A Machine Learning-Based Routing Configuration Method for Key Services in Power Communication

Abstract

The present invention discloses a method for configuring key service routes in power communication based on machine learning, and its content includes: establishing a link state matrix according to the topology of the power communication backbone transmission network, service bearing, and optical cable parameters; using the shortest path algorithm and combining historical data of service route configuration to generate a data set, providing training and testing conditions for the machine learning algorithm; designing the structure of a deep neural network, and training a machine learning model for each node of the power communication backbone transmission network respectively; given a key service route configuration request, using the machine learning model, calculating the path feature vectors of the current node and its adjacent nodes, and selecting the best adjacent node as the next-hop node. The above process is executed cyclically point by point starting from the source node until the destination node. The present invention helps to comprehensively optimize the configuration of key service routes in power communication, improve service reliability, achieve load balancing of power optical cable services, and reduce the operation risks of power communication services.

Description

Technical Field

[0001] The present invention relates to the technical field of power communication, and particularly to a method for configuring critical service routes in power communication based on machine learning. Background Technique

[0002] The critical services in power communication include various types such as line relay protection, security and stability control, dispatching data network, and dispatching program-controlled switching network, etc., which are usually carried by the backbone transmission network of power communication. The optimized configuration of service routes is an important means to ensure the quality of critical services. The basic requirement of the critical service route configuration method is to optimize the configuration of service routes according to the topological structure of the backbone transmission network of power communication and the optical cable resources, so that the overall performance index of the service path reaches the optimum. With the continuous expansion of the scale of the backbone transmission network of power communication and the continuous improvement of network complexity, the optimized configuration method of critical service routes faces many challenges.

[0003] Generally, the local routing of power communication services realizes route configuration with the help of the network management system based on the shortest path algorithm, while the overall routing of services is realized through manual operation and relying on engineering practice experience. This way of service route configuration has low efficiency, and it is difficult to ensure the overall optimum of the configuration result, and it is difficult to meet the service development requirements of the backbone transmission network of power communication. In terms of theoretical research, relevant research work highly abstracts the backbone transmission network of power communication, simplifies many characteristic information of power optical cables, and uses graph theory to configure service routes under highly idealized conditions. Obviously, the research results obtained under this premise are divorced from reality and difficult to play their due role in engineering practice.

[0004] Currently, in the aspect of configuring critical service routes in power communication, no technology related to the technical solution of the present invention has been disclosed. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for configuring critical service routes in power communication based on machine learning, which helps to comprehensively optimize the configuration of critical service routes in power communication, improve service reliability, realize load balancing of power optical cable services, and reduce the operation risk of power communication services, so as to solve the problems raised in the above background technique.

[0006] To achieve the above purpose, the present invention provides the following technical solution: A method for configuring critical service routes in power communication based on machine learning, including the following steps:

[0007] S1: Establish a link state matrix according to the topology of the backbone transmission network of power communication, service bearing, and optical cable parameters;

[0008] S2: Use the shortest path algorithm and combine with the historical data of service route configuration to generate a data set, providing training and testing conditions for the machine learning algorithm;

[0009] S3: Design a deep neural network architecture according to the functional requirements of service routing configuration. Based on the generated dataset, train and test a machine learning model for each node of the power communication backbone transmission network.

[0010] S4: Given a critical service routing configuration request, use the machine learning model to calculate the path feature vectors of the current node and its adjacent nodes.

[0011] S5: Select the best adjacent node as the next-hop node, and loop through the above process point by point starting from the source node until the destination node is reached.

[0012] Furthermore, in S1, the topology of the power communication backbone transmission network refers to the network topology structure information; service bearing refers to the bearing information of key services such as line relay protection, security and stability control, dispatching data network, and dispatching program-controlled switching network on the transmission network; the optical cable parameters include: optical cable type, voltage level, operation years, and optical cable length.

[0013] The link state matrix S = {s ij} n×m represents the network topology information and optical cable attribute parameters required by the critical service routing configuration algorithm. There is a one-to-one correspondence between the link and the optical cable. The number of links is n, and there are m state parameters for each link. Each row of the link state matrix S represents the complete state information of a link. The row vector of the i-th link state is expressed as

[0014] l i = [u i , v i , w i , n b,i , K V,i , K λ,i , λ c,i , t i , t 0,i , L i

[0015] where u i is the starting point of the link; v i is the ending point of the link; w i is the link weight; n b,i is the number of services already carried by the link; K V,i is the optical cable voltage level coefficient; K λ,i is the annual aging coefficient of the optical cable; λ c,i is the failure rate during the stable operation stage of the optical cable; t i is the actual operation years of the optical cable; t 0,i is the stable operation years of the optical cable; L i is the optical cable length.​

[0016] Furthermore, in S2, the shortest path algorithm is used in combination with the historical data of service route configuration to generate a data set, specifically including three types of paths, namely: the path with the highest reliability, the path with the fewest hops, and the path with the best load balance degree;

[0017] Among them, the generation method of the path with the highest reliability is to convert the link reliability into an unreliability summation problem, and then, using the shortest path algorithm, solve the path with the minimum unreliability to approximately represent the path with the highest reliability. According to the link state matrix, the link reliability calculation formula is

[0018] R i = 1 - τ·λ i ·L i

[0019] Among them, τ is the average repair time after the optical cable is interrupted; L i is the length of the optical cable; λ i is the failure rate of the optical cable per kilometer, 1 / hour, and is further expressed as

[0020] λ i = K V,i ·(λ c,i + K λ,i ·max{(t i - t 0,i ), 0})

[0021] All the parameters in the formula are taken from the row vector l of the link state matrix S i , assuming that the power optical cable has very high operating reliability, then the unreliability of the service path is approximately equal to the sum of the unreliabilities of each link in the path, that is

[0022]

[0023] Among them, p is the link set of the path;

[0024] The path with the fewest hops is obtained by setting the link weight w i = 1, and under the condition of a given critical service route configuration request, using the shortest path algorithm to solve the path with the fewest hops;

[0025] The path with the best load balance degree is obtained by calculating the link load balance index and using the shortest path algorithm to solve the path with the minimum sum of the link load balance indices. Let the existing number of services of the optical cable be n b,i , and use the following formula to calculate the link load balance index

[0026] b i = 1 - exp(-c·(n b,i + ε)), n b,i ≥ 0

[0027] Among them, c is a proportionality coefficient, c ∈ [0.004, 0.006]; ε is an offset, ε ∈ (0, 0.001];

[0028] Under the condition that the network topology remains unchanged, the initial sample data is adjusted and improved by using the shortest path algorithm and combining with the historical data of service route configuration to obtain a formal data set. The formal sample data set is decomposed into node sample data sets according to the network node characteristics. The node sample data sets are used to train and test the machine learning models of individual nodes, where the training sample data accounts for 80% and the test sample data accounts for 20%.

[0029] Furthermore, the designed deep neural network architecture described in S3 includes:

[0030] It is composed of 1 input layer, multiple hidden layers and 1 output layer. The neurons between each layer adopt a fully connected method. The number of hidden layers of the deep neural network and the number of neurons in each hidden layer are optimized to ensure the learning efficiency and prediction accuracy of the deep neural network model;

[0031] The input data of the deep neural network is the link state matrix S and the destination node number, a total of (n × m + 1) variables; the output is the distance of the path with the highest reliability from the current node to the destination node and the distance of the path with the best load balance. The deep neural network implements the following function operations

[0032]

[0033] Among them, S is the link state matrix; u t is the destination node; f R is the distance mapping function of the path with the highest reliability from the current node to u t ; f B is the distance mapping function of the path with the best load balance from the current node to u t ; y R is the predicted value of the distance of the path with the highest reliability; y B is the predicted value of the distance of the path with the best load balance; the path information with the fewest hops from the current node to the destination node is only related to the topological structure and has nothing to do with the change of link state parameters. Therefore, it does not need to be learned by the deep neural network and can be directly calculated by using the shortest path algorithm.

[0034] Furthermore, the training of the machine learning model described in S3 includes:

[0035] Each node of the power communication backbone transmission network corresponds to a separate deep neural network model. The deep neural networks of each node have the same architecture, the same learning and testing methods, but each node has its own independent node sample data set;

[0036] The node model adopts a supervised learning method. The input data is the link state matrix S and the destination node number, and the output label is the reliability - highest path distance and the load - balance - best path distance from this node to the destination node. The learning error and the number of iteration steps are determined comprehensively according to the actual requirements and the convergence speed.

[0037] Furthermore, in S4, a machine - learning model is used to calculate the path feature vectors of the current node and its adjacent nodes, including:

[0038] The node path feature vector contains three elements, namely the reliability - highest path distance u from the current node to the destination node, the load - balance - best path distance b from the node to the destination node, and the least - hop - number path distance h from the node to the destination node. The expression form is

[0039] P k =[u k ,b k ,h k T

[0040] Among them, P k represents the path feature vector of the k - th node; u k , b k and h k respectively represent the predicted values of the reliability - highest path distance, the load - balance - best path distance, and the least - hop - number path distance from this node to the destination node. u k and b k are calculated through the trained deep neural network model of the k - th node, and h k is obtained by using the shortest - path algorithm;

[0041] Suppose the current node is v1, and its adjacent nodes are v2, v3, v4, and v5 respectively. Then, the 5 path feature vectors of the current node and its adjacent nodes can be calculated by using the deep neural network model and the shortest - path algorithm of each node.

[0042] Furthermore, the selection of the best adjacent node as the next - hop node in S5 includes:

[0043] First, judge whether the set of adjacent nodes of the current node contains the destination node. If it contains, directly select the destination node as the next - hop node; otherwise, proceed to the next step of estimating the node path feature vector;

[0044] Suppose the current node is v k , and estimate the path feature vector of v k from the feature vectors of the adjacent node i and the corresponding link. The expression form is

[0045] ​

[0046] Among them, the link feature vector is Q i =[u q,i , b q,i , 1] T , where u q,i represents the unreliability of link i, and b q,i represents the load balancing index of link i. Let v k be the estimated value of the path feature vector Then

[0047]

[0048] According to the element values of, this method can select which adjacent node as the best next-hop node;

[0049] The adjacent node of the best reliability path is

[0050]

[0051] The adjacent node of the best load balancing path is

[0052]

[0053] The adjacent node of the basic shortest path

[0054]

[0055] Among them, N(k) is the set of adjacent nodes of the current node k;

[0056] The selection of the best next-hop node depends on the preference and tendency of the power communication key service routing configuration policy. A possible selection strategy is that if two or more of the three types of nodes are the same adjacent node, then select this node as the best next-hop node of the service routing; otherwise, select the neighbor node with the shortest path distance as the next-hop node.

[0057] Furthermore, the process of looping point by point from the source node as described in S5 until the destination node is reached includes:

[0058] First, according to the given key service routing configuration request, determine the source node and destination node of the power communication key service routing configuration, and use the source node as the current node;

[0059] Then, determine whether the set of adjacent nodes of the current node contains the destination node. If it contains, directly jump to the destination node and end the routing configuration; otherwise, perform the next-hop node selection;

[0060] The node selection process includes: according to the link feature vector Q of the connection linki and the path feature vector P of adjacent nodes i , calculate the estimated value of the path feature vector of each adjacent node Follow the relevant selection strategy to select the best next-hop node of the current node, use the selected next-hop node as the current node, and repeat the judgment of adjacent nodes and the estimation of path feature vectors;

[0061] Finally, loop the above process until the destination node is found.

[0062] Compared with the prior art, the beneficial effects of the present invention are:

[0063] 1. A key service routing configuration method for power communication based on machine learning provided by the present invention uses the topology of the power communication backbone transmission network, service bearing, and optical cable parameters, takes the link state matrix as an input variable, and uses the trained deep neural network model for key service routing configuration under the given service routing configuration request and specific link state matrix conditions. It simplifies the deep neural network structure by adopting the method of corresponding one deep neural network model to each network node, and improves the training performance and prediction accuracy.

[0064] 2. A key service routing configuration method for power communication based on machine learning provided by the present invention calculates the path feature vectors of the current node and its adjacent nodes using a machine learning model, and selects the best adjacent node as the next-hop node according to the strategy, making full use of the network topology information and enhancing the flexibility of routing selection under multiple constraint conditions.

[0065] 3. A key service routing configuration method for power communication based on machine learning provided by the present invention fully considers the historical data of service routing configuration during the construction of the sample data set, integrates the artificial key service routing configuration experience into the machine learning model, and improves the effectiveness and usability of the service routing configuration results.

[0066] 4. A key service routing configuration method for power communication based on machine learning provided by the present invention helps to comprehensively optimize the configuration of key services for power communication, improve service reliability, achieve load balancing of power optical cable services, and reduce the operation risks of power communication services. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 is the overall flowchart of the method of the present invention;

[0068] Figure 2 is a schematic diagram of the service routing configuration principle in the method of the present invention;

[0069] Figure 3 is the flowchart of service path calculation in the method of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0070] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0071] The basic idea of the embodiments of the present invention is to realize the configuration process of the key service routing in power communication through machine learning technology. First, according to the topology of the power communication backbone transmission network, service bearing and optical cable parameters, a link state matrix is established, and the shortest path algorithm is used to generate a data set in combination with the historical data of service routing configuration. Then, according to the functional requirements of service routing configuration, an appropriate deep neural network architecture is designed, and machine learning models are trained for each node of the power communication backbone transmission network respectively. Finally, under the condition of a given key service routing configuration request, the machine learning model is used to calculate the path feature vectors of the current node and its adjacent nodes, and the best adjacent node is selected as the next-hop node according to the policy. The above process is executed point by point in a loop starting from the source node until the destination node, and the configuration process of the key service routing is completed. The method of the present invention helps to comprehensively optimize the configuration of the key service routing in power communication, improve service reliability, realize load balancing of power optical cable services, and reduce the operation risk of power communication services.

[0072] As Figure 1 shown, in order to further and better explain the embodiments of the present invention, the present invention provides a method for configuring key service routing in power communication based on machine learning, including the following steps:

[0073] S1: According to the topology of the power communication backbone transmission network, service bearing and optical cable parameters, establish a link state matrix; the topology of the power communication backbone transmission network in the embodiments of the present invention refers to the network topology structure information; service bearing refers to the bearing information of key services such as line relay protection, security and stability control, dispatching data network and dispatching program-controlled switching network on the transmission network; the optical cable parameters include: optical cable type, voltage level, operation years and optical cable length, etc.

[0074] The link state matrix S in the embodiments of the present invention = {s ij} n×m represents the network topology information and optical cable attribute parameters required by the key service routing configuration algorithm. Each row of the link state matrix S represents the complete state information of a link, denoted as l i = [u i , v i , w i , n b,i , K V,i , Kλ,i , λ c,i , t i , t 0,i , L i . Among them, u i is the starting point of the link; v i is the ending point of the link; w i is the link weight; n b,i is the number of services already carried by the link; K V,i is the optical cable voltage level coefficient; K λ,i is the annual aging coefficient of the optical cable; λ c,i is the failure rate in the stable operation stage of the optical cable; t i is the actual operating years of the optical cable; t 0,i is the stable operating years of the optical cable; L i is the length of the optical cable (km).

[0075] S2: Use the shortest path algorithm and combine with the historical data of service route configuration to generate a data set, providing training and testing conditions for the machine learning algorithm; In the embodiment of the present invention, the shortest path algorithm is used and combined with the historical data of service route configuration to generate a data set. The best path includes three aspects, namely, the path with the highest reliability, the path with the fewest hops, and the path with the best load balancing degree;

[0076] The generation method of the path with the highest reliability is to convert the link reliability into an unreliability summation problem, and then, use the shortest path algorithm to solve the path with the minimum unreliability to approximately represent the path with the highest reliability. The link unreliability is calculated according to the optical cable parameters provided by the link state matrix. The best path information is obtained by using the shortest path algorithm in combination with the historical data of service route configuration.

[0077] The path with the fewest hops is obtained by using the shortest path algorithm according to the network topology structure represented by the link state matrix.

[0078] The path with the best load balancing degree is obtained by calculating the link load balancing index and using the shortest path algorithm to solve the minimum sum of the link load balancing indexes in the path.

[0079] Under the condition that the network topology structure remains unchanged, this method is repeated multiple times (for example, 10,000 times), randomly generating the link state matrix and key service route configuration requests, and using the shortest path algorithm to calculate the path with the highest reliability, the path with the fewest hops, and the path with the best load balancing degree, so as to form a sufficient number of initial sample data sets for machine learning. Then, in combination with the historical data of service route configuration, the initial sample data is adjusted and improved to obtain the formal data set. The formal sample data set is decomposed into node sample data sets according to the network node characteristics. The node sample data sets are used to train and test the machine learning model of a single node. Among them, the training sample data accounts for 80%, and the test sample data accounts for 20%.

[0080] S3: According to the functional requirements of service routing configuration, design a deep neural network architecture. Based on the generated dataset, train and test a machine learning model for each node of the power communication backbone transmission network to ensure the effectiveness of the model. In the embodiments of the present invention, a deep neural network architecture is selected as the appropriate machine learning model. The deep neural network architecture consists of 1 input layer, multiple hidden layers, and 1 output layer. Neurons between each layer are fully connected. According to the functional requirements of service routing configuration, optimize and determine the number of hidden layers of the deep neural network and the number of neurons in each hidden layer to ensure the learning efficiency and prediction accuracy of the deep neural network model. The input data of the deep neural network is the link state matrix S and the destination node number. The output is the reliability highest path distance and the load balance best path distance from the current node to the destination node, and the path information with the fewest hops from the current node to the destination node, which is directly obtained using the shortest path algorithm.

[0081] The node model adopts a supervised learning method. For each node, train and test a deep neural network model based on its respective sample dataset. The input data is the link state matrix S and the destination node number, and the output label is the reliability highest path distance and the load balance best path distance from the node to the destination node. The learning error and the number of iterations can be comprehensively determined according to actual requirements and convergence speed.

[0082] S4: Given a critical service routing configuration request, use the machine learning model to calculate the path feature vectors of the current node and its adjacent nodes. In the embodiments of the present invention, under the condition of a given critical service routing configuration request, use the machine learning model to calculate the path feature vectors of the current node and its adjacent nodes.

[0083] The path feature vector P of the kth node k is P k = [u k , b k , h k , T , where u k represents the reliability highest path distance from the node to the destination node; b k represents the load balance best path distance from the node to the destination node; h k represents the path distance with the fewest hops from the node to the destination node. Elements u k and b k are predicted by the trained deep neural network model of the kth node. h k is directly calculated using the shortest path algorithm.

[0084] S5: Select the best neighboring node as the next-hop node, and loop through the above process point by point starting from the source node until the destination node is reached. In the embodiment of the present invention, it is first determined whether the set of neighboring nodes of the current node contains the destination node; if it does, the destination node is directly selected as the next-hop node to end the route configuration; otherwise, the node path feature vector estimation for the next step is performed.

[0085] The current node v k Based on the feature vectors of its neighboring node i and the corresponding link, obtain the k estimated value of the path feature vector of v If the link feature vector is Q i = [u q,i , b q,i , 1] T , then the k estimated value of the path feature vector of v is This method, according to the element values of, selects a neighboring node in the set of neighboring nodes as the best next-hop node;

[0086] The selection strategy for the best next-hop node depends on the preference and tendency of the power communication critical service route configuration strategy. One possible selection strategy is that if two or more of the three types of nodes are the same neighboring node, then this node is selected as the best next-hop node for the service route; otherwise, the neighboring node with the shortest hop count path distance is selected as the best next-hop node.

[0087] The loop through the above process point by point starting from the source node until the destination node is reached in the embodiment of the present invention includes:

[0088] First, according to the given critical service route configuration request, determine the source node and the destination node of the power communication critical service route configuration, and use the source node as the current node.

[0089] Then, determine whether the set of neighboring nodes of the current node contains the destination node. If it does, directly jump to the destination node and end the route configuration; otherwise, perform the next-hop node selection.

[0090] The node selection process includes: According to the feature vector Q i of the connection link and the path feature vector P i of the neighboring node, calculate the estimated value of the path feature vector of each neighboring node, and

[0091] Finally, loop the above process until the destination node is found.

[0092] To further better explain the embodiments of the present invention, the business routing configuration principle in the method of the present invention is as Figure 2 shown as follows:

[0093] The power communication backbone transmission network in the embodiments of the present invention has 10 nodes and 16 links. According to the network topology structure, optical cable parameters, and the current service status, the method establishes a link state matrix, further calculates the unreliability and load balancing index of the links, uses the shortest path algorithm, combines the service configuration historical data, randomly generates a sample data set for machine learning training and testing, determines the deep neural network architecture, and uses the supervised learning method to train 10 deep neural network models for 10 nodes respectively.

[0094] Given the business routing configuration requirement (the source node is 2 and the destination node is 9) and the determined link state matrix, the method starts the routing configuration from the source node s (node 2). The adjacent nodes of node 2 are node 1 and 5. According to the link state matrix and the destination node t (node 9), the path distance vectors P1 and P5 of node 1 and 5 are calculated respectively through the deep neural network model, and then the link feature vectors Q1 and Q5 corresponding to the adjacent nodes are calculated respectively. Then, the estimated value of the node path distance vector of node 2 is calculated and Through comprehensive judgment, node 5 is the best next-hop node.

[0095] Taking node 5 as the current node, the path distance vectors P1, P7, and P8 of the adjacent nodes 1, 7, and 8 are calculated respectively by using the deep neural network model, and the estimated value of the path distance vector of node 5 is further calculated and Through comprehensive judgment, node 7 is selected as the best next-hop node.

[0096] Taking node 7 as the current node, calculate the path distance vectors of the adjacent nodes 6, 8, and 10, and estimate and Through comprehensive judgment, node 6 is selected as the best next-hop node.

[0097] Taking node 6 as the current node, the algorithm finds that the destination node 9 is included in its adjacent node set. So far, the business routing configuration process ends. The business route (2-5-7-6-9) is the best route.

[0098] As Figure 3 shown, to further better explain and illustrate the embodiments of the present invention, the business path calculation flowchart in the method of the present invention includes the following content:

[0099] First, input the topology structure of the power communication backbone transmission network, the historical data of key service route configurations, and the number of samples required for machine learning model training and testing. Randomly generate the link state matrix and the source and destination nodes of the key service routes. Using the shortest path algorithm, obtain the corresponding set of optimal paths, and calculate the reliability - highest path distance, load - balancing optimal path distance, and least - hop path distance from each node to the destination node. Repeat the above process to generate a sample data set with a quantity meeting the requirements. Supplement and improve the sample data set using the historical data of service route configurations to obtain the formal data set. Divide the data set into multiple network node sample data sets according to node characteristics. 80% of each node sample data set is used for training, and 20% is used for testing.

[0100] Then, design a deep neural network architecture suitable for the training of the node machine learning model. For each node, train the node deep neural network model to obtain a machine learning model with a calculation accuracy meeting the requirements.

[0101] Finally, use the trained machine learning model to assist the key service route configuration process. The content includes: input the service route configuration request, specifying the source node and the destination node; taking the source node as the current node, generate the set of adjacent nodes of the current node, and check whether the set of adjacent nodes contains the destination node. If it contains, end the routing process and form the set of nodes of the optimal path for the service route; otherwise, use the machine learning model to calculate the path feature vectors of each adjacent node, and combine with the link feature vectors to calculate the estimated value of the path feature vector of the current node. An estimated value can be obtained for each adjacent node. According to the selection strategy, based on the estimated value of the path feature vector of the current node, select the best next - hop node. Set the best next - hop node as the current node, and repeat the above process until the destination node is found.

[0102] In summary, the power communication key service routing configuration method based on machine learning provided by the present invention uses the topology of the power communication backbone transmission network, service bearing, and optical cable parameters. With the link state matrix as the input variable, the trained deep neural network model is used for key service routing configuration under the given service routing configuration request and specific link state matrix conditions. The present invention simplifies the deep neural network structure by adopting a method where each network node corresponds to a deep neural network model, improving the training performance and prediction accuracy. The present invention calculates the path feature vectors of the current node and its adjacent nodes using the machine learning model, and selects the best adjacent node as the next-hop node according to the strategy, making full use of the network topology information and enhancing the flexibility of routing selection under multiple constraint conditions. In the process of constructing the sample data set, the present invention fully considers the historical data of service routing configuration, integrates the manual key service routing configuration experience into the machine learning model, improves the effectiveness and usability of the service routing configuration result, realizes the load balancing of power optical cable services, and reduces the operation risk of power communication services.

[0103] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

Claims

1. A key service routing configuration method for power communication based on machine learning, characterized in that It includes the following steps: S1: Establish a link state matrix according to the topology of the power communication backbone transmission network, service bearing, and optical cable parameters; S2: Use the shortest path algorithm and combine it with the historical data of service route configuration to generate a data set, providing training and testing conditions for the machine learning algorithm; S3: Design a deep neural network architecture according to the functional requirements of service route configuration. Based on the generated data set, train and test the machine learning model for each node of the power communication backbone transmission network respectively; S4: Given a critical service route configuration request, use the machine learning model to calculate the path feature vectors of the current node and its adjacent nodes, including: The node path feature vector contains 3 elements, namely the distance u of the highest reliability path from the current node to the destination node, the distance b of the best load-balanced path from the node to the destination node, and the distance h of the path with the fewest hops from the node to the destination node. The expression form is P k = [u k , b k , h k T ​ Among them, P k represents the path feature vector of the k-th node; u k , b k and h k respectively represent the predicted values of the path distance with the highest reliability, the path distance with the best load balancing, and the path distance with the fewest hops from this node to the destination node. u k and b k are calculated by the deep neural network model trained by the k-th node, and h k is obtained by using the shortest path algorithm; Suppose the current node is v1, and its adjacent nodes are v2, v3, v4, and v5 respectively. Then, 5 path feature vectors of the current node and its adjacent nodes can be calculated using the deep neural network model and the shortest path algorithm of each node; S5: Select the best adjacent node as the next-hop node, and loop through the above process point by point from the source node until the destination node. Among them, selecting the best adjacent node as the next-hop node includes: First, judge whether the set of adjacent nodes of the current node contains the destination node. If it contains, directly select the destination node as the next-hop node; otherwise, perform the next node path feature vector estimation; Let the current node be v k , estimate the path feature vector of v k from the eigenvectors of adjacent nodes i and the corresponding links, and the expression form is Among them, the link feature vector is Q i = [u q,i , b q,i , 1] T , where u q,i represents the unreliability of link i, and b q,i represents the load balancing index of link i. Let v k be the estimated value of the path feature vector Then According to the element value of, the method can select which adjacent node as the best next-hop node; The adjacent node of the best reliability path is The adjacent node of the best load-balanced path is The basic shortest path adjacent node Among them, N(k) is the set of adjacent nodes of the current node k; The selection of the optimal next-hop node depends on the preferences and inclinations of the power communication critical service routing configuration strategy. A possible selection strategy is that if two or more of the three types of nodes are the same adjacent node, then this node is selected as the optimal next-hop node for the service routing; otherwise, the neighbor node with the shortest hop count path distance is selected as the next-hop node.

2. The method for configuring the key service routing of power communication based on machine learning according to claim 1, characterized in that In S1, the topology of the power communication backbone transmission network refers to the network topology structure information; service bearing refers to the bearing information of key services such as line relay protection, security and stability control, dispatching data network, and dispatching program-controlled switching network on the transmission network; The optical cable parameters include: optical cable type, voltage level, operation years, and optical cable length; The link state matrix S described in S1 = {s ij} n×m represents the network topology information and fiber optic cable attribute parameters required by the critical service routing configuration algorithm. There is a one-to-one correspondence between the link and the fiber optic cable. The number of links is n, and each link has m state parameters. Each row of the link state matrix S represents the complete state information of a link. The row vector of the state of the i-th link is expressed as l i = [u i , v i , w i , n b,i , K V,i , K λ,i , λ c,i , t i , t 0,i , L i ​ Among them, u i is the starting point of the link; v i is the ending point of the link; w i is the link weight; n b,i is the number of services already carried by the link; K V,i is the voltage level coefficient of the optical cable; K λ,i is the annual aging coefficient of the optical cable; λ c,i is the failure rate during the stable operation stage of the optical cable; t i is the actual operation years of the optical cable; t 0,i is the stable operation years of the optical cable; L i is the length of the optical cable.

3. A method for configuring a key service route in power communication based on machine learning according to claim 2, characterized in that, In S2, using the shortest path algorithm and combining it with the historical data of service route configuration to generate a data set specifically includes three paths, namely: the path with the highest reliability, the path with the fewest hops, and the path with the best load balance; Among them, the generation method of the path with the highest reliability is to convert the link reliability into an unreliability summation problem, and then use the shortest path algorithm to solve the path with the minimum unreliability to approximately represent the path with the highest reliability. According to the link state matrix, the link reliability calculation formula is R i = 1 - τ·λ i ·L i where τ is the average repair time after the optical cable is interrupted; L i is the length of the optical cable; λ i is the failure rate of the optical cable per kilometer, expressed further as λ i = K V,i ·(λ c,i + K λ,i · max{(t i - t 0,i ), 0}) All parameters in the formula are taken from the row vector l of the link state matrix S i , assuming that the power optical cable has very high operating reliability, the unreliability of the service path is approximately equal to the sum of the unreliabilities of each link in the path, that is Among them, p is the set of links of the path; The path with the fewest hops is obtained by setting the link weight w i = 1 and using the shortest path algorithm to solve for the path with the fewest hops under the given critical service routing configuration request conditions; The optimal path of load balancing degree is obtained by calculating the link load balancing index and using the shortest path algorithm to solve the minimum sum of link load balancing indexes in the path. Let the number of existing services of the optical cable be n b,i , and the link load balancing index is calculated using the following formula b i = 1 - exp(-c·(n b,i + ε)), n b,i ≥ 0 Among them, c is a proportionality coefficient, c ∈ [0.004, 0.006]; ε is an offset, ε ∈ (0, 0.001]; Under the condition that the network topology remains unchanged, the initial sample data is adjusted and improved by using the shortest path algorithm and combining with the historical data of service route configuration to obtain the formal data set. The formal sample data set is decomposed into node sample data sets according to the network node characteristics. The node sample data sets are used to train and test the machine learning models of individual nodes, where the training sample data accounts for 80% and the test sample data accounts for 20%.

4. A method for configuring a key service route in power communication based on machine learning according to claim 1, characterized in that, The designed deep neural network architecture described in S3 includes: It consists of 1 input layer, multiple hidden layers and 1 output layer. The neurons between each layer adopt the full connection method. The number of hidden layers of the deep neural network and the number of neurons in each hidden layer are optimized to ensure the learning efficiency and prediction accuracy of the deep neural network model; The input data of the deep neural network is the link state matrix S and the destination node number, with a total of (n×m + 1) variables; the output is the reliability highest path distance and the load balance best path distance from the current node to the destination node. The deep neural network implements the following function operations Among them, S is the link state matrix; u t is the destination node; f R is the distance mapping function of the path with the highest reliability from the current node to u t ; f B is the distance mapping function of the best load-balanced path from the current node to u t ; y R is the predicted value of the distance of the path with the highest reliability; y B is the predicted value of the distance of the best load-balanced path; The path information with the fewest hops from the current node to the destination node is only related to the topology structure and has nothing to do with the change of link state parameters. Therefore, it does not require deep neural network learning and can be directly calculated using the shortest path algorithm.

5. A method for configuring the routing of key services in power communication based on machine learning according to claim 1, characterized in that, The training of the machine learning model described in S3 includes: Each node of the power communication backbone transmission network corresponds to a separate deep neural network model. The deep neural networks of each node have the same architecture, the same learning and testing methods, but each node has its own independent node sample data set; The node model adopts the supervised learning method. The input data is the link state matrix S and the destination node number, and the output is marked as the reliability highest path distance and the load balance best path distance from this node to the destination node. The learning error and the number of iteration steps are determined comprehensively according to the actual requirements and the convergence speed.

6. A method for configuring the routing of key services in power communication based on machine learning according to claim 1, characterized in that, The process of sequentially and circularly executing the above process from the source node until the destination node described in S5 includes: First, according to the given critical service route configuration request, determine the source node and the destination node of the power communication critical service route configuration, and take the source node as the current node; Then, judge whether the set of adjacent nodes of the current node contains the destination node. If it contains, directly jump to the destination node and end the route configuration; otherwise, perform the next-hop node selection; The node selection process includes: based on the feature vector Q of the connection link i and the path feature vector P of adjacent nodes i , calculate the estimated value of the path feature vector of each adjacent node Follow the relevant selection strategy to select the best next-hop node of the current node, use the selected next-hop node as the current node, and repeat the judgment of adjacent nodes and the estimation of the path feature vector; Finally, loop the above process until the destination node is found.

Images