Power distribution network source-load regulation method and device in weak measurement environment, equipment and medium

Abstract

The application discloses a power distribution network source and load regulation method, device and equipment and medium in a weak measurement environment, and belongs to the field of power distribution network source and load scheduling. The method is: acquiring the node importance and node measurement data of the power distribution network respectively, constructing a complete data set by using the node measurement data and determining a power distribution network source and load scene, wherein the complete data set is complete measurement information containing all observation nodes of the power distribution network; dividing the complete data set into a plurality of hierarchical training subsets according to the node importance and the power distribution network source and load scene; calculating a missing measurement parameter by using the hierarchical training subsets and determining a safety constraint index by using the node measurement data; performing progressive training according to the missing measurement parameter, the safety constraint index and the hierarchical training subsets to obtain a weak measurement environment regulation model; and calling the weak measurement environment regulation model to perform power distribution network source and load regulation processing. The application can call the model to perform power distribution network source and load regulation in a weak measurement environment, and improves the regulation precision.

Description

Technical Field

[0001] This invention relates to the technical field of power distribution network source-load regulation, and in particular to a method, device, equipment and medium for power distribution network source-load regulation in a weak measurement environment. Background Technology

[0002] With the advancement of technology, the scale of distributed energy resources (DERs), represented by distributed photovoltaic, wind power, and electric vehicle charging piles, being connected to the distribution network is increasing daily. The output of DERs is characterized by significant intermittency, volatility, and randomness, causing the power flow direction, voltage distribution, and power balance of the distribution network to exhibit drastic dynamic changes at different time scales (seconds, minutes, hours, seasons), posing a severe challenge to the safe and stable operation of the power grid.

[0003] To ensure the system safety of distribution networks in scenarios with diverse power sources and loads, a commonly used method for power source and load regulation is zoned coordinated control based on voltage sensitivity analysis. This method can be widely applied to voltage control scenarios in distribution networks containing distributed power sources. Specifically, it involves calculating the sensitivity matrix of node voltage to reactive power output, using algorithms such as spectral clustering to divide voltage control zones, and selecting a dominant node for coordinated control within each zone.

[0004] However, the above method has the following technical problems: its operation relies entirely on accurate sensitivity data and complete measurement information. However, when faced with problems such as imperfect measurement configuration and missing key node data, the sensitivity matrix calculation will fail, leading to the collapse of the partition structure, failure of the control strategy, and low control accuracy. Summary of the Invention

[0005] This invention provides a method, device, equipment, and medium for source-load regulation in a distribution network under weak measurement environment, which can solve the technical problems of easy regulation failure, reduced source-load regulation accuracy, and large regulation deviation in the existing technology under weak measurement environment.

[0006] A first aspect of this invention provides a method for source-load regulation of a distribution network in a weakly measurable environment, the method comprising: The importance of nodes and node measurement data of the distribution network are obtained respectively. A complete dataset is constructed using the node measurement data and the source-load scenario of the distribution network is determined. The complete dataset contains complete measurement information of all observed nodes of the distribution network. The complete dataset is divided into several hierarchical training subsets based on the node importance and the power distribution network source-load scenario, wherein each hierarchical training subset corresponds to a degradation level. The missing parameters of the measurement are calculated using the hierarchical training subset, and the safety constraint indicators are determined using the node measurement data. Based on the missing measurement parameters, the safety constraint indicators, and the hierarchical training subset, a weak measurement environment control model is obtained through progressive training. The weak measurement environment control model is then used for power distribution network source-load control processing.

[0007] The progressive multi-stage training of this invention trains the model step by step from low to high according to the level of observation degradation. This not only allows for matching the corresponding source-load scenario dataset to achieve progressive learning from easy to difficult, but also allows the model to gradually adapt to measurement deficiencies, greatly improving the generalization and robustness in weak measurement environments. As a result, the model can be called to perform source-load regulation of the distribution network in weak measurement environments, avoiding the failure of control strategies due to "weak measurement" and improving the accuracy of regulation.

[0008] A second aspect of the present invention provides a power distribution network source-load regulation device for weak measurement environments, the device comprising: The acquisition module is used to acquire the node importance and node measurement data of the distribution network respectively, and to construct a complete dataset and determine the source-load scenario of the distribution network using the node measurement data. The complete dataset contains complete measurement information of all observed nodes of the distribution network. The partitioning module is used to divide the complete dataset into several hierarchical training subsets according to the node importance and the power distribution network source-load scenario, wherein each hierarchical training subset corresponds to a degradation level; The calculation module is used to calculate the missing parameters of the measurement using the hierarchical training subset and to determine the safety constraint index using the node measurement data; The control module is used to perform progressive training based on the missing measurement parameters, the safety constraint indicators, and the hierarchical training subset to obtain a weak measurement environment control model, and then call the weak measurement environment control model to perform power distribution network source-load control processing.

[0009] Compared to existing technologies, the present invention provides a method, apparatus, device, and medium for source-load regulation of a distribution network in a weak measurement environment. The advantages of this invention are as follows: The invention can acquire the node importance and node measurement data of the distribution network, construct a complete dataset using the node measurement data, and determine the source-load scenario of the distribution network. The complete dataset contains complete measurement information for all observed nodes in the distribution network. Based on the node importance and the distribution network source-load scenario, the complete dataset is divided into several hierarchical training subsets. The hierarchical training subsets are used to calculate missing measurement parameters and to determine safety constraint indicators using node measurement data. Progressive training is performed based on the missing measurement parameters, safety constraint indicators, and hierarchical training subsets to obtain a weak measurement environment regulation model. This weak measurement environment regulation model is then used for source-load regulation of the distribution network. The progressive multi-stage training of this invention trains the model step by step from low to high according to the level of observation degradation. This not only allows for matching the corresponding source-load scenario dataset to achieve progressive learning from easy to difficult, but also allows the model to gradually adapt to measurement deficiencies, greatly improving the generalization and robustness in weak measurement environments. As a result, the model can be called to perform source-load regulation of the distribution network in weak measurement environments, avoiding the failure of control strategies due to "weak measurement" and improving the accuracy of regulation. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating a power distribution network source-load regulation method in a weak measurement environment according to an embodiment of the present invention. Figure 2 This is an operation flowchart of a power distribution network source-load regulation method in a weak measurement environment provided by an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a power distribution network source-load control device in a weak measurement environment provided by an embodiment of the present invention. Detailed Implementation

[0011] 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.

[0012] With the advancement of technology, the scale of distributed energy resources (DERs), represented by distributed photovoltaic, wind power, and electric vehicle charging piles, being connected to the distribution network is increasing daily. The output of DERs is characterized by significant intermittency, volatility, and randomness, causing the power flow direction, voltage distribution, and power balance of the distribution network to exhibit drastic dynamic changes at different time scales (seconds, minutes, hours, seasons), posing a severe challenge to the safe and stable operation of the power grid.

[0013] To ensure the system safety of distribution networks in scenarios with diverse power sources and loads, a commonly used method for power source and load regulation is zoned coordinated control based on voltage sensitivity analysis. This method can be widely applied to voltage control scenarios in distribution networks containing distributed power sources. Specifically, it involves calculating the sensitivity matrix of node voltage to reactive power output, using algorithms such as spectral clustering to divide voltage control zones, and selecting a dominant node for coordinated control within each zone.

[0014] However, the above method has the following technical problems: its operation relies entirely on accurate sensitivity data and complete measurement information. However, when faced with problems such as incomplete measurement configuration and missing data at key nodes, the sensitivity matrix calculation will fail, leading to the collapse of the zoning structure and the failure of the control strategy. If a fixed zoning structure and offline tuning parameters are used, the control strategy will also be mismatched or even have a delayed response once the photovoltaic output changes suddenly or the load fluctuates rapidly. This will increase the safety risks of the distribution network and make it difficult to meet the safety requirements of the distribution network.

[0015] To address the aforementioned issues, the following specific embodiments will provide a detailed description and explanation of a power distribution network source-load regulation method, device, equipment, and medium for weak measurement environments provided in this application.

[0016] To address the technical problems of existing technologies that easily lead to control failure, reduced source-load control accuracy, and large control deviations under weak measurement environments, this paper refers to... Figure 1 The diagram shows a schematic flowchart of a power distribution network source-load regulation method in a weak measurement environment according to an embodiment of the present invention.

[0017] As an example, the power distribution network source-load regulation method in a weakly measurable environment may include: S11. Obtain the node importance and node measurement data of the distribution network respectively, construct a complete dataset using the node measurement data and determine the source-load scenario of the distribution network, wherein the complete dataset contains complete measurement information of all observed nodes of the distribution network.

[0018] In one embodiment, the distribution network topology, node admittance matrix, power flow data, and photovoltaic or load operation data can be acquired to construct a node electrical importance evaluation system to determine the node importance of the distribution network.

[0019] Simultaneously, it can acquire full-node measurement data of the distribution network (including voltage, active power, reactive power, photovoltaic output, and load power), and construct a complete and missing benchmark dataset by dividing the scenarios according to the source and load operating characteristics.

[0020] Specifically, based on the photovoltaic penetration rate and load level of the distribution area, the source and load scenarios are divided into three categories: conventional scenarios, high penetration scenarios, and peak load scenarios. For extreme scenarios (high penetration / peak), core node measurements should be retained as a priority.

[0021] To simulate the objective law of the dynamic evolution of measurement gaps in real distribution networks under different operating scenarios, this invention constructs a hierarchical training dataset corresponding to the level of observation degradation. The distribution area is defined. a The set of all observation nodes is According to the district a Total photovoltaic output With total load Calculate the photovoltaic penetration rate index of the transformer area. .in accordance with and The source load scenario is divided into several typical categories: when When the penetration threshold is exceeded, it is classified as a high-penetration scenario; when When the peak threshold is exceeded, it is classified as a peak scene; when and When all conditions are within the normal range, it is classified as a normal scenario.

[0022] Define a complete dataset This dataset contains complete measurement information for all observation nodes in the distribution network. Each node... The measurement information includes node voltage amplitude. Node-injected active power Node-injected reactive power Photovoltaic power output Load power Then the complete dataset can be represented as shown in the following formula (1): (1); In the above formula, For scene category indexing, station area a In the scene s Complete subset of measurement data It can be shown in the following formula (2): (2); In the above formula, For time index, Taiwan District a At any moment t The node measurement state vector.

[0023] To quantify the electrical criticality of each node in a distribution network, the node importance of the distribution network is calculated using the electrical criticality. As an example, obtaining the node importance of the distribution network may include the following sub-steps: S111. Obtain node information of the distribution network, wherein the node information includes: the topology of the distribution network, node admittance matrix, power flow data, photovoltaic operation data and load operation data.

[0024] S112. Calculate the electrical key index values ​​of the node using the node information, wherein the electrical key index values ​​include: power flow sensitivity index, voltage stability index, and source load type correction factor.

[0025] S113. Normalize and weightedly fuse several electrical key indicator values ​​in sequence to obtain the node importance.

[0026] In one embodiment, by determining the importance of nodes in the distribution network, the real weak measurement patterns of the distribution network can be simulated, allowing the model to learn step by step from "complete measurement" to "extreme lack", and integrating electrical prior knowledge to improve robustness.

[0027] To determine the importance of nodes in a distribution network, the electrical criticality of each node can be quantified, thereby determining the measurement shielding priority based on the node importance (the more important the node, the later it is shielded, or no shielding at all).

[0028] In one embodiment, node information of the distribution network can be obtained, including: the distribution network topology, node admittance matrix, power flow data, photovoltaic operation data, and load operation data. Then, power flow sensitivity indicators, voltage stability indicators, and source-load type correction factors can be calculated based on the node information.

[0029] Among them, the power flow sensitivity index PTDF can be calculated using a simplified DC power flow model. Specifically, it can calculate the impact of unit active power injected at node i on the power flow of line r. It is used to reflect the influence of nodes on the power flow regulation of the grid, and the larger the value, the more important it is.

[0030] The voltage stability index L is calculated based on the HELM power flow method, which calculates the sensitivity of node voltage to power changes. The closer the value is to 1, the closer the node is to the voltage collapse threshold and the higher its safety vulnerability.

[0031] The source-load type correction factor can be calculated separately for different scenarios. For photovoltaic nodes, it can be corrected according to photovoltaic penetration rate, with higher penetration nodes having a higher weight. For load nodes, it can be corrected according to load peak-valley characteristics, with higher peak load nodes having a higher weight.

[0032] Assume that the number of observation nodes in the distribution network is defined as follows: The set of observation nodes is denoted as The number of branch roads is R The set of branches is The number of transformer substations is A The area is assembled as For each node i An evaluation index for electrical importance is constructed, which is comprehensively based on the following three dimensions: 1. Power Flow Sensitivity PTDF: Computational Node i The sensitivity of injected power changes to power flow in critical branches or at slack nodes describes the impact of changes in control variables on line power flow. This can be calculated using the Power Transfer Distribution Factor (PTDF) method, reflecting the effect of a change in injected power at a particular node on the power flow of a specific line. When describing the impact on node voltage changes, it is calculated as the ratio of the change in active power at the node to the change in node voltage; when describing the impact on branch power flow changes, it is calculated as the ratio of the change in current-carrying capacity of the branch containing the node to the change in injected power at the node.

[0033] For branch power flow sensitivity, the PTDF method establishes a quantitative relationship between nodal injected power and line power flow under a linearized model. The calculation process is as follows: First, construct the system's node admittance matrix. Inverse the matrix to obtain the nodal impedance matrix. For connected nodes m and nodes n The route r Define nodes i Power transmission distribution factor of the line This value represents the node. i When injecting unit active power, the line r The change in the tidal current can be calculated using the following formula (3): (3) In the above formula, , For nodes in the impedance matrix i With nodes m , n mutual impedance, For nodes m and nodes n The mutual impedance.

[0034] In engineering, simplified models based on DC power flow are often used, neglecting reactive power and resistance, and only considering the relationship between active power and voltage phase angle. Definition The system node susceptance matrix, For the line r The susceptance. Let the node phase angle be... With injection power satisfy Line flow satisfy ,in , For nodes m and nodes n The phase angle. Based on the above conditions, equation (3) can be simplified as shown in the following formula (4): (4); In the above formula, Inverse matrix Middle m Line 1 i Column elements, Inverse matrix Middle n Line 1 i The elements of the column.

[0035] When the injected power at a certain node changes At that time, the change in power flow along the line is... Taking engineering applications as an example, in a certain regional power grid, the node can be determined through PTDF calculation. i When the injected power increases by 1MW, the power flow of line L-1 increases by 0.8MW, and the power flow of line L-2 decreases by 0.3MW. Dispatchers can then precisely adjust the nodes accordingly. i The output effectively avoids line overload while meeting system safety constraints, thus optimizing the power generation plan.

[0036] 2. Voltage stability index (L) Voltage stability indices can be based on the Holomorphic Embedding Load-flow Method (HELM) to calculate the partial derivatives of node voltage magnitude / phase angle with respect to node injected power (active P, reactive Q), characterizing the sensitivity of voltage to power changes, and used for voltage stability analysis, reactive power optimization, and weak node identification. The closer a node is to 1, the closer it is to the critical stability point, and the higher its importance. The calculation process is as follows: First, construct the system's node admittance matrix. Determine the node i self-guided nano ,node i and nodes j mutual conductance Perform power flow calculations for the current source-load operation scenario to obtain node voltage amplitude and voltage phase angle. For distribution network nodes... i The voltage stability index L can be calculated as shown in the following formula (5): (5); In the above formula, , For nodes i and nodes j voltage amplitude, For nodes i and nodes j The voltage phase angle difference.

[0037] 3. Source Load Type Correction Factor To differentiate the electrical importance of different types of nodes and achieve dynamic adaptation between node importance evaluation and source load operating scenarios, this invention introduces a source load type correction factor based on the source load attributes and operating characteristics of the nodes. Differentiate the importance of different types of nodes. Define nodes. i The photovoltaic penetration rate is ,node i The set of nodes in the corresponding transformer area The photovoltaic penetration rate in this area is .

[0038] For photovoltaic grid connection nodes i The source-load type correction factor is set according to its photovoltaic penetration rate. Specifically, it can be shown in the following formula (6): (6); In the above formula, This is an adjustable parameter used to balance the photovoltaic penetration rate term. With uniform distribution of the transformer area Weights in the correction factor; This is the baseline correction factor for photovoltaic nodes; For nodes i The total number of nodes in the corresponding area.

[0039] For load nodes i The source load type correction factor is set according to its load peak and valley characteristics, as shown in the following formula (7): (7); In the above formula, This is the baseline correction factor for the load node; , For nodes i Peak and trough load values ​​across all samples; For load nodes i The reference power.

[0040] 4. Index normalization: In order to eliminate the differences in physical dimension and numerical magnitude between the power flow sensitivity PTDF and voltage stability L index, and to ensure the physical consistency of weighted fusion, this invention performs normalization processing on the above indexes and uniformly maps them to the [0,1] interval.

[0041] The tidal current sensitivity was standardized by deviation and a normalized index was obtained. Specifically, it can be shown in the following formula (8): (8); In the above formula, For nodes i A set of connected lines.

[0042] The voltage stability index L is truncated and normalized to eliminate extreme value interference and obtain a normalized index, as shown in the following formula (9): (9); In the above formula, , This is the critical threshold for voltage stability and can be adjusted according to engineering requirements.

[0043] The normalized power flow sensitivity, voltage stability L index, and source-load type correction factor index are objectively weighted, and the comprehensive importance score of each node is calculated. Specifically, it can be shown in the following formula (10): (10); In the above formula, The weights of the power flow sensitivity PTDF and voltage stability L index can be adjusted according to the actual system operation requirements. The observation priority of physical constraint binding was clarified. The higher the score, the more critical the node, and it should be reserved in the measurement shielding.

[0044] Overall importance score of nodes The result is obtained by objectively weighting and fusing three indices: normalized power flow sensitivity, voltage stability L index, and source-load type correction factor.

[0045] Among them, the normalized power flow sensitivity and voltage stability L index accurately characterize the irreplaceable role of nodes in the safe control of the distribution network from two complementary physical dimensions: the controllability of power flow regulation and the vulnerability of voltage security, while taking into account both the efficiency of regulation execution and the safety of grid operation. The source-load type correction factor can adapt to extreme conditions such as high photovoltaic penetration and peak load, and dynamically adjusts the importance of nodes according to the source-load scenario. The hierarchical measurement shielding strategy formulated in this way fits the actual operating law of the distribution network and provides core physical constraint support for the robust training of the regulation model under weak measurement environment.

[0046] The electrical importance of nodes in this invention comprises three dimensions: power flow sensitivity (PTDF), voltage stability (L) index, and source-load type correction factor. PTDF characterizes the node's ability to regulate line power flow, the L index characterizes the node's voltage security vulnerability, and the source-load correction factor dynamically adjusts the node's importance to suit photovoltaic or load scenarios. The higher the normalized weighted score of these three factors, the more critical the node, and the more preferentially it is retained in measurement shielding, which can solve the problem that traditional random shielding does not conform to actual operating conditions.

[0047] S12. The complete dataset is divided into several hierarchical training subsets according to the node importance and the power distribution network source-load scenario, wherein each hierarchical training subset corresponds to a degradation level.

[0048] In one embodiment, nodes can be sorted according to their importance to dynamically classify the observation degradation level of nodes. Then, the complete dataset corresponding to different power distribution network source-load scenarios can be divided according to the observation degradation level to obtain several hierarchical training subsets, where each hierarchical training subset corresponds to a degradation level.

[0049] In an optional embodiment, dividing the complete dataset into several hierarchical training subsets based on the node importance and the power distribution network source-load scenario may include the following sub-steps: S121. Based on the node importance and the preset masking threshold, several degradation levels are divided.

[0050] S122. Selectively mask the complete dataset corresponding to each of the power distribution network source-load scenarios according to the size of the degradation level to obtain several hierarchical training subsets with different degradation levels. The selective data masking is to set the measurement data corresponding to the nodes that do not match the degradation level to zero.

[0051] In one embodiment, the present invention can simulate a graded weak measurement environment ranging from "minor missing" to "extreme missing" by dynamically dividing the observation degradation level and generating graded training subsets according to the source load scenario.

[0052] Based on real-time source-load scenarios and node importance scores, selective data masking is performed on the complete dataset to construct training subsets under different observation degradation levels. , This represents the degradation level; the higher the level value, the more severe the missing measurement data. The specific construction process is as follows: Based on node importance score Set the shielding thresholds respectively and Degradation level k corresponding training subset It is generated by applying a masking operation to the complete dataset, as shown in the following formulas (11) and (12): (11); (12); In the above formula, For the first k The set of masked nodes corresponding to the first-order degradation satisfies This means that as the degradation level increases, the set of masked nodes increases incrementally. Mask represents the set of nodes masked in the dataset. Set of shielded nodes The included nodes are masked, that is, the measurement data corresponding to the masked nodes are set to zero, simulating a weak measurement scenario where node measurements are missing.

[0053] To implement the hierarchical masking logic, node importance scores are used. With shielding threshold , For each degradation level k Define the corresponding set of shielded nodes, and further divide the nodes into three levels, as shown in the following formula (13): (13); At low degradation level ( When ), only the less important ones are masked. ) nodes As the degradation level increases ( Gradually expand the shielding range to nodes of medium importance. At the highest degradation level ( Under this setting, importance can be masked. nodes ,reserve The core load and hub nodes. Through the above methods, the distribution pattern of measurement shielding is matched with the measurement loss patterns caused by communication interruptions, sensor failures, etc., in the actual operation of the distribution network, ensuring that core nodes still preferentially retain their observation data even in extreme scenarios.

[0054] The observation degradation level is set to K. The larger the level k, the more serious the measurement loss. There are K levels of progressive training. Two levels of masking thresholds are set according to importance, and the nodes are divided into high importance, medium importance and low importance.

[0055] Hierarchical masking rules (Formulas 11-12): Low degradation level (small k): only low importance nodes are masked; Medium degradation level: masking is expanded to medium importance nodes; High degradation level (large k): only core high importance nodes are retained, and the rest are masked.

[0056] For masking operations, the measurement data of the corresponding node can be set to 0 to simulate the actual measurement loss.

[0057] When generating hierarchical training subsets, each level of degradation corresponds to one training subset, and the missing distribution conforms to the real power grid fault or communication interruption patterns.

[0058] As analyzed above, the source-load scenario s is divided according to the photovoltaic penetration rate of the transformer area. Total load The operating scenarios are divided into three categories, as follows: s = standard scenario: , All are within the normal range, and the power grid is operating smoothly; s = High-penetration scenario: Exceeding the high penetration threshold, photovoltaic power output fluctuates greatly; s = Peak load scenario: When the peak load threshold is exceeded, the power supply pressure on the load is high.

[0059] Assume the degradation levels are: k=1,2,...,K (a total of K levels).

[0060] Node importance can be divided into 3 levels (Formula 13), and scores are given according to node importance. and shielding threshold and All nodes can be divided into 3 groups: High importance nodes Node importance score ≥ (Core hubs / source load nodes will be prioritized and never easily blocked). Medium importance nodes : ≤ Node Importance Score < (Secondary critical nodes, only shielded when degradation is medium to high); Low importance nodes Node importance score < (Edge nodes should be prioritized for masking).

[0061] For high-penetration or peak-load scenarios, core nodes must be preserved first, and only low-importance nodes should be shielded; for normal scenarios, shielding can be gradually expanded according to the level.

[0062] (1) Normal scenario s = normal.

[0063] The shielding is expanded progressively according to the degradation level, strictly adhering to the node importance: k=1 (low degeneracy): (Only low importance is masked); k=2 (degenerate): (Mask low + medium importance); k=K (high degeneracy): M_K = all nodes - (Only high importance is retained, the rest are all hidden).

[0064] (2) High-permeability scenario s = high-permeability.

[0065] Photovoltaic power generation is highly volatile and poses significant safety risks, requiring extremely conservative shielding. All levels k: Only contains C lo w (Only low importance is masked, never medium / high importance is masked); Objective: To retain core node measurements and ensure the security of control.

[0066] (3) Peak load scenario s = peak.

[0067] High load pressure and voltage easily exceeding limits necessitate conservative shielding: All levels k: Only contains (Only low importance is hidden); Objective: To retain measurements of critical loads / hub nodes and prevent control failures.

[0068] When generating tiered training subsets, all source load scenarios can be traversed: s∈{normal, high penetration, peak}; complete data for the current scenario is retrieved: load D s (Full measurement, no missing data), and generate a masking set according to the level. : For typical scenarios: As k increases, from None of them are core components; High penetration or peak: Fixed = (Low importance only); Execute masking to generate subsets: (Set the masked node to 0) to complete all combinations: obtain all graded training subsets of 3 types of scenarios × K levels of degradation.

[0069] The progressive multi-stage training of this invention trains the model step by step from low to high according to the observation degradation level. At the low level, only low-importance nodes are masked, and at the high level, the masking range is gradually expanded. After the model of the previous level converges, the parameters are transferred to the next level, and only the parameters of the dual-path electrical attention, feedforward network and output layer are finely adjusted. At the same time, the model is matched with the corresponding source and load scene dataset to achieve progressive learning from easy to difficult, allowing the model to gradually adapt to the lack of measurement and greatly improving the generalization and robustness in weak measurement environment.

[0070] S13. Calculate the missing parameters using the hierarchical training subset and determine the safety constraint index using the node measurement data.

[0071] Next, the missing parameters can be calculated using the hierarchical training subset, and the safety constraint indicators can be determined using the node measurement data. The sparsity of the data can be quantified by measuring the missing parameters, and the risks can be identified in advance by the safety constraint indicators, so as to improve the accuracy of subsequent model training.

[0072] In one embodiment, the step of calculating the missing measurement parameters using the hierarchical training subset and determining the security constraint index using the node measurement data may include the following sub-steps: S131. Calculate the global missing value and the local missing value using the hierarchical training subset to obtain the measurement missing parameters.

[0073] S132. Using the node measurement data, calculate the voltage over-limit index, branch power over-limit index and comprehensive risk value respectively to obtain the safety constraint index.

[0074] Before model forward propagation and training, two key indicators can be determined to quantify the intensity of the dual demand for core feature extraction and safety control. Among them, the safety constraint risk prediction is a pre-emptive risk quantification step before the execution of control actions. It is calculated based on the inherent electrical characteristics of the node and the current measurement status, and is used to guide the feature extraction network to focus on high-risk nodes in advance.

[0075] For the current input sample, define the first... k Global measurement missingness at degradation level With nodes i Local measurement missing degree Specifically, it can be shown in the following formula (14): (14); In the above formula, For nodes i A local set of nodes, containing nodes i itself, and its relationship with nodes i All adjacent nodes that have direct electrical connection or electrical coupling relationship; For nodes i The total number of local nodes, This is an indicator function that takes the value 1 when node measurement data is missing, and 0 otherwise.

[0076] Based on the current operating conditions, the inherent electrical characteristics of the nodes, and the lack of measurement data, the following safety indicators are defined to predict potential over-limit trends: (1) Node voltage over-limit index Specifically, it can be shown in the following formula (15): (15); In the above formula, This is the reference voltage. This indicator is used to determine in advance how much the voltage deviates from the safety boundary.

[0077] (2) Branch power over-limit index Specifically, it can be shown in the following formula (16): (16); In the above formula, branch road r active power, This represents the upper limit of active power in the branch circuit. This indicator is used to determine in advance how close the power is to the overload boundary.

[0078] Define nodes based on the potential over-limit risks of both nodes and associated branches. i Safety constraint risk prediction indicators Specifically, it can be shown in the following formula (17): (17); In the above formula, These are the weighting coefficients. For nodes i The total number of connected lines. The larger the value, the stronger the node. i The higher the risk of safe operation, the higher the index will be incorporated into the attention weight calculation of the electrically coupled attention network as a bias term, so that the model can adaptively focus on high-risk nodes during feature extraction and improve the safety assurance capability of the control strategy under weak measurement conditions.

[0079] S14. Based on the missing measurement parameters, the safety constraint indicators, and the hierarchical training subset, progressive training is performed to obtain a weak measurement environment control model, and the weak measurement environment control model is called to perform power distribution network source-load control processing.

[0080] After determining the missing measurement parameters and safety constraints, the input layer of the model can be built using the missing measurement parameters. Then, the model can be trained using a hierarchical training subset with the safety constraints as constraints, until the model meets the safety constraints, thus obtaining a weak measurement environment control model.

[0081] After obtaining the weak measurement environment control model, it can be used to perform source and load control processing of the distribution network.

[0082] In one embodiment, the step of progressively training based on the missing measurement parameters, the safety constraint indicators, and the hierarchical training subset to obtain a weak measurement environment control model, and then calling the weak measurement environment control model for distribution network source-load control processing, may include the following sub-steps: S141. The measurement missing attention branch and the electrical coupling attention branch of the model are constructed using the measurement missing parameters respectively.

[0083] S142. The measurement-deficient attention branch and the electrical coupling attention branch are fused and used as inputs to the policy output layer of the model to obtain a dual-path electrical attention network.

[0084] S143. The dual-path electrical attention network is progressively trained using the safety constraint index and the hierarchical training subset to obtain a weak measurement environment control model.

[0085] S144. Call the weak measurement environment control model to perform power distribution network source and load control processing based on the real-time measurement data of the power distribution network.

[0086] In one embodiment, the present invention constructs a dual-path electrical attention network dedicated to the power distribution network as the core feature extraction module of the control agent, and its structure includes two parallel attention mechanisms.

[0087] (1) Measuring the lack of attention network: This network will have nodes i In the k Local measurement missingness at the degradation level Introducing the attention score as a bias term enables the model to perceive the degree of data loss in the current scene. The parameters of the missing attention network can be specifically measured as shown in the following formula (18): (18); In the above formula, For the first k Weight coefficients for measuring the lack of attention in a network at a degradation level. For the first k The bias baseline term for measuring the lack of attention network at the first level of degradation. The weights for the degree of missing data in local measurements.

[0088] (2) Electrically Coupled Attention Network: This network will constrain the risk of safety. The parameters of the electrical coupling attention network are added as weights in the calculation. The specific parameters of the electrical coupling attention network can be shown in the following formula (19): (19); In the above formula, For the first kThe weighting benchmark term of the electrically coupled attention network at the degradation level. For the first k Bias terms of the electrically coupled attention network at the degradation level.

[0089] To achieve dynamic feature adaptation under different operating scenarios, the outputs of the two attention channels are modulated using learnable modulation coefficients. The fusion can be performed as shown in the following formula (20): (20); In the above formula, the learnable modulation coefficients It can adaptively adjust the fusion weights of the two features according to the degree of measurement loss and changes in the source load scenario, so as to achieve collaborative optimization of data quality perception and physical security perception.

[0090] Features after fusion As input to the strategy output layer, it is used to generate the final control action, thus obtaining the dual-path electrical attention network.

[0091] Next, the dual-path electrical attention network can be trained using a progressive multi-stage training mechanism with safety constraint indicators and tiered training subsets to obtain a control model for weak measurement environments. By employing a progressive multi-stage training mechanism and combining the adaptive focusing capability of the dual-path electrical attention network, progressive learning from easy to difficult can be achieved, enabling the model to maintain robust control performance even under extreme operating conditions where measurements are gradually lost.

[0092] Finally, the weak measurement environment control model can be invoked to perform source-load control processing of the distribution network based on real-time measurement data. It should be noted that the trained and solidified model can include: a dual-path electrical attention network (measuring missing + electrical coupling), a source-load scenario discriminator, observation degradation level matching rules, a rigid safety constraint judgment module, and a control strategy output network.

[0093] In one specific operational approach, the trained model can be deployed to the distribution network dispatch master station or edge gateway, enabling it to access distribution network weak measurement data (voltage, active / reactive power, photovoltaic, load) in real time and load pre-trained safety constraint thresholds (voltage 0.93~1.07 times, line power limit, equipment adjustment range). Its scenario configuration can be divided into three categories of source-load scenario discrimination thresholds: pre-set conventional, high-penetration, or peak load.

[0094] When scheduling and invoking online, the operation may include the following steps: Step 1: Real-time weak measurement data acquisition and preprocessing.

[0095] Specifically, it can collect node voltage, injected power, photovoltaic output, load power, and branch power. The missing node measurement is set to 0 (completely consistent with the Mask operation during training) and the same normalization scale is maintained with the training set to ensure the consistency of model input. Finally, a standardized real-time weak measurement state vector can be obtained.

[0096] Step 2: Real-time source-load scene identification.

[0097] Specifically, it can calculate the current photovoltaic penetration rate and total load level of the distribution area, and automatically identify the current scenario (including normal scenario, high penetration scenario (high photovoltaic output) or peak load scenario (high load ratio)) to obtain the current source-load scenario label s.

[0098] Step 3: Determine the degradation level through online observation.

[0099] Specifically, based on real-time data, global measurement missing degree and node local missing degree can be calculated, and the observation degradation level k during training can be matched according to the missing degree. Specifically, the observation degradation level k during training can be matched by combining scene labels (high penetration or peak scenes are fixed with low degradation level, and regular scenes are matched according to missing degree) to obtain the current observation degradation level k.

[0100] Step 4: Extraction of dual-path electrical attention features.

[0101] Specifically, the weak measurement state vector can be input into the dual-path electrical attention layer of the model. The measurement missing attention can perceive the current degree of data missing and weight the features; the electrical coupling attention can focus on high-risk nodes and strengthen safety features, thereby outputting the fused electrical feature vector (no training, only forward propagation). This electrical feature vector is the core electrical feature of the adaptive weak measurement scenario.

[0102] Step 5: Model inference generates the original control strategy.

[0103] Specifically, the fused features can be input into the strategy output layer, and the model can directly infer and output source-load collaborative regulation instructions, including: 1. Source-side command: Active or reactive power regulation of each photovoltaic inverter; 2. Load-side instructions: Reduce or restore power for each controllable load; 3. Reactive power side command: switching or adjustment of reactive power compensation device.

[0104] Step 6: Online rigid safety constraint verification.

[0105] Specifically, it allows for pre-implementation safety simulation verification of the original instructions (without actual execution). It can rigorously verify three rigid constraints (consistent during the training phase), including: 1. Voltage constraint: All node voltages ∈ 0.93~1.07 times the rated voltage; 2. Power constraint: Branch power ≤ Line thermal stability limit; 3. Equipment constraints: The adjustment range is within the rated range of the equipment.

[0106] Release is only granted when all constraints are 100% met; otherwise, a policy correction is triggered. Ultimately, a final control command for safety and compliance can be output.

[0107] Step 7: Issuance and execution of dispatch instructions on site.

[0108] Specifically, the master station can issue compliance instructions to distributed photovoltaic inverters, controllable load terminals, and reactive power compensation devices (SVG, capacitors). The field equipment then executes control actions to smooth voltage fluctuations and balance source and load power.

[0109] Subsequently, the distribution network's operational status can be updated in real time, and post-execution measurement data can be collected to verify the control effect, record safety compliance rate, voltage fluctuation rate, and photovoltaic absorption rate, and retain lightweight logs for model iteration and optimization. Optionally, a dispatch execution report can also be output.

[0110] In one embodiment, the step of progressively training the dual-path electrical attention network using the safety constraint index and the hierarchical training subset to obtain a weak measurement environment control model may include the following sub-steps: S1431. Sort the hierarchical training subsets according to their degradation levels to obtain the training sorting results.

[0111] S1432. Based on the training ranking results, the dual-path electrical attention network is trained step by step using the hierarchical training subset and the safety constraint index, and the model parameters are adjusted to obtain a pre-trained model. The model parameters include: dual-path electrical attention parameters, feedforward network weight bias, and policy output layer.

[0112] S1433. After the pre-trained model has reached convergence according to the preset training loss, a security check is performed on the pre-trained model. When the security check passes, a weak measurement environment control model is obtained.

[0113] In one embodiment, the present invention can divide the training process into categories corresponding to the observed degradation levels. K In each stage, the agent is trained and regulated progressively, from low to high degradation levels. Let the first stage be... k The set of parameters for the regulation model corresponding to the grade degradation is This parameter set contains input layer parameters. Layer-Norm scaling offset parameters Measurement missing - Electrically coupled dual-path electrical attention network layer parameters The weight bias parameters of the two layers of the feed-forward network (FFN) and output layer parameters Specifically, it can be shown in the following formula (21): (twenty one); In the above formula, This can be further subdivided into measuring the parameters of attention-deficient networks. Electrically Coupled Attention Network Parameters and dual-channel modulation coefficient The specific construction method is described in S1.4-S1.5.

[0114] Once the previous training stage reaches convergence, its model parameters are used as the initial parameters for the next training stage. Let the first stage be... Level model after After the nth iteration, it converges, and the parameters at convergence are... Then the first k Initial parameters of the level model It can be shown in the following formula (22): (twenty two); In the above formula, This represents the number of iterations required for convergence in the previous stage, determined by monitoring changes in the loss function and physical constraint metrics on the validation set.

[0115] In the next training stage, only the adjustable parameters in the dual-path electrical attention network, the weight bias parameters of the two layers of the feedforward network, and the policy output layer are fine-tuned. Then, the model parameters involved in fine-tuning at the k-th degradation level are... It can be shown in the following formula (23): (twenty three); In the above formula, It represents the union of sets.

[0116] To support the mathematical logic of progressive multi-stage training, the definition is given at the first stage. k Level 1 Degradation Loss function during the training process The loss function integrates the physical constraints and control precision of the control strategy. The attention parameter update adopts the gradient descent method, as shown in the following formula (24): (twenty four); In the above formula, These are the parameters after the y-th training iteration at the k-th degradation level; The learning rate is the learning rate at the k-th degradation level. Characterization loss function Attention parameters The partial derivatives are used to quantify the rate of change of the loss function as the parameters change.

[0117] In one approach, the training order can be determined first (fixed and irreversible), and training can proceed from low to high according to the observed degradation level, for example: k=1 (low missing) → k=2 (medium missing) → … → k=K (extremely weak measurement).

[0118] The smaller the level k, the fewer missing measurements and the simpler the training; the larger the level k, the more missing measurements and the more difficult the training.

[0119] For the definition of single-level model parameters, as shown in Equation 21, the degenerate model parameter θ_k at level k consists of 5 parts: θ_k = {θ_input, θ_norm, θ_att, θ_ffn, θ_out}.

[0120] θ_input: Input layer parameters; θ_norm: Layer-Norm normalization parameter; θ_att: Dual-path electrical attention network parameters (core); θ_ffn: Weights and biases of two layers of the feedforward network; θ_out: Parameters for the policy output layer.

[0121] Next, parameter transfer across levels can be performed (Equation 22). Once the previous level (k-1) training has fully converged, all the converged parameters from the previous level can be directly used as the initial parameters for the current level. The purpose of this operation is to retain the learned regulatory knowledge, avoid training from scratch, and improve generalization in weak measurements.

[0122] Throughout the model training process, the aforementioned safety constraint index can be used. During progressive training, the safety constraint index must be used to calculate each sample. Through the safety constraint index, safety risk features can be provided for training, enabling the attention network to learn to focus on high-risk nodes.

[0123] Next, we can perform single-level training parameter fine-tuning (Formula 23). In the current level (k) training, we only fine-tune 3 types of parameters, and freeze the rest: ; Freezing includes: input layer and normalization layer parameters (to ensure the stability of the underlying features); Fine-tuning includes: dual-path attention, feedforward network, and output layer (adapting to the current missing scenarios).

[0124] Finally, gradient descent parameter updates can be performed (Equation 24). Iteratively update and fine-tune the parameters to minimize the overall loss function: ; Where η_k is the learning rate for the k-th level (decreasing slightly at each level). This is the partial derivative of the loss function with respect to the parameters.

[0125] The iterative method can be batch training + forward propagation + back propagation.

[0126] When the maximum number of iterations is reached or the loss decreases significantly and slows down, training is paused and preliminary convergence verification is performed to obtain the pre-trained model for the current degradation level k.

[0127] Next, the pre-trained model can be converged using the preset training loss. The core objective of verification is to determine whether the model loss has reached basic convergence. The preset training loss includes training set loss and validation set loss. Two convergence criteria must be met simultaneously: First, training set loss convergence: after 20-50 consecutive iterations, the loss decreases by less than 1×10⁻⁻⁻⁻⁶. 4 ~ 1×10⁻³.

[0128] Second, the validation set loss is stable: the validation set loss does not show a continuous decreasing or increasing trend, but tends to be stable.

[0129] After initial convergence, the pre-trained model can be subjected to a security check. If the security check passes, a weak measurement environment control model is obtained. If initial convergence fails, iterative training continues.

[0130] When performing security verification, you can prepare all samples of the verification set, the pre-trained model, and the power grid security parameters. The core objective of the verification is to determine whether the output of the rigid verification model 100% meets the physical security constraints of the distribution network.

[0131] In this embodiment, the verification criteria include three rigid constraints: (1) Node voltage constraints. All node voltages are between 0.93 and 1.07 times the rated voltage.

[0132] (2) Branch power constraint. Branch active power / apparent power ≤ Line thermal stability limit + Rated transmission capacity (3) Adjustable equipment constraints. The adjustment range of the photovoltaic inverter and reactive power compensation device is within the rated adjustment range of the equipment. During verification, the entire weakly measured sample set of the validation set can be input into the pre-trained model. The model outputs control actions, calculates the power grid operation results, and verifies the three major constraints sample by sample, node by node, and branch by branch. Finally, the number of samples, nodes, and branches exceeding the limits is counted. If 100% of the samples in the validation set satisfy all constraints and there are no exceedances, the safety verification is considered successful. Conversely, if any one exceedes the limit (voltage / power / equipment), the safety verification fails.

[0133] If training at the current level is completed and passes the safety check, training can proceed to the next higher degradation level, k+1.

[0134] In one embodiment, the step of progressively training the dual-path electrical attention network using the safety constraint index and the hierarchical training subset to obtain a weak measurement environment control model may further include the following sub-steps: S1435. If the security check fails, an optimized training subset is constructed based on the re-acquired optimized shielded node set, and the dual-path electrical attention network is trained using the optimized training subset until the security constraint index is met. The optimized shielded node set is a training subset obtained by selectively shielding the measured data after measuring the critical nodes that exceed the limits. The critical nodes that exceed the limits are obtained by calculating the contribution of distribution network nodes to voltage or power exceeding the limits.

[0135] In one embodiment, a physical constraint convergence condition is set. During the training process of each degradation level, in addition to monitoring the convergence of the loss function, the degree to which the control strategy meets physical constraints such as node voltage deviation and branch power over-limit is evaluated on the validation set. Secondly, when the validation results do not meet the grid operation constraints, the measurement shielding list for the current level is optimized in reverse. By analyzing the key nodes that cause the constraint over-limit, the set of shielded nodes is dynamically adjusted to achieve reverse correction of the measurement shielding strategy by physical errors. Then, the training subset is regenerated using the adjusted shielding strategy, and the model is retrained with the current model parameters as the initial values. This process is iterated repeatedly until both the convergence condition and safety constraints are met simultaneously. Finally, a closed loop of "training-physical validation-shielding adjustment-retraining" is formed, achieving synergistic optimization of dual-path electrical attention modulation and physical error weight correction, enabling the model to continuously adapt to the safety control requirements under low observability. The specific implementation process is as follows: Safety constraints are rigid operating boundaries that cannot be breached by the control strategy, specifically including: 1. Node voltage constraints: Strictly adhere to the national standard for power supply voltage deviation, and the node voltage must be maintained within the range of 0.93 to 1.07 times the rated voltage; 2. Branch power constraints: The active power and apparent power of each branch must not exceed the line thermal stability limit and rated transmission capacity; 3. Adjustable resource constraints: The adjustment range of control equipment such as photovoltaic inverters and reactive power compensation devices must be within the rated adjustment range of the equipment.

[0136] The convergence criteria are the stopping of a single-level training iteration and the judgment standard for parameter migration to the next degradation level. Both must be met simultaneously: 1. Loss function convergence: The overall loss function of the training set decreases by less than 1×10⁻ for 20-50 consecutive iterations. 41. ~1×10⁻³, the validation set loss shows no continuous downward trend; 2. Safety constraints are fully met: all samples in the validation set meet the above safety constraints 100% without exceeding the limits; 3. The regulation performance is stable and convergent: the core regulation indicators such as voltage fluctuation rate and photovoltaic absorption rate of the validation set are continuously optimized by multiple rounds of iterations with the magnitude less than the set threshold, achieving the preset engineering performance target.

[0137] Assume the total number of validation set samples is For index is When the node voltage deviation index and branch power over-limit index do not meet the grid operation constraints, the measurement masking list of the current level is optimized in reverse.

[0138] The specific execution process of the reverse optimization mechanism is as follows: First, analyze the key nodes that lead to the constraint exceeding the limit, calculate the contribution of each node to the limit exceeding, and identify the set of nodes related to the limit exceeding. Secondly, dynamically adjust the set of shielded nodes for the current level. Specifically, it can be shown in the following formula (25): (25); In the above formula, In order to be in The set of restored measurement nodes is randomly selected from the nodes in the data, which means reducing its shielding level or completely restoring its measurement.

[0139] Subsequently, the training subset was regenerated using the adjusted masking strategy, and the current model parameters were used. Retraining with initial values ​​can be performed as shown in the following formula (26): (26); In the next training stage, the adjustable parameters in the dual-path electrical attention network and the weight bias parameters of the two layers of the feedforward network are fine-tuned. Then the... k Model parameters involved in fine-tuning at the degradation level It can be shown in the following formula (27): (27); Finally, the attention parameters are iteratively updated using the same method as in formula (24) until both the convergence condition and the safety constraint requirements are met. The calculation method can be shown in the following formula (28): (28).

[0140] After passing the security verification, the model can meet the security requirements of the distribution network and reduce the security risks of the distribution network.

[0141] Reference Figure 2The diagram illustrates an operation flowchart of a power distribution network source-load regulation method for weak measurement environments provided by an embodiment of the present invention.

[0142] Specifically, the operation of the power distribution network source-load regulation method in a weak measurement environment may include the following steps: The first step is to start training and build a node electrical importance system.

[0143] The process involves acquiring node information such as distribution network topology, node admittance matrix, power flow data, and photovoltaic / load operation data; calculating power flow sensitivity indicators, voltage stability indicators, and source-load type correction factors; and obtaining node importance through normalization and weighted fusion. A node electrical importance system is then constructed based on these node importance scores. The hierarchical shielding of node importance closely reflects the actual missing data patterns, and the dual-path attention adaptive focus on key features significantly improves the robustness of weak measurements.

[0144] The second step is to dynamically classify the observation degradation levels and generate graded training subsets.

[0145] Execution steps: Based on node importance and a preset masking threshold, several observation degradation levels (low, medium, and high) are defined (higher levels indicate more severe measurement loss). A complete dataset is constructed using node measurement data, and source-load scenarios are categorized into three types: normal, high penetration, and peak load, based on the photovoltaic penetration rate and load level of the distribution area. Selective data masking (setting masked node measurements to 0) is performed on the complete dataset according to the degradation level, generating graded training subsets corresponding to different degradation levels. Dynamically classifying observation degradation levels adapts to extreme conditions such as high penetration and peak load, enhancing the adaptive capability of source-load scenarios.

[0146] The third step is to predict the measurement gaps and safety constraints.

[0147] Actions performed: Global measurement missing degree and local measurement missing degree are calculated using hierarchical training subsets to obtain measurement missing parameters; node voltage over-limit index and branch power over-limit index are calculated using node measurement data and fused to obtain safety constraint risk prediction index.

[0148] The fourth step is to construct a dual-path electrical attention network.

[0149] Execution actions: Using missing measurement parameters as input, a missing measurement attention branch is built to realize the perception of the degree of missing data; using safety constraint risk indicators as input, an electrical coupling attention branch is built to realize the focusing of high-risk nodes; the two attention branches are merged, and the fused features are input to the strategy output layer to complete the construction of a dual-path electrical attention network.

[0150] The fifth step is progressive, multi-stage training.

[0151] Actions: Train the dual-path electrical attention network by using hierarchical training subsets in order of degradation level from low to high; after the previous level model converges, transfer the parameters to the next level, and only fine-tune the dual-path electrical attention parameters, feedforward network weight bias, and policy output layer parameters; based on the preset training loss, determine whether the model has reached preliminary convergence.

[0152] Step 6: Physical constraint security verification.

[0153] Execution action: Perform rigid safety verification on the pre-trained model that has initially converged. The verification dimensions include: Node voltage constraint (0.93~1.07 times rated voltage); Branch power constraint (not exceeding the line thermal stability limit); Adjustable equipment constraints (adjustment amount within the rated range).

[0154] The model security verification process can distinguish between "safe" and "unsafe" branches.

[0155] Step 7, dual-branch determination: meets safety control requirements.

[0156] Branch 1: Satisfy safety control requirements → Enter convergence condition judgment.

[0157] Execution action: Further determine whether the model simultaneously satisfies the convergence condition: The training set loss decreases by less than a threshold with each consecutive iteration. The validation set loss is stable and without fluctuation. The optimization of the control performance indicators has met the standards.

[0158] The process is determined as follows: Convergence condition met: Directly output the robust control model for weak measurement environment, and the process ends; Convergence condition not met: Return to the progressive multi-stage training module and continue iterative training.

[0159] Branch 2: Does not meet security control requirements → Reverse optimization of shielded nodes.

[0160] Perform the following actions: Analyze the key nodes that cause voltage or power limits to exceed the limits, and calculate the contribution of each node to the limit exceedance. Dynamically adjust the set of masked nodes at the current degradation level, restore key node measurements, and generate an optimized training subset; The model is retrained using the current model parameters as initial values ​​and an optimized training subset.

[0161] Through the above process, a closed loop of "training-validation-optimization-retraining" can be formed until the safety constraints are met. Physical constraint closed-loop optimization can 100% satisfy the safety boundary, ensuring the rigidity of the safety constraints.

[0162] Step 8, process endpoint: Output a safe and robust control model.

[0163] When the model simultaneously meets the safety constraints and convergence conditions, the process ends, and the final trained source-load control model for distribution networks in weak measurement environments is output, which can be directly used for real-time source-load control of distribution networks.

[0164] By implementing physical safety constraints upfront, progressive training, and reverse closed-loop optimization, the technical problems of insufficient robustness and lack of safety guarantee in distribution network control models under weak measurement environments are solved.

[0165] In this embodiment, the present invention provides a method for source-load regulation of a distribution network in a weak measurement environment. Its advantages are as follows: the present invention can acquire the node importance and node measurement data of the distribution network, construct a complete dataset using the node measurement data, and determine the source-load scenario of the distribution network. The complete dataset contains complete measurement information of all observed nodes in the distribution network. Based on the node importance and the distribution network source-load scenario, the complete dataset is divided into several hierarchical training subsets. The hierarchical training subsets are used to calculate missing measurement parameters and to determine safety constraint indicators using node measurement data. Progressive training is performed based on the missing measurement parameters, safety constraint indicators, and hierarchical training subsets to obtain a control model for a weak measurement environment. This weak measurement environment control model is then used for source-load regulation of the distribution network. The progressive multi-stage training of this invention trains the model step by step from low to high according to the level of observation degradation. This not only allows for matching the corresponding source-load scenario dataset to achieve progressive learning from easy to difficult, but also allows the model to gradually adapt to measurement deficiencies, greatly improving the generalization and robustness in weak measurement environments. As a result, the model can be called to perform source-load regulation of the distribution network in weak measurement environments, avoiding the failure of control strategies due to "weak measurement" and improving the accuracy of regulation.

[0166] This invention also provides a power distribution network source-load regulation device for weak measurement environments, see [link to relevant documentation]. Figure 3 The diagram shows a schematic of the structure of a power distribution network source-load control device in a weak measurement environment according to an embodiment of the present invention.

[0167] As an example, the power distribution network source-load control device in the weak measurement environment may include: The acquisition module 201 is used to acquire the node importance and node measurement data of the distribution network respectively, and to construct a complete dataset and determine the source-load scenario of the distribution network using the node measurement data. The complete dataset contains complete measurement information of all observed nodes of the distribution network. The partitioning module 202 is used to partition the complete dataset into several hierarchical training subsets according to the node importance and the power distribution network source-load scenario, wherein each hierarchical training subset corresponds to a degradation level; Calculation module 203 is used to calculate the missing parameters of the measurement using the hierarchical training subset and to determine the safety constraint index using the node measurement data; The control module 204 is used to perform progressive training based on the missing measurement parameters, the safety constraint indicators, and the hierarchical training subset to obtain a weak measurement environment control model, and to call the weak measurement environment control model to perform power distribution network source-load control processing.

[0168] Optionally, obtaining the node importance of the distribution network includes: Obtain node information of the distribution network, wherein the node information includes: the topology of the distribution network, node admittance matrix, power flow data, photovoltaic operation data and load operation data; The node information is used to calculate the electrical key index values ​​of the node, wherein the electrical key index values ​​include: power flow sensitivity index, voltage stability index and source load type correction factor; The importance of a node is obtained by sequentially normalizing and weighting the values ​​of several key electrical indicators.

[0169] Optionally, the step of dividing the complete dataset into several hierarchical training subsets based on the node importance and the power distribution network source-load scenario includes: Based on the node importance and the preset masking threshold, several degradation levels are defined; Based on the magnitude of the degradation level, selective data masking is performed on the complete dataset corresponding to each of the power distribution network source-load scenarios to obtain several hierarchical training subsets with different degradation levels. The selective data masking involves setting the measurement data corresponding to nodes that do not match the degradation level to zero.

[0170] Optionally, the step of calculating the missing measurement parameters using the hierarchical training subset and determining the security constraint index using the node measurement data includes: The global missing value and local missing value are calculated using the hierarchical training subset to obtain the measurement missing parameters; The voltage over-limit index, branch power over-limit index, and comprehensive risk value are calculated using the node measurement data to obtain the safety constraint index.

[0171] Optionally, the step of progressively training based on the missing measurement parameters, the safety constraint indicators, and the hierarchical training subset to obtain a weak measurement environment control model, and then calling the weak measurement environment control model for distribution network source-load control processing, includes: The measurement missing parameters are used to construct the measurement missing attention branch and the electrical coupling attention branch of the model, respectively. By fusing the measurement-deficient attention branch and the electrical coupling attention branch as inputs to the policy output layer of the model, a dual-path electrical attention network is obtained. The dual-path electrical attention network is progressively trained using the safety constraint index and the hierarchical training subset to obtain a weak measurement environment control model. The weak measurement environment control model is invoked to perform source-load control processing of the distribution network based on real-time measurement data of the distribution network.

[0172] Optionally, the step of progressively training the dual-path electrical attention network using the safety constraint index and the hierarchical training subset to obtain a weak measurement environment control model includes: The hierarchical training subsets are sorted according to their degradation levels to obtain the training ranking results; Based on the training ranking results, the dual-path electrical attention network is trained step by step using the hierarchical training subset and the safety constraint index, and the model parameters are adjusted to obtain a pre-trained model. The model parameters include: dual-path electrical attention parameters, feedforward network weight bias, and policy output layer. After the pre-trained model reaches convergence based on the preset training loss, a safety check is performed on the pre-trained model. When the safety check passes, a weak measurement environment control model is obtained.

[0173] Optionally, the step of progressively training the dual-path electrical attention network using the safety constraint index and the hierarchical training subset to obtain a weak measurement environment control model further includes: If the security check fails, an optimized training subset is constructed based on the re-acquired optimized shielded node set, and the dual-path electrical attention network is trained using the optimized training subset until the security constraint index is met. The optimized shielded node set is a training subset obtained by selectively shielding the measured data after measuring the critical nodes that exceed the limits. The critical nodes that exceed the limits are obtained by calculating the contribution of distribution network nodes to voltage or power exceeding the limits.

[0174] Those skilled in the art will understand that, for ease of description and brevity, the specific working process of the device described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0175] Furthermore, this application also provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the power distribution network source-load regulation method for weak measurement environments as described in the above embodiments.

[0176] Furthermore, embodiments of this application also provide a computer-readable storage medium storing a computer-executable program, the computer-executable program being used to cause a computer to execute the power distribution network source-load regulation method in a weak measurement environment as described in the above embodiments.

[0177] In the description of the embodiments of the present invention, it should be noted that the terms "above," "below," etc., indicating the orientation or positional relationship are based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the embodiments of the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. When an element such as a layer, region, or substrate is referred to as being "above" or "on top of" another element, it may be directly on the other element, or there may be an intermediate element. Conversely, when an element is referred to as being "directly on" or "above" another element, there is no intermediate element. It should also be understood that when an element is referred to as being "below" or "under" another element, it may be directly below or under the other element, or there may be an intermediate element. Conversely, when an element is referred to as being "directly below" or "under" another element, there is no intermediate element. Unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to the internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0178] Those skilled in the art will understand that embodiments of this application may also include computer program products. Therefore, this application may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application may take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0179] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), devices, and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0180] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0181] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0182] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for source-load regulation in a distribution network under weak measurement conditions, characterized in that, The method includes: The importance of nodes and node measurement data of the distribution network are obtained respectively. A complete dataset is constructed using the node measurement data and the source-load scenario of the distribution network is determined. The complete dataset contains complete measurement information of all observed nodes of the distribution network. The complete dataset is divided into several hierarchical training subsets based on the node importance and the power distribution network source-load scenario, wherein each hierarchical training subset corresponds to a degradation level; The missing parameters of the measurement are calculated using the hierarchical training subset, and the safety constraint indicators are determined using the node measurement data. Based on the missing measurement parameters, the safety constraint indicators, and the hierarchical training subset, a weak measurement environment control model is obtained through progressive training. The weak measurement environment control model is then used for power distribution network source-load control processing.

2. The power supply and load regulation method for a distribution network in a weakly measurable environment according to claim 1, characterized in that, The process of obtaining the node importance of the distribution network includes: Obtain node information of the distribution network, wherein the node information includes: the topology of the distribution network, node admittance matrix, power flow data, photovoltaic operation data and load operation data; The node information is used to calculate the electrical key index values ​​of the node, wherein the electrical key index values ​​include: power flow sensitivity index, voltage stability index and source load type correction factor; The importance of a node is obtained by sequentially normalizing and weighting the values ​​of several key electrical indicators.

3. The power distribution network source-load regulation method in a weakly measurable environment according to claim 1, characterized in that, The complete dataset is divided into several hierarchical training subsets based on the node importance and the power distribution network source-load scenario, including: Based on the node importance and the preset masking threshold, several degradation levels are defined; Based on the magnitude of the degradation level, selective data masking is performed on the complete dataset corresponding to each of the power distribution network source-load scenarios to obtain several hierarchical training subsets with different degradation levels. The selective data masking involves setting the measurement data corresponding to nodes that do not match the degradation level to zero.

4. The power supply and load regulation method for a distribution network in a weakly measurable environment according to claim 1, characterized in that, The process of calculating missing parameters using the hierarchical training subset and determining security constraint indicators using the node measurement data includes: The global missing value and local missing value are calculated using the hierarchical training subset to obtain the measurement missing parameters; The voltage over-limit index, branch power over-limit index, and comprehensive risk value are calculated using the node measurement data to obtain the safety constraint index.

5. The power distribution network source-load regulation method for weak measurement environments according to any one of claims 1-4, characterized in that, The process of progressively training based on the missing measurement parameters, the safety constraint indicators, and the hierarchical training subset to obtain a weak measurement environment control model, and then using the weak measurement environment control model for power distribution network source-load control processing, includes: The measurement missing parameters are used to construct the measurement missing attention branch and the electrical coupling attention branch of the model, respectively. By fusing the measurement-deficient attention branch and the electrical coupling attention branch as inputs to the policy output layer of the model, a dual-path electrical attention network is obtained. The dual-path electrical attention network is progressively trained using the safety constraint index and the hierarchical training subset to obtain a weak measurement environment control model. The weak measurement environment control model is invoked to perform source-load control processing of the distribution network based on real-time measurement data of the distribution network.

6. The power distribution network source-load regulation method in a weakly measurable environment according to claim 5, characterized in that, The method of progressively training the dual-path electrical attention network using the safety constraint index and the hierarchical training subset to obtain a weak measurement environment control model includes: The hierarchical training subsets are sorted according to their degradation levels to obtain the training ranking results; Based on the training ranking results, the dual-path electrical attention network is trained step by step using the hierarchical training subset and the safety constraint index, and the model parameters are adjusted to obtain a pre-trained model. The model parameters include: dual-path electrical attention parameters, feedforward network weight bias, and policy output layer. After the pre-trained model reaches convergence based on the preset training loss, a safety check is performed on the pre-trained model. When the safety check passes, a weak measurement environment control model is obtained.

7. The power distribution network source-load regulation method in a weakly measurable environment according to claim 6, characterized in that, The step of progressively training the dual-path electrical attention network using the safety constraint index and the hierarchical training subset to obtain a weak measurement environment control model further includes: If the security check fails, an optimized training subset is constructed based on the re-acquired optimized shielded node set, and the dual-path electrical attention network is trained using the optimized training subset until the security constraint index is met. The optimized shielded node set is a training subset obtained by selectively shielding the measured data after measuring the critical nodes that exceed the limits. The critical nodes that exceed the limits are obtained by calculating the contribution of distribution network nodes to voltage or power exceeding the limits.

8. A power distribution network source-load control device for weak measurement environments, characterized in that, The device includes: The acquisition module is used to acquire the node importance and node measurement data of the distribution network respectively, and to construct a complete dataset and determine the source-load scenario of the distribution network using the node measurement data. The complete dataset contains complete measurement information of all observed nodes of the distribution network. The partitioning module is used to divide the complete dataset into several hierarchical training subsets according to the node importance and the power distribution network source-load scenario, wherein each hierarchical training subset corresponds to a degradation level; The calculation module is used to calculate the missing parameters of the measurement using the hierarchical training subset and to determine the safety constraint index using the node measurement data; The control module is used to perform progressive training based on the missing measurement parameters, the safety constraint indicators, and the hierarchical training subset to obtain a weak measurement environment control model, and then call the weak measurement environment control model to perform power distribution network source-load control processing.

9. An electronic device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, when the processor executes the program, it implements the power distribution network source-load regulation method for weak measurement environments as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions for causing a computer to perform the power distribution network source-load control method for weak measurement environments as described in any one of claims 1-7.

Images