Power distribution network load transfer auxiliary decision-making method based on partition autonomous model

Abstract

The invention discloses a power distribution network load transfer auxiliary decision-making method based on a partition autonomous model, and the method comprises the steps: carrying out the modeling of a power grid simulation example through offline power grid data, and carrying out the partitioning of the simulation example, and obtaining an autonomous region range; constructing an autonomous model according to the offline power grid partition data, carrying out model training on the offline fault data, and obtaining a converged auxiliary decision-making model; obtaining online power grid data, and preprocessing the power grid data to obtain an online power grid data partitioning result; and according to the actual fault scene of the power grid, calling the trained model to carry out load transfer online auxiliary decision making. A power grid partition autonomy and interval cooperation method is adopted, partition standards are constructed through multiple pieces of power grid information such as electrical distance and inter-bus voltage sensitivity, cooperation reward functions of different autonomous region models are constructed according to power grid operation specifications, and a preprocessing method of power grid online data is designed based on data specifications of an offline scene. And finally, the auxiliary decision-making of the load transfer after the power grid fault is realized.

Description

Technical Field

[0001] The invention belongs to the technical field of power grid operation dispatching control, and in particular to a distribution network load transfer auxiliary decision method based on a partitioned autonomous model. Background Art

[0002] As the "last mile" in the power supply process, the stable operation of the distribution network is crucial to ensuring safe, reliable and continuous power supply to users. In recent years, with the intensification of climate change and the frequent occurrence of extreme weather events, the risk of large-scale power outages faced by the distribution network has increased significantly; in addition, with the gradual increase in the proportion of renewable energy power generation, especially the large-scale access of intermittent power sources such as wind power and solar energy, the uncertainty on the source side of the power grid has increased significantly. These factors have posed many challenges to the load transfer and recovery tasks after a power outage in the distribution network, especially the load transfer tasks after a large-scale power outage in the distribution network.

[0003] Traditional distribution network load transfer methods rely on a lot of manual experience and are difficult to cope with the high uncertainty of the system. The operation plan based on empirical methods cannot ensure the optimality of the plan, and when there are power outages in multiple places, load transfer cannot be coordinated. When modeling and solving, due to the large and complex structure of the distribution network, which contains a large number of switch combinations, it is very easy to cause the selection of load transfer paths after the fault to present a combinatorial explosion problem, resulting in a time-consuming solution process and difficulty in meeting the actual requirements of engineering applications. Summary of the invention

[0004] The purpose of the present invention is to provide a distribution network load transfer auxiliary decision method based on a partitioned autonomous model to solve the problems raised in the background technology.

[0005] The present invention solves the technical problem by the following technical solutions:

[0006] A distribution network load transfer auxiliary decision method based on a partitioned autonomous model comprises the following steps:

[0007] S1, obtain the offline distribution network model and fault information, divide the area according to the offline samples, and obtain the partition result;

[0008] S2, combining the partitioning result of step S1, constructing a multi-agent deep reinforcement learning model based on autonomous partitioning, and training the multi-agent deep reinforcement learning model based on offline fault data until the multi-agent deep reinforcement learning model converges;

[0009] S3, obtaining the online distribution network model data, dividing the area according to the online samples, and obtaining the partitioning results;

[0010] S4, input the partitioning result obtained in step S3 into the multi-agent deep reinforcement learning model that has been trained and converged in step S2 to obtain the fault auxiliary decision-making result of the current online power grid data.

[0011] Furthermore, the method of obtaining the offline distribution network model and fault information in step S1, dividing the offline samples into regions, and obtaining the partition results includes the following steps:

[0012] S1.1, obtain the offline model data file of the power grid, build the busbar branch power grid according to the verified computing nodes and branch information in the file, and build the node switch power grid according to the physical nodes, branch information and switch status in the homologous data, and determine the mapping relationship between the two power grids according to the file data;

[0013] S1.2, for the bus branch power grid, the voltage amplitude sensitivity relationship between any two buses is defined as follows:

[0014] ΔU i =α ij ΔU j

[0015]

[0016] In the formula, ΔU i,j and U i,j Respectively represent the voltage amplitude deviation and voltage amplitude at busbar i and j; α ij It represents the voltage amplitude sensitivity coefficient between busbar i and busbar j, Q i,j represents the reactive injection power of busbar i,j;

[0017] S1.3, define the electrical distance between busbar i and busbar j as d ij , to eliminate differences due to orientation:

[0018] d ij =d ji =--log(α ij α ji )

[0019] S1.4, based on the voltage amplitude sensitivity coefficient α between busbars i and j ij and the electrical distance d between busbar i and busbar j ij The two parameters define the voltage sensitivity matrix and electrical distance matrix of the bus branch network respectively as D = [d ij ],A=[a ij ], define the comprehensive weight matrix G = [g ij ] is used to describe the similarity between nodes, where:

[0020]

[0021] In the formula, α∈[0,1],β∈[0,1] are adjustment parameters used to balance the impact of electrical distance and voltage sensitivity on partitioning. They need to be selected according to the specific conditions of the actual distribution network. At the same time, a comprehensive weight lower limit g is specified according to the actual parameters of the power grid. min , when the combined weight coefficient of the two buses is lower than g min It is not allowed to be divided into the same partition;

[0022] S1.5, in order to limit the number of grid partitions within a reasonable range, the modularity H of the entire grid after partitioning is defined as follows:

[0023]

[0024] Where b n is the total number of grid buses, δ(c i ,c j )satisfy:

[0025]

[0026] In the formula, c i Represents the partition number where bus i is located.

[0027] S1.6, applying the partitioning algorithm to the bus branch grid, after obtaining the bus branch grid partition, mapping the partitioning result to the node switch grid to complete the grid partitioning.

[0028] Moreover, the specific process of the busbar branch grid partitioning algorithm is as follows:

[0029] a Initialize each bus as an independent partition;

[0030] b For each busbar i, calculate and move it to each busbar where the comprehensive weight coefficient is greater than g min Modularity changes in adjacent bus partitions, select the partition with the largest modularity change, and move bus i to the corresponding partition;

[0031] c Repeat step b until there are no nodes that can be moved or the modularity cannot be further improved.

[0032] Furthermore, step S2 constructs a multi-agent deep reinforcement learning model based on autonomous partitioning, and trains the model based on offline fault data until the model converges, specifically including the following steps:

[0033] S2.1, each partition subnetwork is modeled as an intelligent agent, and the observation space and shared observation space are defined. The observation space mainly includes the bus voltage, load demand, branch load rate, and switch disconnection status of the partition, and the shared observation space mainly includes the power supply remaining capacity and power supply load rate;

[0034] For a specific grid partition, the agent input consists of the observation space within the partition and the shared observation space of the partition and neighboring partitions. The final agent output forms the auxiliary decision of the switch action in the partition.

[0035] S2.2,In order to ensure that the decision of the agent neural network can converge, the reinforcement learning method is used to design the reward function to guide the agent. At the same time, in order to ensure that the convergence strategy meets the operation specifications of the distribution network, it is necessary to set constraints for each auxiliary decision. The reward function of the agent can be set as;

[0036]

[0037] Where P recover is the load restored after the agent makes a decision, P loss is the active power loss of the system after the agent makes a decision, and λ is the adjustment coefficient to ensure that the reward function is within a reasonable range;

[0038] S2.3, in addition to the reward function, the system must also comply with safety constraints, where the power flow equation constraint is determined by the return value of the power grid power flow calculation simulation software. In addition, the distribution network mainly includes node voltage constraints, branch capacity constraints and generator output constraints:

[0039]

[0040] Where U i represents the voltage at node i, Represents the voltage upper and lower limits of node i, P i gen represents the active power of generator i, P i min / max represents the upper and lower limits of the active power of generator i, represents the reactive power of generator i, Represents the upper and lower limits of reactive power of generator i, I ij represents the current in the branch between node i and node j, Indicates the upper limit of the current flowing through the branch. If the system does not meet these constraints after a certain auxiliary decision, the agent will be punished by -1 according to the reward function calculation formula;

[0041] S2.4, after completing the setting of the agent model and reward function, the neural network is trained according to the fault scenario derived offline until the reward function converges, which means the training is completed.

[0042] Furthermore, the step S3 of acquiring the online distribution network model data and dividing the online samples into regions specifically includes the following steps:

[0043] S3.1, obtain the online model data file of the power grid. Since the online data source usually does not have the computing node data, we can only build the node switch power grid according to the physical nodes, branch information and switch status in the file. Let the set of power grid nodes be N = {n1, n2, ...n k}, the switch set is S = {s1,s2,...s m}, the branch set is L = {l1,l2,...l z}, obtain the busbar branch equivalent power grid through topological analysis;

[0044] S3.2, the connection relationship of each node in the power grid is calculated through the node switch matrix M∈{0,1} k×m and the node branch matrix W∈{0,1} k×l Description, where:

[0045]

[0046] S3.3, define the state quantity x of each switch j ∈{0, 1}, where:

[0047]

[0048] S3.4, starting from any branch l0 of the power grid, determine the nodes n0, n'0 at both ends of the branch, obtain the node sets N0, N0' connected to the nodes n0, n'0 respectively through the matrix M and the switch state quantity, and replace the nodes of the two sets with the bus b0, b'0, update the matrix M and the matrix W, and repeat the above operations until all the switches retained in the matrix M are disconnected, that is, the bus branch equivalent power grid of the node switch power grid is obtained, and then the bus branch equivalent power grid is partitioned using the partitioning method in step S1 to obtain each power grid autonomous region.

[0049] Furthermore, the fault auxiliary decision result of the current online power grid data obtained in step S4 specifically includes the following steps:

[0050] The trained model is loaded into the neural network of each autonomous region’s intelligent agent, the current grid status is input, and the load transfer auxiliary decision is returned.

[0051] The advantages and beneficial effects of the present invention are:

[0052] This distribution network load transfer auxiliary decision-making method based on the partitioned autonomous model adopts a power grid partitioned autonomous model, divides the large-scale distribution network into multiple small-scale distribution network areas through the topological connection characteristics of the power grid, designs a deep reinforcement learning fault recovery transfer model for a single distribution network partition to realize load transfer in each area, and adopts multi-agent collaborative technology to realize decision-making coordination between the autonomous regions of each distribution network, while accelerating the decision-making efficiency of each autonomous region to improve the emergency response capability of power grid faults. After a large-scale power outage occurs in the distribution network, the autonomous partition is obtained according to the partitioning algorithm. For the partition where the fault occurs, the system state is input into the deep neural network to obtain the load transfer switch action auxiliary decision. This method effectively utilizes the physical characteristics of the distribution network itself, simplifies the model calculation complexity, and improves the calculation efficiency of the operation auxiliary decision. The present invention can finally quickly calculate the load transfer auxiliary decision after a large-scale power outage occurs in the distribution network, thereby improving the power supply reliability of the power grid. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 is a flow chart of the method of the present invention;

[0054] Figure 2 The figure is a flow chart of the partitioning method of the present invention. DETAILED DESCRIPTION

[0055] The present invention is further described in detail below through specific examples. The following examples are only illustrative and not restrictive, and the protection scope of the present invention cannot be limited thereto.

[0056] A distribution network load transfer auxiliary decision method based on a partitioned autonomous model, the innovation of which is that it includes the following steps:

[0057] S1. Obtain offline distribution network model and fault information, and divide the offline samples into regions;

[0058] S2. Combined with the partitioning results of step S1, a multi-agent deep reinforcement learning model based on autonomous partitioning is constructed, and the model is trained based on offline fault data until the model converges;

[0059] S3. Obtain online distribution network model data and divide the online samples into regions;

[0060] S4. Input the partitioning result obtained in step S3 into the trained and converged model in step S2 to obtain the fault auxiliary decision result of the current online power grid data;

[0061] Step S1 obtains the offline distribution network model and fault information, and divides the offline samples into regions, specifically including the following steps:

[0062] S1.1, obtain the offline model data file of the power grid, build the bus branch power grid according to the verified computing nodes and branch information in the file, and build the node switch power grid according to the physical nodes, branch information and switch status in the same source data, and determine the mapping relationship between the two power grids according to the file data.

[0063] S1.2, for the bus branch power grid, the voltage amplitude sensitivity relationship between any two buses is defined as follows:

[0064] ΔU i =α ij ΔU j

[0065]

[0066] In the formula, ΔU i,j and U i,j Respectively represent the voltage amplitude deviation and voltage amplitude at busbar i and j; α ij It represents the voltage amplitude sensitivity coefficient between busbar i and busbar j, Q i,j Represents the reactive injection power of bus i, j.

[0067] S1.3, define the electrical distance between busbar i and busbar j as d ij , to eliminate differences due to orientation:

[0068] d ij =d ji =--log(α ij α ji )

[0069] S1.4, based on the above two parameters, define the voltage sensitivity matrix and electrical distance matrix of the bus branch network as D = [d ij ],A=[a ij ], define the comprehensive weight matrix G = [g ij ] is used to describe the similarity between nodes, where:

[0070]

[0071] In the formula, α∈[0,1],β∈[0,1] are adjustment parameters used to balance the impact of electrical distance and voltage sensitivity on partitioning. They need to be selected according to the specific conditions of the actual distribution network. At the same time, a comprehensive weight lower limit g is specified according to the actual parameters of the power grid. min , when the combined weight coefficient of the two buses is lower than g min It is not allowed to be divided into the same partition.

[0072] S1.5, in order to limit the number of grid partitions within a reasonable range, the modularity H of the entire grid after partitioning is defined as follows:

[0073]

[0074] Where b n is the total number of grid buses, δ(c i ,c j )satisfy:

[0075]

[0076] In the formula, c i Represents the partition number where bus i is located.

[0077] The specific process of the busbar branch grid partitioning algorithm is as follows:

[0078] a Initialize each bus as an independent partition;

[0079] b For each busbar i, calculate and move it to each busbar where the comprehensive weight coefficient is greater than g min Modularity changes in adjacent bus partitions. Select the partition with the largest modularity change and move bus i to the corresponding partition;

[0080] c Repeat step b until there are no nodes that can be moved or the modularity cannot be further improved.

[0081] S1.6, applying the partitioning algorithm in the bus branch power grid, after obtaining the bus branch power grid partition, mapping the partitioning result to the node switch power grid to complete the power grid partitioning.

[0082] Step S2 constructs a multi-agent deep reinforcement learning model based on autonomous partitioning, and trains the model based on offline fault data until the model converges, specifically including the following steps:

[0083] S2.1, each partition subnetwork is modeled as an intelligent agent, and the observation space and shared observation space are defined. The observation space mainly includes the bus voltage, load demand, branch load rate, and switch disconnection status of the partition, and the shared observation space mainly includes the power supply remaining capacity and power supply load rate;

[0084] For a specific power grid partition, the agent input mainly consists of the observation space within the partition and the shared observation space of the partition and the neighboring partitions. The final agent output forms the auxiliary decision of the switch action in the partition.

[0085] S2.2,In order to ensure that the decision of the agent neural network can converge, the reinforcement learning method is used to design the reward function to guide the agent. At the same time, in order to ensure that the convergence strategy meets the operation specifications of the distribution network, it is necessary to set constraints for each auxiliary decision. In general, the reward function of the agent can be set as;

[0086]

[0087] Where P recover is the load restored after the agent makes a decision, P loss is the active power loss of the system after the agent makes a decision, and λ is the adjustment coefficient to ensure that the reward function is within a reasonable range;

[0088] S2.3, in addition to the reward function, the system must also comply with safety constraints. System safety constraints involve multiple aspects, among which the power flow equation constraint is determined by the return value of the power grid power flow calculation simulation software. In addition, the distribution network mainly includes node voltage constraints, branch capacity constraints and generator output constraints:

[0089]

[0090] Where U i represents the voltage at node i, Represents the voltage upper and lower limits of node i, P i gen represents the active power of generator i, P i min / max represents the upper and lower limits of the active power of generator i, represents the reactive power of generator i, Represents the upper and lower limits of reactive power of generator i, I ij represents the current in the branch between node i and node j, Indicates the upper limit of the current flowing through the branch. If the system does not meet these constraints after a certain auxiliary decision, the agent will be punished by -1 according to the reward function calculation formula;

[0091] S2.4, after completing the setting of the agent model and reward function, the neural network is trained according to the fault scenario derived offline until the reward function converges, which means the training is completed.

[0092] Step S3 obtains the online distribution network model data and divides the online samples into regions, specifically including the following steps:

[0093] S3.1, obtain the online model data file of the power grid. Since the online data source usually does not have the computing node data, we can only build the node switch power grid according to the physical nodes, branch information and switch status in the file. Let the set of power grid nodes be N = {n1, n2, ...n k}, the switch set is S = {s1,s2,...s m}, the branch set is L = {l1,l2,...l z}, obtain the busbar branch equivalent power grid through topological analysis;

[0094] S3.2, the connection relationship of each node in the power grid is calculated through the node switch matrix M∈{0,1} k×m and the node branch matrix W∈{0,1} k×l Description, where:

[0095]

[0096] S3.3, at the same time, define the state quantity x of each switch j ∈{0, 1}, where:

[0097]

[0098] S3.4, starting from any branch l0 of the power grid, determine the nodes n0, n'0 at both ends of the branch, obtain the node sets N0, N0' connected to the nodes n0, n'0 respectively through the matrix M and the switch state quantity, and replace the nodes of the two sets with the bus b0, b'0, update the matrix M and the matrix W, and repeat the above operation until the switches retained in the matrix M are all disconnected, that is, the bus branch equivalent power grid of the node switch power grid is obtained. Then the bus branch equivalent power grid is partitioned using the partitioning method in step S1 to obtain each power grid autonomous region.

[0099] The fault auxiliary decision result of the current online power grid data obtained in step S4 specifically includes the following steps:

[0100] The trained model is loaded into the neural network of each autonomous region’s intelligent agent, the current grid status is input, and the load transfer auxiliary decision is returned.

[0101] Although the embodiments and drawings of the present invention are disclosed for illustrative purposes, those skilled in the art will appreciate that various substitutions, changes and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the scope of the present invention is not limited to the contents disclosed in the embodiments and drawings.

Claims

1. A distribution network load transfer auxiliary decision method based on a partitioned autonomous model, characterized by: The following steps are involved: S1, obtain the offline distribution network model and fault information, divide the area according to the offline samples, and obtain the partition result; S2, combining the partitioning result of step S1, constructing a multi-agent deep reinforcement learning model based on autonomous partitioning, and training the multi-agent deep reinforcement learning model based on offline fault data until the multi-agent deep reinforcement learning model converges; S3, obtaining the online distribution network model data, dividing the area according to the online samples, and obtaining the partitioning results; S4, input the partitioning result obtained in step S3 into the multi-agent deep reinforcement learning model that has been trained and converged in step S2 to obtain the fault auxiliary decision-making result of the current online power grid data.

2. The load transfer auxiliary decision method for distribution network based on a partitioned autonomous model according to claim 1 is characterized in that: The method of obtaining the offline distribution network model and fault information in step S1, dividing the offline samples into regions, and obtaining the partition results includes the following steps: S1.1, obtain the offline model data file of the power grid, build the busbar branch power grid according to the verified computing nodes and branch information in the file, and build the node switch power grid according to the physical nodes, branch information and switch status in the homologous data, and determine the mapping relationship between the two power grids according to the file data; S1.2, for the bus branch power grid, the voltage amplitude sensitivity relationship between any two buses is defined as follows: D.U. i =a ij D.U. j In the formula, ΔU i,j and U i,j Respectively represent the voltage amplitude deviation and voltage amplitude at busbar i and j; α ij It represents the voltage amplitude sensitivity coefficient between busbar i and busbar j, Q i,j represents the reactive injection power of busbar i,j; S1.3, define the electrical distance between busbar i and busbar j as d ij , to eliminate differences due to orientation: d ij =d ji =--log(α ij ·a ji ) S1.4, based on the voltage amplitude sensitivity coefficient α between busbars i and j ij and the electrical distance d between busbar i and busbar j ij The two parameters define the voltage sensitivity matrix and electrical distance matrix of the bus branch network respectively as D = [d ij ],A=[a ij ], define the comprehensive weight matrix G = [g ij ] is used to describe the similarity between nodes, where: In the formula, α∈[0,1],β∈[0,1] are adjustment parameters used to balance the impact of electrical distance and voltage sensitivity on partitioning. They need to be selected according to the specific conditions of the actual distribution network. At the same time, a comprehensive weight lower limit g is specified according to the actual parameters of the power grid. min , when the combined weight coefficient of the two buses is lower than g min It is not allowed to be divided into the same partition; S1.5, in order to limit the number of grid partitions within a reasonable range, the modularity H of the entire grid after partitioning is defined as follows: Where b n is the total number of grid buses, δ(c i ,c j )satisfy: In the formula, c i Represents the partition number where bus i is located; S1.6, applying the partitioning algorithm to the bus branch grid, after obtaining the bus branch grid partition, mapping the partitioning result to the node switch grid to complete the grid partitioning.

3. The load transfer auxiliary decision method for distribution network based on a partitioned autonomous model according to claim 2 is characterized in that: The specific process of the busbar branch grid partitioning algorithm is as follows: a Initialize each bus as an independent partition; b For each busbar i, calculate and move it to each busbar where the comprehensive weight coefficient is greater than g min Modularity changes in adjacent bus partitions, select the partition with the largest modularity change, and move bus i to the corresponding partition; c Repeat step b until there are no nodes that can be moved or the modularity cannot be further improved.

4. The method for auxiliary decision-making of load transfer in distribution network based on partitioned autonomous model according to claim 1, characterized in that: Step S2 constructs a multi-agent deep reinforcement learning model based on autonomous partitioning, and trains the model based on offline fault data until the model converges, specifically including the following steps: S2.1, each partition subnetwork is modeled as an intelligent agent, and the observation space and shared observation space are defined. The observation space mainly includes the bus voltage, load demand, branch load rate, and switch disconnection status of the partition, and the shared observation space mainly includes the power supply remaining capacity and power supply load rate; For a specific grid partition, the agent input consists of the observation space within the partition and the shared observation space of the partition and neighboring partitions. The final agent output forms the auxiliary decision of the switch action in the partition. S2.2,In order to ensure that the decision of the agent neural network can converge, the reinforcement learning method is used to design the reward function to guide the agent. At the same time, in order to ensure that the convergence strategy meets the operation specifications of the distribution network, it is necessary to set constraints for each auxiliary decision. The reward function of the agent can be set as; Where P recover is the load restored after the agent makes a decision, P loss is the active power loss of the system after the agent makes a decision, and λ is the adjustment coefficient to ensure that the reward function is within a reasonable range; S2.3, in addition to the reward function, the system must also comply with safety constraints, where the power flow equation constraint is determined by the return value of the power grid power flow calculation simulation software. In addition, the distribution network mainly includes node voltage constraints, branch capacity constraints and generator output constraints: Where U i represents the voltage at node i, Represents the voltage upper and lower limits of node i, P i gen represents the active power of generator i, P i min / max represents the upper and lower limits of the active power of generator i, represents the reactive power of generator i, Represents the upper and lower limits of reactive power of generator i, I ij represents the current in the branch between node i and node j, Indicates the upper limit of the current flowing through the branch. If the system does not meet these constraints after a certain auxiliary decision, the agent will be punished by -1 according to the reward function calculation formula; S2.4, after completing the setting of the agent model and reward function, the neural network is trained according to the fault scenario derived offline until the reward function converges, which means the training is completed.

5. The method for auxiliary decision-making of load transfer in distribution network based on partitioned autonomous model according to claim 1 is characterized in that: Step S3 obtains the online distribution network model data and divides the online samples into regions, specifically including the following steps: S3.1, obtain the online model data file of the power grid. Since the online data source usually does not have the computing node data, we can only build the node switch power grid according to the physical nodes, branch information and switch status in the file. Let the set of power grid nodes be N = {n1, n2, ...n k }, the switch set is S = {s1,s2,...s m }, the branch set is L = {l1,l2,...l z }, obtain the busbar branch equivalent power grid through topological analysis; S3.2, the connection relationship of each node in the power grid is calculated through the node switch matrix M∈{0,1} k×m and the node branch matrix W∈{0,1} k×l Description, where: S3.3, define the state quantity x of each switch j ∈{0, 1}, where: S3.4, starting from any branch l0 of the power grid, determine the nodes n0, n'0 at both ends of the branch, obtain the node sets N0, N0' connected to the nodes n0, n'0 respectively through the matrix M and the switch state quantity, and replace the nodes of the two sets with the bus b0, b'0, update the matrix M and the matrix W, and repeat the above operations until all the switches retained in the matrix M are disconnected, that is, the bus branch equivalent power grid of the node switch power grid is obtained, and then the bus branch equivalent power grid is partitioned using the partitioning method in step S1 to obtain each power grid autonomous region.

6. The method for auxiliary decision-making of load transfer in distribution network based on partitioned autonomous model according to claim 1, characterized in that: The fault auxiliary decision result of the current online power grid data obtained in step S4 specifically includes the following steps: The trained model is loaded into the neural network of each autonomous region’s intelligent agent, the current grid status is input, and the load transfer auxiliary decision is returned.

Images