Closed-loop monitoring method and system for whole-process scene chain of urban power distribution network fused with digital twinning

Abstract

The invention relates to the technical field of intelligent power grids, in particular to an urban power distribution network whole-process scene chain closed-loop monitoring method and system fusing digital twinning. The method comprises the following steps of: fusing a weighted evidence theory and a dynamic time bending algorithm to acquire multi-source data, and constructing a physical mechanism-data driven hybrid enhanced digital twinborn model; based on the model, abnormity is researched and judged through an improved isolated forest algorithm, grading alarms are generated, simulation deduction is conducted through Monte Carlo-Latin hypercube sampling, an optimal strategy is screened through a non-dominated sorting genetic algorithm III with uncertainty, and instructions are executed after graph neural network topology verification. Feedback is formed by the collection result, and the strategy is optimized through meta-learning and Bayesian optimization calibration model. The system comprises six cooperation units, an intelligent algorithm and a digital twinning technology are integrated, the power distribution network abnormity identification precision, decision scientificity and adaptive capacity are improved, and safe and efficient operation is guaranteed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of smart grid, in particular to a city power distribution network whole-process scene chain closed-loop monitoring method and system fusing digital twinning. BACKGROUND

[0002] Under the promotion of new power system construction and energy transformation, the city power distribution network has formed a complex form of "source-grid-load-storage" multi-element cooperation. In order to cope with the challenges of large-scale access of distributed new energy, intensified load fluctuation and frequent extreme weather, the monitoring of power distribution network is upgrading towards digitization and intelligence, and technologies such as digital twinning, multi-source data fusion and artificial intelligence have become the core support. At present, the industry has built a power grid visualization platform through a digital twinning model, integrated multi-source information such as real-time sensor data, historical operation data and weather data, and carried out fault detection, state assessment and simulation deduction combined with PCA and machine learning algorithms. Some technical applications have shortened the fault troubleshooting time, improved the efficiency of basic operation and maintenance, and provided preliminary protection for the safe operation of power grid.

[0003] Multi-source data has problems such as heterogeneous format, timestamp mismatch and noise interference, and existing fusion methods cannot achieve accurate timing alignment and conflict resolution, and the data support accuracy is insufficient. Digital twinning models mostly use single architecture of pure physical mechanism or pure data-driven, lack dynamic error correction mechanism, and cannot realize high-precision mapping of virtual and real systems. The sensitivity of abnormal detection to small faults is low, the misdiagnosis rate is high, the risk assessment relies on fixed index weight, and does not combine dynamic factors such as equipment health status and real-time load, and the rationality of alarm grading is poor. Simulation deduction uses a single sampling method and does not fully consider uncertain factors, the robustness of optimization strategy is insufficient, and there is no effective topology verification before instruction generation, which is easy to cause potential power grid risks. There is no complete closed-loop mechanism, model parameters and strategies cannot be updated adaptively according to actual operation data, and the accuracy may decay after long-term operation. SUMMARY

[0004] In order to completely solve the problems of inaccurate multi-source data fusion, mapping deviation of twinning model, and insufficient abnormal detection and risk assessment in the prior art, the present application proposes a city power distribution network whole-process scene chain closed-loop monitoring method and system fusing digital twinning, which realizes accurate monitoring through multi-technology cooperation and closed-loop mechanism. The specific technical solutions are as follows:

[0005] The city power distribution network whole-process scene chain closed-loop monitoring method fusing digital twinning comprises the following steps:

[0006] S1: Multi-source data acquisition and fusion is realized by using weighted evidence theory combined with dynamic time warping algorithm, and a physical mechanism-data driven hybrid enhanced digital twinning model is constructed;

[0007] S2: Based on the twin model, the isolated forest algorithm is used to identify anomalies and a dynamic risk matrix is ​​embedded to generate hierarchical alarms;

[0008] S3: Based on alarm information, a Monte Carlo combined with Latin hypercube sampling method is used for simulation and deduction, and the optimal strategy is selected by non-dominated sorting genetic algorithm III with uncertainty.

[0009] S4: Analyze the optimal strategy, generate execution instructions after topological verification based on graph neural network, and execute them.

[0010] S5: Collect the execution results after the command is implemented, compare and analyze them with the preset target, and generate a feedback report;

[0011] S6: Using the data in the feedback report, calibrate the twin model and optimize the strategy using a method that combines meta-learning with Bayesian optimization.

[0012] Preferably, in step S1, to achieve accurate collection and fusion of multi-source data, a data fusion mechanism is constructed using weighted evidence theory combined with a dynamic time-warping algorithm. The relevant expressions are as follows:

[0013] ;

[0014] ;

[0015] In the formula, The basic probability allocation after the fusion of weighted evidence theories, where To define the equipment status, For the first Basic probability allocation of each data source; The conflict coefficient, Dynamic weights are assigned to the data source and adjusted based on historical data accuracy and real-time stability. The two time series data calculated by the dynamic time warp algorithm are in the first... , Alignment distance at any moment The Mahalanobis distance, This is the data covariance matrix.

[0016] Preferably, in step S1, when constructing the physical mechanism-data-driven hybrid enhanced digital twin model, the residual compensation results of the physical model output and the data-driven model are fused through an attention gating mechanism to achieve dynamic feature selection and error correction. The relevant expressions are as follows:

[0017] ;

[0018] ;

[0019] In the formula, This is the final output of the twin model; The physical model output is obtained through the distribution network node voltage equations. calculate, Inject current vectors into nodes. Here is the nodal admittance matrix. The node voltage vector; The residuals output by the Bi-LSTM data-driven model. For the input vector, These are the model weight parameters; For element-wise multiplication, It is the Sigmoid activation function. and These are the attention layer weight matrix and bias vector, respectively; This is the weight vector output by the attention gating mechanism.

[0020] Preferably, in step S2, to improve the accuracy of anomaly assessment and enhance the sensitivity of identifying minor anomalies in high-risk equipment, the isolated forest algorithm is improved by introducing equipment health and load fluctuation coefficients to weight and correct the basic anomaly score. The specific calculation process is as follows: first, calculate the basic anomaly score of the sample; then, calculate the equipment health and load fluctuation coefficients separately; finally, fuse them using an exponential function to obtain the corrected anomaly score. The relevant expression is as follows:

[0021] ;

[0022] ;

[0023] ;

[0024] ;

[0025] In the formula, The base anomaly scores of the samples calculated by the Isolation Forest algorithm, where The sample to be tested. The total number of samples, For the sample exist The expected average path length in a 10 isolated trees For the sample size Average path length at that time; For equipment health, For the equipment The aging rate of each component The operating time of this component. This represents the total number of equipment components. This is the load fluctuation coefficient. For load power standard deviation For load power The mean; The final outlier score after correction. and As a preset impact factor, It is an exponential function.

[0026] Preferably, in step S2, during the process of generating hierarchical alarms by embedding a dynamic risk matrix, a comprehensive risk value is calculated by integrating three core indicators: fault impact range, recovery resource cost, and user importance. The indicator weights are then dynamically adjusted based on the real-time operating status of the power grid, and alarm levels are classified. The relevant expressions are as follows:

[0027] ;

[0028] ;

[0029] ;

[0030] ;

[0031] In the formula, This represents the overall risk value of the fault. The dynamic weights of the three types of indicators satisfy... It updates adaptively based on the power grid load status; This is a normalized indicator of the scope of the fault's impact. The fault affects the load. This represents the largest load impact in history. Duration of the fault The system mean recovery time constant; To restore the normalized indicator of resource costs, The resource costs required for fault recovery This represents the highest historical cost in terms of resource recovery. A weighted metric for user importance. For the first The importance weight of user class The proportion of this type of user affected by the fault. This represents the total number of user categories.

[0032] Preferably, in step S3, to ensure the comprehensiveness and compliance of the simulation, a Monte Carlo combined with Latin hypercube sampling method is used for simulation. Latin hypercube sampling improves the coverage of fault parameters. The simulation process strictly follows the N-1 safety constraint and voltage and current amplitude constraints. Finally, the probability of the system's critical state is calculated through kernel density estimation. The relevant expressions are as follows:

[0033] ;

[0034] ;

[0035] ;

[0036] ;

[0037] In the formula, For nodes voltage amplitude, , These are the lower and upper limits of the node voltage, respectively. Voltage amplitude constraint condition; For the line The current amplitude, For the line Maximum allowable current, For current amplitude constraints; For the system to enter a critical state The probability, Here is the kernel density estimation function; This represents the kernel density estimate of the system state index. For the number of samples, For Gaussian kernel function, For bandwidth parameters, For the first The system state indices obtained from the simulation.

[0038] Preferably, in step S3, when selecting the optimal strategy using the non-dominated sorting genetic algorithm III with uncertainty, the core optimization objectives are expected power supply reliability, expected overall cost, and network loss variance. The robust Pareto optimal solution set is solved using the sample-average approximation method, and the relevant expressions are as follows:

[0039] ;

[0040] ;

[0041] In the formula, For multi-objective optimization function vectors, For policy variables, Uncertain parameters include load fluctuations and renewable energy output. For the feasible region of the uncertainty parameter;

[0042] For the desired power supply reliability index, This is the index of the system's average power outage duration. The expected total cost includes emergency repair costs and network loss costs. For network loss variance; These are the system operation constraints. This represents the total number of constraints.

[0043] Preferably, in step S4, during the generation of execution instructions, a topology verification algorithm based on graph neural networks is incorporated. The distribution network topology is modeled as a graph structure, and the core features of nodes and edges are learned through the graph neural network. The algorithm outputs a topology validity score and an operational risk assessment. The execution instruction can only be output after the verification is passed. The relevant expressions are as follows:

[0044] ;

[0045] ;

[0046] In the formula, For GNN Layer nodes Feature representation, This is a diagram of the power distribution network topology. For a set of nodes, Let be the set of edges. For nodes The set of neighboring nodes; , GNN number The weight matrix and bias vector of the layer, For activation function, This represents the total number of layers in the GNN. This is the output vector for topology verification. For topological validity score, For operational risk values, For a multilayer perceptron, the verification must meet the following requirements. and , , This is a preset threshold.

[0047] Preferably, in step S6, when calibrating the Siamese model and optimizing the strategy using the data in the feedback report and employing a method combining meta-learning and Bayesian optimization, the Siamese model parameters are quickly updated through meta-learning to adapt to data distribution drift, and the strategy parameters are efficiently adjusted and computational overhead is reduced through Bayesian optimization. The relevant expressions are as follows:

[0048] ;

[0049] ;

[0050] In the formula, The parameters of the Siamese model are updated after meta-learning. These are the model parameters before the update. The meta-learning rate, The meta-loss function is used to measure the model's performance on feedback data. Generalization ability; To optimize the selection of the next round of policy parameters for evaluation using Bayesian methods, As a performance indicator of the strategy, This refers to the observed strategy performance data; This represents the expected improvement value based on the Gaussian process model.

[0051] The urban power distribution network full-process scenario closed-loop monitoring system integrating digital twins includes a multi-source data acquisition and fusion unit, a digital twin modeling and simulation unit, an anomaly analysis and risk assessment unit, an optimization decision and instruction generation unit, an instruction execution and feedback unit, and a learning and adaptive optimization unit.

[0052] The multi-source data acquisition and fusion unit is used to acquire multi-source heterogeneous data from the power distribution network and perform fusion processing to provide high-quality data input for the digital twin model.

[0053] The aforementioned digital twin modeling and simulation unit is used to construct a physical mechanism-data driven hybrid enhanced digital twin model, supporting the simulation of power distribution network operation status and scenario extrapolation;

[0054] The aforementioned anomaly analysis and risk assessment unit is used to monitor the operation status of the power distribution network in real time based on the twin model, accurately identify anomalies, and dynamically assess the risk level.

[0055] The optimization decision-making and instruction generation unit is used to formulate the optimal control strategy and generate execution instructions after safety verification in response to abnormal events or preset targets, based on simulation results.

[0056] The instruction execution and feedback unit is used to issue execution instructions to the actual power distribution network equipment or control system, and to provide real-time feedback on the instruction execution status and results;

[0057] The learning and adaptive optimization unit is used to continuously calibrate the twin model parameters using feedback results and long-term operating data, and adaptively adjust the anomaly judgment rules, risk assessment model and optimization decision algorithm.

[0058] The beneficial effects of this invention are as follows:

[0059] 1. This invention optimizes the multi-source data fusion mechanism, laying a solid data foundation for power distribution network monitoring. It employs a synergistic effect of weighted evidence theory and dynamic time-warping algorithm to effectively resolve spatiotemporal conflicts and inconsistencies in multi-source heterogeneous data. By dynamically adjusting the weights of data sources, it balances historical accuracy and real-time stability, outputting high-quality fused data.

[0060] 2. This invention constructs a hybrid enhanced digital twin model to ensure simulation accuracy and adaptability. It integrates a physical mechanism model and a data-driven model, leveraging an attention gating mechanism to achieve complementary advantages between the two models. Simultaneously, through meta-learning combined with a Bayesian optimization calibration method, it quickly adapts to data distribution drift and maintains high simulation accuracy over the long term.

[0061] 3. This invention improves the anomaly detection algorithm, enhancing the sensitivity of identifying minor anomalies. Based on the isolated forest algorithm, it introduces equipment health and load fluctuation coefficients to weight and correct the basic anomaly score, significantly enhancing the ability to identify minor anomalies in high-risk equipment and effectively reducing the risk of missed or false positives.

[0062] 4. This invention establishes a dynamic risk assessment model to achieve precise hierarchical alarms. It constructs a dynamic risk matrix that integrates the scope of fault impact, recovery resource costs, and user importance. The indicator weights are adaptively updated according to the real-time operating status of the power grid, generating hierarchical alarm results that meet actual needs and providing accurate basis for subsequent handling.

[0063] 5. This invention employs compliant simulation and multi-objective optimization strategies to ensure the scientific nature of decision-making and the safety of execution. Monte Carlo simulation combined with Latin hypercube sampling achieves comprehensive coverage of fault parameters under safety constraints such as N-1. The NSGA-Ⅲ algorithm, which incorporates uncertainties, takes into account multiple objectives, and graph neural network topology verification further mitigates operational risks.

[0064] 6. This invention constructs a closed-loop mechanism for the entire process scenario chain, enhancing the system's adaptive iteration capability. The complete closed loop of "monitoring-judgment-decision-execution-feedback-optimization" covers all stages. The learning and adaptive optimization unit integrates feedback data, synchronously optimizes model parameters, judgment rules, and decision algorithms, and dynamically adapts to complex working conditions.

[0065] 7. This invention has significant practical application value, contributing to the intelligent upgrading of power distribution networks. It effectively improves power supply reliability, shortens fault duration, reduces operation and maintenance and network loss costs, optimizes operational efficiency, and provides strong technical support for the lean and intelligent operation of urban power distribution networks. Attached Figure Description

[0066] Figure 1 This is a schematic diagram illustrating the steps of the closed-loop monitoring method for the entire process of urban power distribution network that integrates digital twins, as described in this invention. Detailed Implementation

[0067] A closed-loop monitoring method for the entire process of urban power distribution networks integrating digital twins includes the following steps:

[0068] S1: We adopt weighted evidence theory combined with dynamic time warp algorithm to realize multi-source data acquisition and fusion, and construct a physical mechanism-data driven hybrid enhanced digital twin model;

[0069] S2: Based on the twin model, the isolated forest algorithm is used to identify anomalies and a dynamic risk matrix is ​​embedded to generate hierarchical alarms;

[0070] S3: Based on alarm information, a Monte Carlo combined with Latin hypercube sampling method is used for simulation and deduction, and the optimal strategy is selected by non-dominated sorting genetic algorithm III with uncertainty.

[0071] S4: Analyze the optimal strategy, generate execution instructions after topological verification based on graph neural network, and execute them.

[0072] S5: Collect the execution results after the command is implemented, compare and analyze them with the preset target, and generate a feedback report;

[0073] S6: Using the data in the feedback report, calibrate the twin model and optimize the strategy using a method that combines meta-learning with Bayesian optimization.

[0074] In step S1, to achieve accurate collection and fusion of multi-source data, a data fusion mechanism is constructed using weighted evidence theory combined with a dynamic time warp algorithm. The relevant expressions are as follows:

[0075] ;

[0076] ;

[0077] In the formula, The basic probability allocation after the fusion of weighted evidence theories, where To define the equipment status, For the first Basic probability allocation of each data source; The conflict coefficient, Dynamic weights are assigned to the data source and adjusted based on historical data accuracy and real-time stability. The two time series data calculated by the dynamic time warp algorithm are in the first... , Alignment distance at any moment The Mahalanobis distance, This is the data covariance matrix.

[0078] In step S1, when constructing the physical mechanism-data-driven hybrid enhanced digital twin model, the residual compensation results of the physical model output and the data-driven model are fused through an attention gating mechanism to achieve dynamic feature selection and error correction. The relevant expressions are as follows:

[0079] ;

[0080] ;

[0081] In the formula, This is the final output of the twin model; The physical model output is obtained through the distribution network node voltage equations. calculate, Inject current vectors into nodes. Here is the nodal admittance matrix. The node voltage vector; The residuals output by the Bi-LSTM data-driven model. For the input vector, These are the model weight parameters; For element-wise multiplication, It is the Sigmoid activation function. and These are the attention layer weight matrix and bias vector, respectively; This is the weight vector output by the attention gating mechanism.

[0082] In step S2, to improve the accuracy of anomaly assessment and enhance the sensitivity of identifying minor anomalies in high-risk equipment, the Isolation Forest algorithm is improved by introducing equipment health and load fluctuation coefficients to weight and correct the basic anomaly score. The specific calculation process is as follows: first, calculate the basic anomaly score of the sample; then, calculate the equipment health and load fluctuation coefficients separately; finally, fuse them using an exponential function to obtain the corrected anomaly score. The relevant expressions are as follows:

[0083] ;

[0084] ;

[0085] ;

[0086] ;

[0087] In the formula, The base anomaly scores of the samples calculated by the Isolation Forest algorithm, where The sample to be tested. The total number of samples, For the sample exist The expected average path length in a 10 isolated trees For the sample size Average path length at that time; For equipment health, For the equipment The aging rate of each component The operating time of this component. This represents the total number of equipment components. This is the load fluctuation coefficient. For load power standard deviation For load power The mean; The final outlier score after correction. and As a preset impact factor, It is an exponential function.

[0088] In step S2, during the process of generating hierarchical alarms by embedding a dynamic risk matrix, a comprehensive risk value is calculated by integrating three core indicators: fault impact range, recovery resource cost, and user importance. The indicator weights are then dynamically adjusted based on the real-time operating status of the power grid, and alarm levels are classified. The relevant expressions are as follows:

[0089] ;

[0090] ;

[0091] ;

[0092] ;

[0093] In the formula, This represents the overall risk value of the fault. The dynamic weights of the three types of indicators satisfy... It updates adaptively based on the power grid load status; This is a normalized indicator of the scope of the fault's impact. The fault affects the load. This represents the largest load impact in history. Duration of the fault The system mean recovery time constant; To restore the normalized indicator of resource costs, The resource costs required for fault recovery This represents the highest historical cost in terms of resource recovery. A weighted metric for user importance. For the first The importance weight of user class The proportion of this type of user affected by the fault. This represents the total number of user categories.

[0094] In step S3, to ensure the comprehensiveness and compliance of the simulation, a Monte Carlo combined with Latin hypercube sampling method is used for simulation. Latin hypercube sampling improves the coverage of fault parameters. The simulation process strictly follows the N-1 safety constraints and voltage and current amplitude constraints. Finally, the probability of the system's critical state is calculated through kernel density estimation. The relevant expressions are as follows:

[0095] ;

[0096] ;

[0097] ;

[0098] ;

[0099] In the formula, For nodes voltage amplitude, , These are the lower and upper limits of the node voltage, respectively. Voltage amplitude constraint condition; For the line The current amplitude, For the line Maximum allowable current, For current amplitude constraints; For the system to enter a critical state The probability, Here is the kernel density estimation function; This represents the kernel density estimate of the system state index. For the number of samples, For Gaussian kernel function, For bandwidth parameters, For the first The system state indices obtained from the simulation.

[0100] In step S3, when selecting the optimal strategy using the non-dominated sorting genetic algorithm III with uncertainty, the core optimization objectives are expected power supply reliability, expected overall cost, and network loss variance. The robust Pareto optimal solution set is solved using the sample-average approximation method, and the relevant expressions are as follows:

[0101] ;

[0102] ;

[0103] In the formula, For multi-objective optimization function vectors, For policy variables, Uncertain parameters include load fluctuations and renewable energy output. For the feasible region of the uncertainty parameter;

[0104] For the desired power supply reliability index, This is the index of the system's average power outage duration. The expected total cost includes emergency repair costs and network loss costs. For network loss variance; These are the system operation constraints. This represents the total number of constraints.

[0105] In step S4, during the generation of execution instructions, a topology verification algorithm based on graph neural networks is incorporated. The distribution network topology is modeled as a graph structure, and the core features of nodes and edges are learned through the graph neural network. The algorithm outputs a topology validity score and an operational risk assessment. Only after the verification is passed can the execution instruction be output. The relevant expressions are as follows:

[0106] ;

[0107] ;

[0108] In the formula, For GNN Layer nodes Feature representation, This is a diagram of the power distribution network topology. For a set of nodes, Let be the set of edges. For nodes The set of neighboring nodes; , GNN number The weight matrix and bias vector of the layer, For activation function, This represents the total number of layers in the GNN. This is the output vector for topology verification. For topological validity score, For operational risk values, For a multilayer perceptron, the verification must meet the following requirements. and , , This is a preset threshold.

[0109] In step S6, when calibrating the Siamese model and optimizing the strategy using the data from the feedback report and employing a method combining meta-learning and Bayesian optimization, the Siamese model parameters are quickly updated through meta-learning to adapt to data distribution drift, and the strategy parameters are efficiently adjusted and computational overhead is reduced through Bayesian optimization. The relevant expressions are as follows:

[0110] ;

[0111] ;

[0112] In the formula, The parameters of the Siamese model are updated after meta-learning. These are the model parameters before the update. The meta-learning rate, The meta-loss function is used to measure the model's performance on feedback data. Generalization ability; To optimize the selection of the next round of policy parameters for evaluation using Bayesian methods, As a performance indicator of the strategy, This refers to the observed strategy performance data; This represents the expected improvement value based on the Gaussian process model.

[0113] The urban power distribution network full-process scenario chain closed-loop monitoring system integrating digital twins can realize the above-mentioned urban power distribution network full-process scenario chain closed-loop monitoring method integrating digital twins, including multi-source data acquisition and fusion unit, digital twin modeling and simulation unit, anomaly analysis and risk assessment unit, optimization decision and instruction generation unit, instruction execution and feedback unit, and learning and adaptive optimization unit.

[0114] The multi-source data acquisition and fusion unit is used to acquire multi-source heterogeneous data from the distribution network and perform fusion processing to provide high-quality data input for the digital twin model;

[0115] The digital twin modeling and simulation unit is used to construct a hybrid enhanced digital twin model driven by physical mechanisms and data, supporting the simulation of power distribution network operation status and scenario extrapolation.

[0116] The anomaly detection and risk assessment unit is used to monitor the operation status of the distribution network in real time based on the twin model, accurately identify anomalies and dynamically assess the risk level;

[0117] The optimization decision and instruction generation unit is used to formulate the optimal control strategy and generate execution instructions after safety verification in response to abnormal events or preset targets, based on simulation results.

[0118] The instruction execution and feedback unit is used to issue execution instructions to the actual power distribution network equipment or control system, and to provide real-time feedback on the instruction execution status and results;

[0119] The learning and adaptive optimization unit is used to continuously calibrate the twin model parameters using feedback results and long-term operating data, and adaptively adjust the anomaly judgment rules, risk assessment models and optimization decision algorithms.

[0120] The embodiments of the present invention are as follows:

[0121] To verify the effectiveness of the closed-loop monitoring method and system for the entire process of urban power distribution networks integrating digital twins, a 10kV power distribution network was selected as the application object. This power distribution network includes 120 nodes, 85 lines, 32 transformers, and 20 distributed photovoltaic access points. The implementation was carried out throughout the entire process of data acquisition, model building, anomaly analysis, strategy optimization, command execution, and model calibration. The specific content is as follows:

[0122] In terms of system configuration, the multi-source data acquisition and fusion unit deploys 300 smart meters, 85 line fault indicators, 32 transformer condition monitoring devices, and 20 photovoltaic inverter data acquisition modules. The transformer condition monitoring devices can monitor oil temperature, winding temperature, partial discharge, and other conditions. This unit supports the IEC61850 protocol, with a sampling frequency set to 1Hz, and is also equipped with 5 edge gateways for data preprocessing. The digital twin modeling and simulation unit uses two GPU servers, each equipped with NVIDIA A100 graphics cards and 128GB of memory, and builds models based on the Python TensorFlow framework. The anomaly analysis and risk assessment unit relies on an industrial server equipped with an Intel Xeon 8375C processor and 64GB of memory, integrating an improved isolated forest algorithm and a dynamic risk matrix module, with a response time controlled within 200ms. The optimization decision-making and instruction generation unit uses a 4-node distributed computing cluster, running a non-dominated sorting genetic algorithm III with uncertainty and a graph neural network topology verification module. The instruction execution and feedback unit interfaces with the distribution network automation system and SCADA system, with remote control instruction feedback latency controlled within 500ms. The learning and adaptive optimization unit relies on elastic cloud servers to deploy meta-learning and Bayesian optimization modules.

[0123] In terms of software support, CentOS 7.9 operating system is used, and PostgreSQL 14 database is used to store historical running data, model parameters and feedback results. Vue and ECharts are combined to build a visualization platform, which can realize the visualization of twin models, monitoring of running status and alarm display, providing a stable software environment support for the entire monitoring process.

[0124] The implementation of the method first enters the S1 phase, which involves multi-source data acquisition and fusion, and twin model construction. The acquired data includes six categories: node voltage, line current, transformer load, photovoltaic output, equipment operating temperature, and ambient humidity. The acquisition cycle is 1 Hz, and continuous acquisition lasts for 72 hours, ultimately forming approximately... Sample data was collected. During data fusion, dynamic weights were assigned based on the accuracy of the data source, with voltage data weighted at 0.35, current data at 0.3, temperature data at 0.2, and photovoltaic output data at 0.15. The alignment distance of the time-series data was calculated using a dynamic time-bending algorithm, with the alignment error controlled within 0.02. When constructing the twin model, the physical model was based on a 120×120 dimensional nodal admittance matrix, using nodal voltage equations. The calculation output; the data-driven model uses a Bi-LSTM network with a hidden layer dimension of 256 and an iteration count of 100 to output the residual; then the physical model output and the residual compensation result are fused through an attention gating mechanism. The initial weights of this mechanism are set to 64×128 dimensions and the bias is 64 dimensions. The error of the final constructed Siamese model is controlled within 3%.

[0125] Phase S2 focuses on anomaly analysis and tiered alarms, using a 10kV transformer as the monitoring object and selecting 10,000 operational data points as a sample. First, the isolated forest algorithm is used to calculate the basic anomaly score of the samples. Then, the equipment health is calculated based on the parameters of the equipment components, where the core aging rate is 0.002 / year, the winding rate is 0.0015 / year, and the bushing rate is 0.001 / year, and all three have been operating for 5 years. Based on this, the equipment health is calculated to be 0.012. Simultaneously, the load fluctuation coefficient is calculated based on the load power data. The standard deviation of the load power is 120kW, and the mean is 800kW, resulting in a load fluctuation coefficient of 0.15. An influencing factor is defined. 0.8 The base anomaly score is 0.5. By fusing the base anomaly score, equipment health, and load fluctuation coefficient using an exponential function, a corrected anomaly score of 0.78 is obtained. The preset anomaly judgment threshold is 0.6, based on which the transformer is determined to be abnormal. Subsequently, a dynamic risk matrix is ​​embedded to generate tiered alarms. The fault-affected load is 15MW, the historical maximum fault-affected load is 30MW, the fault duration is 10min, and the system average recovery time constant is 30min. Based on these, a normalized index of the fault impact range is calculated. The value is 0.48; the resource cost required for fault recovery is 80,000 yuan, and the historical maximum recovery resource cost is 200,000 yuan. The normalized index of recovery resource cost is calculated as follows: The value is 0.4; the weighted index of user importance is calculated based on the user's importance weight and the proportion of users affected. The dynamic weight allocation for the three indicators is 0.4, 0.3, and 0.3, respectively, resulting in a final comprehensive risk value of 0.47, which is classified as a Level 2 alarm.

[0126] In Phase S3, simulation and optimal strategy selection were conducted using a Monte Carlo combined with Latin hypercube sampling method. The number of samples was set to 5000, and the simulation process strictly adhered to N-1 safety constraints while simultaneously satisfying voltage and current constraints. The voltage constraint ranged from 0.95 to 1.05 pu, and the current constraint was limited to no more than 1.1 times the rated value. The critical state probability of the system was calculated using kernel density estimation with a bandwidth of 0.03, ultimately yielding a critical state probability of 0.02. Based on this, the core optimization objectives were expected power supply reliability, expected comprehensive cost, and network loss variance. The expected power supply reliability index (SAIDI) needed to be controlled within 0.5 hours / household, the expected comprehensive cost within 500,000 yuan / year, and the network loss variance within 0.05. A robust Pareto optimal solution set was obtained using a non-dominated sorting genetic algorithm III with uncertainties and a sampled average approximation method, ultimately selecting the optimal strategy of "outgoing line switch reconfiguration + energy storage charging and discharging regulation."

[0127] In the S4 phase, topology verification and command generation are performed. The distribution network topology is modeled as a graph structure containing 120 nodes and 85 edges. A three-layer graph neural network is used for topology verification. Each layer of the graph neural network has a 64×64-dimensional weight matrix and a 64-dimensional bias vector, with ReLU as the activation function. The topology validity score and operational risk value are calculated and output. The topology validity score is 0.92, with a preset topology validity threshold of 0.8; the operational risk value is 0.15, with a preset operational risk threshold of 0.2. Both indicators meet the verification conditions. Subsequently, the optimal strategy is analyzed to generate switch action timing commands and energy storage power adjustment commands. The switch action timing commands include a 0.5s opening time for outgoing switch 1 and a 1s closing time for outgoing switch 2. The energy storage power adjustment command is a charge / discharge power of ±500kW. After generation, the commands are sent to the field equipment for execution.

[0128] In Phase S5, results are collected and a feedback report is generated. This involves collecting operational data within one hour of the command execution. The average node voltage is 1.02 pu, while the preset target range for node voltage is 1.0 to 1.05 pu. The average line current is 0.8 times the rated value, meeting the constraint of not exceeding 1.1 times the rated value. Network loss is reduced by 8% compared to before the command execution. These operational data are compared and analyzed with the preset targets to generate a feedback report, which includes information on anomaly handling effectiveness and changes in operational indicators.

[0129] Phase S6 completes model calibration and policy optimization. Using data from the feedback report, the Siamese model parameters are updated through meta-learning. The meta-learning rate is set to 0.001, and gradients are calculated based on the meta-loss function, which measures the model's performance on the feedback data. The generalization ability is demonstrated by obtaining updated model parameters through gradient descent. After calibration, the twin model error was reduced to 2.2%. Simultaneously, Bayesian optimization was used to screen the parameters for the next round of policy evaluation. Bayesian optimization uses policy performance metrics With the goal of [target], based on observed policy performance data By calculating the expected improvement value Optimal parameters were determined to reduce computational overhead. After the optimized strategy was implemented, the SAIDI index was further improved, the overall cost was effectively controlled, and both the stability and economy of the distribution network operation were improved.

Claims

1. A closed-loop monitoring method for the entire process of urban power distribution network integrating digital twins, characterized in that: Includes the following steps: S1: We adopt weighted evidence theory combined with dynamic time warp algorithm to realize multi-source data acquisition and fusion, and construct a physical mechanism-data driven hybrid enhanced digital twin model; S2: Based on the twin model, the isolated forest algorithm is used to identify anomalies and a dynamic risk matrix is ​​embedded to generate hierarchical alarms; S3: Based on alarm information, a Monte Carlo combined with Latin hypercube sampling method is used for simulation and deduction, and the optimal strategy is selected by non-dominated sorting genetic algorithm III with uncertainty. S4: Analyze the optimal strategy, generate execution instructions after topological verification based on graph neural network, and execute them. S5: Collect the execution results after the command is implemented, compare and analyze them with the preset target, and generate a feedback report; S6: Using the data in the feedback report, calibrate the twin model and optimize the strategy using a method that combines meta-learning with Bayesian optimization.

2. The method for closed-loop monitoring of the entire process of urban power distribution network integrating digital twins as described in claim 1, characterized in that, In step S1, to achieve accurate collection and fusion of multi-source data, a data fusion mechanism is constructed using weighted evidence theory combined with a dynamic time warp algorithm. The relevant expressions are as follows: ； ； In the formula, The basic probability allocation after the fusion of weighted evidence theories, where To define the equipment status, For the first Basic probability allocation of each data source; The conflict coefficient, Dynamic weights are assigned to the data source and adjusted based on historical data accuracy and real-time stability. The two time series data calculated by the dynamic time warp algorithm are in the first... , Alignment distance at any moment The Mahalanobis distance, This is the data covariance matrix.

3. The method for closed-loop monitoring of the entire process of urban power distribution network integrating digital twins as described in claim 1, characterized in that, In step S1, when constructing the physical mechanism-data-driven hybrid enhanced digital twin model, the residual compensation results of the physical model output and the data-driven model are fused through an attention gating mechanism to achieve dynamic feature selection and error correction. The relevant expressions are as follows: ； ； In the formula, This is the final output of the twin model; The physical model output is obtained through the distribution network node voltage equations. calculate, Inject current vectors into nodes. Here is the nodal admittance matrix. The node voltage vector; The residuals output by the Bi-LSTM data-driven model. For the input vector, These are the model weight parameters; For element-wise multiplication, It is the Sigmoid activation function. and These are the attention layer weight matrix and bias vector, respectively; This is the weight vector output by the attention gating mechanism.

4. The method for closed-loop monitoring of the entire process of urban power distribution network integrating digital twins as described in claim 1, characterized in that, In step S2, to improve the accuracy of anomaly assessment and enhance the sensitivity of identifying minor anomalies in high-risk equipment, the Isolation Forest algorithm is improved by introducing equipment health and load fluctuation coefficients to weight and correct the basic anomaly score. The specific calculation process is as follows: first, calculate the basic anomaly score of the sample; then, calculate the equipment health and load fluctuation coefficients separately; finally, fuse them using an exponential function to obtain the corrected anomaly score. The relevant expressions are as follows: ； ； ； ； In the formula, The base anomaly scores of the samples calculated by the Isolation Forest algorithm, where The sample to be tested. The total number of samples, For the sample exist The expected average path length in a 10 isolated trees For the sample size Average path length at that time; For equipment health, For the equipment The aging rate of each component The operating time of this component. This represents the total number of equipment components. This is the load fluctuation coefficient. For load power standard deviation For load power The mean; The final outlier score after correction. and As a preset impact factor, It is an exponential function.

5. The method for closed-loop monitoring of the entire process of urban power distribution network integrating digital twins as described in claim 1, characterized in that, In step S2, during the process of generating hierarchical alarms by embedding a dynamic risk matrix, a comprehensive risk value is calculated by integrating three core indicators: fault impact range, recovery resource cost, and user importance. The indicator weights are then dynamically adjusted based on the real-time operating status of the power grid, and alarm levels are classified. The relevant expressions are as follows: ； ； ； ； In the formula, This represents the overall risk value of the fault. The dynamic weights of the three types of indicators satisfy... It updates adaptively based on the power grid load status; This is a normalized indicator of the scope of the fault's impact. The fault affects the load. This represents the largest load impact in history. Duration of the fault The system mean recovery time constant; To restore the normalized indicator of resource costs, The resource costs required for fault recovery This represents the highest historical cost in terms of resource recovery. A weighted metric for user importance. For the first The importance weight of user class The proportion of this type of user affected by the fault. This represents the total number of user categories.

6. The method for closed-loop monitoring of the entire process of urban power distribution network integrating digital twins as described in claim 1, characterized in that, In step S3, to ensure the comprehensiveness and compliance of the simulation, a Monte Carlo combined with Latin hypercube sampling method is used for simulation. Latin hypercube sampling improves the coverage of fault parameters. The simulation process strictly follows the N-1 safety constraints and voltage and current amplitude constraints. Finally, the probability of the system's critical state is calculated through kernel density estimation. The relevant expressions are as follows: ； ； ； ； In the formula, For nodes voltage amplitude, , These are the lower and upper limits of the node voltage, respectively. Voltage amplitude constraint condition; For the line The current amplitude, For the line Maximum allowable current, For current amplitude constraints; For the system to enter a critical state The probability, Here is the kernel density estimation function; This represents the kernel density estimate of the system state index. For the number of samples, For Gaussian kernel function, For bandwidth parameters, For the first The system state indices obtained from the simulation.

7. The method for closed-loop monitoring of the entire process of urban power distribution network integrating digital twins as described in claim 1, characterized in that, In step S3, when selecting the optimal strategy using the non-dominated sorting genetic algorithm III with uncertainty, the core optimization objectives are expected power supply reliability, expected overall cost, and network loss variance. The robust Pareto optimal solution set is solved using the sample-average approximation method, and the relevant expressions are as follows: ； ； In the formula, For multi-objective optimization function vectors, For policy variables, Uncertain parameters include load fluctuations and renewable energy output. For the feasible region of the uncertainty parameter; For the desired power supply reliability index, This is the index of the system's average power outage duration. The expected total cost includes emergency repair costs and network loss costs. For network loss variance; These are the system operation constraints. This represents the total number of constraints.

8. The method for closed-loop monitoring of the entire process of urban power distribution network integrating digital twins as described in claim 1, characterized in that, In step S4, during the generation of execution instructions, a topology verification algorithm based on graph neural networks is incorporated. The distribution network topology is modeled as a graph structure, and the core features of nodes and edges are learned through the graph neural network. The algorithm outputs a topology validity score and an operational risk assessment. Only after the verification is passed can the execution instruction be output. The relevant expressions are as follows: ； ； In the formula, For GNN Layer nodes Feature representation, This is a diagram of the power distribution network topology. For a set of nodes, Let be the set of edges. For nodes The set of neighboring nodes; , GNN number The weight matrix and bias vector of the layer, For activation function, This represents the total number of layers in the GNN. This is the output vector for topology verification. For topological validity score, For operational risk values, For a multilayer perceptron, the verification must meet the following requirements. and , , This is a preset threshold.

9. The method for closed-loop monitoring of the entire process of urban power distribution network integrating digital twins as described in claim 1, characterized in that, In step S6, when calibrating the Siamese model and optimizing the strategy using the data from the feedback report and employing a method combining meta-learning and Bayesian optimization, the Siamese model parameters are quickly updated through meta-learning to adapt to data distribution drift, and the strategy parameters are efficiently adjusted and computational overhead is reduced through Bayesian optimization. The relevant expressions are as follows: ； ； In the formula, The parameters of the Siamese model are updated after meta-learning. These are the model parameters before the update. The meta-learning rate, The meta-loss function is used to measure the model's performance on feedback data. Generalization ability; To optimize the selection of the next round of policy parameters for evaluation using Bayesian methods, As a performance indicator of the strategy, This refers to the observed strategy performance data; This represents the expected improvement value based on the Gaussian process model.

10. A closed-loop monitoring system for the entire process of urban power distribution network integrating digital twins, characterized in that: It includes a multi-source data acquisition and fusion unit, a digital twin modeling and simulation unit, an anomaly detection and risk assessment unit, an optimization decision-making and instruction generation unit, an instruction execution and feedback unit, and a learning and adaptive optimization unit; The multi-source data acquisition and fusion unit is used to acquire multi-source heterogeneous data from the power distribution network and perform fusion processing to provide high-quality data input for the digital twin model. The aforementioned digital twin modeling and simulation unit is used to construct a physical mechanism-data driven hybrid enhanced digital twin model, supporting the simulation of power distribution network operation status and scenario extrapolation; The aforementioned anomaly analysis and risk assessment unit is used to monitor the operation status of the power distribution network in real time based on the twin model, accurately identify anomalies, and dynamically assess the risk level. The optimization decision-making and instruction generation unit is used to formulate the optimal control strategy and generate execution instructions after safety verification in response to abnormal events or preset targets, based on simulation results. The instruction execution and feedback unit is used to issue execution instructions to the actual power distribution network equipment or control system, and to provide real-time feedback on the instruction execution status and results; The learning and adaptive optimization unit is used to continuously calibrate the twin model parameters using feedback results and long-term operating data, and adaptively adjust the anomaly judgment rules, risk assessment model and optimization decision algorithm.

Images