An artificial intelligence-based distribution network scheduling decision method

Abstract

The application relates to an artificial intelligence-based distribution network scheduling decision method, belonging to the technical field of power grid scheduling, which comprises the following steps: collecting distribution network topology data to construct a node topology graph composed of power nodes and their connection relationships; collecting node operation data and preprocessing to form node feature vectors and edge feature vectors; inputting the node topology graph and the feature vectors into a region-node double-layer topology perception graph network model to obtain fusion features of each node; inputting the fusion features into a time sequence coding model to obtain comprehensive time sequence features of each node; scoring a preset scheduling action set based on an action decoder and outputting a probability distribution, selecting the scheduling action with the highest probability as the current node execution action, realizing intelligent scheduling decision of the distribution network, and improving the operation safety and economy of the distribution network.

Description

Technical Field

[0001] This invention relates to an artificial intelligence-based distribution network scheduling decision-making method, belonging to the field of power grid scheduling technology. Background Technology

[0002] With the large-scale integration of distributed power sources and energy storage devices into distribution networks, the structure of distribution networks is becoming increasingly complex, and their operation is characterized by multi-source, multi-point, bidirectional power flow, and strong time-varying features. Traditional distribution network dispatching decisions mainly rely on the experience of dispatchers or heuristic algorithms based on simple rules. When dealing with large-scale and complex distribution networks, it is difficult to provide timely and accurate dispatching solutions that balance safety and economy.

[0003] To improve decision-making efficiency, some technical solutions have begun to utilize machine learning or deep learning methods to model distribution network operation data, in order to assist or replace manual dispatching. However, existing data-driven intelligent dispatching methods still suffer from the following common shortcomings:

[0004] Insufficient utilization of complex network structures: Some methods focus on directly inputting node measurement data into a general neural network model without fully combining the electrical connection relationships between buses, feeders, tie lines, transformers, distributed power sources and energy storage devices, and do not accurately characterize the coupling characteristics of multiple nodes and branches in the distribution network.

[0005] The disconnect between physical constraints and data-driven approaches: When generating or recommending scheduling schemes, existing methods often use historical operation records or experience-based schemes as learning targets, making it difficult to effectively integrate electrical constraints such as voltage acceptable range, line thermal stability limits, and the number of times equipment can be interrupted into the model. This results in some "optimal" outputs having the risk of exceeding limits or overloading in actual operation, requiring dispatchers to manually screen them.

[0006] Limited adaptability to changes in operating conditions: Under the conditions of load fluctuations, changes in distributed power generation output, and weather disturbances, the operating conditions of the distribution network exhibit obvious time series characteristics. Existing methods generally do not adequately consider this dynamic process, and the robustness of the models to changes in operating conditions and their ability to predict future states are limited, making it difficult to continuously provide stable and reliable scheduling support. Summary of the Invention

[0007] To address the problems existing in the prior art, this invention proposes an artificial intelligence-based distribution network scheduling decision-making method.

[0008] The technical solution of the present invention is as follows:

[0009] On the one hand, this invention proposes an artificial intelligence-based distribution network scheduling decision-making method, including the following steps:

[0010] The distribution network topology data is collected to construct a distribution network node topology graph, which consists of power nodes and edges formed by the connection relationships between each power node;

[0011] After collecting and preprocessing the running data of each node, feature vectors for each node are constructed, and feature vectors for each edge are also constructed.

[0012] A two-layer topology sensing graph network model for regional nodes is constructed. The topology graph of the distribution network nodes, the feature vectors of all nodes, and the feature vectors of all edges are combined and then input into the two-layer topology sensing graph network model for regional nodes to obtain the fused node features of each node.

[0013] A temporal coding model is constructed, and the fused node features of each node are input into the temporal coding model to obtain the comprehensive temporal features of each node;

[0014] Construct an action decoder, input the comprehensive temporal features of each node into the action decoder, and output the probability distribution of each scheduled action for a preset set of scheduled actions;

[0015] The scheduling action with the highest probability is selected as the execution action for the current node.

[0016] Preferably, the distribution network topology data includes the topological connection relationships of the primary equipment and lines of the distribution network.

[0017] Preferably, the feature vectors of each node include voltage, active power, reactive power, load type, distributed power output, energy storage power, and meteorological characteristics;

[0018] The feature vectors of each side include line impedance, line current, line thermal stability limit, and switching state.

[0019] Preferably, the regional node two-layer topology sensing graph network model includes regional units, node units, and regional node fusion units;

[0020] The regional unit divides the nodes in the distribution network node topology diagram into multiple regions according to the distribution network topology, and constructs a regional feature vector based on the feature vector of the nodes in each region, and constructs a regional-level graph based on the regional feature vector and the regional connection relationship.

[0021] The regional unit includes an attention layer for inter-regional energy mutual assistance relationships. The attention weight of each regional node in the regional-level graph is calculated through the attention layer for inter-regional energy mutual assistance relationships to obtain a weighted regional-level graph.

[0022] The node unit includes an attention layer with dynamic bias of impedance power flow. The attention weight of each node in the distribution network node topology is calculated through the attention layer with dynamic bias of impedance power flow to obtain a weighted distribution network node topology.

[0023] The regional node fusion unit reduces the dimensionality of the weighted regional level graph to the same dimension as the weighted distribution network node topology graph and then combines them to obtain a fused topology graph. The feature vector of each node in the fused topology graph is extracted to obtain the fused node features of each node.

[0024] Preferably, the timing coding model is constructed based on a gated recurrent neural network.

[0025] Preferably, the action decoder includes a fully connected feature transformation layer, a feature fusion layer, and an action scoring output layer;

[0026] The overall temporal features of each node are aggregated to obtain the network aggregated state vector. The network aggregated state vector is then input into the fully connected feature transformation layer to obtain the network aggregated state vector mapped to the latent space.

[0027] The feature fusion layer concatenates the aggregated state vector of the entire network mapped to the latent space with the comprehensive temporal features of each node and then performs secondary feature extraction to obtain the secondary features of each node.

[0028] The action scoring output layer outputs the score of each node for each scheduling action in the preset scheduling action set based on the secondary features of each node.

[0029] Preferably, pre-set scheduling action constraints, and calculate the mask for each scheduling action based on the scheduling action constraints;

[0030] Apply the mask of each scheduling action to the score of each scheduling action of each node to obtain the final score of each scheduling action of the current node;

[0031] The probability distribution of each scheduling action of the current node is calculated based on the final score of each scheduling action of the current node.

[0032] Preferably, the action decoder loss consists of action matching cross-entropy loss and power flow performance loss.

[0033] On the other hand, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in the present invention.

[0034] In another aspect, the present invention also provides a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the method described in the present invention.

[0035] The present invention has the following beneficial effects:

[0036] 1. This invention simultaneously incorporates the topology information of the distribution network and multi-source operational measurement data during the scheduling decision-making process, and jointly models the overall operating status of the distribution network. Compared with schemes that rely solely on local measurements or simple features, this invention can more comprehensively and accurately reflect the current operating conditions, providing a more reliable state basis for subsequent scheduling decisions.

[0037] 2. This invention introduces relevant constraint mechanisms during the generation of scheduling actions to filter or suppress candidate actions that do not meet the conditions for safe operation. This enables the output scheduling scheme to meet the control objectives while avoiding problems such as voltage overruns and line overloads, thereby improving the engineering feasibility and safety of the scheme.

[0038] 3. In the model training phase, this invention incorporates operational indicators related to the scheduling objective into the optimization objective, enabling the model to learn from historical experience while paying more attention to key indicators such as power supply reliability, voltage quality, and network loss. This allows the model to achieve a scheduling effect that balances safety and economy in practical applications and is conducive to improving the overall operation level of the distribution network under the condition of new energy access. Attached Figure Description

[0039] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] It should be understood that the step numbers used in the text are for ease of description only and are not intended to limit the order in which the steps are performed.

[0042] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0043] The terms “comprising” and “including” indicate the presence of the described feature, whole, step, operation, element and / or component, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or collections thereof.

[0044] The term “and / or” refers to any combination of one or more of the associated listed items, as well as all possible combinations, and includes these combinations.

[0045] See Figure 1 In some embodiments, an artificial intelligence-based distribution network scheduling decision-making method is proposed, including the following steps:

[0046] The distribution network topology data is collected to construct a distribution network node topology graph, which consists of power nodes and edges formed by the connection relationships between each power node;

[0047] After collecting and preprocessing the running data of each node, feature vectors for each node are constructed, and feature vectors for each edge are also constructed.

[0048] A two-layer topology sensing graph network model for regional nodes is constructed. The topology graph of the distribution network nodes, the feature vectors of all nodes, and the feature vectors of all edges are combined and then input into the two-layer topology sensing graph network model for regional nodes to obtain the fused node features of each node.

[0049] A temporal coding model is constructed, and the fused node features of each node are input into the temporal coding model to obtain the comprehensive temporal features of each node;

[0050] Construct an action decoder, input the comprehensive temporal features of each node into the action decoder, and output the probability distribution of each scheduled action for a preset set of scheduled actions;

[0051] The scheduling action with the highest probability is selected as the execution action for the current node.

[0052] In some embodiments, the distribution network topology data includes the topological connection relationships of primary equipment and lines in the distribution network.

[0053] In one specific embodiment, the primary equipment of the distribution network includes switches, tie lines, feeders, transformers, distributed power sources, energy storage devices, etc., and each primary equipment of the distribution network is regarded as a node, and the nodes are connected according to the connection relationship between the primary equipment of the distribution network.

[0054] In some embodiments, the feature vectors of each node include voltage, active power, reactive power, load type (different types are assigned different codes), distributed power output, energy storage power, and meteorological characteristics (ambient temperature and ambient humidity).

[0055] The feature vectors of each side include line impedance, line current, line thermal stability limit, and switching state.

[0056] In one specific embodiment, the feature vectors of each node and the feature vectors of each edge are normalized.

[0057] In some embodiments, the regional node two-layer topology sensing graph network model includes regional units, node units, and regional node fusion units;

[0058] The regional unit divides the nodes in the distribution network node topology diagram into multiple regions according to the distribution network topology, and constructs a regional feature vector based on the feature vector of the nodes in each region, and constructs a regional-level graph based on the regional feature vector and the regional connection relationship.

[0059] The regional unit includes an attention layer for inter-regional energy mutual assistance relationships. The attention weight of each regional node in the regional-level graph is calculated through the attention layer for inter-regional energy mutual assistance relationships to obtain a weighted regional-level graph.

[0060] The node unit includes an attention layer with dynamic bias of impedance power flow. The attention weight of each node in the distribution network node topology is calculated through the attention layer with dynamic bias of impedance power flow to obtain a weighted distribution network node topology.

[0061] The regional node fusion unit reduces the dimensionality of the weighted regional level graph to the same dimension as the weighted distribution network node topology graph and then combines them to obtain a fused topology graph. The feature vector of each node in the fused topology graph is extracted to obtain the fused node features of each node.

[0062] In one specific embodiment, the regional unit divides the nodes in the distribution network node topology into multiple regions according to the transformer area, and constructs a regional feature vector based on the feature vector of the nodes in each region, as shown in the following formula:

[0063] ;

[0064] in: Indicates the first Feature vectors of each region; This indicates that average pooling and extreme pooling operations are performed separately, followed by a concatenation operation. Indicates the first A set of nodes in each region; Represents the first node in the set of region nodes. Feature vectors of each node;

[0065] For each node in the distribution network node topology diagram, if it belongs to different regions and any of the node devices is identified as a transformer, it is determined that there is a relationship between the regions to which the two nodes belong. After connecting the regions based on the connection relationship between the regions, a region-level diagram is obtained.

[0066] In one specific embodiment, the attention weight of each region node in the region-level graph is calculated through the attention layer based on the energy mutual assistance relationship between regions, as shown below:

[0067] ;

[0068] ;

[0069] in: Indicates the region To the area Attention weights; Indicates the first Feature vectors of each region; Representing a regional map Central region With the region Feature vectors of the edges connecting them; Representing a regional map Central and regional A set of adjacent regions; Representation and region The first in the set of adjacent regions Feature vectors of each region; Representing a regional map Central region With the region The first in the set of adjacent regions Feature vectors of the edges connecting the regions; The encoding function representing the edge features of a region is used to map the edge feature vectors of a region to the same dimension; The vector mapping function consists of several fully connected layers and non-linear activation layers, and is used to map the concatenated vectors to a scalar. This represents the weighted region-level map. Feature vectors of each region; Indicates the activation function; Representing a regional map The characteristic linear transformation matrix; This indicates a splicing operation.

[0070] In a specific embodiment, the feature vector of the inter-region connection edge is specifically the set of feature vectors of the connection edges between all nodes in the region.

[0071] In one specific embodiment, the attention layer with impedance power flow dynamic bias includes multiple sub-layers. The attention weight of each node in the distribution network node topology graph is calculated through the attention layer with impedance power flow dynamic bias, as shown in the following formula:

[0072] ;

[0073] in: Indicates the first The nodes output by each sub-layer With nodes Attention weights between them; This represents the LeakyReLU activation function; Indicates the first The attention parameter vectors of each sub-layer are used to map the concatenated vectors to a unified scalar; Indicates the transpose operation; Indicates the first Nodes in each sub-layer The hidden feature vector (if the current sublayer is the first layer, then the node) (using the feature vector as input) Indicates the first Nodes in each sub-layer The hidden feature vector; Indicates the first Linear transformation matrices of node features in each sub-layer are used to map hidden feature vectors to the attention space; This represents the linear transformation matrix of edge features, used to map edge feature vectors to the same dimension as node feature vectors; Represents a node With nodes Feature vectors of the edges connecting them; Represents a node With nodes The tidal characteristics between them specifically include the magnitude and direction of the active tidal current; Represents a non-linear coding function; Represents a node The set of adjacent nodes; Indicates the first Nodes in each sub-layer The set of adjacent nodes of the first The hidden feature vectors of each node; Represents a node With nodes Feature vectors of the connecting edges between them; Represents a node With nodes Trend characteristics between;

[0074] The formula for calculating the hidden feature vector of each sub-layer is as follows:

[0075] ;

[0076] in: Indicates the first Nodes in each sub-layer The hidden feature vector;

[0077] Finally, the hidden feature vector of each node output from the last sub-layer is updated to the corresponding node in the distribution network node topology graph to obtain the weighted distribution network node topology graph.

[0078] In a specific embodiment, the nonlinear coding function is as follows:

[0079] ;

[0080] in: Represents a node With nodes The maximum absolute value of the active power flow allowed between them; Represents a sign function, when , This indicates that the current active power flow is along the agreed direction (i.e., the node). Flow to Node ),when hour, , This indicates that the current power flow is approximately zero, meaning there can be no active power transmission on this line; when , This indicates that the current active power flow is in the opposite direction (i.e., at the node). Flow to Node ).

[0081] In a specific embodiment, the regional node fusion unit maps the weighted regional-level graph to the same dimension as the weighted distribution network node topology graph and then combines them to obtain a fused topology graph, as shown in the following formula:

[0082] ;

[0083] in: Represents nodes in the merged topology graph Features of fusion nodes; The node representing the output of the last sublayer of the attention layer with impedance power flow dynamic bias. The hidden feature vector; It represents the linear mapping matrix from the region feature space to the node feature space.

[0084] In some embodiments, the timing coding model is constructed based on a gated recurrent neural network.

[0085] In a specific embodiment, the fused node features of each node are temporally encoded using a gated recurrent neural network to obtain the comprehensive temporal features of each node, as shown in the following formula:

[0086] ;

[0087] in: express Time Node The comprehensive temporal characteristics; This represents a gated recurrent neural network, which internally includes update gates, reset gates, etc. express Time Node The characteristics of the fusion nodes, based on Extraction of fused feature maps at different time points; express Time Node The comprehensive temporal characteristics.

[0088] In some embodiments, the action decoder includes a fully connected feature transformation layer, a feature fusion layer, and an action scoring output layer;

[0089] The overall temporal features of each node are aggregated to obtain the network aggregated state vector. The network aggregated state vector is then input into the fully connected feature transformation layer to obtain the network aggregated state vector mapped to the latent space.

[0090] The feature fusion layer concatenates the aggregated state vector of the entire network mapped to the latent space with the comprehensive temporal features of each node and then performs secondary feature extraction to obtain the secondary features of each node.

[0091] The action scoring output layer outputs the score of each node for each scheduling action in the preset scheduling action set based on the secondary features of each node.

[0092] In a specific embodiment, the aggregated temporal features of each node are combined to obtain the aggregated state vector of the entire network, as shown in the following formula:

[0093] ;

[0094] ;

[0095] ;

[0096] in: express The aggregated state vector of the entire network at any given time; express The mean of the comprehensive temporal characteristics of each time point; express The maximum value of the comprehensive temporal characteristics at each time point; Indicates the total number of nodes;

[0097] The aggregated state vector of the entire network is input into the fully connected feature transform layer to obtain the aggregated state vector of the entire network mapped to the latent space, as shown in the following formula:

[0098] ;

[0099] in: express The time-mapping of the network's aggregated state vector to the latent space; express function; This represents the first weight matrix; This indicates the first bias.

[0100] In one specific embodiment, the feature fusion layer concatenates the aggregated state vector of the entire network mapped to the latent space with the comprehensive temporal features of each node and then performs secondary feature extraction to obtain the secondary features of each node.

[0101] ;

[0102] ;

[0103] in: express Time Node The secondary characteristics; express Time Node Comprehensive time series characteristics and The concatenation result of the aggregated state vectors of the entire network mapped to the latent space at each time step; Represents the ReLU function, which enhances the stability of gradient propagation; This represents the second weight matrix; This indicates the second bias.

[0104] In a specific embodiment, the action scoring output layer outputs the score of each node for each scheduled action in the preset set of scheduled actions based on the secondary features of each node, as shown in the following formula:

[0105] ;

[0106] in: express Time Node For each scheduling action in the preset set of scheduling actions, a score set is provided. This represents the output weight matrix; This indicates the output bias.

[0107] In one specific embodiment, the scheduling actions include switching on and off, interconnection switching, reactive power compensation adjustment, and energy storage charging and discharging.

[0108] In some embodiments, scheduling action constraints are preset, and a mask for each scheduling action is calculated based on the scheduling action constraints.

[0109] Apply the mask of each scheduling action to the score of each scheduling action of each node to obtain the final score of each scheduling action of the current node;

[0110] The probability distribution of each scheduling action of the current node is calculated based on the final score of each scheduling action of the current node.

[0111] In one specific embodiment, the scheduling action constraints include node voltage constraints, transformer hot current constraints, distributed power generation output / ramp constraints, and energy storage device SOC / power constraints.

[0112] For all node constraints, calculate the mask for each scheduling action as follows:

[0113] For each scheduling action, predict the state of the node after the action is executed. If the state of the node after execution is outside the constraint range, set the mask to a sufficiently large positive number, and then subtract the score from the mask to make the score approximately equal to 0.

[0114] Node voltage constraints are applied to all nodes. In particular, transformer hot current constraints are applied to transformer nodes, power output / ramp constraints are applied to distributed power generation nodes, and SOC / power constraints are applied to energy storage device nodes.

[0115] In a specific embodiment, the probability distribution of each scheduling action of the current node is calculated based on the final score of each scheduling action of the current node, as shown in the following formula:

[0116] ;

[0117] in: express Time Node For the preset set of scheduling actions, the first... The probability of a scheduling action; express Time Node For the The final score for each scheduling action.

[0118] In some embodiments, the action decoder loss consists of action matching cross-entropy loss and power flow performance loss.

[0119] In one specific embodiment, the action decoder loss The specific formula is as follows:

[0120] ;

[0121] ;

[0122] ;

[0123] in: This represents the action matching cross-entropy loss; express Time Node For the The actual score of each scheduling action; This indicates a loss in power flow performance; This indicates the total active power loss of the network under the current scheduling action; This represents the baseline active power loss value (such as a typical operating condition or historical average loss). Indicates the node under the current scheduling action. The voltage amplitude; Indicates the voltage reference value; This indicates the maximum permissible voltage deviation (10% of the rated voltage in this embodiment). Indicates the total number of edges; Represents a node With nodes The current amplitude of the line corresponding to the connecting edge; Represents a node With nodes The maximum allowable current amplitude of the line corresponding to the connecting edge; , , The weighting coefficients represent the formula for calculating power flow performance loss.

[0124] When the loss of the action decoder reaches a preset threshold, the trained action decoder is obtained.

[0125] In some embodiments, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method as described in any embodiment of the present invention.

[0126] In some embodiments, a computer-readable storage medium is provided on which a computer program is stored, which, when executed by a processor, implements the method as described in any embodiment of the present invention.

[0127] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, A and B simultaneously, or B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" and similar expressions refer to any combination of these items, including any combination of singular or plural items. For example, at least one of a, b, and c can represent: a, b, c, a and b, a and c, b and c, or a and b and c, where a, b, and c can be single or multiple.

[0128] Those skilled in the art will recognize that the units and algorithm steps described in the embodiments disclosed herein can be implemented using electronic hardware, computer software, or a combination of electronic hardware and software. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0129] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0130] In the several embodiments provided in this application, any function, if implemented as a software functional unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0131] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A distribution network scheduling decision-making method based on artificial intelligence, characterized in that, Includes the following steps: The distribution network topology data is collected to construct a distribution network node topology graph, which consists of power nodes and edges formed by the connection relationships between each power node; After collecting and preprocessing the running data of each node, feature vectors for each node are constructed, and feature vectors for each edge are also constructed. A two-layer topology sensing graph network model for regional nodes is constructed. The topology graph of the distribution network nodes, the feature vectors of all nodes, and the feature vectors of all edges are combined and then input into the two-layer topology sensing graph network model for regional nodes to obtain the fused node features of each node. A temporal coding model is constructed, and the fused node features of each node are input into the temporal coding model to obtain the comprehensive temporal features of each node; Construct an action decoder, input the comprehensive temporal features of each node into the action decoder, and output the probability distribution of each scheduled action for a preset set of scheduled actions; Select the scheduling action with the highest probability as the execution action for the current node; The regional node two-layer topology sensing graph network model includes regional units, node units, and regional node fusion units. The regional unit divides the nodes in the distribution network node topology diagram into multiple regions according to the distribution network topology, and constructs a regional feature vector based on the feature vector of the nodes in each region, and constructs a regional-level graph based on the regional feature vector and the regional connection relationship. The regional unit includes an attention layer for inter-regional energy mutual assistance relationships. The attention weight of each regional node in the regional-level graph is calculated through the attention layer for inter-regional energy mutual assistance relationships to obtain a weighted regional-level graph. The node unit includes an attention layer with dynamic bias of impedance power flow. The attention weight of each node in the distribution network node topology is calculated through the attention layer with dynamic bias of impedance power flow to obtain a weighted distribution network node topology. The regional node fusion unit reduces the dimensionality of the weighted regional level graph to the same dimension as the weighted distribution network node topology graph and then combines them to obtain a fused topology graph. The feature vector of each node in the fused topology graph is extracted to obtain the fused node features of each node.

2. The distribution network scheduling decision-making method based on artificial intelligence according to claim 1, characterized in that, The distribution network topology data includes the topological connection relationships of primary equipment and lines in the distribution network.

3. The distribution network scheduling decision-making method based on artificial intelligence according to claim 1, characterized in that, The feature vectors of each node include voltage, active power, reactive power, load type, distributed power output, energy storage power, and meteorological characteristics. The feature vectors of each side include line impedance, line current, line thermal stability limit, and switching state.

4. The distribution network scheduling decision-making method based on artificial intelligence according to claim 1, characterized in that, The time-series coding model is constructed based on a gated recurrent neural network.

5. The distribution network scheduling decision-making method based on artificial intelligence according to claim 1, characterized in that, The action decoder includes a fully connected feature transformation layer, a feature fusion layer, and an action scoring output layer; The overall temporal features of each node are aggregated to obtain the network aggregated state vector. The network aggregated state vector is then input into the fully connected feature transformation layer to obtain the network aggregated state vector mapped to the latent space. The feature fusion layer concatenates the aggregated state vector of the entire network mapped to the latent space with the comprehensive temporal features of each node and then performs secondary feature extraction to obtain the secondary features of each node. The action scoring output layer outputs the score of each node for each scheduling action in the preset scheduling action set based on the secondary features of each node.

6. The distribution network scheduling decision-making method based on artificial intelligence according to claim 5, characterized in that, Pre-set scheduling action constraints, and calculate the mask for each scheduling action based on the scheduling action constraints; Apply the mask of each scheduling action to the score of each scheduling action of each node to obtain the final score of each scheduling action of the current node; The probability distribution of each scheduling action of the current node is calculated based on the final score of each scheduling action of the current node.

7. The distribution network scheduling decision-making method based on artificial intelligence according to claim 5, characterized in that, The action decoder loss consists of action matching cross-entropy loss and power flow performance loss.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 7.

Images