Intelligent routing decision-making system and method for power distribution communication network, and storage medium

Abstract

The invention relates to the technical field of intelligent networking of power distribution communication networks, in particular to an intelligent routing decision-making system and method for a power distribution communication network and a storage medium. The real-time intelligent routing decision-making strategy is generated by combining the time sequence generation large model and the routing decision-making model. According to the method, the dynamic rule of the time sequence data can be fully learned, the future link state and the network load are accurately predicted through the historical data, a reliable routing scheme is generated, and the stability and the flexibility of the network are effectively improved.

Description

Technical Field

[0001] This application belongs to the technical field of intelligent networking for distribution communication networks, and specifically relates to an intelligent routing decision-making system, method, and storage medium for distribution communication networks. Background Art

[0002] The distribution communication network is an important part of the distribution automation system. The task of the distribution communication network is to provide efficient and stable data transmission support for the equipment in the distribution network, ensuring that the distribution network can perceive, diagnose, and control grid equipment in real time. With the rapid development of new distribution network services, the complexity of distribution network management has been continuously increasing, specifically reflected in: the demand for information collection has increased explosively, resulting in a large increase in the number of collection points and the amount of collection, and the requirements for collection frequency and real-time performance have been significantly improved. Facing a more complex distribution network, distribution control has expanded from traditional centralized control to end control, the number of control points has increased from the hundred-thousand level to the million level, the control delay has been improved from quasi-real-time to real-time, and the control frequency has been increased from low frequency to high frequency.

[0003] In order to improve the digital level of the distribution network, building an intelligent distribution network has become an inevitable trend in the development of the power system.

[0004] In the prior art, the distribution network is still in the stage of "blind adjustment" and "blind control", and the routing decisions it makes are mainly based on static or predefined rules, making it difficult to fully cope with the changes in the real-time state of communication links and complex dynamic environments. For example, common routing protocols (such as IPV6) usually use single indicators such as hop count, delay, or link quality to select parent nodes, and cannot adapt to the frequent changes in link quality, dynamic changes in node load, and interference from environmental noise in the power communication network. Moreover, traditional distribution network decision-making generation models mainly focus on classification or regression tasks, and fail to fully utilize the complex temporal features in the distribution communication network, such as the change trend of link state over time and the dynamic interaction relationship between nodes. Summary of the Invention

[0005] This application provides an intelligent routing decision-making system and method for a distribution communication network, which can fully learn the dynamic laws of time series data, accurately predict future link states and network loads through historical data, generate reliable routing schemes, and effectively improve the stability and flexibility of the network.

[0006] The technical solution of this application is as follows: An intelligent routing decision-making system for a distribution communication network, comprising: Distributed sensors and communication nodes, which jointly collect link state data of each node with temporal features in the distribution communication network, including: hop count, load information, remaining energy, bit error rate, and communication delay between nodes; A time series generation large model, which takes network topology data with time series characteristics as input and outputs newly synthesized network topology data with time series characteristics; A routing decision model, which takes the output of the time series generation large model as input and outputs the optimal parent node; the routing decision model includes an optimal parent node selection module, and the optimal parent node selection module calculates the cost of a node reaching its parent node and the cost of reaching the destination simultaneously, and selects the optimal parent node for the node based on the total cost of the two costs.

[0007] This application also provides an intelligent routing decision method for a distribution communication network, including the following steps: S1. Obtain the network topology data of each node with original time series characteristics through distributed sensors and communication nodes, and the network topology data is represented by the original time series; S2. Input the network topology data in S1 into the time series generation large model, and the time series generation large model synthesizes and outputs a new time series, denoted as the predicted time series, and represents the network topology data of each node under the new time series with the predicted time series; S3. After preprocessing the predicted time series, input it into the routing decision model, and the routing decision model outputs the optimal parent node of each node, and the optimal parent nodes of a node at each layer form a routing decision.

[0008] Further, in step S1, the original time series is a set of time series, and the original time series is expressed as: ; Among them, represents the original time series, represents the time series, and: ; In the formula, represents the dimension of, represents the network topology data corresponding to the time stamp in, and: ; In the formula, represents the metric parameter of the node and its link, represents the serial number of the metric parameter of the node and its link.

[0009] Further, in step S2, the time series generation large model is a TimeVAE model, and the encoder of the TimeVAE model includes a CNN, an RNN and a fully connected layer, and the decoder includes a CNN and a fully connected layer; S2 specifically includes the following steps: S201. Input the network topology data in S1 into the encoder. The CNN of the encoder extracts the local features of the network topology data, and the RNN of the encoder captures the long-term dependencies of the time series of the network topology data; S202. The encoder outputs the mean and variance of the latent space variables, obtaining the data distribution of the latent space variables in the latent space. The latent space variables are the representations of the predicted time series data in the low-dimensional latent space; S203. Perform a resampling operation on the data distribution in the latent space to generate the latent space representation of the predicted time series; S204. Input the latent space representation of the predicted time series into the decoder. The decoder learns the inverse mapping from the latent space variables of the predicted time series to the time series through multi-layer convolution operations, and reconstructs and outputs a predicted time series in the same format as the original time series.

[0010] Furthermore, in the TimeVAE model, the reconstruction error and the KL divergence are used as the loss function to optimize the generation ability of the predicted time series. The loss function is as follows: ; In the formula, represents the total loss function, represent the generator parameters and the encoder parameters respectively, represents the encoder The expectation of sampling the distribution of the latent variable z output by represents the conditional probability distribution of the latent variable z when the input is x, represents the conditional probability distribution of generating the input x given the latent variable z, represents the KL divergence, which is used to measure and The difference between represents the prior distribution of the latent variable z.

[0011] Furthermore, the TimeVAE model is trained with the goal of maximizing the variational lower bound.

[0012] Furthermore, step S3 specifically includes the following steps: S301. Set the event that the data packet sent by the node arrives at the parent node as the class label , and set the event that the data packet sent by the node does not arrive at the parent node as the class label ; S302. Calculate the path cost for the node to reach each parent node. The formula is as follows: ; In the formula, represents the child node, represents the parent node, represents an eigenvalue, represents the probability that a sample of feature x belongs to class 1; S303. Calculate the path costs from the child nodes to the destination in sequence, and form a routing decision path with the parent nodes in the obtained minimum path costs. The path cost formula for the child nodes to reach the destination is as follows: + ); In the formula, represents the optimal parent node of node z.

[0013] Furthermore, in step S303, ; In the formula, n represents.

[0014] Furthermore, in step S303, set a parent node switching threshold. When the newly calculated path cost for the child node to reach the destination is less than the current path cost, switch the parent node.

[0015] A computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, it implements an intelligent routing decision method for a distribution communication network as described above.

[0016] Due to the adoption of the above technical solutions, the beneficial effects of the present application are as follows: 1. The present application innovatively combines a time series generation large model and a routing decision model. Using the feature representation and generation ability of the time series generation large model, it predicts future link states and network loads based on historical data, and uses the prediction results as the basis for the subsequent routing decision model to give routing decisions. The time series generation large model can flexibly perceive the dynamic changes of the network environment, and the routing decision model gives play to its advantages in processing categorical features and missing values, fully coping with the complex dynamic changes in the power line communication environment, and realizing the accurate selection of parent nodes. The combination of the two can significantly shorten the end-to-end delay of node networking and improve the balance of network loads.

[0017] 2. The present application calculates the total path cost from the child node to the destination from a global perspective, effectively avoiding the problem of unbalanced costs in the path decision process. The present application can capture the dynamic changes of each node and make predictions, so the present application is suitable for a distribution communication network with complex dynamic characteristics. Description of the Drawings

[0018] The drawings described herein are used to provide a further understanding of the present application, and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation of the present application.

[0019] Figure 1An architecture diagram of an intelligent routing decision system for a distribution communication network provided by this application; Figure 2 A topology diagram of an optimal path for STA22 in an embodiment of this application; Figure 3 Another topology diagram of an optimal path for STA22 in an embodiment of this application; Figure 4 A comparison diagram of the traffic fairness indices of each first-level and second-level node in a simulation network under different routing policies; Figure 5 The average end-to-end delay of the uplink of each level of node in a simulation network under different routing policies. Detailed implementation manners

[0020] Based on what is described in the background art, referring to the appendix Figure 1 , this application provides an intelligent routing decision system for a distribution communication network, including: Distributed sensors and communication nodes, which jointly collect the link state data of each node with time series characteristics in the distribution communication network, including: hop count, load information, remaining energy, bit error rate, and communication delay between nodes; The distributed sensors and communication nodes jointly collect the link state data of each node with time series characteristics in the distribution communication network, including the hop count, load information, remaining energy, bit error rate, and communication delay between nodes with time series characteristics. The communication node can be understood as the previous hop of the distributed sensor, the sensor node is the end node of the network, and the link state data is the measurement value of the node state and the link state. The node hop count measurement value is the number of proxy nodes passed by the data packet during the transmission from this node to the coordinator CCO; the load measurement value is the number of all child nodes of each node; the remaining energy measurement value refers to the current available remaining electric energy or battery energy of the node, representing the ability of the node to continue to perform communication and computing tasks; the link bit error rate is the proportion of the number of error code elements in the data packet transmitted between the node and the previous hop node in the total number of code elements, representing the link quality; the communication delay represents the transmission delay of the data packet transmitted between the node and the previous hop node. The network topology data is formed based on the link state data, specifically as described in the following step S1.

[0021] A time series generation large model, which takes the network topology data with time series characteristics as the input and outputs newly synthesized network topology data with time series characteristics; A routing decision model, which takes the output of the time series generation large model as the input and outputs the optimal parent node; the routing decision model includes an optimal parent node selection module, and the optimal parent node selection module calculates the cost of the node reaching the parent node and the cost of reaching the destination at the same time, and selects the optimal parent node for the node based on the total cost of the two costs.

[0022] Link state data includes metrics such as bandwidth, packet loss rate, and latency. The communication device collects the load information of each node, the communication latency between nodes, and the bit error rate through the communication protocol and the low-latency acquisition device.

[0023] Based on the above model, the present application provides an intelligent routing decision method for a distribution communication network, including the following steps: S1. Obtain the network topology data of each node with original time series characteristics through distributed sensors and communication devices, and the network topology data is represented by the original time series. In step S1, the original time series is a set of individual time series, and the original time series is represented as: ; Wherein, represents the original time series, represents the time series, and: ; In the formula, represents the dimension of, represents the network topology data corresponding to the time stamp in, and: ; In the formula, represents the metric parameter of the node and its link, represents the serial number of the metric parameter of the node and its link.

[0024] In specific implementation, let be the true distribution of the original time series , and the output of the time series generation large model is to create a new synthetic time series , so that the distribution of this time series is close to the true distribution , that is, the mean and variance are the same, and thus the network topology data with new time series characteristics is generated.

[0025] S2. Input the network topology data in S1 into the time series generation large model, and the time series generation large model synthesizes and outputs a new time series, denoted as the predicted time series, which represents the network topology data of each node under the new time series.

[0026] In step S2, the time series generation large model is the TimeVAE model. The encoder of the TimeVAE model includes a CNN, an RNN, and a fully connected layer, and the decoder includes a CNN and a fully connected layer; S2 specifically includes the following steps: S201. Input the network topology data in S1 into the encoder. The CNN of the encoder extracts the local features of the network topology data, and the RNN of the encoder captures the long-term dependencies of the time series of the network topology data. S202. The encoder outputs the mean and variance of the latent space variables, obtaining the data distribution of the latent space variables in the latent space; the latent space variables are the representations of the network topology data in the low-dimensional latent space. S203. Perform a resampling operation on the data distribution in the latent space to generate the latent space representation of the predicted time series. S204. Input the latent space representation of the predicted time series into the decoder. The decoder learns the inverse mapping from the latent space variables of the predicted time series to the time series through multi-layer convolution operations, and reconstructs and outputs a predicted time series in the same format as the original time series.

[0027] In the TimeVAE model, the reconstruction error and the KL divergence are used as the loss functions to optimize the generation ability of the predicted time series. The loss function is as follows: ; In the formula, represents the total loss function, represent the generator parameters and the encoder parameters respectively, represents the encoder the expectation of sampling the distribution of the latent variable z output by represents the conditional probability distribution of the latent variable z when inputting x, represents the conditional probability distribution of generating the input x given the latent variable z, represents the KL divergence, which is used to measure and the difference between represents the prior distribution of the latent variable z.

[0028] During model training, the TimeVAE model is trained with the goal of maximizing the variational lower bound.

[0029] The reconstruction error measures the difference between the generated data and the real data. The KL divergence term constrains the choice of the latent variable distribution, making the generated latent space have a good structure and avoiding overfitting of the model. The goal of this embodiment is to minimize the losses of these two parts. Therefore, the model is trained by maximizing the variational lower bound. In the TimeVAE model, the latent space can capture the temporal patterns of the data and cooperate with the reconstruction error to optimize the temporal generation ability. When the model training is completed, new time series data can be generated.

[0030] S3. Preprocess the predicted time series and input it into the routing decision model. The routing decision model outputs the optimal parent node of each node, and the optimal parent nodes of a node at each layer constitute the routing decision.

[0031] In specific implementation, the real dataset used for training the time series generation large model comes from the automatic meter reading system in the actual production scenario, and the real dataset reflects the monitoring results of the power line carrier communication network. As shown in Table 1, the real dataset includes the power line carrier communication networks on three different substations.

[0032] Table 1 Real Dataset In the real dataset, the bit error rate and communication delay between each pair of connected nodes, as well as the hop count, load information, remaining energy, and queue utilization rate of each node, are recorded at each timestamp. According to Table 1, the number of nodes in the L1 substation is the largest, and the average number of edges within each timestamp is also larger, indicating that the connections in the network are closer. In contrast, the number of nodes on the L2 substation is the smallest, and the number of edges is significantly sparser than that of L1 and L3. Before inputting into the routing decision model, the data in the dataset is preprocessed by maximum-minimum normalization, and the positive event that the data packet sent by the node successfully reaches the parent node is set as the class label , otherwise the class label is . Thus, the parent node selection is modeled as a binary classification problem, and the preprocessed dataset is shown in Table 2.

[0033] Table 2 Preprocessed Dataset During model training, 80% of the preprocessed dataset is divided as the training set, and 20% as the test set. The CatBoost algorithm is used to train the model to train the routing decision model. During the model training process, grid search is used to search for the optimal parameter values in the specified parameter grid, that is, a value range is assigned to each hyperparameter, and it is evaluated through grid search to determine the most suitable parameter combination, such as adjusting the depth of the tree, learning rate, and number of iterations, etc. In this embodiment, the positive and negative ratios of the sample data are very unbalanced, so the ROC curve is used to characterize the test results of the model under the optimal parameter values. The ROC curve plots the true positive rate (TPR) and false positive rate (FPR) at different classification thresholds. The closer it is to the upper left corner, the better the classification performance of the model.

[0034] Step S3 specifically includes the following steps: S301. This step is the setting of the aforementioned label type.

[0035] S302. Calculate the path cost from the computing node to each parent node. The formula is as follows: ; In the formula, represents the child node, represents the parent node, represents the eigenvalue, that is, the metric parameter of the aforementioned node and its link, represents the probability that the sample of feature x belongs to class 1; S303. Calculate the path cost from the child node to the destination in sequence. The parent nodes in the obtained minimum path cost are composed of the routing decision path. The formula for the path cost from the child node to the destination is as follows: + ); In the formula, represents the optimal parent node of node z.

[0036] In step S303, ; In the formula, n represents the number of parent nodes of child node z.

[0037] In step S303, set the parent node switching threshold. When the newly calculated path cost from the child node to the destination is less than the current path cost, switch the parent node.

[0038] Example: This example is a three-layer network composed of 30 nodes. Refer to Appendix Figure 2 and Appendix Figure 3 . The first-layer nodes include STA1, STA2, STA3, STA4, and STA5. The second-layer nodes include STA6~STA15. The third-layer nodes include STA16~STA30. In the network of this example, there are link connections between multiple nodes of STA2, STA3, STA4, STA22 and STA8~STA10, and there are link connections between multiple nodes of STA8~STA10 and multiple nodes of STA23~STA27.

[0039] First, through distributed sensors and intelligent devices, the link state data of each node with temporal characteristics in the network is collected in real time, including: hop count, load information, remaining energy, bit error rate, and communication delay between nodes. These data are transmitted to the central control system in real time through an efficient communication protocol and low-latency acquisition devices. The collected node data is input into the encoder of the TimeVAE model to generate new network topology data. Then, the CatBoost algorithm is used to calculate the optimal parent node of each node to construct the optimal path. Specifically, after the CCO broadcasts a beacon message, STA1, STA2, STA3, STA4, and STA5, upon receiving it, reply to the CCO to request network access and use the CCO as the optimal parent node, that is, the destination. After STA1, STA2, STA3, STA4, and STA5 complete network access, they become relay node PCO. The CCO broadcasts the beacon message again, and after the first-level nodes process it, it is forwarded to other nodes with which they have a link relationship. As second-level nodes, STA6~STA15 select the optimal parent node according to the output result of the CatBoost model after receiving the beacon message. In the network of this embodiment, STA7 has only STA2 as a parent node. Therefore, STA7 can only use STA2 as its own optimal parent node; while STA11 will receive three beacon messages from STA3, STA4, and STA5 in three time slots. Then, STA11 selects a parent node from STA3, STA4, and STA5 as the optimal parent node according to the temporal generation result of TimeVAE. In this embodiment, STA11 selects STA3 as the optimal parent node according to the output result. STA11 will use STA3 as the optimal parent node for the uplink, and STA4 and STA5 as alternative parent nodes. Then, the uplink route of STA11→STA3→CCO is constructed. Further, the construction process of the uplink route of the three-layer network is roughly similar. Refer to Appendix Figure 2 , STA22 will receive multiple beacon messages within STA7 and STA8~STA10. Assume At time, the bit error rate of the link STA22→STA7 is small, while the bit error rates of multiple links of STA7→STA8~STA10 are large, and the connection load situation of STA7 is better than that of STA8~STA10. Then, the output result of the STA22 model is to use STA7 as the optimal parent node, and STA22→STA7→STA2→CCO is used as the best path for the uplink route. Refer to Appendix Figure 3 , The link quality from STA22 to STA7 deteriorates due to noise interference, and the bit error rate is relatively high; or the number of child nodes STA16 - STA20 of STA7 increases while the number of child nodes of STA8 decreases. After STA22 obtains the updated parent node information and link quality data, it switches the optimal parent node to STA8 through model algorithm processing. Therefore, the best path of the uplink route of STA22 becomes STA22→STA8→STA2→CCO.

[0040] To compare the end - to - end delay and load - balancing performance of different routing strategies in the example, the results of integrating the routing decision model test on the OMNET++ network emulator are used to conduct uplink routing delay analysis and load - balancing analysis on a simulation network of 30 nodes, and compare with the hop - based routing decision mechanism OF0 and the ETX - based routing decision mechanism MRHOF.

[0041] To characterize the load fairness, the Jain fairness index of traffic load is used to represent the degree of load balancing. The definition of the Jain fairness index is as shown in the formula: where, represents the number of each first - level node or second - level node, represents the traffic load of each first - level node or second - level node, The closer the value of is to 1, the fairer the traffic load. Using the number of bytes uploaded per unit time to represent the traffic load, the traffic fairness indices of each first - level and second - level node in the simulation network under different routing strategies are calculated respectively. As shown in the appendix Figure 4 In the first - level nodes and second - level nodes of the simulation network, the traffic load fairness indices of the intelligent routing decision system based on the large model generated by time series are higher than those of the hop - based routing decision mechanism OF0 and the ETX - based routing decision mechanism MRHOF. Among them, the intelligent routing decision system based on the large model generated by time series can make the traffic load fairness indices of the first - level and second - level nodes reach 0.813 and 0.732 respectively, indicating that the intelligent routing decision system based on the large model generated by time series has better load - balancing performance.

[0042] The quality of the link affects the number of packet retransmissions, which in turn affects the end - to - end delay of the uplink route. The average end - to - end delay of the uplink links of each level of nodes in the simulation network is as Figure 5As shown in the figure, the intelligent routing decision-making system based on the time series generation large model makes the end-to-end delay of the upstream routing of each level of nodes close to MRHOF. The end-to-end delay of some nodes is higher than MRHOF, but the end-to-end delay of each level of nodes is lower than OF0. For example, for the average end-to-end delay of the upstream link of the secondary nodes, OF0 is 87.36 ms, MRHOF is 74.32 ms, and the intelligent routing decision-making system based on the time series generation large model is 73.92 ms. This is because the intelligent routing decision-making system based on the time series generation large model not only considers the link quality, but also considers the load balance, and the link quality is the main factor affecting the end-to-end delay.

[0043] The embodiment of the present disclosure also provides a computer-readable storage medium storing computer-executable instructions, and the computer-executable instructions are configured to execute the intelligent routing decision-making method for the distribution communication network described above.

[0044] The above-mentioned computer-readable storage medium may be a transient computer-readable storage medium or a non-transient computer-readable storage medium.

[0045] The technical solution of the embodiment of the present disclosure may be embodied in the form of a software product. The computer software product is stored in a storage medium and includes one or more instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in the embodiment of the present disclosure. The foregoing storage medium may be a non-transient storage medium, including: various media such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs that can store program codes, or may also be a transient storage medium.

[0046] What is not described in this application can be realized by adopting or referring to the existing technologies.

[0047] The above are only the embodiments of the present application and are not intended to limit the present application. For those skilled in the art, various changes and modifications can be made to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included within the scope of the claims of the present application.

Claims

1. An intelligent routing decision system for a power distribution communication network, characterized in that: include: Distributed sensors and communication nodes, which jointly collect link status data of each node in the power distribution communication network with timing characteristics, including: number of hops, load information, remaining energy, bit error rate, and communication delay between nodes; A time series generation model, which takes network topology data with time series characteristics as input and takes newly synthesized network topology data with time series characteristics as output; A routing decision model, which takes the output of the time-series generation model as input and outputs an optimal parent node; the routing decision model includes an optimal parent node selection module, which simultaneously calculates the cost of the node reaching the parent node and the cost of reaching the destination, and selects the optimal parent node for the node based on the total cost of the two costs.

2. A method for intelligent routing decision-making of a power distribution communication network based on the system of claim 1, characterized in that: The following steps are involved: S1. Obtain network topology data of each node with original time series characteristics through distributed sensors and communication nodes, and the network topology data is represented by the original time series; S2, input the network topology data in S1 to the time series generation model, the time series generation model synthesizes and outputs a new time series, recorded as the predicted time series, and uses the predicted time series as the network topology data representation of each node under the new time series; S3. After preprocessing the predicted time series, the model is input into the routing decision model. The routing decision model outputs the optimal parent node of each node. The optimal parent node of a node in each layer constitutes the routing decision.

3. The intelligent routing decision method for a power distribution communication network according to claim 2, characterized in that: In step S1, the original time series is a collection of time series, and the original time series is expressed as: ； in, represents the original time series, represents a time series, and: ； In the formula, express The dimension of express The timestamp corresponding to The network topology data is: ； In the formula, represents the metric parameters of nodes and their links, Indicates the metric parameter sequence number of the node and its links.

4. The intelligent routing decision method for a power distribution communication network according to claim 3, characterized in that: In step S2, the time series generation model is a TimeVAE model, the encoder of the TimeVAE model includes CNN, RNN and a fully connected layer, and the decoder includes CNN and a fully connected layer; S2 specifically includes the following steps: S201, input the network topology data in S1 into the encoder, the CNN of the encoder extracts the local features of the network topology data, and the RNN of the encoder captures the long-term dependency of the time series of the network topology data; S202, the encoder outputs the mean and variance of the latent space variable to obtain the data distribution of the latent space variable in the latent space, where the latent space variable is the representation of the predicted time series data in the low-dimensional latent space; S203, resampling the data distribution of the latent space to generate a latent space representation of the predicted time series; S204. Input the latent space representation of the predicted time series into the decoder. The decoder learns the inverse mapping of the latent space variables of the predicted time series to the time series through multi-layer convolution operations, and reconstructs the predicted time series in the same format as the original time series.

5. The method for intelligent routing decision-making in a power distribution communication network according to claim 4, characterized in that: In the TimeVAE model, reconstruction error and KL divergence are used as loss functions to optimize the generation ability of the predicted time series. The loss function is as follows: ； In the formula, represents the total loss function, denote the generator parameters and encoder parameters respectively, Indicates the encoder The expected sampling distribution of the output latent variable z, represents the conditional probability distribution of the latent variable z when input x is given, represents the conditional probability distribution of generating input x given the latent variable z, Represents KL divergence, which is used to measure and The difference represents the prior distribution of the latent variable z.

6. A method for intelligent routing decision-making in a power distribution communication network according to claim 5, characterized in that: The TimeVAE model is trained with the goal of maximizing the variational lower bound.

7. The method for intelligent routing decision-making in a power distribution communication network according to claim 4, characterized in that: Step S3 specifically includes the following steps: S301, set the event that the data packet sent by the node arrives at the parent node as the category label , set the event that the data packet sent by the node does not reach the parent node as the category label ; S302, calculate the path cost from the node to each parent node, the formula is as follows: ； In the formula, Represents a child node, Represents the parent node, represents the eigenvalue, Indicates the probability that the sample of feature x belongs to category 1; S303, calculate the path cost of the child node to the destination in turn, and use the parent node with the minimum path cost to form a routing decision path. The path cost formula of the child node to the destination is as follows: + )； In the formula, Represents the optimal parent node of node z.

8. A method for intelligent routing decision-making in a power distribution communication network according to claim 7, characterized in that: In step S303, ； In the formula, n represents.

9. A method for intelligent routing decision-making in a power distribution communication network according to claim 8, characterized in that: In step S303, a parent node switching threshold is set, and when the newly calculated path cost of the child node to the destination is less than the current path cost, the parent node is switched.

10. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the program is executed by a processor, the method as claimed in claim 8 is implemented.

Images