Power distribution network grid target network frame planning method and system based on artificial intelligence

Abstract

The application discloses a power distribution network grid target network framework planning method and system based on artificial intelligence, establishes a power distribution network multi-modal semantic atlas, defines power grid equipment as nodes and edges and gives structural, operation and geographical modal characteristic vectors. The network framework operation coupling encoder is used to extract features, the topological attention mechanism is used to learn the electrical connection relationship, and the problem sensitivity attention mechanism is used to dynamically adjust the node weight according to the overload rate, voltage deviation, three-phase imbalance and frequency outage number. The network framework generator based on dynamic constraint reinforcement learning is constructed, the action space containing newly-built lines, adjusted switch positions, upgraded device capacity and load cutting is defined, and a composite reward function is designed. The agent interacts and executes actions in the dynamic simulation environment and updates the policy network parameters until the target network framework planning scheme is output. The application realizes the deep integration of the network framework structure and the operation problem, and the generated scheme meets the standard liaison mode and has high robustness.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of power system planning and artificial intelligence, and relates to a method and system for planning the target grid structure of distribution network based on artificial intelligence. Background Technology

[0002] The distribution network is a crucial component of the power system, and its structure directly impacts power supply reliability and power quality. With the large-scale integration of distributed generation and the rapid growth of electricity load, the operating environment of the distribution network is becoming increasingly complex. Existing distribution network planning methods primarily rely on traditional mathematical optimization algorithms or conventional machine learning-assisted methods. Traditional mathematical optimization algorithms typically establish mixed-integer linear programming models, setting the minimum investment cost or minimum network loss as the objective function, and solving them by combining power flow equations and radial operating constraints. Such methods are computationally inefficient when dealing with large-scale nodes and struggle to flexibly adapt to changing operating scenarios. Conventional machine learning-assisted methods often utilize historical data to train regression or classification models for predicting load growth or recommending simple line modification schemes. Some studies have attempted to introduce graph neural networks to extract grid topology features, but significant limitations remain.

[0003] Existing technologies have several shortcomings in addressing the needs of grid-based transformation of distribution networks. First, existing models lack sufficient representation of the network structure. Most methods treat the power grid as a general graph structure, lacking the ability to distinguish specific engineering semantic structures such as single-radial, standard interconnections, and multi-segment, multi-interconnection structures. This often leads to planning schemes that do not conform to power engineering specifications, easily resulting in non-radial operational risks or unreasonable interconnection point settings. Second, existing methods have weak ability to handle the coupling of multi-dimensional operational problems. Heavy overload, low voltage, three-phase imbalance, and frequent outages are interconnected in real power grids. Existing technologies typically treat these indicators as independent constraints or transform them into a single objective through weighted summation, ignoring the strong coupling mechanisms between problems and making precise management difficult. Third, existing technologies have poor adaptability to dynamic scenarios. Traditional planning is mostly based on typical days or peak load times, ignoring the risks of frequent outages under distributed generation output fluctuations and fault scenarios. Existing static mapping models cannot simulate the dynamic evolution process under future high-proportion renewable energy integration, resulting in insufficient robustness of planning schemes in extreme scenarios. Finally, existing technical solutions have poor interpretability. Deep learning models are often regarded as black boxes, and the output grid structure schemes lack physical meaning and rarely consider construction feasibility and corridor resource constraints, making it difficult to implement the schemes directly.

[0004] Therefore, there is an urgent need for a distribution network grid target grid planning method that can deeply integrate grid structure characteristics with multi-dimensional operational problem characteristics and has dynamic scenario adaptability. Summary of the Invention

[0005] To address the problems existing in the background technology, this invention proposes a method and system for planning the target grid structure of power distribution networks based on artificial intelligence.

[0006] The first aspect of this application provides an artificial intelligence-based method for planning the target grid structure of a power distribution network, including:

[0007] A multimodal semantic graph of the distribution network is constructed, defining substations, switching stations, distribution transformers and user access points as nodes, and lines, tie switches and sectional switches as edges, and assigning feature vectors containing structural modes, operational modes and geographical modes to nodes and edges;

[0008] The network operation coupled encoder based on the attention mechanism is used to extract features from the multimodal semantic map of the distribution network. The electrical connection relationship between nodes is learned through the topology attention mechanism. The node weight is dynamically adjusted according to the heavy overload rate, voltage deviation, three-phase imbalance and frequency of outages through the problem sensitivity attention mechanism. The output is a node implicit state vector that integrates local topology and global operation problem information.

[0009] A grid generator based on dynamic constraint reinforcement learning is constructed, and the grid planning process is modeled as a Markov decision process. An action space including new line construction, switch position adjustment, equipment capacity upgrade and load cutover is defined. A composite reward function including safety reward, grid structure reward, economic reward and robustness reward is designed.

[0010] The agent interacts with the current state of the network structure in a dynamic simulation environment to perform actions, updates the policy network parameters according to the composite reward function, and outputs the target network structure planning scheme after the model converges.

[0011] Optionally, the feature vector of the structural mode includes node type identifier, edge physical impedance, edge rated capacity and current switch state, wherein the node type identifier includes single-radial mode label, standard interconnection mode label and multi-segment multi-interconnection mode label; the feature vector of the operating mode includes normalized heavy overload index, low voltage index, three-phase imbalance index and annual frequency outage index; the feature vector of the geographic mode includes geographic information system coordinates and corridor resource constraint factor.

[0012] Optionally, the calculation process of the attention coefficient in the problem sensitivity attention mechanism includes: calculating the unnormalized attention score between adjacent nodes, which is obtained by activating the transpose of the learnable attention vector and the concatenation vector through a linear rectifier unit, wherein the concatenation vector is formed by concatenating the hidden state vector of the starting node, the hidden state vector of the ending node, and the difference vector of the severity of the running problem between the two nodes; performing an exponential operation on the unnormalized attention score and normalizing it to obtain the final attention coefficient, so that adjacent nodes with large differences in the severity of the running problem have higher attention weights.

[0013] Optionally, the calculation formula of the composite reward function is as follows: the reward value equals the first safety weight coefficient multiplied by the safety reward item plus the second weight coefficient multiplied by the grid structure reward item minus the third weight coefficient multiplied by the investment cost item minus the fourth weight coefficient multiplied by the network loss power item plus the fifth weight coefficient multiplied by the robustness score item; wherein the safety reward item is the sum of the safety indication functions of all monitoring points, and the safety indication function takes a positive value when the monitoring point has no overload, no low voltage, no three-phase imbalance exceeding the standard and no frequency outage risk; otherwise, it takes a negative penalty value; the grid structure reward item is the sum of the structure indication functions of all grid units, and the structure indication function takes a positive value when the grid units form a preset standard connection mode; otherwise, it takes a zero value.

[0014] Optionally, the dynamic simulation environment is built based on the DistFlow power flow equation and supports second-level time-series simulation. During the training process, the dynamic simulation environment randomly disturbs the load curve and the output of the distributed power source to simulate the real operating scenario, and injects random faults in each scenario to evaluate the power supply recovery capability of the grid structure. The evaluation results are used as robustness scoring terms and input into the composite reward function.

[0015] Optionally, the method further includes a scheme generation and adaptive verification step: parsing the current status data of the area to be planned input by the converged model into an action sequence to generate a preliminary target network scheme containing the specifications of new lines, switch status adjustment instructions and equipment upgrade list; using the physical engine verification layer to perform full-time power flow calculation and safety verification on the preliminary target network scheme; if the verification fails, the verification result is used as a negative feedback signal to fine-tune the strategy network parameters until the final planning scheme that meets all engineering constraints is output.

[0016] Optionally, the load cut-off action in the action space is specifically for areas where the three-phase imbalance exceeds a preset threshold. The action execution includes adjusting the node phase sequence connection or redistributing the load carried by the node to different phases to reduce the ratio of the effective value of the negative sequence current to the effective value of the positive sequence current at the node.

[0017] A second aspect of this application provides an artificial intelligence-based distribution network grid target structure planning system, comprising:

[0018] The semantic module is used to construct a multimodal semantic graph of the distribution network. It defines substations, switching stations, distribution transformers and user access points as nodes, lines, tie switches and sectionalizing switches as edges, and assigns feature vectors containing structural modes, operational modes and geographical modes to nodes and edges.

[0019] The encoding module is used to extract features from the multimodal semantic graph of the distribution network using a grid operation coupled encoder based on an attention mechanism. It learns the electrical connection relationship between nodes through a topology attention mechanism and dynamically adjusts the node weights based on the heavy overload rate, voltage deviation, three-phase imbalance and frequency of outages through a problem sensitivity attention mechanism. It outputs a node implicit state vector that integrates local topology and global operation problem information.

[0020] The module is used to build a grid generator based on dynamic constraint reinforcement learning. It models the grid planning process as a Markov decision process, defines the action space including new lines, adjustment of switch positions, upgrading of equipment capacity and load cutover, and designs a composite reward function including safety reward, grid structure reward, economic reward and robust reward.

[0021] The output module is used to perform actions by interacting with the current network state in a dynamic simulation environment through an intelligent agent, update the policy network parameters according to the composite reward function, and output the target network planning scheme after the model converges.

[0022] Optionally, when executing the problem sensitivity attention mechanism, the encoding module is configured to calculate the unnormalized attention score between adjacent nodes. This score is obtained by activating the transpose of the learnable attention vector and the concatenation vector through a linear rectifier unit. The concatenation vector is formed by concatenating the hidden state vector of the starting node, the hidden state vector of the ending node, and the difference vector of the severity of the running problem between the two nodes. The unnormalized attention score is then subjected to exponential operation and normalized to obtain the final attention coefficient.

[0023] Optionally, the output module further includes a physics engine verification unit, which is used to parse the action sequence output by the converged model into a preliminary target network scheme, and use a dynamic simulation environment based on the DistFlow power flow equation to perform full-time power flow calculation and safety verification on the preliminary target network scheme. If the verification fails, the verification result is transmitted to the construction module as a negative feedback signal to fine-tune the strategy network parameters until the final planning scheme that meets all engineering constraints is output.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] This invention provides an AI-based method and system for planning the target network structure of a power distribution network. By constructing a multimodal semantic graph of the power distribution network, it achieves refined modeling of equipment such as substations, switching stations, and lines. The graph embeds structural modal labels such as single-radial, standard interconnection, and multi-segment multi-interconnection structures. This structure enables the model to accurately identify the engineering semantic features of the current network structure. The system utilizes a network operation coupling encoder based on an attention mechanism for feature extraction. The topology attention mechanism effectively learns the electrical connection relationships between nodes. The problem sensitivity attention mechanism dynamically adjusts node weights based on heavy overload rate, voltage deviation, three-phase imbalance, and frequency of outages. This mechanism ensures that the model can focus on areas with serious operational problems. The node latent state vectors output by the model integrate local topology and global operational problem information. This provides a precise data foundation for subsequent planning.

[0026] This invention constructs a network structure generator based on dynamic constraint reinforcement learning. This generator models the network structure planning process as a Markov decision process. The action space encompasses various operations such as constructing new lines, adjusting switch positions, upgrading equipment capacity, and load cutover. The composite reward function includes safety rewards, mesh structure rewards, economic rewards, and robustness rewards. The agent interacts with the current network structure state in a dynamic simulation environment to execute actions. The system updates the policy network parameters according to the composite reward function until the model converges. The planning scheme generated by this method strictly conforms to the requirements of standard interconnection modes. The scheme effectively eliminates heavy overload and low voltage problems. The scheme significantly reduces three-phase imbalance and frequent power outage risks. The planning results exhibit strong robustness and can adapt to the uncertainties of load fluctuations and distributed power source integration.

[0027] This invention also includes a scheme generation and adaptive verification process. The physical engine verification layer performs full-time power flow calculations and safety checks on the preliminary target network structure scheme. If the verification fails, the system uses the verification result as a negative feedback signal to fine-tune the strategy network parameters. This closed-loop process ensures that the final output planning scheme meets all engineering constraints. This invention significantly improves the automation and intelligence level of distribution network planning. The generated scheme has clear physical meaning and high engineering feasibility. This method effectively solves the technical problems of insufficient topology representation capability and poor coupling of multi-source heterogeneous problems in existing technologies. Attached Figure Description

[0028] Figure 1 This is a flowchart of the distribution network grid target structure planning method based on artificial intelligence in an embodiment of the present invention;

[0029] Figure 2 This is a schematic diagram of a power distribution network grid target structure planning system based on artificial intelligence in an embodiment of the present invention. Detailed Implementation

[0030] 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.

[0031] In one embodiment, such as Figure 1 As shown, an artificial intelligence-based method for planning the target grid structure of a power distribution network is provided, and this method is applied to... Figure 1 Taking China as an example, the following specific steps will be used:

[0032] S10: Construct a multimodal semantic graph of the distribution network, defining substations, switching stations, distribution transformers and user access points as nodes, and lines, tie switches and sectionalizing switches as edges, and assigning feature vectors containing structural modes, operational modes and geographical modes to nodes and edges.

[0033] Specifically, the process of constructing a multimodal semantic graph of the distribution network first involves the abstract modeling of the physical power grid. The system abstracts substations, switching stations, distribution transformers, and user access points as nodes in the graph. Simultaneously, the lines connecting these devices, tie switches, and sectionalizing switches are abstracted as edges in the graph. This definition method ensures the consistency between the graph topology and the real physical power grid.

[0034] For each node and edge, the system assigns a three-dimensional feature vector containing structural, operational, and geographical modes. The structural mode feature vector describes the inherent properties and topological role of the equipment. For example, a node's structural features include a node type identifier, clearly distinguishing between single-radial, standard interconnection, or multi-segment multi-interconnection modes. An edge's structural features include physical impedance values, rated capacity values, and the current open / closed state of the switch. The operational mode feature vector characterizes the electrical performance of the equipment under real-time or historical operating conditions. This vector includes normalized overload indices, undervoltage indices, three-phase imbalance indices, and annual outage indices. These indices reflect the current health and potential risks of the power grid. The geographical mode feature vector incorporates spatial constraint information, including geographic information system coordinate data and corridor resource constraint factors. Corridor resource constraint factors quantify the construction difficulty or land constraints along the transmission line route.

[0035] By fusing the aforementioned multimodal features, the distribution network multimodal semantic graph not only preserves the topological connections of the power grid but also embeds rich engineering semantics and operational status information. This graph structure enables subsequent neural network models to simultaneously perceive the physical connections, operational bottlenecks, and geographical limitations of the power grid. This invention overcomes the shortcomings of traditional graph models that only focus on topological connections while ignoring operational status. The model can accurately identify nodes under heavy overload or low voltage conditions and understand the difficulty of their modification in specific geographical environments. This lays a solid data foundation for generating grid planning schemes that conform to engineering specifications and possess high feasibility.

[0036] S20: The network operation coupled encoder based on the attention mechanism is used to extract features from the multimodal semantic map of the distribution network. The electrical connection relationship between nodes is learned through the topology attention mechanism. The node weight is dynamically adjusted according to the heavy overload rate, voltage deviation, three-phase imbalance and frequency of outages through the problem sensitivity attention mechanism. The node implicit state vector is output, which integrates local topology and global operation problem information.

[0037] Specifically, the grid operation coupled encoder based on the attention mechanism is the core module for processing the multimodal semantic graph of the distribution network. The encoder first receives the graph data constructed in the preceding steps as input. Internally, the encoder deploys two sub-modules in parallel: a topology attention mechanism and a problem-sensitivity attention mechanism. The topology attention mechanism focuses on learning the electrical connectivity relationships between nodes. This mechanism aggregates the structural and geographical modal features of neighboring nodes by calculating the correlation scores between them. For example, for a distribution transformer node, the topology attention mechanism weightedly fuses the feature information of its upstream switchyard nodes and its peer tie line nodes. This process enables each node to perceive its electrical location and connection strength within the local grid structure.

[0038] The problem-sensitivity attention mechanism focuses on uncovering global operational issues. This mechanism reads operational mode data from each node in real time, including overload rate, voltage deviation, three-phase imbalance, and number of outages. The system dynamically generates attention weights based on the severity of these data. When a node's voltage deviation exceeds a preset threshold or its number of outages is high, the problem-sensitivity attention mechanism automatically assigns it a higher weight. This means the model will focus on areas with serious operational problems during feature extraction. Conversely, nodes with good operational status receive lower weights. This dynamic adjustment strategy ensures the encoder can accurately detect weaknesses in the power grid.

[0039] Finally, the encoder concatenates and fuses the local structural feature vector output by the topology attention mechanism and the global operational problem feature vector output by the problem sensitivity attention mechanism. The fused result is the node's hidden state vector. This vector contains both the node's topology connection information and its current operational health status. Examples show that, in complex distribution network scenarios involving multiple heavily overloaded lines, this encoder can accurately highlight the feature representations of high-fault-prone areas.

[0040] This invention effectively solves the problem of separating topology from operational issues in traditional methods. The model no longer treats heavy overload or low voltage as isolated indicators, but rather deeply couples them with the grid structure. This allows the subsequently generated planning schemes to specifically address key operational problems, significantly improving the accuracy and effectiveness of the planning.

[0041] S30: Construct a grid generator based on dynamic constraint reinforcement learning, model the grid planning process as a Markov decision process, define the action space including new line construction, switch position adjustment, equipment capacity upgrade and load cutover, and design a composite reward function including safety reward, grid structure reward, economic reward and robust reward.

[0042] Specifically, this invention constructs a network structure generator based on dynamic constraint reinforcement learning, which is a concrete implementation of applying artificial intelligence technology to power system planning. The network structure generator models the entire network planning process using Markov decision processes as its mathematical foundation. In the Markov decision process model, the state space is composed of the current power grid topology, node load distribution, equipment operating parameters, and line impedance data. The environment updates the power grid state based on the actions performed by the agent and feeds back corresponding reward signals.

[0043] The action space design covers the core operation types in network planning. The "New Line" action allows agents to add transmission channels between any two nodes that are not directly connected. The "Adjust Switch Position" action restructures the network topology by changing the open / closed state of tie switches or sectionalizing switches. The "Upgrade Equipment Capacity" action expands the capacity of transformers or line conductors to increase transmission capacity. The "Load Cutover" action transfers the electrical load of a specific node to an adjacent power supply path to balance the regional load. All actions are subject to dynamic constraints based on physical connection feasibility and the equipment selection library.

[0044] The composite reward function consists of four parts: safety reward, grid structure reward, economic reward, and robustness reward. The safety reward assesses whether the grid operation meets voltage deviation limits, line thermal stability limits, and the N-1 check standard. When the system exceeds these limits, the safety reward outputs a negative value to guide the agent to avoid risks. The grid structure reward encourages the formation of standard wiring patterns such as daisy-chain networks or multi-source radial networks. This reward provides positive incentives for topologies with clear power supply zones and reasonable interconnectivity. The economic reward comprehensively considers line construction costs, equipment investment costs, and the discounted value of network losses. The agent tends to choose the planning scheme with the lowest total investment and the lowest operating cost over its entire life cycle. The robustness reward measures the grid's resilience in the face of load fluctuations or component failures. Highly robust schemes can quickly restore power supply and maintain voltage stability at critical nodes after disturbances.

[0045] For example, the agent's initial state is an existing single-radial distribution network in a certain area. The agent first performs a new line construction action, establishing a connection channel at the ends of two main feeders. Then, the agent performs an adjustment switch position action, closing the connection switch and disconnecting the original sectionalizing switch to achieve load transfer. Next, the agent detects that a certain main transformer has an excessively high load rate and performs an equipment capacity upgrade action, replacing it with a larger capacity transformer. Finally, the agent performs a load cutover action, relocating some users on heavily overloaded branches to lightly loaded lines. During the execution of the above action sequence, a dynamic constraint mechanism verifies the electrical safety of each step in real time. If any action causes a voltage drop exceeding the allowable range, the action is immediately terminated and a high penalty is applied.

[0046] This invention significantly improves the scientific rigor and adaptability of power grid planning schemes. The dynamic constraint mechanism ensures that every intermediate state and final scheme generated complies with power engineering safety standards. The multi-objective guidance of the composite reward function achieves an optimal balance between safety, economy, and reliability in the planning results. The self-learning characteristics based on reinforcement learning enable the power grid generator to learn from historical planning cases and continuously optimize its strategy. This invention effectively avoids the pitfalls of traditional heuristic algorithms that easily get trapped in local optima. The generated power grid scheme has clear physical meaning and is easy to implement in engineering. The planning process realizes a transformation from being driven by human experience to being driven by data intelligence. The final power grid structure has stronger anti-interference capabilities and better resource allocation efficiency.

[0047] S40: The agent interacts with the current state of the network structure in a dynamic simulation environment to perform actions, updates the policy network parameters according to the composite reward function, and outputs the target network structure planning scheme after the model converges.

[0048] Specifically, the interaction between the agent and the current grid state in the dynamic simulation environment is the core of model training. The dynamic simulation environment incorporates power flow calculation and fault analysis modules. This environment can simulate the electrical characteristics of the power grid under different topologies in real time. The agent first reads the current grid state information, including node voltage magnitudes, line power flow direction, and equipment load rates. Based on the probability distribution output by the policy network, the agent selects a specific action from the action space. This action could be building a new transmission line or adjusting the opening / closing state of a switch. Upon receiving this action, the environment immediately updates the grid topology. Subsequently, the environment recalculates the power flow to obtain new system operating data.

[0049] Updating the policy network parameters based on the composite reward function is a key step in achieving intelligent decision-making. The system calculates the immediate reward value immediately after an action is executed. This reward value is obtained by a weighted sum of safety rewards, grid structure rewards, economic rewards, and robustness rewards. If an action leads to voltage exceeding limits or line overload, the safety reward term will have a large negative value. If an action optimizes the network structure and reduces investment costs, the economic reward term will have a positive value. The agent stores the current state, executed action, immediate reward, and next-time state as experience samples. The policy network calculates gradients by sampling these experience samples. The gradient descent algorithm is used to adjust the weight parameters within the network. This process causes the policy network to gradually favor action sequences that accumulate higher total rewards. A dynamic constraint mechanism performs a secondary verification of the action's feasibility before each parameter update. The probability of any action that violates physical constraints is forcibly reduced to zero.

[0050] The convergence criterion for the model is that the planning scheme output by the policy network remains stable over multiple consecutive training cycles. The model is considered converged when the evaluation metric no longer changes significantly with increasing training rounds. At this point, training is stopped and the policy network parameters are frozen. The frozen model is then used as a target grid generator. After inputting the initial power grid data for the area to be planned, the generator directly outputs a series of optimal action sequences. Applying these actions sequentially to the initial power grid yields the final target grid planning scheme. This scheme clearly defines the new line paths, the switch locations to be adjusted, the equipment models to be upgraded, and the load nodes to be cut over.

[0051] For example, the training process began with a test distribution network containing 50 nodes. The agent performed 100,000 iterations in the simulation environment. In the initial stage, the agent randomly tried various line connection methods and frequently triggered safety penalties. As training progressed, the agent gradually learned to prioritize building ring network structures to improve power supply reliability. In the 50,000th iteration, the agent began to actively perform load cutover actions to balance transformer loads. The final model reached convergence in the 80,000th iteration. The output solution showed that two new tie lines were built between three critical load centers. The solution also recommended upgrading two old transformers to high-efficiency models. All switching operation sequences passed the N-1 safety check.

[0052] This invention provides a low-cost and risk-free trial-and-error platform for the dynamic simulation environment. This enables the agent to explore a massive number of grid combination schemes in a very short time. The parameter update mechanism based on the composite reward function ensures the synchronous improvement of multi-objective optimization effects. The generated planning scheme not only meets strict electrical safety constraints but also maximizes investment benefits. The scheme output after model convergence has high interpretability and engineering applicability. This invention effectively solves the problem that traditional planning methods struggle to handle high-dimensional nonlinear constraints. The agent has adaptive learning capabilities and can generate customized schemes for the characteristics of power grids in different regions. The final grid structure has stronger scalability and operational resilience. The entire planning process significantly shortens the time required for manual scheme development.

[0053] In one embodiment, in step S10, the feature vector of the structural mode includes node type identifier, edge physical impedance, edge rated capacity, and current switch state, wherein the node type identifier includes single-radial mode label, standard interconnection mode label, and multi-segment multi-interconnection mode label; the feature vector of the operating mode includes normalized overload index, low voltage index, three-phase imbalance index, and annual frequency outage index; the feature vector of the geographic mode includes geographic information system coordinates and corridor resource constraint factor.

[0054] Specifically, the feature vectors of structural modes are used to accurately describe the physical connection attributes and wiring configurations of the power grid topology. This vector primarily includes node type identifiers to distinguish the architectural characteristics of different power supply areas. Single-radial mode labels are used to mark end nodes powered by a single source and without interconnection channels. Standard interconnection mode labels are used to mark node areas with typical daisy-chain structures or dual-source interconnection capabilities. Multi-segment multi-interconnection mode labels are used to mark complex mesh nodes that have undergone multiple segmentation modifications and have multiple interconnection switches. The physical impedance of the edges, as a key component of the vector, reflects the impact of line resistance and reactance on power flow distribution. The rated capacity of the edges defines the maximum current or power limit allowed to pass through the transmission channel. The current state of the switches records the on / off status of interconnection switches and segment switches in binary form. These elements together constitute a digital mapping of the static structure of the power grid.

[0055] The eigenvectors of the operating modes are used to quantify the current electrical health level and operational risks of the power grid. The heavy overload index, after normalization, eliminates the influence of dimensions, directly reflecting the degree to which transformer or line load rates exceed safe thresholds. The low voltage index characterizes the deviation of node voltage amplitudes from the lower limit standard. The three-phase imbalance index measures the level of power quality degradation caused by inconsistencies in three-phase current or voltage amplitudes. The annual outage frequency index statistically analyzes the average annual number of outages caused by faults or planned maintenance in a specific area. All operating indicators are mapped to a unified numerical range through a standardized function. This approach ensures that operating parameters with different physical meanings can be compared and integrated on the same dimension. The operating mode vectors enable agents to keenly perceive real-time operational bottlenecks in the power grid.

[0056] The eigenvectors of the geographic modal place power grid planning under real-world geographic constraints. Geographic Information System (GIS) coordinates provide precise latitude and longitude information for substations, towers, and transmission line routes. This coordinate data supports the calculation of actual geographic distances between nodes, rather than electrical distances. Corridor resource constraint factors quantify the construction difficulties and environmental limitations along the transmission line routes. This factor comprehensively considers land use, ecological protection zone boundaries, underground pipeline density, and demolition costs. High constraint factor values ​​indicate that new transmission lines in the area face significant difficulties or are prohibited from construction. Low constraint factor values ​​indicate that the area has sufficient overhead or cable corridor resources. The geographic modal vectors ensure the feasibility of the planning scheme at the physical implementation level.

[0057] For example, the system extracts features from a power distribution network in a new urban area. Structural mode analysis identifies three core nodes as belonging to a multi-segment, multi-tie mode. The physical impedance data of the relevant edges shows 0.12 ohms per kilometer. The rated capacity of the edges is marked as 630 amps. The current switch status indicates the tie switch is in the open position. Operational mode analysis detects that a feeder's overload index reaches 0.85. The low voltage index at the end node shows 0.92 per unit. The three-phase imbalance index is 4%. The annual outage index is recorded as 1.5 times per year. Geographic mode analysis reveals that the planned route passes through a basic farmland protection zone. Therefore, the corridor resource constraint factor is set to a high resistance value. These data are collectively input into the network structure generator as the basis for decision-making.

[0058] The implementation of this technical solution significantly improves the perception accuracy and decision reliability of the grid planning model. The multi-label classification mechanism of structural modalities enables the model to identify and inherit excellent wiring experience from existing power grids. Normalization of operational modalities eliminates interference between different electrical magnitudes, accelerating model convergence. The introduction of geographical modalities effectively avoids planning schemes falling into geographically infeasible blind spots. The fusion of three-dimensional modal vectors achieves comprehensive coverage from electrical logic to physical space. The refined definition of feature vectors enhances the reinforcement learning agent's ability to express complex power grid states. The generated planning scheme can simultaneously consider topological rationality, operational safety, and geographical feasibility. This method significantly reduces the workload of subsequent on-site surveys and corrections for the planning scheme. The final output grid scheme has strong engineering guidance significance and practical implementation capability.

[0059] In one embodiment, in step S20, the calculation process of the attention coefficient in the problem sensitivity attention mechanism includes: calculating the unnormalized attention score between adjacent nodes, which is obtained by activating the transpose of the learnable attention vector and the concatenation vector through a linear rectifier unit, wherein the concatenation vector is formed by concatenating the hidden state vector of the starting node, the hidden state vector of the ending node, and the difference vector of the severity of the running problem between the two nodes; performing an exponential operation on the unnormalized attention score and normalizing it to obtain the final attention coefficient, so that adjacent nodes with large differences in the severity of the running problem have higher attention weights.

[0060] Specifically, the problem sensitivity attention mechanism is a key component in the grid generator for accurately identifying weak links in the power grid. This mechanism quantifies the importance correlation between node pairs by calculating the unnormalized attention score between adjacent nodes. The calculation process first constructs a splicing vector. This splicing vector is composed of three parts of data concatenated sequentially. The first part is the implicit state vector of the starting node. This vector encodes the structural characteristics and operational history of the starting node. The second part is the implicit state vector of the ending node. This vector contains the topology information and real-time operating conditions of the ending node. The third part is the difference vector of the severity of operational problems between the two nodes. This vector is obtained by calculating the difference between the starting node and the ending node in operational indicators such as heavy overload and low voltage. If one of the two adjacent nodes is in a state of severe overload while the other has a lighter load, the value of this difference vector will be larger.

[0061] The learnable attention vector is a parameter matrix automatically optimized during model training. This vector is multiplied by the transpose of the concatenated vector. The product reflects the potential correlation strength between the current node pair and the problem-solving process. The result is then activated by a linear rectifier unit. The linear rectifier unit truncates negative values ​​to zero and retains positive values. This step ensures that the unnormalized attention score only includes positive contribution metrics. The activated value is the unnormalized attention score. This score initially characterizes the degree to which the agent should pay attention to the node pair.

[0062] To obtain the final attention coefficient, the system performs an exponential operation on the unnormalized attention score. This exponential operation amplifies the difference between high and low scores, making significantly important node pairs stand out. The system then normalizes the exponential scores of all neighboring nodes. The normalization process converts all scores into a probability distribution that sums to 1. The final attention coefficient represents the weight allocated by the agent to each neighboring node pair during decision-making. Since the severity difference vector of the running problem directly participates in the score calculation, neighboring node pairs with larger differences in running problem severity will receive higher unnormalized scores. After exponential amplification and normalization, these node pairs ultimately have higher attention weights. This means that when planning the network structure, the agent will prioritize areas with significant differences in running states and potential power flow bottlenecks.

[0063] For example, a power distribution network includes nodes A and B. Node A is under heavy load and has low voltage. Node B is under light load and has normal voltage. The difference vector of operational problem severity between the two is significant. The concatenated vector integrates the latent states of nodes A and B and this difference vector. The learnable attention vector, after interacting with this concatenated vector, outputs a high unnormalized score through a linear rectifier unit. In contrast, nodes C and D are both in normal operating condition with minimal difference. Their corresponding unnormalized scores are low. After exponential operation and normalization, the attention coefficient between nodes A and B is much higher than that between nodes C and D. When generating new lines or adjusting switches, the agent assigns higher priority to connection operations between nodes A and B.

[0064] The problem-sensitivity attention mechanism in this invention enables the model to adaptively focus on critical conflict areas in the power grid. By introducing a severity difference vector for operational problems, the model no longer treats all topological connections equally. This mechanism effectively improves the agent's sensitivity to local fault risks and power flow congestion. High attention weights guide the strategy network to prioritize solving the most severe operational limit-breaking problems. This method accelerates the efficiency of reward signal transmission in the reinforcement learning process. The model can learn effective governance strategies for heavy overload or low voltage problems more quickly. The generated grid planning scheme is highly targeted. The scheme prioritizes eliminating transmission bottlenecks between nodes with significantly different operational indicators. This significantly improves the overall power supply balance and security stability of the power grid. The final planning result avoids the ineffective investment of resources in non-critical areas.

[0065] In one embodiment, in step S30, the calculation formula of the composite reward function is as follows: the reward value equals the first safety weight coefficient multiplied by the safety reward item plus the second weight coefficient multiplied by the grid structure reward item minus the third weight coefficient multiplied by the investment cost item minus the fourth weight coefficient multiplied by the network loss power item plus the fifth weight coefficient multiplied by the robustness score item; wherein the safety reward item is the sum of the safety indication functions of all monitoring points, and the safety indication function takes a positive value when the monitoring point has no overload, no low voltage, no three-phase imbalance exceeding the standard and no frequency outage risk; otherwise, it takes a negative penalty value; the grid structure reward item is the sum of the structure indication functions of all grid units, and the structure indication function takes a positive value when the grid unit forms a preset standard connection mode; otherwise, it takes a zero value.

[0066] Specifically, the composite reward function is the core evaluation criterion guiding the optimization decisions of the grid generator. The calculation formula for this function consists of a linear combination of five key components. The first component is the safety reward, multiplied by a primary safety weight coefficient. This coefficient is set to its maximum value to reflect the paramount importance of grid operation safety. The second component is the grid structure reward, multiplied by a secondary weight coefficient. The third component is the investment cost component, multiplied by a tertiary weight coefficient and subtracted from it in the formula. The fourth component is the grid loss component, multiplied by a tertiary weight coefficient and subtracted from it in the formula. The fifth component is the robustness score component, multiplied by a tertiary weight coefficient and added to it in the formula. This combination of addition and subtraction aims to maximize safety, structural, and robustness benefits while minimizing construction and operating costs.

[0067] The specific calculation of the safety reward depends on the cumulative sum of the safety indication functions of all monitoring points. The system monitors key nodes and lines in the power grid in real time. When no heavy overload occurs at any monitoring point and the voltage is within the normal range, the safety indication function of that point takes a positive value. When no three-phase imbalance exceeds the limit and there is no risk of frequent power outages, the safety indication function of that point continues to accumulate a positive value. Once a heavy overload or low voltage exceeds the limit occurs at a monitoring point, the safety indication function of that point immediately converts to a negative penalty value. If a three-phase imbalance exceeds the limit or there is a risk of frequent power outages, a negative penalty value is also triggered. The sum of the indication function values ​​of all monitoring points yields the final safety reward. This mechanism ensures that any violation of safety constraints will lead to a significant decrease in the total reward value, thereby forcing the agent to avoid unsafe states.

[0068] The grid-based structure reward is the sum of the structure indicator functions of all grid cells. The system divides the planning area into several standard grid cells. For each grid cell, the system determines whether its topology conforms to a preset standard interconnection pattern. Standard interconnection patterns include typical reliable wiring forms such as daisy-chain networks and dual-power radial networks. When a grid cell successfully forms the above-mentioned preset standard interconnection pattern, its structure indicator function takes a positive value. If the grid cell still maintains a single radial shape or irregular and messy connections, its structure indicator function takes a zero value. The grid-based structure reward is obtained by summing the structure indicator function values ​​of all grid cells. This reward encourages agents to build standardized and normalized power grid architectures, improving the overall network manageability.

[0069] For example, the agent performed the action of creating a new tie line. After the action was completed, the system detected that the load rate of the original heavily loaded line had dropped to a safe range. All monitoring points showed no overload, no low voltage, no excessive three-phase imbalance, and no risk of frequency outages. At this time, the safety indicator function output positive values ​​at all monitoring points, significantly increasing the safety reward. Simultaneously, this action merged two independent radial grid cells into a standard daisy-chain network. The corresponding structure indicator function changed from zero to positive, increasing the grid structure reward. Although the new line slightly increased the investment cost and network loss, the total reward still showed an upward trend due to the large first and second weighting coefficients for safety. Therefore, the agent judged this action as a high-quality strategy and reinforced it.

[0070] This invention's composite reward function achieves a globally optimal balance between safety, standardization, and economy through fine-tuning of multi-dimensional weights. The negative penalty mechanism of the safety reward item constructs a strict electrical safety defense, preventing the generation of non-compliant schemes. The positive incentive effect of the grid structure reward item promotes the standardization process of power grid wiring patterns. The deduction mechanism of investment cost and power loss items prompts the agent to actively seek low-cost, low-loss planning paths. The introduction of a robustness scoring item enhances the power grid's ability to cope with future uncertainties. This method effectively solves the problem of coordinating multiple conflicting objectives in traditional planning. The generated grid scheme not only meets strict safety operation indicators but also possesses a clear and reasonable topology. The planning results significantly improve power supply reliability and resource utilization efficiency while reducing total life-cycle costs.

[0071] In one embodiment, in step S30, the load cut-off action in the action space is specifically for the region where the three-phase imbalance exceeds a preset threshold. The action execution includes adjusting the node phase sequence connection or redistributing the load carried by the node to different phases to reduce the ratio of the effective value of the negative sequence current to the effective value of the positive sequence current at the node.

[0072] Specifically, the load cutover action in the action space is a specific operational strategy designed to address three-phase imbalance issues. The trigger condition for this action is strictly limited to areas where the three-phase imbalance exceeds a preset threshold. The system first monitors the current waveform data of each node in real time. The calculation module extracts the effective values ​​of the negative-sequence current and the positive-sequence current. The system compares the ratio of these two values ​​with a preset threshold. When the ratio is higher than the threshold, the agent determines that there is a serious risk of three-phase imbalance in that area. At this time, the agent activates the load cutover action option from the action space. If the ratio does not exceed the threshold, the action option remains unavailable. This mechanism ensures that planned resources are concentrated on solving the most prominent power quality problems.

[0073] The actions performed primarily involve two specific physical adjustment methods. The first method is adjusting the phase sequence connection of nodes. The agent instructs the phase-switching switch to change the phase of the load connection. For example, switching the load originally connected to phase A to phase B or C. The second method is redistributing the loads connected to nodes to different phases. The agent transfers high-capacity single-phase loads from heavily loaded phases to lightly loaded phases. Through this redistribution, the total load on each phase tends to be balanced. The adjustment process is simulated in a dynamic simulation environment. The simulation module updates the current distribution of each phase in real time. The system then recalculates the ratio of the effective value of the negative-sequence current to the effective value of the positive-sequence current after adjustment. The goal is to significantly reduce this ratio and bring it back within the safe standard range.

[0074] For example, a severe three-phase imbalance was detected at the outlet of a transformer substation. Monitoring data showed that the load rate of phase A was 80%, phase B was 30%, and phase C was 25%. The calculated ratio of the effective value of the negative-sequence current to the effective value of the positive-sequence current far exceeded the preset threshold. The agent then generated a load cut-off action plan. The plan instructed to cut off two high-power air conditioning loads on phase A to phase B. At the same time, it instructed to switch some lighting loads from phase A to phase C. The simulation environment executed the above phase sequence adjustment and load redistribution operations. After the adjustment, the load rate of phase A decreased to 45%, the load rate of phase B increased to 55%, and the load rate of phase C increased to 40%. The recalculated ratio of the effective value of the negative-sequence current to the effective value of the positive-sequence current decreased significantly. This ratio is now below the preset threshold, indicating that the three-phase imbalance problem has been effectively resolved.

[0075] This invention specifically addresses areas where three-phase imbalance exceeds a preset threshold, enhancing the targeted nature of power grid operation. By adjusting the phase sequence connection of nodes, the phase distribution of single-phase loads is rapidly altered. The redistribution of loads to different phases by nodes achieves dynamic power balance across phases. Reducing the ratio of the effective value of negative-sequence current to the effective value of positive-sequence current at nodes directly improves power quality. This method effectively reduces additional transformer losses caused by three-phase imbalance. It also reduces safety hazards caused by excessive neutral current and extends the service life of power equipment. Load cutover actions, as an important component of the action space, enrich the governance methods of the intelligent agent. The generated planning scheme can solve imbalance problems without significant additional hardware investment. The final grid structure possesses stronger three-phase load adaptive adjustment capabilities.

[0076] In one embodiment, in step S40, the dynamic simulation environment is constructed based on the DistFlow power flow equation and supports second-level time-series simulation. During the training process, the dynamic simulation environment randomly disturbs the load curve and the output of the distributed power source to simulate the real operating scenario, and injects random faults in each scenario to evaluate the power supply recovery capability of the grid structure. The evaluation results are input into the composite reward function as a robustness scoring item.

[0077] Specifically, the dynamic simulation environment is the core interactive platform for training reinforcement learning agents in network planning. The underlying physical model of this environment is rigorously constructed based on the DistFlow power flow equations. The DistFlow equations are a simplified power flow calculation method applicable to radial and weakly looped distribution networks. This set of equations accurately describes the linear recursive relationship between the square of node voltage magnitude, branch power flow, and the square of branch current. Using these equations to build the environment significantly reduces computational complexity while maintaining computational accuracy. The environment architecture supports second-level time-series simulation capabilities. This means the simulator can extrapolate the entire day's operation of the power grid with a time step of seconds. The system can continuously simulate the dynamic changes of the power grid over a span of 24 hours or even longer. This high-time-resolution simulation mechanism ensures that the training data can capture the transient characteristics of load fluctuations and power output.

[0078] During training, the dynamic simulation environment introduces multiple random perturbation mechanisms to simulate the uncertainties of real-world operating scenarios. First, the environment applies random perturbations to the load curve. The system generates various typical daily load curve shapes based on historical statistical patterns. Random noise is superimposed on the baseline load value to simulate the randomness of user electricity consumption behavior. Simultaneously, the environment applies random perturbations to the output of distributed power sources. For photovoltaic power sources, the system simulates output drops caused by cloud cover. For wind power sources, the system simulates power oscillations caused by wind speed variations. These perturbations enable the agent to maintain robust decision-making even when faced with non-ideal data. Each perturbation combination constitutes an independent operating scenario. The agent must validate the adaptability of its planning scheme across thousands of such scenarios.

[0079] To evaluate the power supply recovery capability of the power grid structure, random faults are injected into the environment under each operating scenario. Fault types include single-phase-to-ground short circuits, phase-to-phase short circuits, and open-circuit faults. Fault locations are randomly selected within the power grid topology. Fault occurrence times are randomly triggered during the time-series simulation. Upon fault injection, the dynamic simulation environment immediately initiates the power supply recovery assessment program. The program simulates an agent executing recovery strategies such as network reconfiguration, load transfer, and islanding operation. The system records the total load that successfully restores power after fault isolation. The system calculates the number of operation steps and time cost required for power restoration. The system detects whether new voltage exceedances or line overloads occur during the recovery process. Based on these indicators, the environment calculates a quantitative robustness score. This score directly reflects the current power grid structure's survivability and self-healing level under extreme operating conditions.

[0080] The robustness score obtained from the evaluation is directly input into the composite reward function as a robustness score item. During the reward function calculation, this score item is multiplied by a fifth weighting coefficient and then summed to the total reward value. If the network structure can quickly recover most of the load under random faults without secondary limit exceedances, the robustness score item is a high positive value, significantly increasing the total reward. If the network structure causes a large-scale power outage under faults or cannot restore power through reconfiguration, the robustness score item is a low or negative value, significantly reducing the total reward. This feedback mechanism guides the agent to proactively consider fault defense and recovery strategies during the planning phase. The agent tends to generate network topologies with multiple communication channels, reasonable segmentation, and the potential for islanded operation.

[0081] For example, an agent is being trained on a new distribution network with multiple distributed photovoltaic (PV) connections. A dynamic simulation environment generates a summer afternoon operating scenario. At this time, the load is at its peak, and PV output fluctuates drastically due to cloud cover. The environment then injects a permanent three-phase short-circuit fault in the middle of the main line. The agent attempts to close the tie switch to transfer the load downstream of the fault to an adjacent feeder. Due to the fast computation characteristics of the DistFlow power flow equations, the environment completes the power flow verification after the transfer within milliseconds. Simulation results show that the transfer path does not experience overload and the terminal voltage meets requirements. More than 90% of the lost power load is restored within seconds. Based on this, the environment assigns a high robustness score. This score is included as a robustness score in the composite reward function, providing positive incentives to the agent. Conversely, if the transfer path causes overload on the new line, the score decreases, the agent is penalized, and its strategy is adjusted.

[0082] This invention achieves a perfect balance between computational efficiency and physical accuracy through a dynamic simulation environment built upon the DistFlow power flow equations. The second-level time-series simulation capability allows the agent to fully learn the temporal dynamic characteristics of the power grid. The mechanisms of random disturbance load curves and distributed generation output greatly enrich the diversity of training samples. This effectively avoids the problem of the agent overfitting to specific ideal scenarios. By injecting a random fault evaluation method into each scenario, the agent is forced to incorporate safety and defense thinking from the initial planning stage. The evaluation results are input as a robustness scoring term into the composite reward function, establishing a direct feedback loop from fault consequences to planning decisions. The generated grid structure not only meets normal operation indicators but also possesses strong disaster resilience and fault self-healing capabilities. This method significantly reduces the risk of power outages and economic losses when the power grid faces sudden faults in actual operation. The final planning results can adapt to the complex operational challenges brought about by the high proportion of new energy integration in the future.

[0083] Optionally, in one embodiment, the AI-based distribution network grid target structure planning method further includes a scheme generation and adaptive verification step: parsing the current status data of the area to be planned input by the converged model into an action sequence to generate a preliminary target structure scheme including new line specifications, switch status adjustment instructions and equipment upgrade list; using the physical engine verification layer to perform full-time power flow calculation and safety verification on the preliminary target structure scheme; if the verification fails, the verification result is used as a negative feedback signal to fine-tune the strategy network parameters until the final planning scheme that meets all engineering constraints is output.

[0084] Specifically, the scheme generation and adaptive verification steps are the final implementation steps in the AI-based distribution network grid target network planning method. This step begins in the application phase after model convergence. The system inputs the current status data of the area to be planned into the trained strategy network. The current status data includes geographic information, existing topology, load distribution, and equipment parameters. The strategy network parses this high-dimensional data into specific action sequences. The action sequences consist of a series of discrete decisions, directly mapped to engineering operations in the physical world. The generated preliminary target network scheme includes three core components. The first is the specifications for new lines, specifying the path, conductor type, and number of loops for the new lines. The second is the switch status adjustment instructions, specifying the opening and closing states of sectionalizing switches and tie switches. The third is the equipment upgrade list, listing transformers that need capacity replacement or compensation devices that need to be installed.

[0085] The physics engine verification layer then rigorously verifies the preliminary target network structure scheme. This verification layer incorporates a high-precision power system simulation kernel. The system performs full-time power flow calculations, simulating the operating state of the planned scheme over a typical 24-hour period. The calculation process extrapolates changes in node voltage, branch current, and power loss second by second. Simultaneously, a safety check is performed to examine for heavy overloads, low voltage, or exceeding thermal stability limits. The check result is either pass or fail. If the preliminary target network structure scheme meets the engineering constraints at all times, the check passes, and the scheme is confirmed as the final planned scheme. If the check fails, the system immediately triggers an adaptive fine-tuning mechanism.

[0086] When the verification fails, the physics engine's verification layer converts the specific violation data into a negative feedback signal. This negative feedback signal quantifies the degree to which the solution violates constraints. This signal is then fed back to the policy network. The policy network fine-tunes its internal parameters using a gradient descent algorithm based on the negative feedback signal. The parameter adjustment direction aims to reduce the probability of generating violating actions. The system then re-analyzes the current data based on the updated parameters, generating a revised action sequence and a new preliminary target network structure. The new scheme re-enters the physics engine's verification layer for full-time power flow calculation and safety verification. This process forms a closed-loop iteration until the output planning scheme fully satisfies all engineering constraints. This mechanism ensures that the AI-generated scheme is not only theoretically optimal but also feasible and reliable in practical engineering.

[0087] For example, a planned area in an old urban district has multiple power supply bottlenecks. The converged model first outputs a preliminary target network structure. The scheme suggests building a new tie line between two main lines and adjusting the states of three sectionalizing switches. The physics engine's verification layer performs full-time power flow calculations on this scheme. Simulations reveal that during the summer evening peak hours, the voltage drop along the path of the new tie line exceeds the allowable range, resulting in low voltage for end users. The verification result is unsatisfactory. The system generates and sends back a negative feedback signal containing the voltage exceedance magnitude. After receiving the signal, the policy network fine-tunes the parameters, reducing the probability of selecting the small-section conductor and increasing the weight of actions such as selecting a large-section conductor or adding reactive power compensation. The corrected scheme upgrades the new line specification to a larger cross-section cable and adds two parallel capacitors to the list. A second verification shows that the voltage index returns to normal, and all safety constraints are met. The system outputs this corrected scheme as the final planning scheme.

[0088] This invention's scheme generation and adaptive verification steps achieve a seamless connection from intelligent decision-making to engineering implementation. It parses the current status data of the area to be planned into an action sequence, ensuring the concretization and executability of the planning scheme. It generates a scheme including new line specifications, switch status adjustment instructions, and equipment upgrade lists, providing a complete engineering construction guide. Utilizing a physics engine verification layer for full-time power flow calculation and safety verification eliminates physically infeasible solutions that might arise from purely data-driven models. The verification results are used as negative feedback signals to fine-tune the strategy network parameters, constructing a self-correcting intelligent evolution mechanism. This method automatically avoids engineering risks without repeated intervention from human experts. The final output planning scheme possesses extremely high engineering practicality and safety reliability. This significantly shortens the distribution network planning and design cycle and reduces the manpower and time costs caused by scheme rework. The generated network structure can adapt to complex and ever-changing actual operating conditions, ensuring the long-term safe and stable operation of the power grid.

[0089] In one embodiment, such as Figure 2As shown, an AI-based distribution network grid target structure planning system is provided. This AI-based distribution network grid target structure planning system corresponds one-to-one with the AI-based distribution network grid target structure planning method in the above embodiments. The AI-based distribution network grid target structure planning system includes: a semantic module, an encoding module, a construction module, and an output module. The detailed description of each functional module is as follows:

[0090] The semantic module is used to define substations, switching stations, distribution transformers and user access points as nodes, lines, tie switches and sectionalizing switches as edges, and to assign feature vectors containing structural modes, operational modes and geographical modes to nodes and edges.

[0091] The encoding module is used to extract features from the multimodal semantic graph of the power distribution network using an encoder based on an attention mechanism, learn the electrical connection relationship between nodes through a topological attention mechanism, and dynamically adjust the node weights based on the heavy overload rate, voltage deviation, three-phase imbalance and frequency of outages through a problem sensitivity attention mechanism, and output the node implicit state vector that integrates local topology and global operation problem information.

[0092] The building module is used to build a generator based on dynamic constraint reinforcement learning, model the grid planning process as a Markov decision process, define the action space including new line construction, switch position adjustment, equipment capacity upgrade and load cutover, and design a composite reward function including safety reward, grid structure reward, economic reward and robust reward.

[0093] The output module is used to perform actions by interacting with the current network state in a dynamic simulation environment through an intelligent agent, update the policy network parameters according to the composite reward function, and output the target network planning scheme after the model converges.

[0094] Furthermore, when executing the problem sensitivity attention mechanism, the encoding module is configured to calculate the unnormalized attention score between adjacent nodes. This score is obtained by activating the transpose of the learnable attention vector and the concatenation vector through a linear rectifier unit. The concatenation vector is formed by concatenating the hidden state vector of the starting node, the hidden state vector of the ending node, and the difference vector of the severity of the running problem between the two nodes. The unnormalized attention score is then subjected to exponential operation and normalized to obtain the final attention coefficient.

[0095] Furthermore, the output module also includes a physics engine verification unit, which is used to parse the action sequence output by the converged model into a preliminary target network scheme, and to perform full-time power flow calculation and safety verification on the preliminary target network scheme using a dynamic simulation environment based on the DistFlow power flow equation. If the verification fails, the verification result is transmitted to the construction module as a negative feedback signal to fine-tune the strategy network parameters until the final planning scheme that satisfies all engineering constraints is output.

[0096] Specific limitations regarding the AI-based distribution network grid target structure planning system can be found in the limitations of the AI-based distribution network grid target structure planning method described above, and will not be repeated here. Each module in the aforementioned AI-based distribution network grid target structure planning system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0097] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0098] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for planning the target grid structure of a power distribution network based on artificial intelligence, characterized in that, include: A multimodal semantic graph of the distribution network is constructed, defining substations, switching stations, distribution transformers and user access points as nodes, and lines, tie switches and sectional switches as edges, and assigning feature vectors containing structural modes, operational modes and geographical modes to nodes and edges; The network operation coupled encoder based on the attention mechanism is used to extract features from the multimodal semantic map of the distribution network. The electrical connection relationship between nodes is learned through the topology attention mechanism. The node weight is dynamically adjusted according to the heavy overload rate, voltage deviation, three-phase imbalance and frequency of outages through the problem sensitivity attention mechanism. The output is a node implicit state vector that integrates local topology and global operation problem information. A grid generator based on dynamic constraint reinforcement learning is constructed, and the grid planning process is modeled as a Markov decision process. An action space including new line construction, switch position adjustment, equipment capacity upgrade and load cutover is defined. A composite reward function including safety reward, grid structure reward, economic reward and robustness reward is designed. The agent interacts with the current state of the network structure in a dynamic simulation environment to perform actions, updates the policy network parameters according to the composite reward function, and outputs the target network structure planning scheme after the model converges.

2. The method for planning the target grid structure of a power distribution network based on artificial intelligence according to claim 1, characterized in that, The feature vector of the structural mode includes node type identifier, edge physical impedance, edge rated capacity and current switch status, wherein the node type identifier includes single-radial mode label, standard interconnection mode label and multi-segment multi-interconnection mode label; the feature vector of the operating mode includes normalized heavy overload index, low voltage index, three-phase imbalance index and annual frequency outage index; the feature vector of the geographic mode includes geographic information system coordinates and corridor resource constraint factor.

3. The method for planning the target grid structure of a power distribution network based on artificial intelligence according to claim 1, characterized in that, The calculation process of the attention coefficient in the problem sensitivity attention mechanism includes: calculating the unnormalized attention score between adjacent nodes, which is obtained by activating the transpose of the learnable attention vector and the concatenation vector through a linear rectifier unit, wherein the concatenation vector is formed by concatenating the hidden state vector of the starting node, the hidden state vector of the ending node, and the difference vector of the severity of the running problem between the two nodes; performing an exponential operation on the unnormalized attention score and normalizing it to obtain the final attention coefficient, so that adjacent nodes with large differences in the severity of the running problem have higher attention weights.

4. The method for planning the target grid structure of a power distribution network based on artificial intelligence according to claim 1, characterized in that, The calculation formula for the composite reward function is as follows: the reward value equals the first safety weight coefficient multiplied by the safety reward item plus the second weight coefficient multiplied by the grid structure reward item minus the third weight coefficient multiplied by the investment cost item minus the fourth weight coefficient multiplied by the network loss power item plus the fifth weight coefficient multiplied by the robustness score item; where the safety reward item is the sum of the safety indication functions of all monitoring points. When the monitoring point has no overload, no low voltage, no three-phase imbalance exceeding the standard, and no frequency outage risk, the safety indication function takes a positive value; otherwise, it takes a negative penalty value. The grid structure reward item is the sum of the structure indication functions of all grid units. When the grid units form a preset standard connection mode, the structure indication function takes a positive value; otherwise, it takes a zero value.

5. The method for planning the target grid structure of a power distribution network based on artificial intelligence according to claim 1, characterized in that, The dynamic simulation environment is built based on the DistFlow power flow equation and supports second-level timing simulation. During the training process, the dynamic simulation environment randomly disturbs the load curve and the output of the distributed power source to simulate the real operating scenario. Random faults are injected in each scenario to evaluate the power supply recovery capability of the grid structure. The evaluation results are used as robustness scoring terms and input into the composite reward function.

6. The method for planning the target grid structure of a power distribution network based on artificial intelligence according to claim 1, characterized in that, The method also includes a scheme generation and adaptive verification step: parsing the current status data of the area to be planned input by the converged model into an action sequence, generating a preliminary target grid scheme that includes new line specifications, switch status adjustment instructions and equipment upgrade list; The physical engine verification layer is used to perform full-time power flow calculation and safety verification of the preliminary target network structure scheme. If the verification fails, the verification result is used as a negative feedback signal to fine-tune the strategy network parameters until the final planning scheme that meets all engineering constraints is output.

7. The method for planning the target grid structure of a power distribution network based on artificial intelligence according to claim 1, characterized in that, The load cutover action in the action space is specifically designed for areas where the three-phase imbalance exceeds a preset threshold. The action includes adjusting the phase sequence connection of nodes or redistributing the load carried by the nodes to different phases to reduce the ratio of the effective value of the negative sequence current to the effective value of the positive sequence current at the node.

8. A distribution network grid target structure planning system based on artificial intelligence, characterized in that, include: The semantic module is used to construct a multimodal semantic graph of the distribution network. It defines substations, switching stations, distribution transformers and user access points as nodes, lines, tie switches and sectionalizing switches as edges, and assigns feature vectors containing structural modes, operational modes and geographical modes to nodes and edges. The encoding module is used to extract features from the multimodal semantic graph of the distribution network using a grid operation coupled encoder based on an attention mechanism. It learns the electrical connection relationship between nodes through a topology attention mechanism and dynamically adjusts the node weights based on the heavy overload rate, voltage deviation, three-phase imbalance and frequency of outages through a problem sensitivity attention mechanism. It outputs a node implicit state vector that integrates local topology and global operation problem information. The module is used to build a grid generator based on dynamic constraint reinforcement learning. It models the grid planning process as a Markov decision process, defines the action space including new lines, adjustment of switch positions, upgrading of equipment capacity and load cutover, and designs a composite reward function including safety reward, grid structure reward, economic reward and robust reward. The output module is used to perform actions by interacting with the current network state in a dynamic simulation environment through an intelligent agent, update the policy network parameters according to the composite reward function, and output the target network planning scheme after the model converges.

9. The distribution network grid target structure planning system based on artificial intelligence according to claim 8, characterized in that, When executing the problem sensitivity attention mechanism, the encoding module is configured to calculate the unnormalized attention score between adjacent nodes. This score is obtained by activating the transpose of the learnable attention vector and the concatenated vector through a linear rectifier unit. The concatenated vector is formed by concatenating the hidden state vector of the starting node, the hidden state vector of the ending node, and the difference vector of the severity of the running problem between the two nodes. The unnormalized attention score is then subjected to exponential operation and normalized to obtain the final attention coefficient.

10. The distribution network grid target structure planning system based on artificial intelligence according to claim 8, characterized in that, The output module also includes a physics engine verification unit, which is used to parse the action sequence output by the converged model into a preliminary target network scheme, and to perform full-time power flow calculation and safety verification on the preliminary target network scheme using a dynamic simulation environment based on the DistFlow power flow equation. If the verification fails, the verification result is transmitted to the construction module as a negative feedback signal to fine-tune the strategy network parameters until the final planning scheme that meets all engineering constraints is output.

Images