Large-scale heterogeneous local power communication network routing and resource scheduling method based on hybrid experts

Abstract

The invention discloses a large-scale heterogeneous local power communication network routing and resource scheduling method based on hybrid experts, and the method comprises the following steps: S1, building a large-scale heterogeneous local communication network routing and resource scheduling optimization problem according to the differentiated QoS demands of services for the data transmission rate, time delay and reliability, a transmission path of each service and a transmission medium, a channel and a data modulation mode on the path are jointly optimized, and the minimum difference between the performance required by the service and the network supply performance is achieved. And S2, based on a hybrid expert algorithm, solving the routing and resource scheduling optimization problem established in the step S1, and obtaining a transmission path of each service, a transmission medium on the path, a channel, a data modulation mode scheme and a performance supply and demand difference minimum value.

Description

Technical Field

[0001] The present invention relates to the technical field of communication networks, and in particular to a hybrid expert-based large-scale heterogeneous local power communication network routing and resource scheduling method. Background Art

[0002] Under the background of energy transformation, the high proportion of access to distributed clean energy is leading to a profound change in the dispatching and operation mode of power grids. The power grid has evolved from traditional single dispatching to a complex system of coordinated control of source, grid, load and storage, and multi-level coordination of transmission and distribution microgrids. This transformation has not only spawned a series of innovative power system application scenarios such as distributed new energy grid connection and consumption, vehicle-grid interaction, and virtual power plants, but also greatly enriched the connotation of power business. Today, power business not only covers traditional fields such as distribution automation, but also widely involves new businesses such as new energy output forecasting and charging pile assistance for peak shaving and valley filling. These businesses show significant diversity and differences in terms of generation cycle and QoS requirements, and put forward more stringent requirements on the performance of power communication networks, such as ultra-large bandwidth, ultra-high speed, ultra-low latency and ultra-high reliability. In addition, with the large-scale access of new power electronic equipment, the number of power grid terminals has increased dramatically, and "response" has gradually extended to the end of the distribution network. Information collection and information exchange between platforms will grow explosively, and the demand for high-speed collaborative interaction between devices will become increasingly high. It is necessary to rely on ubiquitous and flexible local power communication networks to provide real-time response speed and wide coverage access capabilities. However, faced with limited communication resources and growing business performance requirements, how to provide on-demand services and improve network capacity while ensuring business QoS has become a huge challenge facing local power communication networks.

[0003] At present, the heterogeneous data communication architecture integrating power line carrier (PLC) and wireless communication is a very promising means to meet the ubiquitous connection, ultra-high reliability and large-capacity communication requirements of large-scale heterogeneous power local communication networks. Based on this architecture, many scholars have proposed corresponding routing and resource scheduling schemes. However, both optimization-based or heuristic resource allocation and scheduling methods require precise mathematical models and parameters. With the increase of network scale and optimization variables, both methods have problems of difficulty and time consumption in solving, and therefore cannot adapt well to the increasing application scenarios of electricity, the complexity and uncertainty of network topology and network resources.

[0004] In recent years, deep reinforcement learning (DRL) has attracted widespread attention due to its outstanding performance in dealing with the state space dimensional disaster and the dynamic change of the environment. It is achieved by the agent constantly exploring and trial-and-error in the environment, and adjusting its behavior strategy according to the immediate rewards obtained, in order to obtain the maximum cumulative return. However, due to the large number of local power communication network terminals, the action space of the existing DRL method is too large, the computational complexity is too high, the method is difficult to converge, and it is difficult to apply to actual scenarios. In addition, for large-scale multi-service problems, training all services together in the same DRL model may lead to negative transfer and reduce the performance of the method.

[0005] Recent advances in the Mixture of Experts (MoE) framework provide an effective solution that significantly reduces the need for model training specific to a single business by employing a collection of AI models as specialized experts. The Mixture of Experts (MoE) framework is a simple yet effective ensemble architecture that can effectively switch between AI models in the ensemble based on their online performance. Summary of the invention

[0006] To solve the above problems, the present invention discloses a large-scale heterogeneous local power communication network routing and resource scheduling method based on hybrid experts, which can meet the differentiated QoS requirements of power services as much as possible in the resource-constrained local power communication network; it can process diversified power services in parallel and improve the effectiveness of resource allocation.

[0007] The technical solution adopted by the present invention is:

[0008] Step S1: Based on the differentiated QoS requirements of the services for data transmission rate, latency, and reliability, a large-scale heterogeneous local communication network routing and resource scheduling optimization problem is established, and the transmission path of each service and the transmission medium, channel, and data modulation method on the path are jointly optimized 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 and the transmission medium, channel and data modulation scheme on the path as well as the minimum performance supply and demand difference.

[0010] In the aforementioned large-scale heterogeneous local communication network routing and resource scheduling method based on hybrid experts, in step S1, the large-scale heterogeneous local communication network includes three types of nodes, namely central nodes, aggregation nodes and end nodes. The three types of nodes include high-speed power line carrier (High PLC, HPLC) links, wireless links (Wireless Radio Frequency, RF) and low-power links (such as Lora, WiFi, Zigbee, etc.). The central node and the aggregation node communicate through dual-mode communication (HPLC link + RF link), the aggregation nodes communicate through dual-mode communication or single-mode communication (only HPLC link or only RF link), and the end nodes and the aggregation nodes communicate through low-power links. Model the large-scale heterogeneous power local communication network as a graph structure. in Represents a collection of nodes. represents the link set, ε is composed of the HPLC link set and wireless link set Assume that the set of all available channels in the network is in represents the HPLC channel set, represents the set of RF channels. Assume that the set of all available data modulation modes in the network is Including Binary Phase Shift Keying (BPSK), Quadrature Phase Shift Keying (QPSK) and Hexadecimal Quadrature Amplitude Modulation (16QAM).

[0011] make Indicates a business set, 5-tuple represents the differentiated requirements of business l, where and are the source node and destination node of service l, D l , R l , T l and Rel l represents the size (bits), data transmission rate requirement, end-to-end delay requirement, and end-to-end reliability requirement of service l. From the source node To the destination node The routing service aims to find an end-to-end path and the transmission medium, channel and data modulation method on the path. The forwarding node sequence of the path can be expressed as Let vector Where Λ∈{PLC,RF}, 0-1 variable It is used to indicate whether service l selects the Λ-type link between nodes i and j. If yes, it is 1, otherwise it is 0. Let vector Among them, 0-1 variables It is used to indicate whether the Λ-type link between nodes i and j uses channel c. If yes, it is 1, otherwise it is 0. At the same time, define the vector Among them, 0-1 variables Indicates whether the channel c on the Λ-type link between nodes i and j selects the modulation mode M. According to the difference between the channel models of HPLC and wireless communication, represents the signal-to-noise ratio of the HPLC link between nodes i and j, represents the signal-to-noise ratio of the wireless link between nodes i and j, and the data transmission rate of service l in the Λ-type link channel c between nodes i and j is

[0012]

[0013] in, is the bandwidth of the Λ-type link channel c. When the transmission link, transmission channel and data modulation method of service l between nodes i and j are determined, its transmission delay for

[0014]

[0015] Among them, D l represents the size of service l (bits), and M represents the data modulation mode. When a service is transmitted from node i to node j, node j is called the parent node of node i. At this time, the level of node j is higher than that of node i. The queuing delay of service l at parent node j is

[0016]

[0017] in V represents the size of the business waiting to be processed in the cache queue of parent node j when business l arrives at the parent node j. j Represents the service processing rate of parent node j. For computing-intensive services, define a binary variable μ l,v Indicates whether the computation-intensive service l selects node v for computation. If yes, it is 1, otherwise it is 0. Assume that F v is the available computing power of node v (such as the number of CPU cycles per second), then the computing delay of service l at node v is

[0018] in is a set of nodes with computing resources. Then business l passes through the path End-to-end delay for

[0019]

[0020] If business l does not require computing resources, Usually, the bit error distribution in reality is close to the Poisson distribution, so the probability of f bit errors when a service l is transmitted from node i to node j is

[0021]

[0022] Where P M represents the bit error rate under modulation mode M. Then the reliability of service l when transmitted between nodes i and j is Rel l,ij for

[0023]

[0024] Business l Passing Path End-to-end reliability for

[0025]

[0026] The optimization goal of the large-scale heterogeneous local communication network routing and resource scheduling optimization problem is to minimize the performance supply and demand difference, that is, the gap between the performance required by the business and the performance provided by the network, which consists of three parts: data transmission rate, end-to-end delay and end-to-end reliability. l for

[0027]

[0028] Among them, w1, w2 and w3 are adjustable weight coefficients. When service l is more sensitive to delay, w2 is increased, and when service l has higher requirements on reliability, w3 is increased. w1, w2, w3∈[0,1], w1+w2+w3=1.

[0029] The transmission paths, transmission media, transmission channels and data modulation methods of all services in the large-scale heterogeneous local power communication network are jointly optimized. The routing and resource scheduling optimization problem of the large-scale heterogeneous local power communication network is proposed as follows:

[0030] P1:

[0031] stC1:

[0032] C2:

[0033] C3:

[0034] C4:

[0035] C5:

[0036] C6:

[0037] C7:

[0038] C8:

[0039] C9: Among them, C1, C3 and C4 represent the link data transmission rate, end-to-end delay and reliability requirements of service 1. C2 indicates that the total amount of services transmitted on the Λ-type link cannot exceed the maximum data transmission rate of the link. C5 indicates that the computing resource requirements of the business calculated on node v cannot exceed the maximum computing capacity F of the node. v,max . C6 indicates that service l can only select one link type when transmitting between nodes i and j. C7 indicates that service l can only select one channel when transmitting on a Λ-type link between nodes i and j. C8 indicates that service l can only select one modulation mode when transmitting on channel c of a Λ-type link between nodes i and j. C9 indicates that service l can only select one node for calculation.

[0040] In the aforementioned large-scale heterogeneous local communication network routing and resource scheduling method based on hybrid experts, in step S2, a unified hybrid expert (MoE) model is established to simultaneously realize 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: expert network, gate control network and tower network. These three networks are specifically:

[0041] S21. Shared experts based on graph attention networks (GATs) in expert networks: An optimal next-hop selector is constructed through a shared expert model based on GAT. The optimal next hop needs to be selected from the first-order neighbor nodes that meet the node characteristics. First, three high-quality node features are selected, namely load rate, betweenness centrality, and hierarchy. Due to the different dimensions between features, the feature data cannot be directly input into the GAT model for training. Betweenness centrality is a benefit-type feature, and the larger the value, the better the network performance. Load rate and hierarchy are cost-type features, and the smaller the value, the better the network performance. Therefore, the maximum and minimum normalization method is used to dimensionlessly process the data.

[0042] Secondly, a shared weight matrix Y for all nodes in the graph is trained as the mapping between the input features of the nodes and the output features. In the calculation process, the weight matrix is ​​first used to perform feature mapping on each high-level neighbor node of the central node and the results of the feature mapping of each node are concatenated. Then a feedforward neural network is input to map the concatenated vector to a value. Then, the LeakyReLU activation function is used to obtain the importance of node j to i, ρ ij :

[0043]

[0044] Among them, || is a concatenation operation. is the input of the GAT model, U = 3, and F is the input feature dimension. F×F′ is the node shared weight matrix, is the weight matrix of the feedforward neural network, and F' is the output feature dimension.

[0045] Finally, calculate the probability that node i's neighboring node j is selected as the optimal next hop

[0046]

[0047] in is the high-level first-order neighbor set of node i (including i), ρ ij Indicates the importance of node j to i. The self-attention layer outputs the probability of the neighboring node being selected as the optimal next hop, and selects the neighboring node with the largest probability as the center node of the next self-attention layer. Similarly, each self-attention layer generates the probability that the first-order neighboring node of the center node belongs to the optimal next hop, and the neighboring node with the largest probability value is selected as the center node of the next self-attention layer, until the operation terminates when the first-order neighboring node of a certain center node contains the destination node.

[0048] S22. Task-specific QoS experts based on multi-armed bandit in expert network: Three experts are set to learn the mapping relationship between different service QoS requirements and the transmission medium, channel and data modulation mode allocation between node pairs. These three experts use the same multi-armed bandit structure. The transmission medium between node pairs includes HPLC and RF. They are combined with channel selection. While selecting the optimal next hop, the optional HPLC / RF channels and data modulation modes between node pairs are traversed in turn to construct the rocker. The set of HPLC and RF channels available between node pairs is The selected modulation mode set is Then the rocker arm set Arm and each rocker arm p for

[0049] Arm={arm1,arm2,...,arm P},arm p =(c,m)

[0050] in, Indicates the number of rocker arms, c and m are rocker arm p The channel and data modulation scheme selected between node pairs, In each iteration, one of the P rockers needs to be selected for execution. Assume Represents an end-to-end path The set of rocker arms selected above is used to calculate the data transmission rate, delay and reliability corresponding to each rocker arm, and the optimization target of the routing and resource scheduling optimization problem in claim 2 is set as the reward

[0051]

[0052] Incremental arm p The expected return value Q at the current iteration number t t (arm p ) to update

[0053]

[0054] Where Q t-1 (arm p ) is the expected return at iteration t-1, R(arm p ) is the rocker arm p The return, count(arm p ) represents arm p In addition, the decaying ∈-greedy strategy is used to select the rocker arm p , the exploration probability ∈(t) is updated as

[0055]

[0056] Among them, ∈(0) represents the initial value of the exploration probability, and χ represents the decay coefficient. Assume that the maximum number of iterations is T max , in each iteration arm p The selected methods are:

[0057]

[0058] When T max When it is much larger than P, it is possible to traverse all rocker arms and find the optimal rocker arm.

[0059] Analysis shows that the arm selected for execution in each iteration is closely related to ∈(t); as the number of iterations t increases, ∈(t) gradually decreases, and the exploration rate gradually slows down. The multi-armed bandit model gradually transitions from the initial bias towards exploration to the bias towards using the known optimal arm for execution, but it will still try new arms with a small probability. Therefore, when T max When it is much larger than P, it is basically possible to traverse all rocker arms and find the optimal rocker arm;

[0060] S23, the gating network is a softmax classifier that outputs the probability distribution of each expert being selected. For computationally intensive services with multiple QoS requirements, which require data transmission rate, latency, and reliability, each expert is selected.

[0061] S24, the tower network consists of a fully connected network (FCN) layer. The FCN layer, as the final output module of the model, consists of an independent network layer structure. The results of the collective experts are then forwarded to the FCN network of the specific business. The FCN layer uses the aggregated information to perform business-specific processing and generate the final output. Each business is associated with its own dedicated FCN layer and selects the ReLU function as the activation function.

[0062] Beneficial effects of the present invention:

[0063] 1. Aiming at the difficulties such as the surge in demand for communication applications in local electric power communication networks, differentiated QoS requirements for services, and limited communication resources of various types, the present invention establishes a method for minimizing the performance supply and demand differences in large-scale heterogeneous local electric 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 electric power communication networks, it is expected that the resources allocated to the services will just meet the QoS requirements of the services, so as to provide services for more services under limited resource conditions and increase the total number of successful transmissions of the entire network services.

[0064] 2. At the same time, a hybrid expert model including an expert network and a gated network is designed. Through different expert models, routing and resource scheduling strategies are learned specifically and in parallel to meet the personalized needs of diversified services for data transmission rate, latency and reliability, and achieve faster convergence speed. Compared with various baseline methods, the proposed method has better performance in terms of performance supply and demand differences and business successful transmission ratio. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 It is a hybrid expert-based resource allocation and scheduling framework for large-scale heterogeneous local power communication networks;

[0066] Figure 2 It is the instantiation of LPCN node distribution;

[0067] Figure 3 is the difference between performance supply and demand under different business quantities;

[0068] Figure 4 is the ratio of successful transmission of control services;

[0069] Figure 5 is the ratio of successful transmission of computing-intensive services;

[0070] Figure 6 It is the successful transmission ratio of collection services. DETAILED DESCRIPTION

[0071] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. It should be noted that the words "front", "rear", "left", "right", "upper" and "lower" used in the following description refer to directions in the accompanying drawings, and the words "inner" and "outer" refer to directions toward or away from the geometric center of a specific component, respectively.

[0072] This embodiment proposes a hybrid expert-based large-scale heterogeneous local power communication network routing and resource scheduling method, including the following steps:

[0073] Step S1: Based on the differentiated QoS requirements of the services for data transmission rate, latency, and reliability, a large-scale heterogeneous local communication network routing and resource scheduling optimization problem is established, and the transmission path of each service and the transmission medium, channel, and data modulation method on the path are jointly optimized to minimize the difference between the performance required by the service and the performance provided by the network.

[0074] 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 and the transmission medium, channel and data modulation scheme on the path as well as the minimum performance supply and demand difference.

[0075] Step S3: Evaluate the performance of the proposed method under different service numbers and different service QoS requirements.

[0076] In step S1, the large-scale heterogeneous local communication network includes three types of nodes, namely central nodes, aggregation nodes and end nodes. The three types of nodes include high-speed power line carrier (High PLC, HPLC) links, wireless links (Wireless Radio Frequency, RF) and low-power links (such as Lora, WiFi, Zigbee, etc.). The central node and the aggregation node communicate through dual-mode communication (HPLC link + RF link), the aggregation nodes communicate through dual-mode communication or single-mode communication (only HPLC link or only RF link), and the end nodes and the aggregation nodes communicate through low-power links. The large-scale heterogeneous power local communication network is modeled as a graph structure. in Represents a collection of nodes. represents the link set, ε is composed of the HPLC link set and wireless link set Assume that the set of all available channels in the network is in represents the HPLC channel set, represents the set of RF channels. Assume that the set of all available data modulation modes in the network is Including Binary Phase Shift Keying (BPSK), Quadrature Phase Shift Keying (QPSK) and Hexadecimal Quadrature Amplitude Modulation (16QAM).

[0077] make Indicates a business set, 5-tuple represents the differentiated requirements of business l, where and are the source node and destination node of service l, D l , R l , T l and Rel l represents the size (bits), data transmission rate requirement, end-to-end delay requirement, and end-to-end reliability requirement of service l. From the source node To the destination node The routing service aims to find an end-to-end path and the transmission medium, channel and data modulation method on the path. The forwarding node sequence of the path can be expressed as Let vector Where Λ∈{PLC,RF}, 0-1 variable It is used to indicate whether service l selects the Λ-type link between nodes i and j. If yes, it is 1, otherwise it is 0. Let vector Among them, 0-1 variables It is used to indicate whether the Λ-type link between nodes i and j uses channel c. If yes, it is 1, otherwise it is 0. At the same time, define the vector Among them, 0-1 variables Indicates whether the channel c on the Λ-type link between nodes i and j selects the modulation mode M. According to the difference between the channel models of HPLC and wireless communication, represents the signal-to-noise ratio of the HPLC link between nodes i and j, represents the signal-to-noise ratio of the wireless link between nodes i and j, and the data transmission rate of service l in the Λ-type link channel c between nodes i and j is

[0078]

[0079] in, is the bandwidth of the Λ-type link channel c. When the transmission link, transmission channel and data modulation method of service l between nodes i and j are determined, its transmission delay Expressed as

[0080]

[0081] Among them, D l represents the size of service l (bits), and M represents the data modulation mode. When a service is transmitted from node i to node j, node j is called the parent node of node i. At this time, the level of node j is higher than that of node i. The queuing delay of service l at parent node j is

[0082]

[0083] in V represents the size of the business waiting to be processed in the cache queue of parent node j when business l arrives at the parent node j. j Represents the service processing rate of parent node j. For computing-intensive services, define a binary variable μ l,v Indicates whether the computation-intensive service l selects node v for computation. If yes, it is 1, otherwise it is 0. Assume that F v is the available computing power of node v (such as the number of CPU cycles per second), then the computing delay of service l at node v is

[0084]

[0085] in is a set of nodes with computing resources. Then business l passes through the path End-to-end delay for

[0086]

[0087] If business l does not require computing resources, Usually, the bit error distribution in reality is close to the Poisson distribution, so the probability of f bit errors when a service l is transmitted from node i to node j is

[0088]

[0089] Where P M represents the bit error rate under modulation mode M. Then the reliability of service l when transmitted between nodes i and j is Rel l,ij for

[0090]

[0091] Business l Passing Path End-to-end reliability for

[0092]

[0093] The optimization goal of the large-scale heterogeneous local communication network routing and resource scheduling optimization problem is to minimize the performance supply and demand difference, that is, the gap between the performance required by the business and the performance provided by the network, which consists of three parts: data transmission rate, end-to-end delay and end-to-end reliability. l for

[0094]

[0095] Among them, w1, w2 and w3 are adjustable weight coefficients. When service l is more sensitive to delay, w2 is increased, and when service l has higher requirements on reliability, w3 is increased. w1, w2, w3∈[0,1], w1+w2+w3=1.

[0096] The transmission paths, transmission media, transmission channels and data modulation methods of all services in the large-scale heterogeneous local power communication network are jointly optimized. The routing and resource scheduling optimization problem of the large-scale heterogeneous local power communication network is proposed as follows:

[0097] P1:

[0098] stC1:

[0099] C2:

[0100] C3:

[0101] C4:

[0102] C5:

[0103] C6:

[0104] C7:

[0105] C8:

[0106] C9:

[0107] Among them, C1, C3 and C4 represent the link data transmission rate, end-to-end delay and reliability requirements of service 1. C2 indicates that the total amount of services transmitted on the Λ-type link cannot exceed the maximum data transmission rate of the link. C5 indicates that the computing resource requirements of the business calculated on node v cannot exceed the maximum computing capacity F of the node. v,max . C6 indicates that service l can only select one link type when transmitting between nodes i and j. C7 indicates that service l can only select one channel when transmitting on a Λ-type link between nodes i and j. C8 indicates that service l can only select one modulation mode when transmitting on channel c of a Λ-type link between nodes i and j. C9 indicates that service l can only select one node for calculation.

[0108] In step S2, a unified Mixture of Experts (MoE) model is established to simultaneously implement relay node selection and resource allocation and scheduling between node pairs. Figure 1 As shown in the figure, the blue, purple and pink boxes represent different nodes that perform hybrid experts on the end-to-end path. Taking the source node as an example, the hybrid expert model consists of three sub-modules: expert network, gate network and tower network. These three networks are:

[0109] S21. Shared experts based on graph attention networks (GATs) in expert networks: An optimal next-hop selector is constructed through a shared expert model based on GAT. The optimal next hop needs to be selected from the first-order neighbor nodes that meet the node characteristics. First, three high-quality node features are selected, namely load rate, betweenness centrality, and hierarchy. Due to the different dimensions between features, the feature data cannot be directly input into the GAT model for training. Betweenness centrality is a benefit-type feature, and the larger the value, the better the network performance. Load rate and hierarchy are cost-type features, and the smaller the value, the better the network performance. Therefore, the maximum and minimum normalization method is used to dimensionlessly process the data.

[0110] Secondly, a shared weight matrix Y for all nodes in the graph is trained as the mapping between the input features of the nodes and the output features. In the calculation process, the weight matrix is ​​first used to perform feature mapping on each high-level neighbor node of the central node and the results of the feature mapping of each node are concatenated. Then a feedforward neural network is input to map the concatenated vector to a value. Then, the LeakyReLU activation function is used to obtain the importance of node j to i, ρ ij :

[0111]

[0112] Among them, || is a concatenation operation. is the input of the GAT model, U = 3, and F is the input feature dimension. F×F′ is the node shared weight matrix, is the weight matrix of the feedforward neural network, and F' is the output feature dimension.

[0113] Finally, calculate the probability that node i's neighboring node j is selected as the optimal next hop

[0114]

[0115] in is the high-level first-order neighbor set of node i (including i), ρ ij Indicates the importance of node j to i. The self-attention layer outputs the probability of the neighboring node being selected as the optimal next hop, and selects the neighboring node with the largest probability as the center node of the next self-attention layer. Similarly, each self-attention layer generates the probability that the first-order neighboring node of the center node belongs to the optimal next hop, and the neighboring node with the largest probability value is selected as the center node of the next self-attention layer, until the operation terminates when the first-order neighboring node of a certain center node contains the destination node.

[0116] S22. Task-specific QoS experts based on multi-armed bandit in expert network: Three experts are set to learn the mapping relationship between different service QoS requirements and the transmission medium, channel and data modulation mode allocation between node pairs. These three experts use the same multi-armed bandit structure. The transmission medium between node pairs includes HPLC and RF. They are combined with channel selection. While selecting the optimal next hop, the optional HPLC / RF channels and data modulation modes between node pairs are traversed in turn to construct the rocker. The set of HPLC and RF channels available between node pairs is The selected modulation mode set is Then the rocker arm set Arm and each rocker arm p for

[0117] Arm={arm1,arm2,...,arm P},arm p =(c,m)

[0118] in, Indicates the number of rocker arms, c and m are rocker arm p The channel and data modulation scheme selected between node pairs, In each iteration, one of the P rockers needs to be selected for execution. Assume Represents an end-to-end path The set of rocker arms selected above is used to calculate the data transmission rate, delay and reliability corresponding to each rocker arm, and the optimization target of the routing and resource scheduling optimization problem in claim 2 is set as the reward

[0119]

[0120] Incremental arm p The expected return value Q at the current iteration number t t (arm p ) to update

[0121]

[0122] Where Q t-1 (arm p ) is the expected return at iteration t-1, R(arm p ) is the rocker arm p The return, count(arm p ) represents arm p In addition, the decaying ∈-greedy strategy is used to select the rocker arm p , the exploration probability ∈(t) is updated as

[0123]

[0124] Among them, ∈(0) represents the initial value of the exploration probability, and χ represents the decay coefficient. Assume that the maximum number of iterations is T max , in each iteration arm p The selected methods are:

[0125]

[0126] When T max When it is much larger than P, it is possible to traverse all rocker arms and find the optimal rocker arm.

[0127] S23, the gating network is a softmax classifier that outputs the probability distribution of each expert being selected. For computationally intensive services with multiple QoS requirements, which require data transmission rate, latency, and reliability, each expert is selected.

[0128] S24, the tower network consists of a fully connected network (FCN) layer. The FCN layer, as the final output module of the model, consists of an independent network layer structure. The results of the collective experts are then forwarded to the FCN network of the specific business. The FCN layer uses the aggregated information to perform business-specific processing and generate the final output. Each business is associated with its own dedicated FCN layer and selects the ReLU function as the activation function.

[0129] In step S3, NetLogo3.6 is used to simulate the network. The overall layout of the local power communication network is in an area of ​​150*150m 2 The square area consists of 1 central node and 185 aggregation nodes. The specific network topology is as follows: Figure 2 As shown in the figure. Three types of services in the simulated intelligent distribution network are control services, collection services, and computation-intensive services. The destination nodes of control services and collection services are central nodes, and the source nodes are randomly generated from 185 aggregation nodes. The destination node of computation-intensive services is the central node or the aggregation node with computing resources. 50 nodes are randomly selected from the 185 aggregation nodes to be given computing resources, and the remaining aggregation nodes are used as the source nodes of this type of service. The size, end-to-end delay, and end-to-end reliability requirements of control services are {1kbits, 20ms, 99.9999%}. The size and data transmission rate requirements of collection services are {3Mbits, 1Mbps, 99.999%}. The size, data transmission rate, end-to-end delay, and end-to-end reliability requirements of computation-intensive services are {5kbits, 1Mbps, 50ms, 99.9999%}. Control services have high priority, computation-intensive services have medium priority, and collection services have low priority.

[0130] Network performance data was collected from NetLogo to train the GAT model and the multi-armed bandit model. Specifically, 185 aggregation nodes generated three types of services at the same time. The number of single-type services to be transmitted by each aggregation node was set to 10 to 100 in turn. The performance values ​​of the 185 source-destination node pairs on the end-to-end path were counted and calculated. By combining different node pairs and the number of services sent, six network performance indicators, namely betweenness centrality, hierarchy, load rate, data transmission rate, delay, and reliability, were generated. Based on the 100,000 data collected, the GAT model learned the mapping relationship between the first three performance indicators and the optimal next hop, and the three multi-armed bandit models learned the mapping relationship between the three performance indicators of data transmission rate, delay, and reliability and the link type, channel, and data modulation mode. Both models were implemented based on the deep learning framework Pytorch and Python3.8. During model training, the number of iterations of GAT was set to 2000. We set two self-attention layers in the GAT module to better predict the optimal next hop. The number of attention heads in the first layer was 3, and the activation function was the ELU function. The number of attention heads in the second layer is 1, and the activation function is the Relu function. The number of iterations of the multi-armed bandit machine is set to 1000.

[0131] At present, the routing methods commonly used in large-scale heterogeneous power local communication networks are the shortest path (ST) and the most reliable path (MR). The allocation methods of links, channels and data modulation modes include round-robin (RR) and fixed resources (FR). In RR, various resources are allocated to each node pair in turn in different time slices. If the service is not transmitted in a time slice, the resources are released and the service is placed at the end of the queue. In FR, the links, channels and data modulation modes of the entire network are allocated in advance and remain unchanged during service transmission. Therefore, four benchmark methods are considered: ST-FR, ST-FA, MR-FR, and MR-FA. In the proposed method and the benchmark method, the most concerned performance indicators are the difference between performance supply and demand and the proportion of services successfully transmitted by the network. The proportion of successfully transmitted services is defined as the ratio of the maximum number of services that can be successfully transmitted for each type of service within the corresponding end-to-end delay requirement to the total number of services of each type.

[0132] Figure 3The performance supply and demand differences of the five methods under different numbers of services are compared, and the ratio of the number of three types of services is set to 1:1:1. It can be observed that when the total number of services is small, the performance supply and demand differences of the five methods are relatively close. As the total number of services increases, the value increases. This is because as the total number of services increases, the demand for resources of various types of services is also increasing, but the resources of the entire power communication network, such as transmission paths, transmission media, channels, and data modulation methods, are limited. In the resource allocation and scheduling process, only part of the service QoS requirements can be guaranteed on demand, and resources are also provided for the remaining types of services, but the performance provided by these resources may be quite different from the service QoS requirements. Compared with the four benchmark methods, the performance supply and demand difference value of RASMoE is always smaller. This is because RASMoE can dynamically combine expert models specific to service QoS through a gating network. Experts can adaptively find the resource allocation and scheduling scheme that is closest to the service QoS requirements based on the unique QoS requirements of each service, thereby reducing the difference between supply and demand. The time slice rotation method has strong uncertainty. The resources allocated to a certain type of service in a time slice may just meet the service QoS requirements, or resources in poor condition may be allocated to this type of service. The fixed resource method does not distinguish between business characteristics, resulting in relatively abundant resources being provided to businesses of any priority, causing a waste of resources.

[0133] Figure 4 , Figure 5 and Figure 6 It shows the proportion of three types of services that can be successfully transmitted as the total number of services changes. It can be observed that regardless of the type of service, as the number of services continues to increase, the proportion of services that can be successfully transmitted by the five methods shows a downward trend. The most fundamental reason is that resources such as transmission paths, transmission media, channels, and data modulation methods are limited in large-scale heterogeneous power local communication networks. As the total number of services increases, packet loss, congestion, retransmission, and other conditions that seriously affect the QoS of services will occur in the network, resulting in a decrease in the total number of services that the network can successfully serve. However, compared with the four baseline methods, RASMoE always has a higher proportion of services that can be successfully transmitted. RASMoE quickly extracts node load rate information through a graph attention network, and uses a multi-armed bandit model to interact with the network in real time. It can learn efficient and accurate resource allocation strategies to effectively respond to environmental and resource changes. It is worth noting that various methods first provide resources for control services with higher priority, which is also the main reason why the proportion of successful transmission of control services is higher than that of compute-intensive services. Figure 6 Among them, the proportion of successful transmission of collection services is higher than that of the other two types of services. The transmission delay of this type of service is mostly required to be in the order of hours or days. It is not sensitive to time and only pursues reliability during service transmission. Therefore, the demand for resources is not stringent, and a large number of services can be successfully transmitted. The technical means disclosed in the scheme of the present invention are not limited to the technical means disclosed in the above-mentioned implementation mode, but also include technical schemes composed of any combination of the above technical features.

Claims

1. A large-scale heterogeneous local power communication network routing and resource scheduling method based on hybrid experts, characterized by: The steps are as follows: Step S1: Based on the differentiated QoS requirements of services for data transmission rate, latency, and reliability, a large-scale heterogeneous local communication network routing and resource scheduling optimization problem is established, and the transmission path of each service and the transmission medium, channel, and data modulation method on the path are jointly optimized to minimize the difference between the performance required by the service and the performance provided by the network; 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 and the transmission medium, channel and data modulation scheme on the path as well as the minimum performance supply and demand difference.

2. The hybrid expert-based large-scale heterogeneous local power communication network routing and resource scheduling method according to claim 1, characterized in that: The step S1 includes three types of nodes, namely, a central node, a sink node and an end node; the three types of nodes integrate different communication modes according to actual scenarios; the central node and the sink node communicate via dual-mode, using HPLC link + RF link, the sink nodes communicate via dual-mode or single-mode, using HPLC link or only RF link, and the end node and the sink node communicate via a low-power link.

3. The hybrid expert-based large-scale heterogeneous local power communication network routing and resource scheduling method according to claim 2 is characterized by: 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 in Represents a collection of nodes. represents the link set, ε is composed of the HPLC link set and wireless link set Composition; Assume that the set of all available channels in the network is in represents the HPLC channel set, represents the set of RF channels; assuming that the set of all available data modulation modes in the network is Including binary phase shift keying, quadrature phase shift keying and hexadecimal quadrature amplitude modulation; make Indicates a business set, 5-tuple represents the differentiated requirements of business l, where and are the source node and destination node of service l, D l , R l , T l and Rel l Indicates the size of service l in bits, data transmission rate requirement, end-to-end delay requirement, and end-to-end reliability requirement; make From the source node To the destination node The routing service aims to find an end-to-end path and the transmission medium, channel and data modulation method on the path. The forwarding node sequence of the path can be expressed as Let vector where Λ∈{PLC,RF}, a 0-1 variable It is used to indicate whether service l selects the Λ-type link between nodes i and j. If yes, it is 1, otherwise it is 0. Let vector Among them, 0-1 variables It is used to indicate whether the Λ-type link between nodes i and j uses channel c. If yes, it is 1, otherwise it is 0. At the same time, in order to select a suitable modulation method to keep the link transmission stable, the vector is defined Among them, 0-1 variables Indicates whether the channel c on the Λ-type link between nodes i and j selects the modulation mode M; according to the difference between the channel model of HPLC and wireless communication, represents the signal-to-noise ratio of the HPLC link between nodes i and j, represents the signal-to-noise ratio of the wireless link between nodes i and j, then the data transmission rate of service l in the Λ-type link channel c between nodes i and j is in, is the bandwidth of the Λ-type link channel c; when the transmission link, transmission channel and data modulation method of service l between nodes i and j are determined, its transmission delay Expressed as Among them, D l represents the size of service l in bits, M represents the data modulation mode; during the transmission of service from node i to node j, node j is called the parent node of node i, and the level of node j is higher than that of node i; the queuing delay of service l at parent node j is in V represents the size of the business waiting to be processed in the cache queue of parent node j when business l arrives at the parent node j. j Represents the service processing rate of parent node j; for computing-intensive services, it is necessary to first determine which aggregation node to perform the calculation before transmission; define a binary variable μ l,v Indicates whether the computation-intensive service l selects node v for computation. If yes, it is 1, otherwise it is 0. Assume that F v is the available computing capacity of node v, then the computing delay of service l at node v is expressed as in is a set of nodes with computing resources; then business l passes through the path End-to-end delay for If business l does not require computing resources, then The actual bit error distribution is close to the Poisson distribution, so the probability of f bit errors when a service l is transmitted from node i to node j can be expressed as Where P M represents the bit error rate under modulation mode M; then the reliability of service l when transmitted between nodes i and j is Rel l,ij for Business l Passing Path End-to-end reliability for In order to achieve the purpose of on-demand service while ensuring the QoS of the service, the optimization goal is defined as minimizing the performance supply and demand difference, that is, the gap between the performance required by the service and the performance provided by the network, which consists of three parts: data transmission rate, end-to-end delay and end-to-end reliability; the performance supply and demand difference S of service l l It can be expressed as Among them, w1, w2 and w3 are adjustable weight coefficients. When service l is more sensitive to delay, w2 can be increased. When service l has higher requirements for reliability, w3 can be increased. w1, w2, w3∈[0, 1], w1+w2+w3=1; The transmission paths, transmission media, transmission channels and data modulation methods of all services in the large-scale heterogeneous local power communication network are jointly optimized. The routing and resource scheduling optimization problem of the large-scale heterogeneous local power communication network is proposed as P1: s.t.C1: C2: C3: C4: C5: C6: C7: C8: C9: Among them, C1 means that the data transmission rate of the link allocated to service l must be greater than the transmission rate required by the service; C2 means that the total amount of service transmitted on the Λ-type link cannot exceed the maximum data transmission rate of the link. C3 means that the end-to-end delay actually experienced by the service cannot exceed the end-to-end delay requirement of the service; C4 means that the end-to-end reliability of the path actually experienced by the service must be greater than the service reliability requirement; C5 means that the demand for computing resources of the service calculated on node v cannot exceed the maximum computing capacity F of the node. v,max ; C6 means that service l can only select one link type when it is transmitted between nodes i and j; C7 means that service l can only select one channel when it is transmitted on the Λ-type link between nodes i and j; C8 means that service l can only select one modulation mode when it is transmitted on channel c of the Λ-type link between nodes i and j; C9 means that service l can only select one node for calculation.

4. 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 simultaneously realize 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: expert network, gate control network and tower network; these three networks are specifically: S21, shared experts based on graph attention network in expert network: the relay node selection problem can be expressed as the probability that the first-order neighbor node of a 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 that meet the node characteristics; First, three high-quality node features are selected, namely, load rate, betweenness centrality and hierarchy; the maximum and minimum normalization method is used to make the data dimensionless; Secondly, a shared weight matrix Y for all nodes in the graph is trained as the mapping between the input features of the nodes and the output features. In the calculation process, the weight matrix is ​​first used to perform feature mapping on each high-level neighbor node of the central node and the results of the feature mapping of each node are concatenated. Then a feedforward neural network is input to map the concatenated vector to a value. Then, the LeakyReLU activation function is used to obtain the importance of node j to i, ρ ij : Among them, || is the concatenation operation; is the input of the GAT model, U = 3, F is the input feature dimension; Y∈R F×F’ is the node shared weight matrix, is the weight matrix of the feedforward neural network, and F' is the output feature dimension; Finally, calculate the probability that node i's neighboring node j is selected as the optimal next hop in is the high-level first-order neighbor set of node i (including i), ρ ij Indicates the importance of node j to i; By deeply learning the node information under different network conditions, the self-attention layer will output the probability of the neighboring node being selected as the optimal next hop; the neighboring node with the highest probability is selected as the central node of the next self-attention layer; Similarly, each self-attention layer generates the probability that the first-order neighbor node of the central node belongs to the optimal next hop, and the neighbor node with the largest probability value is selected as the central node of the next self-attention layer, until the operation terminates when the first-order neighbor node of a central node contains the destination node; S22. Task-specific QoS experts based on multi-armed bandit in expert network: Three experts are used to learn the mapping relationship between different service QoS requirements and the transmission medium, channel and data modulation mode allocation between node pairs. These three experts adopt the same multi-armed bandit structure; the transmission medium between node pairs includes HPLC and RF, which are combined with channel selection. While selecting the optimal next hop, the optional HPLC / RF channels and data modulation modes between node pairs are traversed in turn to construct the rocker arm; the set of HPLC and RF channels available between node pairs is The available modulation modes are: Then the rocker arm set Arm and each rocker arm p Expressed as Arm={arm1,arm2,...,arm P },arm p =(c,m) in, Indicates the number of rocker arms, c and m are rocker arm p The channel and data modulation scheme selected between node pairs, In each iteration, one of the P rockers needs to be selected for execution; assuming Represents the end-to-end path The set of rocker arms selected above calculates the data transmission rate, latency, and reliability corresponding to each rocker arm, and sets the optimization goal given in P1 as the reward Incremental arm p The expected return value Q at the current iteration number t t (arm p ) to update efficiently Where Q t-1 (arm p ) is the expected return at iteration t-1, R(arm p ) is the rocker arm p The return, count(arm p ) represents arm p In addition, the decaying ∈-greedy strategy is used to select the rocker arm p , the exploration probability ∈(t) is updated as Among them, ∈(0) represents the initial value of the exploration probability, χ represents the decay coefficient; assuming that the maximum number of iterations is T max , in each iteration arm p The selected methods are: Analysis shows that the rocker arm selected for execution in each iteration is closely related to ∈(t); as the number of iterations t increases, ∈(t) gradually decreases, and the exploration rate gradually slows down; the multi-armed tiger machine model gradually transitions from the initial biased exploration to the biased utilization of the known optimal rocker arm, but it will still try new rockers with a small probability; therefore, when T max When it is much larger than P, it is basically possible to traverse all rocker arms and find the optimal rocker arm; S23, the gated network is a softmax classifier, and the output is a probability distribution, indicating the probability of each expert being selected; S24, the tower network is composed of a fully connected network FCN layer; the FCN layer is composed of an independent network layer structure; The results of the aggregate experts are then forwarded to the business-specific FCN network; the FCN layer uses the aggregated information to perform business-specific processing and generate the final output; each business is associated with its own dedicated FCN layer and chooses the ReLU function as the activation function.

Images