Sparse real-time measurement based power distribution network reactive power optimization method based on load reconstruction and projection feedback

Abstract

The application discloses a sparse real-time measurement power distribution network reactive power optimization method based on load reconstruction and projection feedback. First, based on the operation data and real-time measurement configuration of the power distribution network, a voltage safety contribution weight and a voltage control equivalent edge are constructed, and a mask perception multi-relationship graph attention load reconstruction model is trained; second, the unobservable node load is reconstructed by using the observable node real-time load at the regulation time, and the full-network load state is formed; then, the candidate action of the reactive power regulation device is searched based on the reconstructed load, and the candidate action is corrected by the local cone constraint safety projection; finally, the feedback enhancement search is triggered according to the projection residual, the voltage relaxation and the boundary margin, and the reactive power optimization instruction is output. The application can reduce the reactive power optimization mismatch risk caused by the load error under the sparse real-time measurement, and improve the voltage safety of the sparse real-time measurement power distribution network.

Description

Technical Field

[0001] This invention relates to the field of reactive power optimization technology for distribution networks, specifically to a sparse real-time measurement reactive power optimization method for distribution networks based on load reconfiguration and projection feedback. Background Technology

[0002] With the integration of distributed power sources, flexible loads, and various types of reactive power regulation equipment, the voltage distribution and reactive power demand of the distribution network exhibit significant time-varying characteristics. Due to limitations in construction costs, communication conditions, and equipment layout, real-time measurements typically only cover substation outlets, main feeders, and a few key nodes. A large number of end-load nodes lack real-time load information, making it difficult to directly apply traditional reactive power optimization methods that rely on the status of all nodes.

[0003] Existing load completion methods mostly aim to reduce the average reconfiguration error, without fully distinguishing the impact of different unobservable node errors on voltage safety. Small load errors at feeder ends, weak voltage nodes, or nodes with limited compensation capabilities can be amplified into low-voltage risks through power flow coupling; mapping based solely on physical topology or historical load correlations is also insufficient to reflect the similar impact of nodes on critical voltage constraints.

[0004] In reactive power control, heuristic algorithms can handle mixed discrete-continuous actions consisting of on-load tap-changing transformers, parallel capacitor banks, and continuous reactive power compensation equipment. However, candidate actions generated based on reconfigured loads may still approach the voltage safety boundary. Existing safety checks or corrections are usually treated as independent post-processing, without feeding back information such as action correction magnitude, voltage relaxation, active constraints, and boundary margins to the candidate action search stage.

[0005] Therefore, a closed-loop reactive power optimization method is needed that combines unobservable load reconfiguration, voltage safety contribution modeling, hybrid action search, cone-constrained safety projection, and projection feedback enhanced search to form a reliable network load status and output safer control commands under sparse real-time measurement conditions. Summary of the Invention

[0006] To address the problems of reactive power optimization mismatch and difficulty in accurately controlling voltage safety risks caused by the lack of load at unobservable nodes under sparse real-time measurement conditions in existing technologies, this invention provides a reactive power optimization method for sparse real-time measurement distribution networks based on load reconfiguration and projection feedback.

[0007] To achieve the above objectives, the present invention adopts the following technical solution.

[0008] A reactive power optimization method for sparse real-time measurement distribution networks based on load reconfiguration and projection feedback includes the following steps:

[0009] S1: Obtain the network topology, line parameters, historical load data, real-time measurement node configuration, and reactive power regulation equipment operation boundary of the target distribution network to determine the observable and unobservable node sets; based on the AC power flow results during the training period, the fixed power factor load disturbance power flow response, and the effective compensation effect of reactive power equipment, determine the voltage safety risk, load-voltage sensitivity, and constrained compensation capability, and then construct the voltage safety contribution weight and voltage control equivalent edge of the unobservable nodes; S2: Based on line connection relationships, historical load correlation relationships, and voltage control equivalent relationships, physical topology edges, load semantic edges, and voltage control equivalent edges are formed respectively, and a mask-aware multi-relationship graph attention load reconstruction model is constructed. The node mask load features, node observable identifiers, and node structural features are used as inputs to perform attention information aggregation and adaptive fusion in different relational neighborhoods. The model is jointly trained using ordinary load reconstruction error and voltage safety contribution weighted reconstruction error. S3: At the control moment, acquire the real-time load data of observable nodes, mask the real-time load input of unobservable nodes, and input it into the trained load reconstruction model to obtain the reconstructed load of unobservable nodes; concatenate it with the real-time load of observable nodes, and process the root node or balance node separately to form the network load status at the current control moment, and do not use the real-time load of unobservable nodes during the action generation process. S4: Based on the overall network load status and the operating boundary of reactive power regulation equipment, a hybrid discrete-continuous action space is constructed, which includes the tap positions of on-load tap-changing transformers, the switching positions of parallel capacitor banks, and the reactive power output of continuous reactive power compensation equipment. The frost-ice optimization algorithm is used to search for candidate control actions, and the fitness of candidate actions is evaluated from the aspects of network loss, voltage deviation, voltage over-limit risk, voltage safety margin, and equipment action cost based on the power flow calculation results under reconfigured load. S5: Construct a cone-constrained safety projection model based on the branch power flow model within the local action neighborhood of the candidate control action. Combine the node voltage boundary, branch power flow constraint, reactive power regulation equipment boundary, cone constraint relationship and voltage relaxation constraint to perform safety correction on the candidate action, obtain the projection control action that meets the voltage safety requirements and has a limited correction range, and extract projection feedback information such as projection residual, voltage relaxation amount, active voltage constraint, constraint pressure, equipment correction direction and voltage boundary margin. S6: Determine whether to trigger feedback enhancement search based on the projection feedback information; when the projection control action is near the voltage safety boundary, inject the projection control action and its local disturbance into the frost-ice optimization algorithm, and adjust the fitness evaluation, initial population or search bias according to the projection feedback information, and execute candidate action search and cone-constrained safety projection again; when the triggering condition is not met, use the current projection control action, and finally output the obtained projection control action as a reactive power optimization command.

[0010] Furthermore, the voltage security contribution weight is jointly determined by the low voltage risk during the training period, aggregated load-voltage sensitivity, and compensation weakness. Based on the AC power flow calculation results during the training period, the voltage nodes are obtained. At any moment The voltage amplitude, and based on the allowable lower limit of the node voltage. Calculate voltage nodes Low voltage risk : (1) in, For voltage node index, For time indexes within the training period, voltage node At any moment voltage amplitude, This is the lower limit of the allowable node voltage. express ; Apply a combined active and reactive power disturbance to the unobservable load node while maintaining a constant power factor, and calculate the load-voltage sensitivity based on the AC power flow voltage changes before and after the disturbance. For the training representative scene set The sensitivity within the range is aggregated to obtain the aggregated sensitivity: (2) in, For training representative scenarios Unobservable load nodes For voltage nodes Load-voltage sensitivity, This is a set of representative training scenarios used for sensitivity calculation. To train a number of representative scenes, Unobservable load nodes For voltage nodes Polymer load-voltage sensitivity; Based on reactive power regulation equipment, voltage nodes are adjusted without causing node voltage exceedance. Constrained effective compensation amount Calculate the normalized compensation capability and compensatory weakness : (3) (4) in, voltage node Normalization-constrained effective compensation capability voltage node The constrained effective compensation amount, The 5th percentile of the constrained effective compensation for all voltage nodes. The 95th percentile of the constrained effective compensation for all voltage nodes. Indicates will Cut off to interval , To compensate for the lower limit of weakness, voltage node The compensatory weakness; Calculate the unobservable load nodes based on the low voltage risk at voltage nodes, load-voltage sensitivity, and compensation weakness. At any moment Voltage safety contribution weight: (5) in, Unobservable load nodes At any moment Voltage safety contribution weight, This refers to the set of voltage nodes involved in voltage risk, load-voltage sensitivity, and compensation weakness calculations. Set of unobservable load nodes The summation index in the middle, The set of unobservable load nodes after removing root nodes or balancing nodes; when the denominator is zero, the normalized result of the static voltage safety contribution score during the training period is used as the prior voltage safety contribution weight.

[0011] Furthermore, the voltage control equivalent edge is constructed according to the influence pattern of load nodes on critical voltage constraints. Based on the low voltage risk during the training period, the voltage node is calculated. Importance of training risks : (6) in, voltage node The importance of training risks This is the summation index for traversing all voltage nodes in the normalized denominator. This is the set of training periods used to calculate the importance of training risks. This represents the mean operation. Indicates voltage node The 95th percentile of low voltage risk during the training period This is the percentile risk weighting coefficient; Based on aggregation sensitivity, training risk importance, and compensation weakness, load nodes are constructed. Voltage control influence vector , its first The components are: (7) in, For load nodes The voltage control influence vector, For load nodes For voltage nodes The control effect component; compute nodes With nodes Voltage control equivalence similarity between : (8) in, For nodes With nodes The voltage control equivalence similarity between them For nodes The node index used for similarity calculation. For nodes The voltage control influence vector, This represents the transpose of a vector. Represents the L2 norm; For each source node Choose the one with the highest voltage control equivalent similarity. The target nodes are connected as voltage control equivalent edges, forming a voltage control equivalent edge set. : (9) in, For voltage control equivalent edge set, The number of voltage-controlled equivalent edges selected for each source node. This indicates selecting the one with the highest similarity. There are several target nodes; the voltage control equivalent edge is used to characterize the similarity of the influence patterns of different nodes on the critical voltage constraint.

[0012] Furthermore, the mask-aware multi-relationship graph attention load reconstruction model performs attention aggregation in the neighborhoods corresponding to physical topological relationships, load semantic relationships, and voltage control equivalent relationships, and performs adaptive fusion through node-level relationship gating; the model only calculates training error on unobservable load nodes, specifically including: Suppose a training batch contains One sample, This refers to the sample index in the training batch. For the sample At the corresponding time, For the sample Middle node Normalized reconfiguration load, For the sample Middle node Normalized real load, This is for normalized reconstruction error; For the sample The voltage security contribution weights of unobservable load nodes are normalized to obtain the normalized voltage security contribution weights: (10) in, For the sample Middle node Normalized voltage safety contribution weight, For the sample Corresponding time Unobservable load nodes The original voltage safety contribution weight; Voltage safety contribution weighted error loss Represented as: (11) Ordinary mean square reconstruction error loss Defined as: (12) Total training loss function of the model Represented as: (13) in, The weighting coefficients that contribute to the weighted error loss for voltage safety are used to adjust the proportion of this weighted error in the total training loss; Furthermore, the candidate control actions generated by the frost-ice optimization algorithm are used as reference actions for cone-constrained safety projection. The discrete control variables in the candidate control actions include the tap position of the on-load tap-changing transformer and the switching position of the parallel capacitor bank, and the continuous control variables include the reactive power output of the continuous reactive power compensation equipment. Discrete control actions are enumerated in the local discrete neighborhood of the reference action, and the continuous reactive power output, branch power flow variables, node voltage square variables, branch current square variables, and voltage relaxation variables are optimized to obtain projection control actions with limited correction amplitude relative to the reference action. The target for the safety projection near the candidate action is set as a weighted sum of the action correction, voltage relaxation, predicted network loss, and current regularization term: (14) in, Let be the objective function for projections near the candidate action. and These represent the continuous control variables and discrete control variables in the projection model, respectively. and These are the continuous and discrete candidate actions output by the frost optimization algorithm, respectively. and They are respectively Low-voltage relaxation variables and high-voltage relaxation variables; branch road The square of the current, The resistance of this branch is... For branch set; , , , and These are the weighting coefficients for the continuous control action correction term, discrete control action correction term, voltage relaxation term, predicted network loss term, and branch current regularization term, respectively. Denotes the norm 1; The security projection satisfies the radial distribution network branch power flow balance, voltage drop, second-order cone, voltage boundary, equipment boundary and non-negative relaxation constraints; parallel capacitor banks and continuous reactive power compensation equipment are entered into the model through node reactive power injection, and on-load tap-changing transformers adjust the root node equivalent voltage through turns ratio adjustment. Each local discrete combination is sorted lexicographically according to voltage relaxation, action correction cost, predicted network loss and action distance to obtain the 0th round of projected action, and AC power flow safety is checked under reconfigured load.

[0013] Furthermore, the feedback-enhanced search sets four trigger conditions based on the safety projection results of the 0th cone constraint and the safety verification results of the reconfigured load AC power flow: voltage relaxation trigger condition, post-projection voltage margin trigger condition, action projection residual joint trigger condition, and low-voltage hot spot trigger condition. These four trigger conditions are as follows: (15) (16) (17) (18) Based on the above four types of triggering conditions, construct safety boundary trigger variables: (19) in, This is a voltage relaxation trigger condition. The maximum value of the low-voltage relaxation and high-voltage relaxation at each node obtained from the safety projection of the 0th cone constraint is indicated by the superscript. Indicates round 0. This is the voltage relaxation numerical tolerance, used to eliminate small numerical errors generated by the solver; This is the trigger condition for the projected voltage margin. The superscript represents the minimum voltage margin relative to the lower limit of the projected voltage after AC power flow safety verification under the reconfigured load condition for the 0th round of projection control action. This represents the verification quantity obtained based on the projection control action. This is the voltage margin trigger threshold after projection; For the joint triggering condition of motion projection residual, The L2 norm is calculated as the difference between the candidate control action of the 0th round frost-ice optimization algorithm and the safe projection action of the 0th round cone constraint, after scaling according to the allowable variation range of each control variable. The threshold for motion projection residuals. This is the voltage margin threshold corresponding to the triggering of the action projection residual; This is a low-voltage hotspot triggering condition. For the set of active low-voltage constrained nodes, This is the set of hotspot nodes determined based on the low-voltage risk distribution during the training period. To enhance the search trigger variables for feedback, This is an indicator function; it is triggered when any of the following conditions are met: voltage relaxation trigger condition, projected voltage margin trigger condition, combined action projection residual trigger condition, or low-voltage hot spot trigger condition. ,otherwise ; Among them, the symbol " The symbol "" represents the logical AND, meaning that both conditions connected must be true simultaneously; " represents the logical "OR", meaning that at least one of the conditions it connects is true; the symbol " " represents the intersection of sets, symbol " "" indicates an empty set. Therefore, the non-empty intersection of the set of active low-voltage constrained nodes and the set of hot nodes indicates that they contain at least one common node.

[0014] Furthermore, when the safety boundary triggering condition is met, a second round of feedback-enhanced fitness function is constructed based on the projection feedback information: (20) in, This refers to the candidate reactive power control actions searched in the second round of the frost-ice optimization algorithm. To enhance the fitness function through feedback, Based on the fitness function, , , and These are, respectively, motion projection residual penalty, node safety pressure penalty, voltage relaxation reliability penalty, and equipment bottleneck penalty. , , and These are the corresponding weighting coefficients; Among them, the node safety pressure penalty term can be composed of the margin deficit of low-voltage hot spot nodes and the constraint dual pressure, and the equipment bottleneck penalty term can further include the deviation of equipment correction direction and insufficient near-boundary safety margin; the relaxation reliability information is not used as a constant term that is the same for all candidate actions, but is used to amplify the near-boundary voltage deficit penalty of the second round of candidate actions, so that the term can change the order of candidate actions.

[0015] Furthermore, the projection action, local perturbation action, and projection residual direction offset action from round 0 are injected into the initial population for the second round of candidate action search: (twenty one) in, To provide feedback to enhance the initial population for the frost-ice optimization algorithm, Given a random initial set of individuals, For the 0th cone constraint safety projection action, This refers to the set of local perturbation individuals constructed around the projection action. For local disturbance scale. For the 0th cone constraint safety projection action Candidate moves for Frost Ice Optimization Round 0 The difference; its positive direction indicates the safe correction direction from the candidate action of the 0th round of frost optimization to the safe projection action of the 0th round of cone constraint. The set of search bias individuals formed along the direction of the action projection residual; symbol " "" represents the union of sets, used to merge the random initial set of individuals, the safe projection action of the cone constraint in the 0th round, the set of individuals with local perturbation, and the set of individuals with search bias to form the initial population of the feedback-enhanced frost-ice optimization algorithm; Both local disturbance individuals and search bias individuals are mapped back to the feasible region of the equipment: continuous reactive power output is subject to upper and lower limit truncation, and on-load tap changer taps and parallel capacitor bank taps are subject to integer mapping to ensure that the initial population in the second round satisfies the equipment operating boundary. The final reactive power optimization action is determined based on the feedback-enhanced search trigger results and the AC power flow safety verification results of the reconfigured load: (twenty two) in, To freeze the output action, To provide feedback to enhance the safety projection action of the first round cone constraint obtained after the search, The first round of actions is the accept variable; when the first round of actions passes the AC power flow safety check under the reconstructed load state and its safety indicators do not deteriorate relative to the 0th round of projection actions, the accept variable is set to 1, otherwise it is set to 0; when the feedback enhancement search is not triggered or the first round of actions is not accepted, the 0th round of projection actions are output.

[0016] Compared with the prior art, the present invention has the following beneficial effects: First, this invention incorporates the low voltage risk during the training period, the voltage sensitivity of load nodes to each voltage node that cannot be observed in real time, and the constrained effective compensation capability into the load reconfiguration process. This makes the training weight of the load reconfiguration error correspond to the subsequent voltage safety impact, and can prioritize reducing the load reconfiguration error in high-risk, highly sensitive areas with limited compensation capabilities.

[0017] Second, this invention constructs an equivalent voltage control relationship in addition to physical topological relationships and historical load semantic relationships, and uses a mask-aware multi-relationship graph attention model to perform node-level adaptive fusion of the three types of relationships. This enables the model to simultaneously utilize the similarity of line connections, load change correlations, and node influence patterns on key voltage constraints, thereby improving the accuracy of load reconfiguration of unobservable nodes under sparse real-time measurement conditions.

[0018] Third, the present invention constructs a cone-constrained safety projection near the candidate control action, corrects the mixed discrete-continuous candidate action to a local safety region that satisfies voltage boundary, branch power flow and equipment operation constraints, and uses AC power flow for safety verification, thereby reducing the risk of boundary exceedance caused by relying solely on heuristic search.

[0019] Fourth, the present invention feeds back the motion projection residual, voltage relaxation, active voltage constraint, constraint pressure, equipment correction direction and voltage boundary margin generated by the safety projection to the subsequent search, so that the safety projection is transformed from an independent post-processing link into a closed-loop feedback link that can change the evaluation of candidate actions, the search starting point and the search direction.

[0020] Fifth, the present invention uses a condition-triggered method to perform feedback-enhanced search. In normal operation scenarios, the 0th round of projection actions can be used, and the second round of search is only started when the voltage safety boundary is approaching or when the constraint pressure is large. This can improve voltage safety while reducing unnecessary repetitive calculations. Furthermore, the online action generation process does not use the real-time load of unobservable nodes, which conforms to the actual data conditions of sparse measurement distribution networks. Attached Figure Description

[0021] Figure 1 This is an overall flowchart of the reactive power optimization method for sparse real-time measurement distribution networks based on load reconfiguration and projection feedback in an embodiment of the present invention. Figure 2This is a schematic diagram of the network topology, node observability, and location of reactive power control devices in an IEEE 69-node distribution network according to an embodiment of the present invention. Figure 3 This is a weighted distribution diagram of the voltage security contribution of unobservable nodes in an embodiment of the present invention. Figure 4 This is a load reconfiguration curve of a high-risk unobservable node in an embodiment of the present invention; Figure 5 This is a comparison chart of the load reconstruction ablation experiment errors in the embodiments of the present invention; Figure 6 This is a diagram showing the actual minimum voltage distribution of the load under different control strategies in this embodiment of the invention. Figure 7 This is a diagram showing the feedback branch activation ratio for three types of experimental scenarios in this embodiment of the invention; Figure 8 This is a comparative diagram of residual low voltage nodes at the end of the feeder in the feedback branch ablation experiment of this invention. Figure 9 This is a node voltage correction diagram for the boundary case of the test set at hour 564 in this embodiment of the invention; Figure 10 This is a comparison chart of the safety indicators of the complete closed-loop method in the embodiments of the present invention; Figure 11 This is an empirical cumulative distribution diagram of the lowest voltage under the actual load in a complete closed-loop configuration according to an embodiment of the present invention. Figure 12 This is a comparison diagram of residual low voltage nodes at the end of the system-level feeder in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0023] like Figure 1 The diagram shows a flowchart of a sparse real-time measurement distribution network reactive power optimization method based on load reconfiguration and projection feedback in an embodiment of the present invention, including the following steps: S1: Obtain the network topology, line parameters, historical load data, real-time measurement node configuration, and reactive power regulation equipment operation boundary of the target distribution network to determine the observable and unobservable node sets; based on the AC power flow results during the training period, the fixed power factor load disturbance power flow response, and the effective compensation effect of reactive power equipment, determine the voltage safety risk, load-voltage sensitivity, and constrained compensation capability, and then construct the voltage safety contribution weight and voltage control equivalent edge of the unobservable nodes; S2: Based on line connection relationships, historical load correlation relationships, and voltage control equivalent relationships, physical topology edges, load semantic edges, and voltage control equivalent edges are formed respectively, and a mask-aware multi-relationship graph attention load reconstruction model is constructed. The node mask load features, node observable identifiers, and node structural features are used as inputs to perform attention information aggregation and adaptive fusion in different relational neighborhoods. The model is jointly trained using ordinary load reconstruction error and voltage safety contribution weighted reconstruction error. S3: At the control moment, acquire the real-time load data of observable nodes, mask the real-time load input of unobservable nodes, and input it into the trained load reconstruction model to obtain the reconstructed load of unobservable nodes; concatenate it with the real-time load of observable nodes, and process the root node or balance node separately to form the network load status at the current control moment, and do not use the real-time load of unobservable nodes during the action generation process. S4: Based on the overall network load status and the operating boundary of reactive power regulation equipment, a hybrid discrete-continuous action space is constructed, which includes the tap positions of on-load tap-changing transformers, the switching positions of parallel capacitor banks, and the reactive power output of continuous reactive power compensation equipment. The frost-ice optimization algorithm is used to search for candidate control actions, and the fitness of candidate actions is evaluated from the aspects of network loss, voltage deviation, voltage over-limit risk, voltage safety margin, and equipment action cost based on the power flow calculation results under reconfigured load. S5: Construct a cone-constrained safety projection model based on the branch power flow model within the local action neighborhood of the candidate control action. Combine the node voltage boundary, branch power flow constraint, reactive power regulation equipment boundary, cone constraint relationship and voltage relaxation constraint to perform safety correction on the candidate action, obtain the projection control action that meets the voltage safety requirements and has a limited correction range, and extract projection feedback information such as projection residual, voltage relaxation amount, active voltage constraint, constraint pressure, equipment correction direction and voltage boundary margin. S6: Determine whether to trigger feedback enhancement search based on the projection feedback information; when the projection control action is near the voltage safety boundary, inject the projection control action and its local disturbance into the frost-ice optimization algorithm, and adjust the fitness evaluation, initial population or search bias according to the projection feedback information, and execute candidate action search and cone-constrained safety projection again; when the triggering condition is not met, use the current projection control action, and finally output the obtained projection control action as a reactive power optimization command.

[0024] Furthermore, the voltage security contribution weight is jointly determined by the low voltage risk during the training period, aggregated load-voltage sensitivity, and compensation weakness. Based on the AC power flow calculation results during the training period, the voltage nodes are obtained. At any moment The voltage amplitude, and based on the allowable lower limit of the node voltage. Calculate voltage nodes Low voltage risk : (twenty three) in, For voltage node indexing, For time indexes within the training period, voltage node At any moment voltage amplitude, This is the lower limit of the allowable node voltage. express ; Apply a combined active and reactive power disturbance to the unobservable load node while maintaining a constant power factor, and calculate the load-voltage sensitivity based on the AC power flow voltage changes before and after the disturbance. For the training representative scene set The sensitivity within the range is aggregated to obtain the aggregated sensitivity: (twenty four) in, For training representative scenarios Unobservable load nodes For voltage nodes Load-voltage sensitivity, This is a set of representative training scenarios used for sensitivity calculation. To train a number of representative scenes, Unobservable load nodes For voltage nodes Polymer load-voltage sensitivity; Based on reactive power regulation equipment, voltage nodes are adjusted without causing node voltage exceedance. Constrained effective compensation amount Calculate the normalized compensation capability and compensatory weakness : (25) (26) in, voltage node Normalization-constrained effective compensation capability voltage node The constrained effective compensation amount, The 5th percentile of the constrained effective compensation for all voltage nodes. The 95th percentile of the constrained effective compensation for all voltage nodes. Indicates will Cut off to interval , To compensate for the lower limit of weakness, voltage node The compensatory weakness; Calculate the unobservable load nodes based on the low voltage risk at voltage nodes, load-voltage sensitivity, and compensation weakness. At any moment Voltage safety contribution weight: (27) in, Unobservable load nodes At any moment Voltage safety contribution weight, This refers to the set of voltage nodes involved in voltage risk, load-voltage sensitivity, and compensation weakness calculations. Set of unobservable load nodes The summation index in the middle, The set of unobservable load nodes after removing root nodes or balancing nodes; when the denominator is zero, the normalized result of the static voltage safety contribution score during the training period is used as the prior voltage safety contribution weight.

[0025] Furthermore, the voltage control equivalent edge is constructed according to the influence pattern of load nodes on critical voltage constraints. Based on the low voltage risk during the training period, the voltage node is calculated. Importance of training risks : (28) in, voltage node The importance of training risks This is the summation index for traversing all voltage nodes in the normalized denominator. This is the set of training periods used to calculate the importance of training risks. This represents the mean operation. Indicates voltage node The 95th percentile of low voltage risk during the training period This is the percentile risk weighting coefficient; Based on aggregation sensitivity, training risk importance, and compensation weakness, load nodes are constructed. Voltage control influence vector , its first The components are: (29) in, For load nodes The voltage control influence vector, For load nodes For voltage nodes The control effect component; compute nodes With nodes Voltage control equivalence similarity between : (30) in, For nodes With nodes The voltage control equivalence similarity between them For nodes The node index used for similarity calculation. For nodes The voltage control influence vector, This represents the transpose of a vector. Represents the L2 norm; For each source node Choose the one with the highest voltage control equivalent similarity. The target nodes are connected as voltage control equivalent edges, forming a voltage control equivalent edge set. : (31) in, For voltage control equivalent edge set, The number of voltage-controlled equivalent edges selected for each source node. This indicates selecting the one with the highest similarity. There are several target nodes; the voltage control equivalent edge is used to characterize the similarity of the influence patterns of different nodes on the critical voltage constraint.

[0026] Furthermore, the mask-aware multi-relationship graph attention load reconstruction model performs attention aggregation in the neighborhoods corresponding to physical topological relationships, load semantic relationships, and voltage control equivalent relationships, and performs adaptive fusion through node-level relationship gating; the model only calculates training error on unobservable load nodes, specifically including: Suppose a training batch contains One sample, This refers to the sample index in the training batch. For the sample At the corresponding time, For the sample Middle node Normalized reconfiguration load, For the sample Middle node Normalized real load, This is for normalized reconstruction error; For the sample The voltage security contribution weights of unobservable load nodes are normalized to obtain the normalized voltage security contribution weights: (32) in, For the sample Middle node Normalized voltage safety contribution weight, For the sample Corresponding time Unobservable load nodes The original voltage safety contribution weight; Voltage safety contribution weighted error loss Represented as: (33) Ordinary mean square reconstruction error loss Defined as: (34) Total training loss function of the model Represented as: (35) in, The weighting coefficients that contribute to the weighted error loss for voltage safety are used to adjust the proportion of this weighted error in the total training loss; Furthermore, the candidate control actions generated by the frost-ice optimization algorithm are used as reference actions for cone-constrained safety projection. The discrete control variables in the candidate control actions include the tap position of the on-load tap-changing transformer and the switching position of the parallel capacitor bank, and the continuous control variables include the reactive power output of the continuous reactive power compensation equipment. Discrete control actions are enumerated in the local discrete neighborhood of the reference action, and the continuous reactive power output, branch power flow variables, node voltage square variables, branch current square variables, and voltage relaxation variables are optimized to obtain projection control actions with limited correction amplitude relative to the reference action. The target for the safety projection near the candidate action is set as a weighted sum of the action correction, voltage relaxation, predicted network loss, and current regularization term: (36) in, Let be the projection objective function near the candidate action. and These represent the continuous control variables and discrete control variables in the projection model, respectively. and These are the continuous and discrete candidate actions output by the frost optimization algorithm, respectively. and They are respectively Low-voltage relaxation variables and high-voltage relaxation variables; branch road The square of the current, The resistance of this branch is... For branch set; , , , and These are the weighting coefficients for the continuous control action correction term, discrete control action correction term, voltage relaxation term, predicted network loss term, and branch current regularization term, respectively. Denotes the norm 1; The security projection satisfies the radial distribution network branch power flow balance, voltage drop, second-order cone, voltage boundary, equipment boundary and non-negative relaxation constraints; parallel capacitor banks and continuous reactive power compensation equipment are entered into the model through node reactive power injection, and on-load tap-changing transformers adjust the root node equivalent voltage through turns ratio adjustment. Each local discrete combination is sorted lexicographically according to voltage relaxation, action correction cost, predicted network loss and action distance to obtain the 0th round of projected action, and AC power flow safety is checked under reconfigured load.

[0027] Furthermore, the feedback-enhanced search sets four trigger conditions based on the safety projection results of the 0th cone constraint and the safety verification results of the reconfigured load AC power flow: voltage relaxation trigger condition, post-projection voltage margin trigger condition, action projection residual joint trigger condition, and low-voltage hot spot trigger condition. These four trigger conditions are as follows: (37) (38) (39) (40) Based on the above four types of triggering conditions, construct safety boundary trigger variables: (41) in, This is a voltage relaxation trigger condition. The maximum value of the low-voltage relaxation and high-voltage relaxation at each node obtained from the safety projection of the 0th cone constraint is indicated by the superscript. Indicates round 0. This is the voltage relaxation numerical tolerance, used to eliminate small numerical errors generated by the solver; This is the trigger condition for the projected voltage margin. The superscript represents the minimum voltage margin relative to the lower limit of the projected voltage after AC power flow safety verification under the reconfigured load condition for the 0th round of projection control action. This represents the verification quantity obtained based on the projection control action. This is the voltage margin trigger threshold after projection; For the joint triggering condition of motion projection residual, The L2 norm is calculated as the difference between the candidate control action of the 0th round frost-ice optimization algorithm and the safe projection action of the 0th round cone constraint, after scaling according to the allowable variation range of each control variable. The threshold for motion projection residuals. This is the voltage margin threshold corresponding to the triggering of the action projection residual; This is a low-voltage hotspot triggering condition. For the set of active low-voltage constrained nodes, This is the set of hotspot nodes determined based on the low-voltage risk distribution during the training period. To enhance the search trigger variables for feedback, This is an indicator function; it is triggered when any of the following conditions are met: voltage relaxation trigger condition, projected voltage margin trigger condition, combined action projection residual trigger condition, or low-voltage hot spot trigger condition. ,otherwise ; Among them, the symbol " The symbol "" represents the logical AND, meaning that both conditions connected must be true simultaneously; " represents the logical "OR", meaning that at least one of the conditions it connects is true; the symbol " " represents the intersection of sets, symbol " "" indicates an empty set. Therefore, the non-empty intersection of the set of active low-voltage constrained nodes and the set of hot nodes indicates that they contain at least one common node.

[0028] Furthermore, when the safety boundary triggering condition is met, a second round of feedback-enhanced fitness function is constructed based on the projection feedback information: (42) in, This refers to the candidate reactive power control actions searched in the second round of the frost-ice optimization algorithm. To enhance the fitness function through feedback, Based on the fitness function, , , and These are, respectively, motion projection residual penalty, node safety pressure penalty, voltage relaxation reliability penalty, and equipment bottleneck penalty. , , and These are the corresponding weighting coefficients; Among them, the node safety pressure penalty term can be composed of the margin deficit of low-voltage hot spot nodes and the constraint dual pressure, and the equipment bottleneck penalty term can further include the deviation of equipment correction direction and insufficient near-boundary safety margin; the relaxation reliability information is not used as a constant term that is the same for all candidate actions, but is used to amplify the near-boundary voltage deficit penalty of the second round of candidate actions, so that the term can change the order of candidate actions.

[0029] Furthermore, the projection action, local perturbation action, and projection residual direction offset action from round 0 are injected into the initial population for the second round of candidate action search: (43) in, To provide feedback to enhance the initial population for the frost-ice optimization algorithm, Given a random initial set of individuals, For the 0th cone constraint safety projection action, This refers to the set of local perturbation individuals constructed around the projection action. For local disturbance scale. For the 0th cone constraint safety projection action Candidate moves for Frost Ice Optimization Round 0 The difference; its positive direction indicates the safe correction direction from the candidate action of the 0th round of frost optimization to the safe projection action of the 0th round of cone constraint. The set of search bias individuals formed along the direction of the action projection residual; symbol " "" represents the union of sets, used to merge the random initial set of individuals, the safe projection action of the cone constraint in the 0th round, the set of individuals with local perturbation, and the set of individuals with search bias to form the initial population of the feedback-enhanced frost-ice optimization algorithm; Local disturbance individuals and search bias individuals are all mapped back to the feasible region of the equipment: continuous reactive power output is subject to upper and lower limit truncation, and on-load tap changer taps and parallel capacitor bank taps are subject to integer mapping to ensure that the initial population in the second round satisfies the equipment operating boundary. The final reactive power optimization action is determined based on the feedback-enhanced search trigger results and the AC power flow safety verification results of the reconfigured load: (44) in, To freeze the output action, To provide feedback to enhance the safety projection action of the first round cone constraint obtained after the search, The first round of actions is the accept variable; when the first round of actions passes the AC power flow safety check under the reconstructed load state and its safety indicators do not deteriorate relative to the 0th round of projection actions, the accept variable is set to 1, otherwise it is set to 0; when the feedback enhancement search is not triggered or the first round of actions is not accepted, the 0th round of projection actions are output.

[0030] The following uses an IEEE 69-node radial distribution network as an example to illustrate the implementation effect of this invention. All control actions in the experiment were based on reconfigured load generation. To evaluate the impact of reconfiguration errors on actual operational safety, the offline-saved real load was substituted into the AC power flow for posterior evaluation only after the control actions were frozen. This posterior real load did not participate in online model input, action search, projection solution, or feedback triggering.

[0031] like Figure 2 As shown, the solid black circles represent observable nodes, totaling 26: nodes 2, 3, 4, 6, 9, 11, 14, 17, 20, 23, 26, 28, 31, 34, 38, 41, 44, 47, 50, 53, 56, 58, 59, 62, 65, and 68. The hollow white circles represent nodes that are not observable in real time. Except for node 1, the remaining 42 non-observable load nodes all undergo input masking during the online phase and are estimated by the load reconfiguration model. Node 1 is the balancing node. An on-load tap-changing transformer (OLTC) is installed between Node 1 and Node 2, with a tap range of -4 to 4, a change in transformer ratio of 1.25% per tap, and an initial tap of 0. Nodes 61 and 64 are equipped with parallel capacitor banks CB61 and CB64, respectively, with selectable reactive power injection values ​​of 0, 0.15, 0.30, 0.45, and 0.60 MVAr for each bank. Node 65 is equipped with a continuous reactive power compensation device (SVG), with a reactive power output range of -0.30 to 0.80 MVAr, where a positive value indicates the provision of capacitive reactive power to the network.

[0032] This embodiment uses hourly historical load sequences, dividing the data into training, validation, and test data in chronological order. The training data is used to determine load normalization parameters, load semantic edges, VSC weights, VCE edges, and model parameters; the test data is not involved in the offline parameter determination. The test set includes 1314 control times. The frost-free optimization population size can be set to 16, and the maximum number of iterations can be set to 30. In the feedback branch initiation ratio experiment, the results for random seeds 2026, 2027, and 2028 are averaged. The normal range for node voltage is set to 0.95–1.05 pu.

[0033] like Figure 3As shown, the VSC weights calculated offline from the training set are relatively high at nodes 64, 63, 61, 60, and 57, with corresponding weights of approximately 0.0812, 0.0782, 0.0766, 0.0718, and 0.0528, respectively. These nodes are all located in the low-voltage risk region at the end of the feeder, indicating that low-voltage risk, load-voltage sensitivity, and constrained compensation capability are spatially aligned. Nodes 27, 25, and 24, which are not at the end of the feeder, still retain some weight, suggesting that this weight is not solely determined by physical distance.

[0034] like Figure 4 As shown, within a continuous 48-hour window from hour 541 to hour 588 of the test set, the reconstruction loads of nodes 60 and 64 follow the main trends of the offline ground truth labels. The root mean square error of the window for node 60 is 59.93 kW, with a correlation coefficient of 0.966; the root mean square error of the window for node 64 is 25.48 kW, with a correlation coefficient of 0.988. The ground truth labels in the figure are only used for offline evaluation, and hour 564 of the test set serves as a time marker for subsequent boundary cases.

[0035] like Figure 5 As shown, the overall root mean square error (RMSE) of the physical topology baseline, the VSC-weighted RMSE, and the RMSE of high-risk nodes are 41.2 kW, 38.8 kW, and 53.3 kW, respectively. After adding load semantic edges, these decrease to 22.5 kW, 23.8 kW, and 33.9 kW, respectively. With only VCE edges added, the RMSEs are 26.8 kW, 31.9 kW, and 47.3 kW, respectively. With only VSC-weighted losses added, the RMSEs are 38.7 kW, 35.7 kW, and 48.3 kW, respectively. The complete VSC-VCE model, using physical topology edges, load semantic edges, VCE edges, and VSC-weighted losses simultaneously, further reduces these three errors to 20.3 kW, 20.8 kW, and 28.7 kW, all the lowest values ​​among the compared configurations.

[0036] like Figure 6 As shown, compared to the uncontrolled state, the frost-ice optimization candidate actions shift the overall a posteriori minimum voltage distribution of the actual load to the right. After performing cone-constrained safety projection near the candidate actions and combining it with AC power flow safety verification, the low-voltage tail further shrinks. As the projection safety margin increases, the minimum voltage distribution further shifts to a higher voltage range, indicating that improving the projection safety margin can enhance voltage safety in boundary scenarios.

[0037] like Figure 7As shown, in 120 time points of the normal mainline, 50 mixed difficult scenarios, and 24 pressure boundary scenarios, the average activation rates of the triggered feedback branch were 9.17%, 21.57%, and 77.78%, respectively. Pressure boundary scenarios included those where the load was amplified to 1.10 times and the maximum reactive power output of the SVG was reduced to 0.50 MVAr, as well as those where the low-voltage safety requirement was tightened to 0.96 pu. The increased activation rate of the feedback branch with increasing load pressure, equipment derating, or tightened voltage boundaries indicates that the conditional triggering mechanism can focus the second round of search on scenarios with higher safety pressure.

[0038] like Figure 8 As shown, under the condition that the VSC-VCE load reconfiguration, the first round of frost optimization, and the zeroth round of cone constraint projection are all the same, the number of residual low-voltage nodes at nodes 60-65 when the feedback branch is closed is 1, 2, 2, 2, 3, and 1, respectively, for a total of 11 times; after enabling the triggered feedback, the number is 1, 1, 1, 1, 2, and 1, for a total of 7 times. The feedback branch reduces the number of residual over-limits by 1 at nodes 61, 62, 63, and 64, indicating that the additional benefits of the second round of feedback are concentrated at the feeder end nodes near the voltage boundary.

[0039] like Figure 9 As shown, in the 564th hour of the test set, the lowest voltage under uncontrolled conditions was 0.913245 pu, with 22 nodes below 0.95 pu. When only a single round of frost optimization and cone constraint projection was performed, the final action was OLTC gear 2, CB61 gear 0, CB64 gear 0, and SVG output 0.449208 MVAr, increasing the lowest voltage to 0.948951 pu, but nodes 61-64 still had 4 low-voltage nodes. After triggering the second round of feedback, the final action was OLTC gear 3, CB61 gear 1, CB64 gear 1, and SVG output 0.262957 MVAr, increasing the lowest voltage to 0.964110 pu, and reducing the number of low-voltage nodes to 0. The second round simultaneously increased the discrete device gears and decreased the continuous SVG output, indicating that the feedback search corrected the voltage through the coordinated operation of discrete and continuous devices, rather than simply pushing all devices to their limits.

[0040] like Figure 10As shown, under the same 1314 test times and with the same equipment, power flow, and optimization parameters, the ordinary closed-loop baseline uses load reconstruction without VSC weighting and VCE edges, followed by frost optimization and cone constraint projection; the complete method uses VSC-VCE load reconstruction and includes conditional triggering feedback. The voltage feasibility rate of the ordinary closed-loop baseline is 98.935%, with 14 residual low-voltage scenarios and 47 residual low-voltage node occurrences, and the 5th percentile of the lowest voltage is 0.953452 pu; the corresponding metrics for the complete method are 99.848%, 2, 7, and 0.955801 pu. The AC power flow posterior convergence of both methods is shown, and neither method exhibits high-voltage overshoot.

[0041] like Figure 11 As shown, the empirical cumulative probability of the lowest voltage near 0.95 pu is lower than that of the normal closed-loop baseline, indicating a reduction in test times falling into the low-voltage range. Its lower quantile voltage tail shifts to the right, consistent with the increase in the 5th percentile of the lowest voltage from 0.953452 pu to 0.955801 pu. The two empirical cumulative distribution curves intersect in parts of the medium-to-high voltage range; therefore, this result can be used to illustrate an improvement in the safety-critical low-voltage tail.

[0042] like Figure 12 As shown, the residual low-voltage node counts at nodes 60-65 in the ordinary closed-loop baseline are 3, 6, 7, 8, 14, and 9, respectively, totaling 47 times; the complete method reduces these counts to 1, 1, 1, 1, 2, and 1, totaling 7 times. Node 64 decreases from 14 times to 2 times. The residual risks and improvements are concentrated in the aforementioned feeder end risk zone, and... Figure 3 The spatial locations of the high-weight nodes in the VSC shown correspond to each other.

[0043] The simulation results show that VSC weights, VCE edges, and multi-relationship load reconfiguration help reduce load reconfiguration errors at safety-critical unobservable nodes; candidate action safety projection can reduce the low-voltage tail; and conditional triggering feedback can further reduce residual low voltage in boundary scenarios. A cone constraint model is used for safety projection near candidate actions, and the safety of the final action is accepted based on the AC power flow verification of the reconfigured load.

[0044] It should be noted that the above are merely specific embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, adjustments can be made to the measurement configuration, reactive power equipment, graph relationships, model parameters, projection safety margin, and trigger thresholds without departing from the spirit and principles of the present invention; any equivalent substitutions, modifications, or improvements resulting therefrom should be included within the scope of protection of the present invention.

Claims

1. A reactive power optimization method for sparse real-time measurement distribution networks based on load reconfiguration and projection feedback, characterized in that, The method includes the following steps: S1: Obtain the network topology, line parameters, historical load data, real-time measurement node configuration, and reactive power regulation equipment operation boundary of the target distribution network to determine the observable and unobservable node sets; based on the AC power flow results during the training period, the fixed power factor load disturbance power flow response, and the effective compensation effect of reactive power equipment, determine the voltage safety risk, load-voltage sensitivity, and constrained compensation capability, and then construct the voltage safety contribution weight and voltage control equivalent edge of the unobservable nodes; S2: Based on line connection relationships, historical load correlation relationships, and voltage control equivalent relationships, physical topology edges, load semantic edges, and voltage control equivalent edges are formed respectively, and a mask-aware multi-relationship graph attention load reconstruction model is constructed. The node mask load features, node observable identifiers, and node structural features are used as inputs to perform attention information aggregation and adaptive fusion in different relational neighborhoods. The model is jointly trained using ordinary load reconstruction error and voltage safety contribution weighted reconstruction error. S3: At the control moment, acquire the real-time load data of observable nodes, mask the real-time load input of unobservable nodes, and input it into the trained load reconstruction model to obtain the reconstructed load of unobservable nodes; concatenate it with the real-time load of observable nodes, and process the root node or balance node separately to form the network load status at the current control moment, and do not use the real-time load of unobservable nodes during the action generation process. S4: Based on the overall network load status and the operating boundary of reactive power regulation equipment, a hybrid discrete-continuous action space is constructed, which includes the tap positions of on-load tap-changing transformers, the switching positions of parallel capacitor banks, and the reactive power output of continuous reactive power compensation equipment. The frost-ice optimization algorithm is used to search for candidate control actions, and the fitness of candidate actions is evaluated from the aspects of network loss, voltage deviation, voltage over-limit risk, voltage safety margin, and equipment action cost based on the power flow calculation results under reconfigured load. S5: Construct a cone-constrained safety projection model based on the branch power flow model within the local action neighborhood of the candidate control action. Combine the node voltage boundary, branch power flow constraint, reactive power regulation equipment boundary, cone constraint relationship and voltage relaxation constraint to perform safety correction on the candidate action, obtain the projection control action that meets the voltage safety requirements and has a limited correction range, and extract projection feedback information such as projection residual, voltage relaxation amount, active voltage constraint, constraint pressure, equipment correction direction and voltage boundary margin. S6: Determine whether to trigger feedback enhancement search based on the projection feedback information; when the projection control action is near the voltage safety boundary, inject the projection control action and its local disturbance into the frost-ice optimization algorithm, and adjust the fitness evaluation, initial population or search bias according to the projection feedback information, and execute candidate action search and cone-constrained safety projection again; when the triggering condition is not met, use the current projection control action, and finally output the obtained projection control action as a reactive power optimization command.

2. The method according to claim 1, characterized in that, The voltage safety contribution weight is determined jointly by the low voltage risk during the training period, aggregated load-voltage sensitivity, and compensation weakness, specifically including: Based on the AC power flow calculation results during the training period, voltage nodes are obtained. At any moment The voltage amplitude, and based on the allowable lower limit of the node voltage. Calculate voltage nodes Low voltage risk : (1) in, For voltage node indexing, For time indexes within the training period, voltage node At any moment voltage amplitude, This is the lower limit of the allowable node voltage. express ; Apply a combined active and reactive power disturbance to the unobservable load node while maintaining a constant power factor, and calculate the load-voltage sensitivity based on the AC power flow voltage changes before and after the disturbance. For the training representative scene set The sensitivity within the range is aggregated to obtain the aggregated sensitivity: (2) in, For training representative scenarios Unobservable load nodes For voltage nodes Load-voltage sensitivity, This is a set of representative training scenarios used for sensitivity calculation. To train a number of representative scenes, Unobservable load nodes For voltage nodes Polymer load-voltage sensitivity; Based on reactive power regulation equipment, voltage nodes are adjusted without causing node voltage exceedance. Constrained effective compensation amount Calculate the normalized compensation capability and compensatory weakness : (3) (4) in, voltage node Normalization-constrained effective compensation capability voltage node The constrained effective compensation amount, The 5th percentile of the constrained effective compensation for all voltage nodes. The 95th percentile of the constrained effective compensation for all voltage nodes. Indicates will Cut off to interval , To compensate for the lower limit of weakness, voltage node The compensatory weakness; Calculate the unobservable load nodes based on the low voltage risk at voltage nodes, load-voltage sensitivity, and compensation weakness. At any moment Voltage safety contribution weight: (5) in, Unobservable load nodes At any moment Voltage safety contribution weight, This refers to the set of voltage nodes involved in voltage risk, load-voltage sensitivity, and compensation weakness calculations. Set of unobservable load nodes The summation index in the middle, The set of unobservable load nodes after removing root nodes or balancing nodes; when the denominator is zero, the normalized result of the static voltage safety contribution score during the training period is used as the prior voltage safety contribution weight.

3. The method according to claim 1, characterized in that, The voltage control equivalent edge is constructed according to the influence mode of the load node on the critical voltage constraint, specifically including: Based on the low voltage risk during the training period, voltage nodes are calculated. Importance of training risks : (6) in, voltage node The importance of training risks This is the summation index for traversing all voltage nodes in the normalized denominator. This is the set of training periods used to calculate the importance of training risks. This represents the mean operation. Indicates voltage node The 95th percentile of low voltage risk during the training period This is the percentile risk weighting coefficient; Based on aggregation sensitivity, training risk importance, and compensation weakness, load nodes are constructed. Voltage control influence vector , its first The components are: (7) in, For load nodes The voltage control influence vector, For load nodes For voltage nodes The control effect component; compute nodes With nodes Voltage control equivalence similarity between : (8) in, For nodes With nodes The voltage control equivalence similarity between them For nodes The node index used for similarity calculation. For nodes The voltage control influence vector, This represents the transpose of a vector. Represents the L2 norm; For each source node Choose the one with the highest voltage control equivalent similarity. The target nodes are connected as voltage control equivalent edges, forming a voltage control equivalent edge set. : (9) in, For voltage control equivalent edge set, The number of voltage-controlled equivalent edges selected for each source node. This indicates selecting the one with the highest similarity. There are several target nodes; the voltage control equivalent edge is used to characterize the similarity of the influence patterns of different nodes on the critical voltage constraint.

4. The method according to claim 1, characterized in that, The mask-aware multi-relationship graph attention load reconstruction model performs attention aggregation in the neighborhoods corresponding to physical topological relationships, load semantic relationships, and voltage control equivalent relationships, and performs adaptive fusion through node-level relationship gating. The model calculates training error only on unobservable load nodes, specifically including: Suppose a training batch contains One sample, This refers to the sample index in the training batch. For the sample At the corresponding time, For the sample Middle node Normalized reconfiguration load, For the sample Middle node Normalized real load, This is for normalized reconstruction error; For the sample The voltage security contribution weights of unobservable load nodes are normalized to obtain the normalized voltage security contribution weights: (10) in, For the sample Middle node Normalized voltage safety contribution weight, For the sample Corresponding time Unobservable load nodes The original voltage safety contribution weight; Voltage safety contribution weighted error loss Represented as: (11) Ordinary mean square reconstruction error loss Defined as: (12) Total training loss function of the model Represented as: (13) in, The weighting coefficients that contribute to the weighted error loss for voltage safety are used to adjust the proportion of this weighted error in the total training loss.

5. The method according to claim 1, characterized in that, The candidate control actions generated by the frost-ice optimization algorithm are used as reference actions for cone-constrained safe projection. The discrete control variables in the candidate control actions include the tap position of the on-load tap-changing transformer and the switching position of the parallel capacitor bank, and the continuous control variables include the reactive power output of the continuous reactive power compensation equipment. Discrete control actions are enumerated in the local discrete neighborhood of the reference action, and the continuous reactive power output, branch power flow variables, node voltage square variables, branch current square variables, and voltage relaxation variables are optimized to obtain projection control actions with limited correction amplitude relative to the reference action. The target for the safety projection near the candidate action is set as a weighted sum of the action correction, voltage relaxation, predicted network loss, and current regularization term: (14) in, Let be the projection objective function near the candidate action. and These represent the continuous control variables and discrete control variables in the projection model, respectively. and These are the continuous and discrete candidate actions output by the frost optimization algorithm, respectively. and They are respectively Low-voltage relaxation variables and high-voltage relaxation variables; branch road The square of the current, The resistance of this branch is... For branch set; , , , and These are the weighting coefficients for the continuous control action correction term, discrete control action correction term, voltage relaxation term, predicted network loss term, and branch current regularization term, respectively. Denotes the norm 1; The security projection satisfies the radial distribution network branch power flow balance, voltage drop, second-order cone, voltage boundary, equipment boundary and non-negative relaxation constraints; parallel capacitor banks and continuous reactive power compensation equipment are entered into the model through node reactive power injection, and on-load tap-changing transformers adjust the root node equivalent voltage through turns ratio adjustment. Each local discrete combination is sorted lexicographically according to voltage relaxation, action correction cost, predicted network loss and action distance to obtain the 0th round of projected action, and AC power flow safety is checked under reconfigured load.

6. The method according to claim 1, characterized in that, The feedback enhancement search sets four trigger conditions based on the safety projection results of the 0th cone constraint and the safety verification results of the reconfigured load AC power flow: voltage relaxation trigger condition, post-projection voltage margin trigger condition, action projection residual joint trigger condition, and low-voltage hot spot trigger condition. These four trigger conditions are as follows: (15) (16) (17) (18) Based on the above four types of triggering conditions, construct safety boundary trigger variables: (19) in, This is a voltage relaxation trigger condition. The maximum value of the low-voltage relaxation and high-voltage relaxation at each node obtained from the safety projection of the 0th cone constraint is indicated by the superscript. Indicates round 0. This is the voltage relaxation numerical tolerance, used to eliminate small numerical errors generated by the solver; This is the trigger condition for the projected voltage margin. The superscript represents the minimum voltage margin relative to the lower limit of the projected voltage after AC power flow safety verification under the reconfigured load condition for the 0th round of projection control action. This represents the verification quantity obtained based on the projection control action. This is the voltage margin trigger threshold after projection; For the joint triggering condition of motion projection residual, The L2 norm is calculated as the difference between the candidate control action of the 0th round frost-ice optimization algorithm and the safe projection action of the 0th round cone constraint, after scaling according to the allowable variation range of each control variable. The threshold for motion projection residuals. This is the voltage margin threshold corresponding to the triggering of the action projection residual; This is a low-voltage hotspot triggering condition. For the set of active low-voltage constrained nodes, This is the set of hotspot nodes determined based on the low-voltage risk distribution during the training period. To enhance the search trigger variables for feedback, This is an indicator function; it is triggered when any of the following conditions are met: voltage relaxation trigger condition, projected voltage margin trigger condition, combined action projection residual trigger condition, or low-voltage hot spot trigger condition. ,otherwise ; Among them, the symbol " " " represents the logical AND, meaning that both conditions connected must be true simultaneously; the symbol " "" represents the logical "OR", meaning that at least one of the conditions it connects is true; the symbol " "Represents the intersection of sets, symbol " "" indicates an empty set. Therefore, the non-empty intersection of the set of active low-voltage constrained nodes and the set of hot nodes indicates that they contain at least one common node.

7. The method according to claim 1, characterized in that, When the safety boundary triggering condition is met, a second-round feedback-enhanced fitness function is constructed based on the projection feedback information: (20) in, This refers to the candidate reactive power control actions searched in the second round of the frost-ice optimization algorithm. To enhance the fitness function through feedback, Based on the fitness function, , , and These are, respectively, motion projection residual penalty, node safety pressure penalty, voltage relaxation reliability penalty, and equipment bottleneck penalty. , , and These are the corresponding weighting coefficients; Among them, the node safety pressure penalty term can be composed of the margin deficit of low-voltage hot spot nodes and the constraint dual pressure, and the equipment bottleneck penalty term can further include the deviation of equipment correction direction and insufficient near-boundary safety margin; the relaxation reliability information is not used as a constant term that is the same for all candidate actions, but is used to amplify the near-boundary voltage deficit penalty of the second round of candidate actions, so that the term can change the order of candidate actions.

8. The method according to claim 1, characterized in that, The projection action, local perturbation action, and projection residual direction offset action from round 0 are injected into the initial population for the second round of candidate action search: (21) in, To provide feedback to enhance the initial population for the frost-ice optimization algorithm, Given a random initial set of individuals, For the 0th cone constraint safety projection action, This refers to the set of local perturbation individuals constructed around the projection action. For local disturbance scale. For the 0th cone constraint safety projection action Candidate moves for Frost Ice Optimization Round 0 The difference; its positive direction indicates the safe correction direction from the candidate action of the 0th round of frost optimization to the safe projection action of the 0th round of cone constraint. The set of search bias individuals formed along the direction of the action projection residual; symbol " "" represents the union of sets, used to merge the random initial set of individuals, the safe projection action of the cone constraint in the 0th round, the set of individuals with local perturbation, and the set of individuals with search bias to form the initial population of the feedback-enhanced frost-ice optimization algorithm; Local disturbance individuals and search bias individuals are all mapped back to the feasible region of the equipment: continuous reactive power output is subject to upper and lower limit truncation, and on-load tap changer taps and parallel capacitor bank taps are subject to integer mapping to ensure that the initial population in the second round satisfies the equipment operating boundary. The final reactive power optimization action is determined based on the feedback-enhanced search trigger results and the AC power flow safety verification results of the reconfigured load: (22) in, To freeze the output action, To provide feedback to enhance the safety projection action of the first round cone constraint obtained after the search, The first round of actions is the accept variable; when the first round of actions passes the AC power flow safety check under the reconstructed load state and its safety indicators do not deteriorate relative to the 0th round of projection actions, the accept variable is set to 1, otherwise it is set to 0; when the feedback enhancement search is not triggered or the first round of actions is not accepted, the 0th round of projection actions are output.

Images