and end nodes communicate with aggregation nodes via low-power links. The large-scale heterogeneous power line local communication network is modeled as a graph structure. ,in Represents a set of nodes. Represents a set of links. HPLC link set and wireless link set Composition. Assume the set of all available channels in the network is... ,in Represents the HPLC channel set. Let represent the set of RF channels. Assume the set of all available data modulation schemes in the network is . These include Binary Phase Shift Keying (BPSK), Quadrature Phase Shift Keying (QPSK), and 16-ary Quadrature Amplitude Modulation (16QAM).

[0079] make Represents a set of business items, a 5-tuple. Indicates business Differentiated needs, among which and For business The source node and the destination node, , , and Indicates business Size (bits), data transmission rate requirements, end-to-end latency requirements, and end-to-end reliability requirements. Let Indicates from the source node to the destination node The routing service aims to find an end-to-end path, including the transmission medium, channel, and data modulation scheme along that path. The sequence of forwarding nodes along this path can be represented as follows: Let vector ,in 0-1 variables Used to represent business Select node? and between A type of link; if it is, then the value is 1, otherwise it is 0. Let the vector... , where 0-1 variables Used to represent nodes and between Does the link use a channel? If yes, then the value is 1; otherwise, it is 0. Simultaneously, a vector is defined. , where 0-1 variables Represents a node and between Channel on the link Select modulation method Based on the differences between the channel models of HPLC and wireless communication, using... Represents a node and Signal-to-noise ratio of inter-HPLC links, Represents a node and Inter-wireless link signal-to-noise ratio, services At the node and between Type link channel The data transmission rate is

[0080]

[0081] in, for Type link channel The bandwidth. When the service... At the node and Once the transmission link, transmission channel, and data modulation method are determined, the transmission delay... Represented as

[0082]

[0083] in, Indicates business Size (bits) Indicates the data modulation method. Service slave node Transmit to node During the process, the nodes Called a node The parent node, at this time the node The hierarchy is higher than the node High. Business In the parent node Queuing delay at the location is

[0084]

[0085] in Indicates business Reaching the parent node The size of the business waiting to be processed in its cache queue at that time. Indicates the parent node The business processing rate. For compute-intensive businesses, define binary variables. Indicates computationally intensive business Select node? Perform the calculation; if yes, return 1; otherwise, return 0. Assume... For nodes The available computing power (e.g., CPU cycles per second) then the business At the node The computation delay at that location is

[0086]

[0087] in , This refers to a set of nodes with computing resources. Then the business... Path end-to-end delay for

[0088] +

[0089] If the business No computing resources are required. Typically, the error rate distribution in reality approximates a Poisson distribution, therefore business... From node Transmit to node Appear The probability of a bit error is

[0090]

[0091] in Indicating the modulation method The bit error rate is below that. Then the service... At the node and Reliability during inter-transmission for

[0092]

[0093] business Path end-to-end reliability for

[0094]

[0095] The optimization objective of routing and resource scheduling optimization in large-scale heterogeneous local communication networks is to minimize the performance supply-demand gap, i.e., the difference between the performance required by the service and the performance supplied by the network. This gap consists of three parts: data transmission rate, end-to-end latency, and end-to-end reliability. Performance supply and demand differences for

[0096]

[0097] in , and It is an adjustable weighting coefficient, when the business... Increase when more sensitive to latency When business Increase when higher reliability requirements are required , , .

[0098] This paper proposes a routing and resource scheduling optimization problem for large-scale heterogeneous local power communication networks, which jointly optimizes the transmission paths, transmission media, transmission channels, and data modulation methods of all services in the network.

[0099] :

[0100] st

[0101]

[0102]

[0103]

[0104]

[0105]

[0106]

[0107]

[0108]

[0109] C1, C3, and C4 represent business functions. The requirements for link data transmission rate, end-to-end latency, and reliability. C2 represents... The total amount of traffic transmitted on a single link cannot exceed the maximum data transmission rate of that link. C5 indicates that at node The computing resources required by the services being computed must not exceed the maximum computing capacity of that node. C6 indicates business. At the node and Only one link type can be selected for inter-port transmission. C7 indicates the service. At the node and between Only one channel can be selected for transmission via a type-2 link. C8 indicates the service. At the node and between Channel of type link Only one modulation scheme can be selected during uplink transmission. C9 indicates the service. Only one node can be selected for calculation.

[0110] In step S2, a unified Mixture of Expert (MoE) model is established to simultaneously achieve relay node selection and resource allocation and scheduling between node pairs. For example... Figure 1 As shown, the blue, purple, and pink boxes represent different nodes executing the hybrid expert along the end-to-end path. Taking the source node as an example, the hybrid expert model consists of three sub-modules: an expert network, a gating network, and a tower network. These three networks are specifically:

[0111] S21. Shared Experts Based on Graph Attention Networks (GATs) in Expert Networks: An optimal next-hop selector is constructed using a shared expert model based on GATs. The optimal next hop needs to be selected from first-order neighbor nodes that satisfy the node's characteristics. First, three high-quality node features are selected: load rate, betweenness centrality, and level. Due to the different dimensions of the features, the feature data cannot be directly input into the GAT model for training. Betweenness centrality is a benefit-type feature; a larger value indicates better network performance. Load rate and level are cost-type features; a smaller value indicates better network performance. Therefore, a maximum-minimum normalization method is used to perform dimensionless processing on the data.

[0112] Secondly, train a shared weight matrix for all nodes in the graph. This serves as a mapping between the input and output features of a node. During the computation, the weight matrix is ​​first used to map the features of each high-level neighbor node of the central node, and the results of these feature mappings are concatenated. Then, a feedforward neural network is input, mapping the concatenated vector to a single value. Finally, through... Activation function, obtain nodes right Importance :

[0113]

[0114] in This is for concatenation operations. For nodes to share a weight matrix, Here is the weight matrix of the feedforward neural network. This is the output feature dimension.

[0115] Finally, compute nodes Neighboring nodes The probability of being selected as the optimal next hop :

[0116]

[0117] in It is a node (include The high-level first-order nearest neighbor set of ) Represents a node right The importance of each node is considered. The self-attention layer outputs the probability that a neighboring node is selected as the optimal next hop, and selects the neighbor with the highest probability as the center node of the next self-attention layer. Similarly, each self-attention layer generates the probability that the first-order neighbors of the center node are the optimal next hop, and the neighbor with the highest probability is selected as the center node of the next self-attention layer. The operation terminates when the first-order neighbors of a center node contain the destination node.

[0118] S22. Task-Specific QoS Experts Based on Multi-Arm Slot Machines in Expert Networks: Three experts are set up to learn the mapping relationship between different service QoS requirements and the allocation of inter-node pair transmission media, channels, and data modulation schemes. These three experts use the same multi-arm slot machine structure. Inter-node pair transmission media include HPLC and RF, which are combined with channel selection. While selecting the optimal next hop, the available HPLC / RF channels and data modulation schemes between node pairs are traversed sequentially to construct the swing arm. The set of HPLC and RF channels available between node pairs is as follows: The selected modulation scheme set is as follows Then the rocker arm assembly and each of the rocker arms for

[0119]

[0120] in, Indicates the number of rocker arms. and It is a rocker arm The channel and data modulation scheme selected between node pairs , In each iteration, it is necessary to start from... Choose one of the rocker arms to execute. Assume... Indicates end-to-end path The selected set of rocker arms is used to calculate the data transmission rate, latency, and reliability for each rocker arm, and the optimization objective of the routing and resource scheduling optimization problem in claim 2 is set as the reward.

[0121]

[0122] Using incremental form At the current iteration number Expected return valuation at that time Update

[0123]

[0124] in It is the number of iterations. Expected return valuation at that time It is a rocker arm The return express The number of executions. Furthermore, decay is employed. - Greedy strategy for selecting the rocker arm Exploring Probability The update method is

[0125]

[0126] in, This represents the initial value of the exploration probability. This represents the decay coefficient. Assume the maximum number of iterations is... In each iteration, the rocker arm The selection method is as follows:

[0127]

[0128] when Much larger At that time, it can traverse all rocker arms and find the optimal rocker arm.

[0129] S23. A gated network is a softmax classifier that outputs the probability distribution of each expert being selected. For computationally intensive services with multiple QoS requirements, where data transmission rate, latency, and reliability are all critical, then each expert is selected.

[0130] S24. The tower network consists of a Fully Connected Network (FCN) layer. The FCN layer, as the final output module of this model, is composed of independent network layer structures. The results from the aggregated experts are then forwarded to the service-specific FCN network. The FCN layer uses the aggregated information to perform service-specific processing and generate the final output. Each service is associated with its own dedicated FCN layer and selects the ReLU function as its activation function.

[0131] In step S3, network simulation is performed using NetLogo 3.6, and the overall layout of the local power communication network is located in an area of... Within a square area, it consists of one central node and 185 aggregation nodes. The specific network topology is as follows: Figure 2As shown, the simulation depicts three types of services in a smart distribution network: control services, data acquisition services, and computationally intensive services. For control and data acquisition services, the destination node is the central node, and the source node is randomly selected from 185 aggregation nodes. For computationally intensive services, the destination node is either the central node or an aggregation node with computing resources. Fifty nodes are randomly selected from the 185 aggregation nodes and assigned computing resources; the remaining aggregation nodes serve as the source nodes for this type of service. The size, end-to-end latency, and end-to-end reliability requirements for control services are as follows: The size and data transmission rate requirements for data acquisition services are as follows: The requirements for the size, data transmission rate, end-to-end latency, and end-to-end reliability of compute-intensive services are as follows: Control-related services have high priority, computationally intensive services have medium priority, and data acquisition services have low priority.

[0132] Network performance data was collected from NetLogo to train the GAT model and the multi-armed slot machine model. Specifically, 185 aggregation nodes simultaneously generated three types of services, with each aggregation node transmitting between 10 and 100 services of a single type. Performance values ​​on the end-to-end path between the 185 source and destination nodes were statistically analyzed and calculated. By combining different node pairs and the number of services sent, six network performance metrics were generated: betweenness centrality, hierarchy, load factor, data transmission rate, latency, and reliability. Based on 100,000 collected data points, the GAT model learned the mapping relationship between the first three performance metrics and the optimal next hop. The three multi-armed slot machine models learned the mapping relationship between data transmission rate, latency, and reliability, respectively, and link type, channel, and data modulation method. Both models were implemented using the deep learning framework PyTorch and Python 3.8. During model training, the GAT iteration count was set to 2000. Two self-attention layers were set in the GAT module to better predict the optimal next hop; the first layer had three attention heads and used the ELU function as the activation function. The second layer has one attention head and uses the ReLU function as the activation function. The number of iterations for the multi-armed slot machine is set to 1000.

[0133] Currently, the commonly used routing methods in large-scale heterogeneous power local communication networks are Shortest Path (ST) and Most Reliable Path (MR). Link, channel, and data modulation allocation methods include Round-Robin (RR) and Fixed Resource (FR). In RR, various resources are allocated to each node pair in turn across different time slices. If a service is not fully transmitted within a time slice, the resource is released and the service is placed at the end of the queue. In FR, the links, channels, and data modulation schemes of the entire network are pre-allocated and remain unchanged during service transmission. Therefore, four benchmark methods are considered: ST-FR, ST-FA, MR-FR, and MR-FA. Among the proposed methods and benchmark methods, the most important performance indicators are the difference between performance supply and demand and the proportion of services successfully transmitted. The proportion of services successfully transmitted is defined as the ratio of the maximum number of services that can be successfully transmitted within the corresponding end-to-end delay requirement to the total number of services of all types.

[0134] Figure 3 The performance supply-demand differences of five methods were compared under different service quantities, with the ratio of the three service types set at 1:1:1. It can be observed that when the total number of services is small, the performance supply-demand differences of the five methods are relatively close. As the total number of services increases, this value increases. This is because as the total number of services increases, the resource demands of various services also increase. However, the resources of the entire power communication network, such as transmission paths, transmission media, channels, and data modulation methods, are limited. During resource allocation and scheduling, only a portion of the service QoS requirements can be guaranteed on demand. Resources are also provided for other types of services, but the performance provided by these resources may differ significantly from the service QoS requirements. Compared with the four benchmark methods, RASMoE consistently has a smaller performance supply-demand difference. This is because RASMoE can dynamically combine service-specific QoS expert models through a gating network. Experts adaptively find the resource allocation and scheduling scheme that best matches the service QoS requirements based on the unique QoS requirements of each service, thereby reducing the supply-demand difference. The time-slice round-robin method has strong uncertainty. The resources allocated to a certain type of service in a time slice may perfectly meet the service QoS requirements, or they may allocate resources with poor performance to that type of service. The fixed-resource approach does not differentiate between business characteristics, resulting in abundant resources being provided to businesses of any priority, leading to resource waste.

[0135] Figure 4 , Figure 5 and Figure 6The diagram illustrates the successful transmission rates of three service types as the total number of services increases. It can be observed that, regardless of the service type, the successful transmission rate of all five methods decreases with the increase in the number of services. The fundamental reason is the limited resources in large-scale heterogeneous power local communication networks, including transmission paths, transmission media, channels, and data modulation methods. As the total number of services increases, packet loss, congestion, and retransmissions occur, severely impacting service QoS and leading to a decrease in the total number of services the network can successfully serve. However, compared to the four benchmark methods, RASMoE consistently achieves a higher successful transmission rate. RASMoE rapidly extracts node load information through graph attention networks and interacts with the network in real time using a multi-armed slot machine model, learning efficient and accurate resource allocation strategies to effectively cope with environmental and resource changes. Notably, all methods prioritize resources for higher-priority control services, which is the main reason why the successful transmission rate of control services is higher than that of computationally intensive services. Figure 6 Among them, the success rate of data acquisition services is higher than that of the other two types of services, and the transmission latency requirements for this type of service are mostly on the order of hours or days.

[0136] The technical means disclosed in this invention are not limited to those disclosed in the above embodiments, but also include technical solutions composed of any combination of the above technical features.

Claims

1. A routing and resource scheduling method for a large-scale heterogeneous local power communication network based on hybrid experts, characterized in that: The specific steps are as follows: Step S1, according to the differentiated QoS requirements of data transmission rate, delay and reliability of services, establish large-scale heterogeneous local communication network routing and resource scheduling optimization problem, jointly optimize the transmission path of each service and the transmission medium, channel and data modulation mode on the path, realize the minimum difference between the required performance of the service and the supply performance of the network; the step S1 includes 3 types of nodes, which are center node, convergence node and end node; 3 types of nodes integrate different communication modes according to actual scene; the center node and the convergence node are communicated through double mode, using HPLC link+RF link, the convergence nodes are communicated through double mode or single mode, using HPLC link or only RF link, the end node and the convergence node are communicated through low power consumption link; in order to form the large-scale heterogeneous local communication network routing and resource scheduling optimization problem described in step S1, the large-scale heterogeneous power local communication network is modeled as a graph structure wherein represents a node set, represents a link set, consisting of HPLC link set and wireless link set ; it is assumed that the set of all available channels in the network is wherein represents the HPLC channel set, represents the RF channel set; it is assumed that the set of all available data modulation modes in the network is ; Including binary phase shift keying, quadrature phase shift keying and 16 quadrature amplitude modulation; Let denote a set of services, 6-tuple denote a service with differentiated requirements, where and are the source node and the destination node of the service , , , and denote the size of the service in bits, the data transfer rate requirement, the end-to-end latency requirement and the end-to-end reliability requirement, respectively. make Indicates from the source node to the destination node The routing service aims to find an end-to-end path, including the transmission medium, channel, and data modulation scheme along that path. The sequence of forwarding nodes along this path is represented as follows: Let vector ,in 0-1 variables Used to represent business Select node? and between For a link of type , set the value to 1 if it is a link, otherwise set it to 0; let the vector... , where 0-1 variables Used to represent nodes and between Does the link use a channel? If yes, it is 1; otherwise, it is 0. Meanwhile, to select a suitable modulation scheme to maintain stable link transmission, a vector is defined. , where 0-1 variables Represents a node and between Channel on the link Select modulation method Based on the differences between the channel models of HPLC and wireless communication, using Represents a node and Signal-to-noise ratio of inter-HPLC links, Represents a node and The signal-to-noise ratio of the inter-wireless link, then the service At the node and between Type link channel The data transmission rate is ;in, for Type link channel Bandwidth; when services At the node and Once the transmission link, transmission channel, and data modulation method are determined, the transmission delay... Represented as ； in, Indicates business Size bits, Indicates the data modulation method; service slave node Transmit to node During the process, the nodes Called a node The parent node, at this time the node The hierarchy is higher than the node High; Business In the parent node Queuing delay at the location is ； where denotes the service arriving at the parent node the size of the service waiting for processing in its buffer queue, denotes the service processing rate of the parent node ; for compute-intensive services, it is first needed to determine which aggregation node to perform the computation before transmission; define a binary variable denotes whether the compute-intensive service chooses the node to perform the computation, 1 if yes, 0 otherwise; assume is the available computation capacity of the node , then the computation latency of the service at the node is expressed as ； wherein , is a set of nodes with computing resources; then the service end-to-end latency of the path is + ； where if the service does not require computational resources then ; therefore the service is transmitted from the node to the node the probability of occurrence of an error of one bit is expressed as ； in Indicating the modulation method The bit error rate is below; then the service At the node and Reliability during inter-transmission for ； Business End-to-end reliability of the path is ； To achieve the purpose of on-demand service under the premise of guaranteeing service QoS, the optimization target is defined as minimizing the performance supply-demand difference, i.e. the gap between the required performance of service and the supplied performance of network, which consists of three parts: data transmission rate, end-to-end delay and end-to-end reliability. The performance supply-demand difference is represented as ； wherein , and are adjustable weight coefficients, , ; The joint optimization of all services in the transmission path, transmission medium, transmission channel and data modulation mode in the large-scale heterogeneous local power communication network, the large-scale heterogeneous local power communication network routing and resource scheduling optimization problem is proposed : ； s.t. ; ； ； ； ； ； ； ； ； Where C1 represents the allocation to the business The data transmission rate of the link must be greater than the transmission rate required by the service; C2 indicates The total amount of traffic transmitted on a single link cannot exceed the maximum data transmission rate of that link. C3 indicates that the actual end-to-end latency experienced by the service cannot exceed the service's end-to-end latency requirement; C4 indicates that the end-to-end reliability of the path actually traversed by the service must be greater than the service's reliability requirement; C5 indicates that at the node... The computing resources required by the services being computed must not exceed the maximum computing capacity of that node. C6 indicates business At the node and Only one link type can be selected for inter-port transmission; C7 indicates the service. At the node and between Only one channel can be selected for transmission via a type-type link; C8 indicates the service. At the node and between Channel of type link Only one modulation scheme can be selected during uplink transmission; C9 indicates the service. Only one node can be selected for calculation; Step S2, based on the hybrid expert algorithm, the routing and resource scheduling optimization problem established in step S1 is solved, and the transmission path of each service and the path transmission medium, channel and data modulation mode scheme and the minimum value of the performance supply and demand difference are obtained.

2. The hybrid expert-based large-scale heterogeneous local power communication network routing and resource scheduling method according to claim 1, characterized in that: In step S2, a unified hybrid expert model is established to realize the relay node selection and the resource allocation and scheduling between node pairs; wherein taking the source node as an example, the hybrid expert model is composed of three sub-modules of expert network, gating network and tower network; the three networks are specifically: S21, the shared expert in the expert network based on graph attention network: the relay node selection problem is expressed as the probability that the first-order neighbor node of a certain node is selected as the optimal next hop; the shared expert model based on GAT actually constructs an optimal next hop selector, and the optimal next hop needs to be selected from the first-order neighbor nodes meeting the node characteristics; First, select 3 high-quality node features, namely load rate, betweenness centrality and level; the maximum and minimum normalization method is used for dimensionless processing of the data; Secondly, train a shared weight matrix for all nodes in the graph. This serves as a mapping between the input and output features of a node. During the computation, the weight matrix is ​​first used to map the features of each high-level neighbor node of the central node, and the results of these mappings are concatenated. Then, a feedforward neural network is input, mapping the concatenated vector to a single value. Finally, through... Activation function, obtain nodes right Importance : ; wherein is a concatenation operation; is an input of the GAT model, , is an input feature dimension; is a node shared weight matrix, is a weight matrix of the feed-forward neural network, is an output feature dimension; Finally, compute nodes Neighboring nodes The probability of being selected as the optimal next hop : ; wherein is a high-level first-order neighbor set of node , represents the importance of node to ; by learning the node information under different network conditions in depth, the self-attention layer outputs the probability that the neighbor node is selected as the optimal next hop; the neighbor node with the maximum probability is selected as the center node of the next self-attention layer; similarly, each self-attention layer generates the probability that the first-order neighbor node of the center node belongs to the optimal next hop, and the neighbor node with the maximum probability value is selected as the center node of the next self-attention layer, and the operation terminates when the first-order neighbor node of a certain center node contains the destination node. S22, the task-specific QoS expert in the expert network based on multi-armed bandit: three experts are used to learn the mapping between different service QoS requirements and the allocation of transmission media, channels and data modulation methods between node pairs, and the three experts use the same multi-armed bandit structure; wherein the transmission media between node pairs includes two kinds of HPLC and RF, which are combined with channel selection together, and the selected HPLC / RF channels and data modulation methods between node pairs are traversed in turn to construct the rocker arm while selecting the optimal next hop; the set of available HPLC and RF channels between node pairs is , the set of available modulation methods is , the set of available modulation methods is , and each rocker arm is represented as ； wherein, denotes the number of arms, and is an arm a channel and data modulation mode selected between node pairs, , In each iteration, 1 needs to be selected from arms to perform; suppose denotes the end-to-end path selected arm set, calculate the data transmission rate, delay and reliability corresponding to each arm, and set the optimization target given in as the return ; ； Using incremental form At the current iteration number Expected return valuation at that time Perform efficient updates ； in It is the number of iterations. Expected return valuation at that time It is a rocker arm The return express The number of executions; in addition, decay is used. - Greedy strategy for selecting the rocker arm Exploring Probability The update method is ； wherein, denotes an initial value of the exploration probability, denotes a decay coefficient; it is assumed that the maximum number of iterations is , the rocker arm is shaken in each iteration is chosen in the following way: ； S23, the gating network is a softmax classifier, and the output is a probability distribution representing the probability of each expert being selected; S24, the tower network is composed of fully connected network FCN layers; the FCN layer is composed of an independent network layer structure; then, the results of the experts are forwarded to the FCN network of a specific service; the FCN layer uses the aggregated information to perform service-specific processing and generate the final output; each service is associated with its own dedicated FCN layer, and the ReLU function is selected as the activation function.

Images