A data-driven based power grid voltage reactive power control optimization method and system

Abstract

The present application relates to the technical field of power system grid optimization control, in particular to a data-driven grid voltage and reactive power control optimization method and system, which comprises constructing a voltage and reactive power control mathematical model of a three-phase unbalanced grid, taking the minimum node voltage deviation rate, the lowest grid line loss and the optimal three-phase unbalance degree of bus voltage as the optimization goal; using Latin hypercube sampling to cover the photovoltaic-load typical scene of the whole operation condition, combining with the historical operation data of the grid to generate sample data, training the highway neural network combined with the improved NSGA-II algorithm, and obtaining the globally optimal grid voltage and reactive power control strategy through population iteration solution. The present application adopts a data-driven method, replaces the traditional physical model with a proxy model, effectively adapts the uncertainty of new energy access, shortens the single time period decision time to seconds, stabilizes the node voltage and three-phase unbalance degree within the safety limit, and at the same time realizes the minimization of the whole grid loss.

Description

Technical Field

[0001] This invention relates to the field of power grid optimization control technology, and in particular to a data-driven method and system for optimizing power grid voltage and reactive power control. Background Technology

[0002] Currently, the global energy transition towards low-carbon energy is accelerating, with renewable energy sources such as wind and solar power being massively integrated into the power system, gradually building a new power system dominated by new energy sources. This has become the current state of development in the power industry. Against this backdrop, the randomness, volatility, and intermittency of renewable energy significantly exacerbate the uncertainty and complexity of power system operation, making traditional voltage control methods that rely on physical models of the power system inadequate for current development needs.

[0003] Currently, although existing centralized reactive power control technology can achieve overall coordination and control, it relies on precise system physical models, has a lagging control response, is difficult to quickly cope with the instantaneous fluctuations of renewable energy, and lacks flexibility.

[0004] Whether centralized or distributed control, voltage regulation is mainly achieved by switching capacitor banks. The reactive power regulation is a discrete value. Not only does the switching operation have a large impact on the system and the number of operations is limited, but the voltage regulation accuracy is also difficult to meet the requirements and cannot adapt to the requirements of new power systems for precise and efficient voltage control. Summary of the Invention

[0005] To overcome the problems mentioned above in the prior art, this invention provides a data-driven method for optimizing grid voltage reactive power control, comprising the following steps:

[0006] S1. Construct a VVC mathematical model for a three-phase unbalanced power grid. The model aims to minimize the node voltage deviation rate, minimize the line loss of the entire network, and optimize the three-phase voltage imbalance of the bus. At the same time, power balance constraints and controllable equipment operation constraints are set.

[0007] S2. Based on the VVC mathematical model, Latin hypercube sampling is used to cover typical photovoltaic-load scenarios under all operating conditions. Combined with historical grid operation data, sample input data and corresponding sample label data are generated to construct a sample database.

[0008] S3. Train HWN using the sample database, and obtain the HWN surrogate model after verification by the fitting accuracy index.

[0009] S4. Improved NSGA-II algorithm: Set performance index over-limit degradation rules in the non-dominated sorting stage of the NSGA-II algorithm to obtain the improved NSGA-II algorithm;

[0010] S5. The HWN proxy model is embedded into the iterative optimization process of the improved NSGA-II algorithm. The optimization objective of the VVC mathematical model is taken as the optimization direction. The globally optimal power grid voltage reactive power control strategy is obtained through population iterative solution.

[0011] Preferably, in step S1, the node voltage offset rate The expression is:

[0012] ;

[0013] in, For nodes exist Voltage offset over a period of time To part ways, , , They are phase A, phase B, and phase C, respectively. The number of observable nodes. , They are nodes exist The minimum and maximum voltage values ​​during the time period;

[0014] Total network line loss The expression is:

[0015] ;

[0016] in, These are collections of all network branches. , , They are respectively Time-of-day branch Positive-sequence, negative-sequence, and zero-sequence currents, branch road The resistance;

[0017] Three-phase bus voltage imbalance The expression is;

[0018] ;

[0019] in, For the collection of all network busbars, , They are respectively Time bus Negative-sequence and positive-sequence voltages.

[0020] Preferably, in step S1, the controllable equipment operation constraints include the charging and discharging power and state of charge constraints of the distributed energy storage system (DESS), the tap position and daily adjustment frequency constraints of the on-load tap changer (SVR), the reactive power output and daily switching frequency constraints of the capacitor bank (CBs), and the reactive power output range constraints of the static var compensator (SVC).

[0021] Preferably, in step S2, the input data set of the sample database The expression is:

[0022] ;

[0023] in, For daily equipment adjustment plan, This refers to the daily average load data for measurable nodes.

[0024] Preferably, in step S3, HWN includes conversion gate units and portable gate units, and its output network satisfies:

[0025] ;

[0026] in, This is a nonlinear transformation of a typical feedforward neural network layer. , These are the weight parameters for a regular neural network and a transition gate unit, respectively. , For HWN Layer inputs and outputs, For the output of the conversion gate unit, satisfying , For the output of the portable gate unit;

[0027] The model fitting accuracy is verified by the root mean square error and the correlation coefficient. When the root mean square error is less than or equal to the preset threshold and the correlation coefficient is greater than or equal to the preset threshold, the HWN surrogate model is deemed to have met the fitting accuracy standard.

[0028] Preferably, in step S4, the improvement to the NSGA-II algorithm specifically involves setting a performance index threshold:

[0029] ;

[0030] in, , , These are the node voltage offset threshold, the line loss threshold, and the three-phase unbalance threshold, respectively.

[0031] In the non-dominated sorting stage of the NSGA-II algorithm, if any performance index of any individual in the population exceeds the preset performance index threshold, the non-dominated sorting level of that individual is reduced by two levels, and the non-dominated sorting is completed again.

[0032] Preferably, the method further includes step S6: distributing the grid voltage reactive power control and regulation strategy to the grid controllable equipment for execution, while collecting real-time grid operation data, updating the sample database and HWN proxy model, and using the updated HWN proxy model to execute step S5 to obtain the latest grid voltage reactive power control and regulation strategy.

[0033] This invention also provides a data-driven power grid voltage reactive power control optimization system, comprising:

[0034] The data acquisition and preprocessing module is used to collect historical power grid operation data, node measurement data, and real-time operation data.

[0035] The power flow calculation module is used to construct a VVC mathematical model for a three-phase unbalanced power grid. It uses Latin hypercube sampling to cover typical photovoltaic-load scenarios under all operating conditions. Combined with historical power grid operating data, it generates sample input data and corresponding sample label data to build a sample database.

[0036] The model training and optimization module is used to train HWN using a sample database and obtain the HWN surrogate model after verification by the fitting accuracy index.

[0037] The NSGA-II population initialization module is used to generate the improved NSGA-II algorithm.

[0038] The multi-objective iterative optimization module is used to embed the HWN surrogate model into the iterative optimization process of the improved NSGA-II algorithm, and to perform population iteration with the optimization objective of the VVC mathematical model as the optimization direction.

[0039] The verification and output module is used to distribute the globally optimal grid voltage and reactive power control strategy to the grid controllable equipment for execution.

[0040] The beneficial effects of this invention are as follows:

[0041] This invention uses the High-Speed ​​Neural Network (HWN) to construct a proxy model for evaluating system performance indicators. With its adaptive gating unit's cross-layer information transmission mechanism, it effectively solves the gradient vanishing problem during deep training of traditional deep networks. Under different numbers of hidden layers and different numbers of observable nodes, it can achieve high-precision nonlinear mapping between local measurement information, control strategies and global performance indicators, providing accurate proxy model support for reactive power optimization of non-fully observable power grids.

[0042] This invention makes adaptive improvements to the NSGA-II algorithm. Through an adaptive inferior solution screening mechanism, it effectively solves the extreme solution problem of the traditional algorithm, avoids the defect of excellent optimization of a single index while other indexes exceed the limit, realizes multi-objective collaborative optimization of voltage deviation rate, line loss and three-phase imbalance, and improves the comprehensive control effect of power grid reactive power optimization.

[0043] This invention embeds the HWN surrogate model into the improved NSGA-II algorithm, constructing the HWN-NSGA-II optimization algorithm framework. During the algorithm iteration process, there is no need to perform frequent online power flow calculations. The fitness function value can be quickly calculated using only the surrogate model. In the IEEE 123-node example, the time for reactive power optimization solution within a single day is reduced from the thousands of seconds of the traditional algorithm to the 10-second level, improving the solution efficiency by more than two orders of magnitude, which fully meets the timeliness requirements of real-time power grid control.

[0044] This invention is based on a data-driven technical approach. It does not rely on real-time measurement data of all nodes in the power grid. It can achieve high-precision reactive power optimization decisions only through load information of locally observable nodes. It is effectively adapted to the current engineering status of incomplete distribution network measurement infrastructure and has strong field applicability and promotion value. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of the overall process of an embodiment of the present invention;

[0046] Figure 2 This is a schematic diagram of the structure of the highway neural network HWN according to an embodiment of the present invention;

[0047] Figure 3 This is a schematic diagram of the execution flow of the HWN proxy model embedding the improved NSGA-II algorithm according to an embodiment of the present invention. Detailed Implementation

[0048] Embodiments of the present invention provide a data-driven method for optimizing grid voltage and reactive power control, such as... Figure 1 As shown, it includes the following steps:

[0049] S1. Construct a mathematical model for voltage and reactive power control (VVC) of a three-phase unbalanced power grid. The model aims to minimize the node voltage deviation rate, minimize the line loss of the entire network, and optimize the three-phase voltage imbalance of the bus. At the same time, power balance constraints and controllable equipment operation constraints are set.

[0050] Collect historical power grid operation data, controllable equipment control scheme data, node measurement data, and real-time operation data.

[0051] The test system in this embodiment is an improved IEEE 123-node three-phase power grid. The system reference voltage is 12.66kV, and the peak load is increased to 5.584MW. The load composition is 48% constant power load + 40% constant impedance load + 12% constant current load, where the constant impedance load is set as a three-phase unbalanced load. The controllable equipment of the system includes: 3 sets of CBs with a rated capacity of 0.5Mvar, 3 sets of SVRs with a tap range of 1±5×2%, 3 sets of SVCs with a capacity of [-0.2, 0.2]Mvar, 5 single-phase photovoltaics with a total rated output of 3MW, and 3 sets of distributed energy storage (DESS) with a rated capacity of 300kW / 12000kWh.

[0052] The optimization cycle is 24 hours per day, and the system performance indicators are node voltage deviation rate, total network line loss, and bus voltage three-phase imbalance.

[0053] Node voltage offset The expression is:

[0054] ;

[0055] in, For nodes exist Voltage offset over a period of time To part ways, , , They are phase A, phase B, and phase C, respectively. The number of observable nodes. , They are nodes exist The minimum and maximum voltage values ​​during the time period;

[0056] Voltage allowable range set to ,Right now ;

[0057] Total network line loss The expression is:

[0058] ;

[0059] in, These are collections of all network branches. , , They are respectively Time-of-day branch Positive-sequence, negative-sequence, and zero-sequence currents, branch road The resistance;

[0060] Three-phase bus voltage imbalance The expression is;

[0061] ;

[0062] in, For the collection of all network busbars, , They are respectively Time bus The negative sequence and positive sequence voltages, and the three-phase imbalance are controlled within 2% according to national standards.

[0063] Controllable equipment operation constraints include the charging and discharging power and state of charge constraints of the distributed energy storage system (DESS), the tap position and daily adjustment frequency constraints of the on-load tap changer (SVR), the reactive power output and daily switching frequency constraints of the capacitor bank (CBs), and the reactive power output range constraints of the static var compensator (SVC).

[0064] Power balance constraints satisfy the conservation of active and reactive power at nodes in a three-phase unbalanced power grid, as well as improved branch power flow equation constraints, specifically:

[0065]

[0066] in, For nodes , The impedance of the p-phase branch between them, for Time period nodes The active power of the middle p phase, for Time period nodes Output power when the PV device is connected to phase p. for Time period nodes Power when DESS is fully operational and for Compensation capacity during the switching of CBs and SVC devices in the p phase of the time period. for Time period nodes The power factor of the p-phase. and for Time period nodes The p-phase voltage phasor value in the equation.

[0067] DESS operating constraints include charging and discharging power limited to [-300kW, 300kW], state of charge (SOC) constrained to [0.3, 0.8], and charging and discharging efficiency set to 0.9. Specifically:

[0068]

[0069] in, and They are respectively Time of the first The charge / discharge power and state of charge of each DESS , and , These are the DESS charge / discharge power and state of charge limits, respectively. and The first The charging and discharging power of each DESS.

[0070] SVR operation constraints: tap position range is [1, 10], maximum number of adjustments per day does not exceed 10 times, specifically:

[0071]

[0072] in, and Upper and lower limits for tap position constraints. for Time of the first The tap positions of each SVR. For the first Maximum permissible number of adjustments per day for each SVR tap. This represents the number of SVRs.

[0073] CBs (Board Switching) constraints: A single CB group contains 10 switching groups, and the maximum number of switching operations per day cannot exceed 12. Specifically:

[0074]

[0075] in, for Time of the first The number of input groups for each CBs The reactive power of each capacitor stage. and The upper and lower limits of reactive power output by CBs. This represents the maximum number of switching operations allowed per day.

[0076] SVC operation constraints: reactive power output range is [-0.2Mvar, 0.2Mvar], specifically:

[0077]

[0078] in, and These are the upper and lower limits of the reactive power output of the SVC.

[0079] S2. Construct an offline sample database for the power grid: Latin hypercube sampling is used to cover typical photovoltaic-load scenarios under all operating conditions. Sampling is performed based on the variable range, constraint boundary and feasible region defined by the VVC mathematical model constructed in step S1 to ensure that the generated scenarios are all operating conditions for the safe and stable operation of the power grid. Combined with historical power grid operating data, sample input data and corresponding sample label data are generated to complete the construction of the offline sample database.

[0080] Input data set of offline sample database The expression is:

[0081] ;

[0082] in, The daily adjustment plan for the equipment includes the location of the SVR tap changer, the number of CBs switched on and off, the charging and discharging power of the DESS, and the reactive power output of the SVC. This refers to the daily average load data for measurable nodes.

[0083] For each scenario, a particle swarm optimization algorithm is used for multi-objective reactive power optimization, generating 20,000 sets of daily controllable equipment control schemes. Simultaneously, daily average load information from 34 measurable nodes was extracted. ,by As a sample input data set.

[0084] For the optimization results of each scenario, calculate the corresponding system performance metrics. This serves as a dataset of 20,000 sample labels. The sample database is then divided into a 16,000-set training set and a 4,000-set test set, thus completing the construction of the offline sample database.

[0085] S3. Construct a proxy model for evaluating system performance indicators based on Highway Networks (HWN): Using the offline sample database constructed in step S2 as the basis for training and validation, perform offline training on the HWN. Verify the model's fitting accuracy through root mean square error and correlation coefficient, thus obtaining an HWN proxy model that can replace traditional power flow calculations, such as... Figure 2 As shown.

[0086] The HWN structure includes conversion gate units and portable gate units, and its output network satisfies:

[0087] ;

[0088] in, This is a nonlinear transformation of a typical feedforward neural network layer. , These are the weight parameters for a regular neural network and a transition gate unit, respectively. , For HWN Layer inputs and outputs, For the output of the conversion gate unit, satisfying , This is the output of the portable gate unit.

[0089] In this embodiment, the network has 10 hidden layers, with 64 neurons in each layer. The activation function is the tanh function, and the transition gate unit is the sigmoid function.

[0090] Using the sample input data as the input of HWN and the sample label data as the output of HWN, the initial learning rate is set to 0.0001, the number of iterations is 500, the mean squared error (MSE) is used as the loss function, and L2 regularization is added to prevent overfitting, thus completing the offline training of the model.

[0091] The model fitting accuracy is verified by using root mean square error (RMSE) and correlation coefficient. The HWN surrogate model is considered to have met the accuracy standard if the RMSE is less than or equal to a preset threshold and the correlation coefficient is greater than or equal to a preset threshold. The preset threshold is set according to actual needs; this implementation does not impose any restrictions on it.

[0092] The model fitting accuracy was verified using a test set. Calculations showed that the root mean square error (ERMSE) of the HWN model in this embodiment was 0.0102, and the correlation coefficient (r) was 0.9930. The model's predicted values ​​and the actual values ​​showed a strong correlation, and the fitting accuracy met the requirements for engineering applications. The construction of the proxy model for evaluating system performance indicators was thus completed.

[0093] S4. Adaptive improvements are made to the non-dominated sorting genetic algorithm NSGA-II to obtain the improved NSGA-II algorithm;

[0094] The specific adaptive improvements to the NSGA-II algorithm are as follows:

[0095] To address the problem of extreme inferior solutions in algorithm iteration, a performance metric degradation rule is set in the non-dominated sorting stage to improve the algorithm.

[0096] The performance metric degradation rule is as follows: Set the performance metric threshold:

[0097] ;

[0098] in, , , These are the node voltage offset threshold, the line loss threshold, and the three-phase unbalance threshold, respectively.

[0099] This implementation pre-sets a safety threshold for performance indicators: voltage offset rate threshold. Line loss threshold Three-phase imbalance threshold .

[0100] In the non-dominated sorting stage of the NSGA-II algorithm, if any performance index of any individual in the population exceeds the preset performance index threshold, the non-dominated sorting level of that individual is reduced by two levels, and then the non-dominated sorting, crowding calculation and elite strategy selection are re-completed to obtain the improved NSGA-II algorithm, so as to reduce the impact of inferior solutions on population evolution.

[0101] S5. Construct the HWN-NSGA-II optimization algorithm framework, such as Figure 3 As shown, the HWN proxy model obtained in step S3 is embedded into the iterative optimization process of the improved NSGA-II algorithm obtained in step S4. The optimization objective of the VVC mathematical model constructed in step S1 is taken as the optimization direction, and the globally optimal grid voltage reactive power control strategy is obtained through population iterative solution.

[0102] The improved NSGA-II algorithm is set to a population size of 500 and a maximum number of iterations of 50. When calculating the fitness of individual individuals, online power flow calculation is not required. Instead, real-time acquisition of observable node load information is input into the pre-trained surrogate model. System performance indicators serve as individual fitness values, supporting the rapid iterative optimization of the NSGA-II algorithm. It executes iterative steps such as non-dominated sorting, crossover mutation, and elite retention until the maximum number of iterations is reached, with a single-period decision-making time ≤ 11 seconds.

[0103] S6. The grid voltage reactive power control and regulation strategy obtained in step S5 is sent to the grid controllable equipment for execution. At the same time, the hourly load information of the 34 observable nodes of the test system and the real-time photovoltaic output data are collected in real time. The offline sample database in step S2 and the HWN proxy model in step S3 are updated to realize the closed-loop optimization of grid voltage reactive power control.

[0104] The globally optimal grid voltage and reactive power control strategy includes SVR tap changer adjustment plan, CBs switching strategy, SVC reactive power output plan, and DESS charging and discharging strategy for each time period of the day within 24 hours.

[0105] Embodiments of the present invention also provide a data-driven power grid voltage reactive power control optimization system, comprising:

[0106] The data acquisition and preprocessing module is used to collect historical power grid operation data, controllable equipment control scheme data, node measurement data and real-time operation data, and complete data filtering, timestamp alignment, outlier removal and normalization preprocessing, providing basic data support for offline sample database construction and online optimization solution respectively.

[0107] The power flow calculation module, connected to the data acquisition and preprocessing module, is used to perform calculations on each photovoltaic-load scenario in step S2 and output three types of system performance indicators under the corresponding scenario: node voltage offset rate, total network line loss, and bus voltage three-phase imbalance, which serve as sample labels for training the HWN model.

[0108] The model training and optimization module is connected to the power flow calculation module and the data acquisition and preprocessing module. It is used to complete the offline training, accuracy verification and iterative update of HWN based on the offline sample database, and output an HWN proxy model that can replace power flow calculation.

[0109] The NSGA-II population initialization module, connected to the data acquisition and preprocessing module, is used to generate an improved NSGA-II algorithm initial population based on real-time power grid operation data, power grid power balance constraints, and controllable equipment operation constraints.

[0110] The multi-objective iterative optimization module is connected to the model training and optimization module and the NSGA-II population initialization module. It is used to call the HWN proxy model to complete the calculation of individual fitness in the population, and execute the non-dominated sorting, crowding calculation, crossover mutation and elite strategy iteration of the improved NSGA-II algorithm to gradually approach the global optimum.

[0111] The verification and output module, connected to the multi-objective iterative optimization module, is used to verify the compliance of the optimal solution set output by the iteration, select the optimal grid voltage reactive power control and regulation strategy according to engineering requirements, and send it to the grid controllable equipment for execution.

[0112] The method proposed in this embodiment was compared with the full-information NSGA-II algorithm, the MOPSO algorithm, the traditional GA algorithm, and the ACO algorithm. The results are as follows:

[0113] The method of this invention achieves an observable node voltage deviation rate of 1.5692%, a line loss of 112.3802kW, and a three-phase imbalance of 1.0475%. The optimization effect is basically consistent with the full-information NSGA-II algorithm and significantly better than traditional algorithms such as MOPSO, GA, and ACO. After optimization, the peak node voltage deviation of the system decreased by 68.70% year-on-year, the average daily line loss decreased by 39.34% year-on-year, and the three-phase imbalance of all buses was controlled within the national standard limit of 2%.

[0114] The optimization solution time of the method of this invention is only 10.6750s per day, while the full information NSGA-II algorithm takes 1115.3956s, which improves the solution efficiency by more than 100 times and fully meets the timeliness requirements of real-time power grid control.

[0115] This invention has been described through embodiments. Those skilled in the art will understand that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of this invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, this invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of this invention.

Claims

1. A data-driven optimization method for power grid voltage and reactive power control, characterized in that, Includes the following steps: S1. Construct a VVC mathematical model for a three-phase unbalanced power grid. The model aims to minimize the node voltage deviation rate, minimize the line loss of the entire network, and optimize the three-phase voltage imbalance of the bus. At the same time, power balance constraints and controllable equipment operation constraints are set. S2. Based on the VVC mathematical model, Latin hypercube sampling is used to cover typical photovoltaic-load scenarios under all operating conditions. Combined with historical grid operation data, sample input data and corresponding sample label data are generated to construct a sample database. S3. Train HWN using the sample database, and obtain the HWN surrogate model after verification by the fitting accuracy index. S4. Improved NSGA-II algorithm: Set performance index over-limit degradation rules in the non-dominated sorting stage of the NSGA-II algorithm to obtain the improved NSGA-II algorithm; S5. The HWN proxy model is embedded into the iterative optimization process of the improved NSGA-II algorithm. The optimization objective of the VVC mathematical model is taken as the optimization direction. The globally optimal power grid voltage reactive power control strategy is obtained through population iterative solution.

2. The data-driven power grid voltage reactive power control optimization method according to claim 1, characterized in that, In step S1, the node voltage offset rate The expression is: ； in, For nodes exist Voltage offset over a period of time To part ways, , , They are phase A, phase B, and phase C, respectively. The number of observable nodes. , They are nodes exist The minimum and maximum voltage values ​​during the time period; The total network line loss The expression is: ； in, These are collections of all network branches. , , They are respectively Time-of-day branch Positive-sequence, negative-sequence, and zero-sequence currents, branch road The resistance; The three-phase imbalance of the bus voltage The expression is; ； in, For the collection of all network busbars, , They are respectively Time bus Negative-sequence and positive-sequence voltages.

3. The data-driven power grid voltage reactive power control optimization method according to claim 1, characterized in that, In step S1, the controllable equipment operation constraints include the charging and discharging power and state of charge constraints of the distributed energy storage system (DESS), the tap position and daily adjustment number constraints of the on-load tap changer (SVR), the reactive power output and daily switching number constraints of the capacitor bank (CBs), and the reactive power output range constraints of the static var compensator (SVC).

4. The data-driven power grid voltage reactive power control optimization method according to claim 1, characterized in that, In step S2, the input data set of the sample database The expression is: ； in, For daily equipment adjustment plan, This refers to the daily average load data for measurable nodes.

5. The data-driven power grid voltage reactive power control optimization method according to claim 1, characterized in that, In step S3, HWN includes conversion gate units and portable gate units, and its output network satisfies: ； in, This is a nonlinear transformation of a typical feedforward neural network layer. , These are the weight parameters for a regular neural network and a transition gate unit, respectively. , For HWN Layer inputs and outputs, For the output of the conversion gate unit, satisfying , For the output of the portable gate unit; The model fitting accuracy is verified by the root mean square error and the correlation coefficient. When the root mean square error is less than or equal to the preset threshold and the correlation coefficient is greater than or equal to the preset threshold, the HWN surrogate model is deemed to have met the fitting accuracy standard.

6. The data-driven power grid voltage reactive power control optimization method according to claim 1, characterized in that, In step S4, the NSGA-II algorithm is improved by setting a performance metric threshold. ； in, , , These are the node voltage offset threshold, the line loss threshold, and the three-phase unbalance threshold, respectively. In the non-dominated sorting stage of the NSGA-II algorithm, if any performance index of any individual in the population exceeds the preset performance index threshold, the non-dominated sorting level of that individual is reduced by two levels, and the non-dominated sorting is completed again.

7. The data-driven power grid voltage reactive power control optimization method according to claim 1, characterized in that, It also includes step S6, which involves distributing the grid voltage and reactive power control strategy to the grid controllable equipment for execution, while simultaneously collecting real-time grid operation data, updating the sample database and the HWN proxy model, and using the updated HWN proxy model to execute step S5 to obtain the latest grid voltage and reactive power control strategy.

8. A data-driven power grid voltage reactive power control optimization system, characterized in that, include: The data acquisition and preprocessing module is used to collect historical power grid operation data, node measurement data, and real-time operation data. The power flow calculation module is used to construct a VVC mathematical model for a three-phase unbalanced power grid. It uses Latin hypercube sampling to cover typical photovoltaic-load scenarios under all operating conditions. Combined with historical power grid operating data, it generates sample input data and corresponding sample label data to build a sample database. The model training and optimization module is used to train HWN using a sample database and obtain the HWN surrogate model after verification by the fitting accuracy index. The NSGA-II population initialization module is used to generate the improved NSGA-II algorithm. The multi-objective iterative optimization module is used to embed the HWN surrogate model into the iterative optimization process of the improved NSGA-II algorithm, and to perform population iteration with the optimization objective of the VVC mathematical model as the optimization direction. The verification and output module is used to distribute the globally optimal grid voltage and reactive power control strategy to the grid controllable equipment for execution.

Images