Transient overvoltage suppression optimization dispatching method and system for high-proportion new energy power grid

Abstract

The application relates to the technical field of high-proportion new energy power grids, and provides a transient overvoltage suppression optimization scheduling method and system for a high-proportion new energy power grid, which comprises the following steps: acquiring new energy power generation prediction data in a day-ahead stage and a day-in stage, obtaining an optimization scheduling scheme through a day-ahead-day-in two-stage optimization scheduling model, wherein the day-ahead-day-in two-stage optimization scheduling model comprises the following steps: in the day-ahead stage, generating an optimal scheduling scheme and a group of standby scheduling schemes by minimizing operation cost and maximizing minimum neighborhood distance; adopting a transient overvoltage evaluation model to evaluate the optimal scheduling scheme, if the evaluation value exceeds a transient overvoltage safety constraint, triggering day-in rescheduling; in the day-in stage, taking the group of standby scheduling schemes as an initial solution set, generating a day-in updated optimization scheduling scheme by minimizing operation cost and control measure adjustment amount. The calculation efficiency meets the online application requirement.

Description

Technical Field

[0001] This invention belongs to the field of high-proportion renewable energy power grid technology, and particularly relates to a transient overvoltage suppression and optimized scheduling method and system for high-proportion renewable energy power grids. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] High-proportion renewable energy (RCA) power grids contain a large number of CIG (converter-connected) power sources, including wind turbines (WT), photovoltaic (PV), and energy storage systems (ESS). In the absence of sufficient synchronous generator support, commutation failure (CF) in the thyristor-commutated converter-HVDC (LCC-HVDC) link can trigger transient overvoltages at the grid connection points (PCC) of various renewable energy power plants. Due to the insufficient overvoltage tolerance of renewable energy generation equipment, transient overvoltages exceeding the safety threshold may trigger protection actions, even causing large-scale grid disconnection of renewable energy generation equipment and interruption of power supply to load centers. Therefore, accurately assessing transient overvoltages under given fault conditions and effectively suppressing overvoltages through optimized operation is crucial to ensuring the safe operation of high-proportion RCA power grids.

[0004] Existing transient overvoltage assessment methods are mainly divided into two categories: mechanism-driven and data-driven. Mechanism-driven methods rely on detailed time-domain simulation models, resulting in high computational costs and difficulty in online applications. With the deployment of Wide Area Measurement Systems (WAMS) and the development of artificial intelligence technology, data-driven methods have shown great potential in online assessment. Convolutional Neural Networks (CNNs), Graph Neural Networks (GNNs), and Transformer-based models have been used for voltage stability assessment. The accuracy of transient overvoltage assessment models is highly dependent on the quality and distribution of training data. Scheduling decisions are typically concentrated near the transient overvoltage safety boundary, and assessment errors near this boundary dominate decision reliability. Therefore, samples near the safety boundary are crucial for scheduling guidance.

[0005] Traditional sample set generation mainly relies on typical operating modes or uniform sampling. These methods require extensive simulations but can only obtain limited samples near the safety boundary. With the increasing penetration rate of new energy power generation equipment, the operating mode space has significantly expanded, leading to higher time-domain simulation annotation costs and even scarcer boundary samples. Directional walk methods can approximate the safety boundary using stability sensitivity, but overvoltages caused by commutation failures lack an analytical gradient relative to the operating mode, making sensitivity-based exploration strategies unsuitable. Boundary-aware data augmentation, by enriching boundary samples through interpolation, can improve model performance and has proven effective in image recognition and industrial fault diagnosis. However, power system optimal scheduling involves strict operational constraints, and the feasible region is highly nonlinear. Interpolation-based data augmentation cannot explicitly satisfy these constraints, thus generating infeasible samples and unreliable labels.

[0006] Prediction errors in renewable energy generation can render day-ahead dispatch schemes infeasible during intraday operation, necessitating rapid intraday rescheduling. Solution methods for Safety-Constrained Economic Dispatch (SCED) models include mathematical programming, heuristic algorithms, and artificial intelligence methods. Mathematical programming offers strong interpretability and high computational efficiency, but it is prone to non-convexity issues under non-convex constraints; furthermore, it struggles to handle black-box constraints based on data-driven models. Artificial intelligence methods respond rapidly, but their black-box nature weakens their reliability under multiple physical constraints. Physical Information Neural Networks (PINNs) and related technologies attempt to introduce physical constraints, but their current capabilities remain limited. Heuristic algorithms can handle non-convex constraints arising from both data-driven evaluation models and complex physical constraints, but they typically only provide a near-optimal solution within a finite timeframe, rather than a globally optimal one. Therefore, improving the computational performance of heuristic algorithms is crucial.

[0007] In summary, in high-proportion renewable energy power grids, commutation failure in thyristor-commutated high-voltage direct current (LCC-HVDC) links can trigger transient overvoltages and even large-scale disconnection of renewable energy from the grid. Therefore, accurate transient overvoltage assessment and online safety-constrained optimized scheduling are urgently needed. However, the uncertainty of renewable energy generation expands the power system's operational mode space and increases the computational burden of training data generation and online scheduling. Summary of the Invention

[0008] To address the technical problems mentioned above, this invention provides a transient overvoltage suppression and optimization scheduling method and system for high-proportion renewable energy power grids. It establishes a two-stage safety constraint optimization scheduling model for day-ahead and intraday periods to identify and suppress transient overvoltages caused by commutation failures. Furthermore, a diverse set of day-ahead backup schemes can provide hot-start for intraday online rescheduling, thereby reducing changes in control measures during the intraday period. The computational efficiency meets the requirements for online applications.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: The first aspect of the present invention provides a transient overvoltage suppression and optimized scheduling method for a high-proportion renewable energy power grid, comprising: Acquire power grid topology information, equipment parameters, and historical operating data to establish a time-domain simulation model of the power grid; Based on the time-domain simulation model and historical operating data of the power grid, a sample of the power grid's operating modes is generated, and the transient overvoltage assessment model is optimized using the boundary-aware active data augmentation method. The system acquires new energy power generation forecast data for the day-ahead and intraday phases, and obtains an optimized scheduling scheme through a two-stage day-ahead-intraday optimization scheduling model. The two-stage day-ahead-intraday optimization scheduling model includes: in the day-ahead phase, generating an optimal scheduling scheme and a set of backup scheduling schemes by minimizing operating costs and maximizing the minimum neighborhood distance; evaluating the optimal scheduling scheme using a transient overvoltage assessment model, and triggering intraday rescheduling if the assessed value exceeds the transient overvoltage safety constraint; and in the intraday phase, using the set of backup scheduling schemes as the initial solution set, generating an updated intraday optimized scheduling scheme by minimizing operating costs and control measure adjustments.

[0010] Furthermore, in the day-ahead phase, a two-stage NSGA-III algorithm is used to solve the problem: the first stage solves for the optimal scheduling scheme with the minimum operating cost; the second stage generates a set of backup scheduling schemes by minimizing the operating cost and maximizing the minimum neighborhood distance around the optimal scheduling scheme.

[0011] Furthermore, during the intraday phase, the NSGA-III algorithm with a hierarchical sorting mechanism is used for solving the problem. The hierarchical sorting mechanism is as follows: for operating modes that violate transient overvoltage safety constraints, the merits of the operating mode are evaluated based on the minimum sum of squares of energy storage output; for operating modes that do not violate transient overvoltage safety constraints, the merits of the operating mode are evaluated according to auxiliary criteria.

[0012] Furthermore, the steps of the boundary-aware active data augmentation method include: Step 301: Based on power grid topology information, equipment parameters and historical operating data, generate a training set and train the transient overvoltage assessment model; Step 302: Based on the transient overvoltage assessment model, use the two-stage NSGA-III algorithm to generate the operating mode located near the transient overvoltage safety boundary; Step 303: Calculate the transient overvoltage of the generated operating mode using the time-domain simulation model, and include it as an enhanced sample in the training set to retrain the transient overvoltage evaluation model; Step 304: Repeat steps 302 and 303. If the evaluation error of the transient overvoltage assessment model for the enhanced sample is less than the threshold, then the transient overvoltage assessment model is the final transient overvoltage assessment model.

[0013] Furthermore, when generating operating modes near the transient overvoltage safety boundary, the two-stage NSGA-III algorithm includes: the first stage adopts a single-objective optimization model, which generates operating modes located near the transient overvoltage safety boundary by minimizing the distance from the operating mode to the decision boundary, and the generated set of operating modes serves as the initial solution set; based on the initial solution set, the second stage adopts a multi-objective optimization model, which maximizes the minimum neighborhood distance between samples by minimizing the distance from the operating mode to the decision boundary, and diffuses the operating modes.

[0014] Furthermore, when generating the operating mode located near the transient overvoltage safety boundary, the objective function of the second stage is: ; Where w represents the distance between the running point vector X and the decision boundary; This represents the transient overvoltage assessment model in the k-th iteration; u0 is the decision boundary; Operational Constraints are the operational constraints; For X i The normalized form of X; i F represents the operating mode of the i-th power system; spread (i) is The minimum neighborhood Euclidean distance; q is the slack variable; μ is the approximation factor.

[0015] Furthermore, the approximation factor is: ; among them, tanh( ) is the hyperbolic tangent function; n is the current iteration number of NSGA-III; N is the maximum iteration number; c is a constant that controls the growth rate of μ.

[0016] A second aspect of the present invention provides a transient overvoltage suppression and optimized dispatching system for a high-proportion renewable energy power grid, comprising: The time-domain simulation model calling module is configured to: acquire power grid topology information, equipment parameters and historical operating data, and establish a time-domain simulation model of the power grid; The deep learning offline training module is configured to: generate power grid operation mode samples based on the power grid time-domain simulation model and historical operating data, and optimize the transient overvoltage assessment model using the boundary-aware active data augmentation method. The optimized scheduling module is configured to: acquire new energy power generation forecast data for the day-ahead and intraday periods, and obtain an optimized scheduling scheme through a two-stage day-ahead-intraday optimized scheduling model; wherein, the two-stage day-ahead-intraday optimized scheduling model includes: in the day-ahead period, generating an optimal scheduling scheme and a set of backup scheduling schemes by minimizing operating costs and maximizing the minimum neighborhood distance; evaluating the optimal scheduling scheme using a transient overvoltage assessment model, and triggering intraday rescheduling if the assessment value exceeds the transient overvoltage safety constraint; in the intraday period, using the set of backup scheduling schemes as the initial solution set, generating an intraday updated optimized scheduling scheme by minimizing operating costs and control measure adjustments.

[0017] Furthermore, the optimized scheduling module includes a day-ahead optimized scheduling module, which is configured to: in the day-ahead stage, use a two-stage NSGA-III algorithm to solve the problem: the first stage solves for the optimal scheduling scheme with the minimum operating cost; the second stage generates a set of backup scheduling schemes by minimizing the operating cost and maximizing the minimum neighborhood distance around the optimal scheduling scheme.

[0018] Furthermore, the optimized scheduling module includes an intraday optimized scheduling module, which is configured to: during the intraday phase, use the NSGA-III algorithm with an added hierarchical sorting mechanism for solving the problem. The hierarchical sorting mechanism is as follows: for operating modes that violate transient overvoltage safety constraints, the merits of the operating mode are evaluated based on the minimum sum of squares of energy storage output; for operating modes that do not violate transient overvoltage safety constraints, the merits of the operating mode are evaluated according to auxiliary criteria.

[0019] Compared with the prior art, the beneficial effects of the present invention are: This invention establishes a two-stage safety constraint optimization scheduling model for day-ahead and intraday periods to identify and suppress transient overvoltages caused by commutation failures. Furthermore, a diverse set of day-ahead backup schemes can provide a hot start for intraday online rescheduling, thereby reducing changes in control measures during the intraday period. The computational efficiency meets the requirements for online applications.

[0020] This invention constructs an active data augmentation framework for training transient overvoltage assessment models, which improves the assessment accuracy of the model near the safety boundary of transient overvoltage, thereby providing guidance for scheduling.

[0021] This invention derives the relationship between the operating state and transient overvoltage of a grid-type energy storage system based on a simplified model, and embeds this relationship into a metaheuristic search through a hierarchical sorting mechanism to improve convergence efficiency and enhance computational performance. Attached Figure Description

[0022] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0023] Figure 1 This is a boundary-aware active data enhancement framework diagram according to Embodiment 1 of the present invention; Figure 2 This is a boundary exploration strategy diagram based on NSGA-III according to Embodiment 1 of the present invention; Figure 3 This is a schematic diagram illustrating the relationship between μ and n / N in Embodiment 1 of the present invention; Figure 4 This is a simplified model diagram of a high-proportion renewable energy power grid according to Embodiment 1 of the present invention; Figure 5 This is a two-stage optimized scheduling framework diagram of Embodiment 1 of the present invention; Figure 6 This is a test system topology diagram according to Embodiment 1 of the present invention; Figure 7 This is an overvoltage distribution diagram of the sample set in Embodiment 1 of the present invention; Figure 8 This is a visualization of the generated sample set distribution according to Embodiment 1 of the present invention; Figure 9 These are the PCC voltage response diagrams after CF in three scenarios according to Embodiment 1 of the present invention; Figure 10 This is a diagram showing the proportion of scheduling schemes that have been verified as feasible through time-domain simulation in Embodiment 1 of the present invention. Figure 11 This is a comparison chart of feasibility and operating cost with and without rule-based guidance in Embodiment 1 of the present invention; Figure 12 This is a comparison diagram of the sample set distribution during the initialization of the backup scheme set with and without prior date in Embodiment 1 of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0025] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0026] Example 1 This embodiment provides a transient overvoltage suppression and optimized scheduling method for high-proportion renewable energy power grids.

[0027] In recent years, researchers have improved the performance of heuristic optimization through mechanism embedding and hot-start initialization. For example, on the one hand, rule-guided heuristics have improved the computational performance of large-scale safety-constrained unit portfolios (SCUC); the Alternating Direction Algorithm (ADA) has been used to solve transmission network expansion planning problems considering multi-year characteristics. On the other hand, support vector machine classifiers are used to predict unit start-up and shutdown states, and the prediction results are used as hot starts for mixed-integer unit portfolios; similarly, predicted optimal points are used to hot start the optimal power flow (OPF) solver in real-time inverter control; and hot-start initialization is introduced into a distributed scheduler based on the alternating direction multiplier method. These studies show that mechanism embedding and hot-start initialization for heuristic algorithms help improve solution quality. How to embed transient overvoltage suppression mechanisms and hot-start calculations into heuristic scheduling to improve computational performance still requires further research.

[0028] This embodiment proposes a transient overvoltage suppression optimization scheduling method based on boundary-aware active data enhancement, which is used to accurately assess and suppress transient overvoltages at the grid connection points of power plants in a high-proportion renewable energy power grid. The main contributions are as follows: First, an active data augmentation framework is constructed for training a transient overvoltage assessment model to improve the model's assessment accuracy near the transient overvoltage safety boundary, thereby providing guidance for scheduling. A boundary exploration strategy based on a two-stage third-generation non-dominated genetic algorithm is proposed to improve the feasibility of augmented samples and the boundary coverage. Secondly, the relationship between the operating state and transient overvoltage of the grid-type energy storage system is derived based on a simplified model, and this relationship is embedded into the metaheuristic search through a hierarchical sorting mechanism to improve convergence efficiency and enhance computational performance. Finally, a two-stage safety constraint optimization scheduling model is established for day-ahead and intraday periods to identify and suppress transient overvoltages caused by commutation failures. The diverse day-ahead backup scheme set can provide hot start for intraday online rescheduling, thereby reducing the adjustment amount of intraday stage control measures.

[0029] Based on a real-world example of a high-proportion renewable energy power grid, the effectiveness of the proposed method in suppressing transient overvoltages was verified.

[0030] This embodiment proposes a transient overvoltage suppression optimization scheduling method based on boundary-aware active data augmentation, including: acquiring grid topology information, equipment parameters, and historical operating data; using a two-stage third-generation non-dominated genetic algorithm and the acquired data, employing a boundary-aware active data augmentation method to directionally identify and calibrate the decision safety boundary of the deep learning transient overvoltage assessment model, improving prediction accuracy near the transient overvoltage safety boundary; establishing a day-ahead and intraday two-stage optimization scheduling model considering transient overvoltage safety constraints based on the deep learning transient overvoltage assessment model; solving the day-ahead optimization scheduling model based on the predicted data of base load, wind power, and photovoltaic output in the day-ahead stage, obtaining the day-ahead optimization scheduling scheme and a set of backup schemes; evaluating the transient overvoltage value of the day-ahead optimization scheduling scheme based on the predicted data of base load, wind power, and photovoltaic output in the intraday stage; if the transient overvoltage value of the day-ahead optimization scheduling scheme is higher than the safety threshold, using the set of backup schemes in the day-ahead stage as the initial solution set, solving the intraday optimization scheduling model, obtaining the intraday updated scheduling scheme.

[0031] The transient overvoltage suppression and optimized scheduling method for high-proportion renewable energy power grids provided in this embodiment specifically includes the following steps: Step 1: Obtain power grid topology information, equipment parameters, and historical operating data.

[0032] Step 2: Based on the acquired data, establish a time-domain simulation model of the power grid.

[0033] Step 3: Based on the time-domain simulation model and historical operating data of the power grid, generate a sample of the power grid's operating mode. Use the boundary-aware active data augmentation method to improve the deep learning transient overvoltage assessment model's ability to assess samples located at the transient overvoltage safety boundary, and obtain the final transient overvoltage assessment model.

[0034] The two-stage third-generation non-dominated genetic algorithm is used to directionally generate samples near the safety boundary of transient overvoltage. Iterative retraining is then used to improve the deep learning transient overvoltage assessment model's ability to evaluate samples near the safety boundary of transient overvoltage. The specific steps are as follows: Step 301: Obtain power grid topology information, equipment parameters and historical operating data, generate an initial training set, and train an initial transient overvoltage assessment model; Step 302: Based on the initial transient overvoltage assessment model, the two-stage NSGA-III algorithm is used to actively generate the operating mode located near the transient overvoltage safety boundary; Step 303: Use the time-domain simulation model to calculate the transient overvoltage of the newly generated operating mode, and include it as an enhanced sample in the training set of the deep learning transient overvoltage evaluation model to retrain the evaluation model and improve the evaluation accuracy of the evaluation model near the transient overvoltage safety boundary. Step 304: Repeat steps 302 and 303. If the evaluation error of the evaluation model for the enhanced sample is less than the threshold, the obtained evaluation model is the final deep learning transient overvoltage evaluation model.

[0035] The two-stage NSGA-III algorithm uses power system operating modes as a population and employs heuristic algorithms to directionally generate dispersed operating modes located near the transient overvoltage safety boundary, including: The first stage involves the targeted generation of operating modes located at the safety boundary of transient overvoltage, shortening the distance to the safety boundary, and the generated set of operating modes serves as the initial solution set. Based on the initial solution set generated in the first stage, the second stage operates in a diffusion mode around the safety boundary of transient overvoltage, increasing the coverage of the safety boundary.

[0036] The first stage employs a single-objective optimization model, including: Based on the final transient overvoltage assessment model, the difference between the transient overvoltage value and the safety threshold under the current operating mode is calculated, with the minimum 1-norm of all differences as the optimization objective. The constraints are system operation constraints, including: system power flow constraints; node voltage constraints; transformer capacity constraints; upper and lower limits of wind power plant output constraints; upper and lower limits of photovoltaic power plant output constraints; upper and lower limits of grid-connected energy storage output constraints; and upper and lower limits of thermal power plant output constraints. The first-stage optimization model is computed using a heuristic algorithm to obtain the initial solution set.

[0037] The second stage employs a multi-objective optimization model, including: The objective functions include: minimizing the 1-norm of the difference between the transient overvoltage value and the safety threshold in the current operating mode; and maximizing the normalized Euclidean distance between the current operating modes. The constraints include: all constraints considered in the first phase; The difference between the transient overvoltage value and the safety threshold in the current operating mode is less than a dynamic threshold, which is dynamically calculated based on the objective function value of the initial population and the calculation progress in the second stage, thereby increasing the coverage of the safety boundary.

[0038] Specifically, the boundary-aware proactive data augmentation method includes: a boundary-aware proactive data augmentation framework and its boundary exploration strategy based on the third-generation non-dominated genetic algorithm (NSGA-III). The boundary-aware proactive data augmentation framework improves evaluation performance by alternately executing boundary sample generation and iterative retraining of the evaluation model; the NSGA-III-based boundary exploration strategy is used to ensure that augmented samples meet operational constraints and improve boundary coverage.

[0039] 1. Boundary-aware proactive data augmentation framework.

[0040] In classification problems, the decision boundary is typically defined as the threshold in the feature space that distinguishes different classes. For binary classification problems, a common decision boundary can be represented as the set p(y=1|x)=0.5, where x represents the input of the machine learning model; y represents the output of the machine learning model, which is 0 or 1; and p represents the probability of y=1. For regression problems, since the output is a continuous value, there is usually no explicit decision boundary. Transient overvoltage assessment can define the acceptable operating range through a safety threshold (e.g., 1.2pu or 1.3pu). The prediction error near the threshold dominates the dispatch reliability, so the safety boundary can be regarded as the decision boundary. Specifically, for a given transient overvoltage safety threshold u0, the decision boundary can be represented as {X|f(X)=u0}, where X is the power system operating mode, f(X) is the transient overvoltage assessment value output by the machine learning model, and u0 is the given transient overvoltage safety threshold. Samples located near this boundary are defined as boundary samples and are used to improve the model's assessment accuracy near the boundary.

[0041] The paper "Decision Boundary-Aware Data Augmentation for Adversarial Training" uses interpolation to generate samples close to the decision boundary to improve model performance. However, under strict operational constraints and highly nonlinear feasible regions, this interpolation method cannot guarantee the feasibility of the augmented samples or the accuracy of the labels. Therefore, this embodiment proposes a boundary-aware active data augmentation framework based on NSGA-III to generate feasible boundary samples with labels verified by time-domain simulation, such as... Figure 1 As shown, the framework includes the following four steps: Step (1) Initial dataset generation and model training evaluation: First, uniform sampling is used to generate an initial dataset D0={(X) containing N samples in the running mode space. i ,u i )} N i=1 Then, an initial transient overvoltage assessment model is trained based on D0. Where D0 is the initial dataset, X i Indicates the operating mode of the i-th power system; u i X represents i The corresponding transient overvoltage value; N indicates that there are a total of N data points; Step (2) Boundary Sample Search: To identify the decision boundary for evaluating the agent model, the objective function is defined as minimizing in the k-th iteration. Subsequently, a boundary exploration strategy based on NSGA-III is used to generate a candidate set {X} that satisfies the operational constraints and covers the decision boundary. boundary}; where X represents the operating mode of the power system; k represents the k-th iteration; This represents the transient overvoltage value of X evaluated by the machine learning evaluation model at the current time (at the kth iteration); u0 is the given transient overvoltage safety threshold. Step (3) Simulation verification and data feedback: For the candidate set {X} boundary Time-domain simulation was performed to obtain the verified overvoltage value u. true As a label; to store data pairs (X) boundary ,u true ) are added to the training set D as augmentation samples, and the evaluation model is retrained. ; Step (4) Iteration and convergence criterion: Repeat steps (2)-(3) until the prediction error ε satisfies ε≤threshold ε th .

[0042] The proposed framework expresses the safety boundary as the decision boundary through an explicit safety threshold. This modeling approach is applicable to safety assessment problems with preset safety limits; for different fault scenarios or different safety indicators, only the sample labels and corresponding thresholds need to be adjusted. Furthermore, this framework effectively complements existing data augmentation methods by supplementing samples near the safety boundary.

[0043] To achieve efficient convergence, step (2) must generate boundary samples with good dispersion. Local clustering reduces data quality and slows down the improvement of the evaluation model in boundary-aware proactive data augmentation. However, metaheuristic algorithms typically emphasize optimality rather than sample diversity, which can easily lead to local clustering of the solution set and slow down the convergence of the overall framework. Step 302 will propose a boundary exploration strategy to improve the diversity of samples distributed along the boundary.

[0044] 2. Boundary exploration strategy based on NSGA-III (i.e., two-stage NSGA-III).

[0045] In the active data augmentation framework, the wider the sample coverage in step (2), the higher the data quality is usually. This step focuses on the boundary exploration strategy.

[0046] Boundary sample search in boundary-aware active data augmentation needs to simultaneously consider two objectives: boundary localization and boundary coverage. Boundary localization aims to minimize the deviation between the evaluation model's predictions and the decision boundary; boundary coverage aims to maintain sample diversity on the boundary. Third-generation non-dominated genetic algorithms are well-suited for this multi-objective scenario because their reference-point-based non-dominated sorting mechanism can balance both objectives and maintain population diversity. However, traditional NSGA-III is sensitive to the initial population and is prone to premature convergence in narrow and highly nonlinear feasible regions, leading to incomplete boundary coverage. Therefore, this embodiment divides the boundary exploration strategy into two stages, such as... Figure 2 As shown.

[0047] Phase I: Boundary Localization. This phase introduces a random encounter rate mechanism to conduct global exploration, generating a more widely distributed initial population for NSGA-III. The objective function is to minimize the distance from the operation mode vector X to the decision boundary, i.e.: (1); Where w represents the distance between the running point vector X and the decision boundary. Let represent the transient overvoltage assessment model in the k-th iteration, with u0 as the decision boundary. Operational constraints include AC power flow constraints, various output upper and lower limits, and related operational constraints. Finally, a candidate set S={X1,X2,...,X...} containing m samples can be obtained. m} and the optimal objective value w min .

[0048] The random encounter rate mechanism enhances global exploration capabilities but weakens local search and constraint feasibility. Therefore, some samples in the candidate set S are still far from the decision boundary. Using S as the initial population, stage II further advances the population to the vicinity of the feasible region and spreads it along the decision boundary.

[0049] Phase II: Boundary Coverage. In Phase II, the objectives and constraints are modified to distribute the samples along the decision boundary. The candidate operating mode set S obtained in Phase I is used as the initial solution set. An additional objective is introduced: maximizing the minimum neighborhood distance between samples. This objective constructs a repulsion mechanism to promote sample dispersion and avoid excessive clustering. The corresponding optimization model is: (2); in, For X i The normalized form of F; spread (i) is The minimum neighborhood Euclidean distance; q>0 is a slack variable used to define the acceptable deviation relative to the decision boundary; μ is an approximation factor used to scale the boundary positioning constraints and achieve gradual tightening, its expression is: (3); Among them, tanh( ) is the hyperbolic tangent function; n is the current iteration number of NSGA-III; N is the maximum iteration number; c is a constant that controls the growth rate of μ.

[0050] As the search progresses, μ will gradually increase from a small value to 1, thus gradually tightening the constraint, such as... Figure 3 As shown. In stage 1, μ is small, such that (w min When +q) / μ is relatively large, the constraint is relatively loose, which is conducive to exploration; in stage 2, μ gradually increases to 1, driving the samples to move closer to the feasible region and the boundary neighborhood; in stage 3, μ approaches 1, and the constraint approaches w. min The addition of q encourages more samples to fall into the feasible region. This approximation factor μ can dynamically adjust the constraints and improve sample feasibility.

[0051] In summary, Phase I provides a globally distributed initial solution set for Phase II, while the NSGA-III and adjustable constraints in Phase II further improve the feasibility of augmented samples and boundary coverage. This boundary exploration strategy is also applicable to optimization scheduling problems, which will be explained in detail later.

[0052] Step 4: Conduct a correlation analysis between the operating status of grid-type energy storage and transient overvoltage.

[0053] Step 4 is based on a simplified model of a high-proportion new energy power grid. It analyzes the impact of the operation status of grid-type energy storage on transient overvoltage, and embeds the derived mechanism into the metaheuristic algorithm through a hierarchical sorting mechanism to guide the search direction and improve computational performance.

[0054] Step 401: Simplified analysis model.

[0055] To analyze the impact of the operating status of multiple converters and grids on transient overvoltages, a system is constructed as follows: Figure 4 The simplified model of a high-proportion renewable energy power grid is shown.

[0056] A simplified model of a high-proportion renewable energy power grid includes two power plants and one synchronous generator. The synchronous generator is used to represent the local AC power grid, and its internal potential remains constant. For example... Figure 4 As shown, U1, U2 and U r X1, X2, X3 represent the voltage amplitudes of the busbars of each power station and the LCC-HVDC busbar, respectively. r and X e E represents the line reactance. gand X g P represents the internal potential amplitude and internal reactance of the synchronous generator, respectively; r and Q r These represent the U injected into each power station. r Total active and reactive power; P d and Q d Q represents the active and reactive power delivered by the LCC-HVDC link; c This indicates the reactive power compensation of the LCC-HVDC station; P1, Q1, P2, Q2, P e and Q e These represent the active and reactive power injected into each power station and the local power grid, respectively; j represents the imaginary unit. Ignoring transverse voltage components and line losses, the power quantity and bus voltage satisfy the following relationship: (4); (5); (6); Step 402: Correlation analysis between energy storage output and transient overvoltage.

[0057] This step analyzes how the energy storage operating status affects transient overvoltage. After a commutation failure in the LCC-HVDC, power supply is interrupted briefly, therefore P... d and Q d It can be approximated as zero. Reactive power compensation Q c When injected into the AC network, a reactive power surplus is created, which in turn triggers a transient overvoltage. For ease of analytical derivation, Q is defined as... et =-Q e And Q rt =-Q r Q et and Q rt With Q e and Q r The meaning is the same, only the direction is opposite, so it is a negative value. Substituting the above definition into formula (5), we get: (7); Similarly, formula (6) can be rewritten as: (8); Find Q using formula (8) et The derivative can be obtained as follows: (9); Among them, X g and X e All are positive values, E g Maintain a constant and slightly above 1.0 pu during commutation failure U r Higher than 1.0 PU, therefore 2Ur -E g >0, therefore, dU r / dQ et A value greater than 0 indicates that the greater the reactive power injected into the local power grid, the higher the reactive power U. r The higher the value, the higher U1 and U2 will be.

[0058] Define the total residual reactive power as Q. sur When the wind turbine generator is not equipped with high voltage ride-through (HVRT) control, (10); Among them, Q ess,i Let i represent the reactive power absorbed by the energy storage at site i, where i∈{1,2}.

[0059] Furthermore, (11); For the energy storage device at site i, the apparent power limit is satisfied as follows: (12); in, P represents the peak value (maximum value) of the transient overvoltage at bus k after a commutation failure fault. ess,i and Q ess,i S represents the active and reactive power of energy storage at site i, respectively. ess,i This represents the upper limit of the apparent power of energy storage at site i. In many power plants, the energy storage capacity is typically less than 30% of the installed capacity of new energy sources. Therefore, during transient overvoltage periods, reactive power regulation often reaches its upper limit, resulting in: (13); in, This represents the maximum reactive power that the energy storage at site i can absorb at the instant of a fault; when S ess,i When the active power output is constant, the maximum reactive power that can be absorbed is smaller.

[0060] right Find P ess,i The derivative yields: (14); Reactive power compensation Q c By P d Decision, when P d When Q is fixed, c Since it remains unchanged, combining with the previous formula (10), we can obtain: (15); Combining formulas (11) and (15), we obtain: (16); Where r represents the busbar.

[0061] Therefore, the higher the active power output of energy storage, the greater the transient overvoltage amplitude U. k The larger the reactive power output, the better; conversely, reducing the active power output of energy storage can improve the reactive power absorption capacity, thereby suppressing transient overvoltage. Step 403 will further analyze the impact of power sharing among multiple energy storage units on reactive power absorption capacity.

[0062] Step 403: Correlation analysis between energy storage power allocation and reactive power regulation capability.

[0063] This step analyzes the impact of power sharing among multiple energy storage units on reactive power regulation capability. Considering... Figure 4 Two energy storage units with the same apparent power upper limit, i.e., S1=S2=S, have a total active power P. ess,tot When fixed, (17); Where S1 represents the apparent power limit of energy storage at site 1, and S2 represents the apparent power limit of energy storage at site 2; M represents the number of energy storage units; in Figure 4 In this case, M=2.

[0064] The total reactive power absorbed by the energy storage is: (18); Substituting formulas (13) and (17) into formula (18), we get: (19); Q ess,tot Find P ess,1 The derivative can be obtained as follows: (20); Define an auxiliary function: (twenty one); Among them, φ(x) is strictly increasing within the feasible interval.

[0065] Then, formula (20) can be rewritten as: (twenty two); Formula (22) shows that the derivative dQ ess,tot / dP ess,1 It decreases monotonically within the feasible interval. When P ess,tot When Q ≠ 0, the derivative has a unique zero point within the interval, as shown in formula (23). ess,tot Reaching the maximum value: (twenty three); Among them, Pess,tot The total active power contribution for energy storage and Energy storage at site 1 and site 2 respectively in Q ess,tot The optimal active power output value corresponding to the attainment of the maximum value.

[0066] In summary, under the conditions of the same energy storage capacity and fixed total active power, the more balanced the output of each energy storage unit, the stronger the total reactive power absorption capacity. In actual systems, lateral components and line losses can cause a shift in the optimal point, but the overall pattern remains unchanged. Step 404 will embed this mechanism into a metaheuristic algorithm to improve computational performance.

[0067] Step 404: Rule-guided metaheuristic framework (i.e., the NSGA-III algorithm with added hierarchical sorting mechanism).

[0068] This step constructs a rule-guided metaheuristic framework, embedding the above mechanism into the metaheuristic algorithm through a hierarchical sorting mechanism to guide the convergence direction and improve computational performance.

[0069] (1) For feasible samples that meet the transient overvoltage safety constraints, sort them according to the objective function value (such as minimum total operating cost) to drive the algorithm to converge to the economic optimal solution.

[0070] (2) For infeasible samples that violate transient overvoltage safety constraints, they are ranked according to auxiliary criteria, which are defined as follows: (twenty four); Where E is the population ranking value under the auxiliary criterion; the smaller the value, the better the population. This represents the active power output P of each energy storage unit. ess,i The mean of the values. This criterion guides the search in infeasible regions toward smaller total energy storage output and higher output balance, thereby pushing the samples closer to the feasible region with smaller changes in control measures.

[0071] During normal operation, energy storage units undertake functions such as peak shaving and valley filling, and smoothing of renewable energy output. The rule-guided metaheuristic framework, by reserving reactive power absorption capacity for transient overvoltage suppression, introduces a certain increase in operating costs compared to purely economic dispatch. Therefore, this framework is only used for intraday emergency rescheduling when the day-ahead scheme is determined to violate transient overvoltage safety constraints. Step 5 will introduce its integration method in the two-stage dispatch model.

[0072] Step 5: Based on the final transient overvoltage assessment model, establish a day-ahead and intraday two-stage optimal scheduling model that considers transient overvoltage safety constraints; based on the new energy power generation forecast data of the day-ahead and intraday stages, solve the day-ahead and intraday two-stage optimal scheduling model to obtain an optimal scheduling scheme that can take into account both the economic efficiency of system operation and the safety of transient overvoltage.

[0073] (1) Based on the day-ahead forecast data of new energy power generation, solve the day-ahead optimal scheduling model, including: The optimization objectives include: minimizing the overall operating cost of the system; maximizing the diversity of operating modes; The constraints include: system operation constraints; the transient overvoltage value evaluated by the transient overvoltage assessment model is less than the safety threshold; The solution algorithm is a two-stage NSGA-III algorithm. The first stage obtains the optimal running scheme with the lowest operating cost, and the second stage generates a set of alternative schemes by expanding around the optimal running scheme.

[0074] (2) A two-stage optimization scheduling model for day-ahead and intraday scheduling is used to assess the transient overvoltage risk of the day-ahead scheduling scheme based on the ultra-short-term renewable energy generation forecast data updated intraday. If there is a risk of transient overvoltage exceeding the limit, the intraday optimization scheduling model is solved, including: The optimization objectives include: minimizing the overall operating cost of the system; and minimizing the adjustment of control measures. The constraints include: system operation constraints; the transient overvoltage value evaluated by the transient overvoltage assessment model is less than the safety threshold; An initial solution set is generated based on the set of alternative solutions from the current phase. The solution algorithm is to add a hierarchical sorting mechanism to the NSGA-III algorithm to accelerate the convergence speed and calculate the optimized scheduling scheme for intraday updates.

[0075] The hierarchical ranking mechanism is as follows: for operating modes that violate transient overvoltage safety constraints, the merits of the operating mode are evaluated based on the minimum sum of squares of energy storage output; for operating modes that do not violate transient overvoltage safety constraints, the merits of the operating mode are evaluated according to the traditional NSGA-III ranking method.

[0076] Step 501: Optimize scheduling of safety constraints in the two-stage phases from the day-to-day to the intraday.

[0077] This step proposes an evaluation model-assisted two-stage safety constraint optimization scheduling model for day-ahead and intraday periods to address the uncertainty of new energy output.

[0078] (1) Method framework.

[0079] The inherent uncertainty in renewable energy output can lead to forecasting errors during the day-ahead phase. Even if the day-ahead dispatching scheme meets the transient overvoltage safety constraints, it may become unfeasible during the intraday phase due to forecast updates. Therefore, online rescheduling is required based on the updated renewable energy forecasts to restore the feasibility of the dispatching scheme.

[0080] Two-stage optimization scheduling framework such as Figure 5As shown. During the day-ahead phase, the economically optimal solution and its surrounding diverse backup solution set are obtained; this backup solution set can provide a warm start for intraday rescheduling, thereby shortening the search distance and reducing changes in control measures relative to the day-ahead solution. An evaluation model is used to quickly assess the transient overvoltage levels of candidate scheduling schemes, and the final selected scheme is verified through time-domain simulation to ensure transient overvoltage safety.

[0081] During the daytime phase, it is necessary to simultaneously solve for an optimal scheduling scheme and a set of backup scheduling schemes. Two objective functions are set for this purpose: the first objective is to minimize operating costs to obtain the economically optimal scheduling scheme; the second objective is to maximize the dispersion relative to the optimal point to generate a set of backup schemes that can be used for intraday warm-ups.

[0082] During the intraday phase, the updated ultra-short-term renewable energy forecasts are first used to refresh the evaluation model input. Subsequently, the evaluation model performs an online assessment of the day-ahead scheme (optimal scheduling scheme). If the predicted value exceeds the transient overvoltage safety constraint, online rescheduling is triggered. At this point, the secondary objective in the day-ahead phase is replaced with minimizing the change in control measures. Through day-ahead-intraday coordination, online response to renewable energy uncertainties can be achieved.

[0083] Based on the above analysis, both the boundary exploration problem and the optimal scheduling problem in step 3 essentially require simultaneous consideration of multiple objectives and sample diversity. Therefore, the exploration strategy proposed in step 3 is also applicable to optimal scheduling. The scheduling models for the day-ahead and intraday phases will be introduced below.

[0084] (2) The scheduling model was optimized in advance.

[0085] The current dispatching system considers two objectives simultaneously: first, to minimize operating costs while meeting the safety constraints of transient overvoltage; and second, to maximize the minimum neighborhood distance to generate a diverse set of backup options near the economic optimum. This objective is consistent with the definition in the aforementioned formula (2). The model can be expressed as: (25); in, This represents the transient overvoltage corresponding to power system operating mode X. This equation represents the transient overvoltage corresponding to power system operation mode X as evaluated by the machine learning evaluation model. The implication is that the transient overvoltage corresponding to power system operation mode X evaluated by the machine learning evaluation model is considered the recognized (correct) transient overvoltage value; C is the total operating cost; P... re P es and P th These represent the active power output of WT / PV, energy storage, and synchronous generator, respectively; C re C es and C thThe corresponding cost coefficient is used. This optimization model employs a boundary exploration strategy based on NSGA-III to obtain the optimal scheduling scheme X. and a set of alternative solutions X′, wherein set X′ covers X The surrounding area can provide a warm start for intraday rescheduling. Since the accuracy of intraday ultra-short-term forecasts is higher than that of day-ahead forecasts, the transient overvoltage safety of day-ahead scheduling schemes needs to be re-verified during the intraday phase.

[0086] (3) Intraday optimized scheduling model.

[0087] During the intraday phase, the ultra-short-term output forecasts for WT and PV are first updated, followed by online verification of safety constraints using a transient overvoltage assessment model. If constraint exceedances occur, intraday rescheduling is triggered. To maintain consistency between day-ahead and intraday scheduling schemes, the second objective function is changed to minimize the change in control measures, i.e.: (26); in, The normalized day-ahead optimal scheduling scheme; F dist (i) is and The Euclidean distance between them is used to characterize the change in control measures relative to the day-ahead scheduling scheme. All candidate scheduling schemes must be verified through time-domain simulation before being put into application.

[0088] On the one hand, intraday rescheduling utilizes the backup scheme set X′ generated in the previous day stage for hot-start initialization, enabling the metaheuristic search to start near the previous day's optimal scheduling scheme, thereby reducing changes in control measures. On the other hand, the rules proposed in step 4 are introduced in the intraday stage to guide the metaheuristic framework, guiding the search to converge toward the safe and feasible region and suppressing transient overvoltages.

[0089] Step 6: Verify the optimized scheduling scheme based on the power grid time-domain simulation model to check the security of the scheduling scheme.

[0090] Because the transient overvoltage values ​​obtained by the machine learning evaluation model may contain errors, a more accurate time-domain simulation model is needed to perform a final simulation verification of the optimized scheduling scheme to ensure that the transient overvoltage is below the safe threshold.

[0091] Example analysis.

[0092] This embodiment uses a simplified actual power transmission system in China to verify the effectiveness of the proposed method.

[0093] A. Introduction to the test system and generation of the sample set.

[0094] Based on a real power grid, the following was constructed: Figure 6The high-proportion renewable energy grid is shown. The test system comprises 8 generation clusters, each with 3 substations; each substation is connected to 320MW of wind power, 213MW of photovoltaic power, and 50MVAr distributed synchronous condensers (DSC). Energy storage capacity is configured at 30% of the total installed capacity of wind and photovoltaic power, with a duration of 2 hours. The total installed capacity of wind and photovoltaic power in the system is 12800MW. Two 1000MW synchronous generators (SGs) are configured near the ±800kV LCC-HVDC substation, and the rated capacity of the LCC-HVDC link is 8000MW. The considered fault is a three-phase short circuit near bus 13 of the load center. This fault will cause a voltage drop and trigger LCC-HVDC commutation failure (CF), thereby inducing transient overvoltages at the grid connection points of the 8 generation clusters.

[0095] The operating mode is generated from 45 input variables, including the LCC-HVDC transmission power, WT and PV active power output, active power and reference voltage of energy storage and synchronous generators, and reference voltage of DSC. The output is the maximum transient overvoltage caused by commutation failure in the eight generator clusters. The initial dataset is generated using uniform sampling, and the specific settings are shown in Table 1.

[0096] Table 1. Operation Mode Generation Settings

[0097] Among them, P LCC-HVDC P represents the active power transmitted via DC. WT P represents the active power of wind power generation. PV P represents the active power of photovoltaic power generation; SG P represents the active power generated by the synchronous generator. ESS Indicates the active power of energy storage operation; U ref This indicates the voltage reference value for the synchronous machine and energy storage.

[0098] In addition, to examine the impact of differences in output distribution among different clusters on transient overvoltages, nine spatially differentiated operating conditions were also considered, as detailed in Table 2. The final dataset contains a total of 4050 samples, with the training set, validation set, and test set accounting for 60%, 20%, and 20%, respectively. The sub-training set contains 810 samples (i.e., one-third of the training set).

[0099] As shown in Table 2, the spatially differentiated operating conditions are composed of two independent factors, each of which contains three subtypes. Therefore, the Cartesian product forms a total of 3 × 3 = 9 operating conditions. For each factor, subtype 1 and subtype 2 apply scaling factors of 1.1 and 0.9 respectively to the clusters listed in Table 9 and the remaining clusters, while keeping the total WT / PV output and total energy storage output unchanged; subtype 3 uses uniform scaling. If a cluster is selected by both factors, its final scaling factor is the product of the two factors, provided that the constraints are satisfied (e.g., 1.1 × 1.1 = 1.21).

[0100] Table 2. Nine Spatially Differentiated Chemical Conditions

[0101] B. Performance of boundary-aware proactive data augmentation.

[0102] First, the feasibility of the enhanced samples generated by the proposed active data augmentation method and the interpolation-based data augmentation method is compared; then, the dispersion of the solution set obtained by the boundary exploration strategy based on NSGA-III is demonstrated.

[0103] (1) Enhance the feasibility of the sample.

[0104] Table 3 compares the feasibility rates of augmented samples generated by the interpolation-based data augmentation method and the proposed active data augmentation method. The feasibility rate represents the proportion of augmented samples that satisfy all operational constraints.

[0105] Table 3. Feasibility of Enhanced Samples (%)

[0106] Table 3 shows that the feasibility rate of augmented samples generated by interpolation-based data augmentation methods is very low, indicating that traditional data augmentation methods struggle to handle complex constraints. The proposed method embeds runtime constraints into NSGA-III, thus ensuring the feasibility of augmented samples generated in each iteration. This result verifies that the proposed method can guarantee the feasibility of augmented samples under complex constraints.

[0107] (2) Decision boundary coverage.

[0108] To verify the effectiveness of the proposed boundary exploration strategy, the proposed method is compared with three other methods, and the specific settings are shown in Table 4.

[0109] Table 4. Sample augmentation settings for different methods

[0110] Set w in formula (2) minWith +q=0.0028pu, 1.3±0.0028pu can be defined as the feasible decision boundary band. The population size is set to 200, and the maximum number of iterations is set to 3000. Figure 7 The transient overvoltage distribution of the generated samples is given. Figure 8 Principal component analysis (PCA) is then used to visualize the sample distribution. PCA can preserve the main variance information in a low-dimensional orthogonal subspace, thus facilitating the comparison of the dispersion and clustering of each solution set; both axes in the figure represent dimensionless linear combinations of normalized decision variables.

[0111] from Figure 7 It can be seen that the samples generated by the proposed method are concentrated within the boundary band near u0=1.3pu; the distributions obtained by methods 2 and 3 are wider, and the deviation from the boundary is also greater. The transient overvoltage distribution of method 1 is the most concentrated, but from... Figure 8 As can be seen, there is obvious local clustering in its sample set, indicating that the samples are highly similar.

[0112] Table 5 uses two indicators to quantify and compare the above results. The in-band rate represents the proportion of samples falling within the 1.3±0.0028pu boundary band; the set diameter represents the maximum Euclidean distance between any two feasible samples, and the larger the value, the wider the sample dispersion range.

[0113] Table 5. Comparison of Boundary Exploration Results

[0114] Table 5 shows that the proposed method achieves 100% in-band efficiency and obtains the largest set diameter. The computation time increases slightly due to the introduction of additional targets and enhanced initialization, but the additional time is on the order of approximately 200 s. Compared to the significant time-domain simulation time required for traditional uniform sampling, this additional overhead is acceptable in offline sample generation scenarios.

[0115] C. Performance of two-stage optimized scheduling.

[0116] First, the training process of the evaluation model is introduced. Then, the transient overvoltage suppression effect of the proposed two-stage optimized scheduling model is evaluated under the condition of uncertain WT and PV output, and its improved computational performance is demonstrated.

[0117] (1) Training of transient overvoltage assessment model.

[0118] This paper employs XGBoost (Limited Gradient Boosting) as the transient overvoltage assessment model and trains two types of models for comparison. The baseline model is trained on the complete training set; the initial model is trained on a sub-training set, and the training samples are expanded round by round through boundary-aware active data augmentation. Both models use the same hyperparameter search space, and the configuration with the lowest validation error is selected through grid search. Table 6 shows the hyperparameters of the initial model.

[0119] Table 6. XGBoost Hyperparameters

[0120] Where n_estimators is the upper limit of the total number of decision trees; learning_rate is the learning rate; max_depth is the maximum depth; subsample is the proportion of randomly sampled samples; colsample_bytree is the proportion of randomly selected features; reg_alpha is the coefficient of the L1 regularization term (L1 norm of the weights); and reg_lambda is the coefficient of the L2 regularization term (L2 norm of the weights).

[0121] Table 7 summarizes the model training process. Root mean square error (RMSE) is expressed in kV. Performance on the test set is given for round 0; prediction errors on newly generated augmented samples are given for rounds 1-6 before each model update. Termination threshold ε. th Set to 2.5kV.

[0122] Table 7. Training status of the transient overvoltage assessment model

[0123] (2) Transient overvoltage suppression effect.

[0124] The control measures for optimized scheduling include adjusting the active power output of WT and PV, the active power and reference voltage of energy storage and synchronous generators, and the reference voltage of DSC. In day-ahead and intraday scheduling, the LCC-HVDC power transmission remains unchanged. The total number of control variables is 44. The day-ahead and intraday WT / PV prediction data are taken from the Belgian public dataset. Table 8 lists three representative scenarios, in which the total WT and PV output of the system are randomly distributed among eight power generation clusters; in addition, the cost coefficient in formula (25) is taken from the open-source survey report "LEVELIZED COST OF ENERGY+". The day-ahead stage is based on the day-ahead prediction to obtain the scheduling scheme and its backup scheme set; the intraday stage is based on the updated prediction for evaluation, and intraday optimized scheduling is triggered when the predicted value of the evaluation model exceeds the safety boundary u0=1.3pu.

[0125] Table 8. Three selected scenarios

[0126] Figure 9 The PCC voltage response of the largest transient overvoltage cluster is presented. Due to WT and PV prediction errors, the day-ahead scheme experienced a safety overrun during the intraday phase, with the voltage peak exceeding 1.3 pu. After intraday rescheduling, the PCC voltage peak was reduced to below 1.3 pu, verifying the transient overvoltage suppression capability of the proposed day-ahead and intraday two-stage optimized scheduling model.

[0127] In the scheduling solution, both the baseline model and the enhanced model are used to assess transient overvoltages, and the intraday rescheduling problem is solved under their respective assessment models. Subsequently, the obtained scheduling scheme is verified through time-domain simulation to ensure that u max Scheduling schemes with ≤1.3 PU account for the following percentages: Figure 10 As shown.

[0128] like Figure 10 As shown, the transient overvoltage assessment model, after retraining with active data augmentation, increased the proportion of scheduling schemes that remained feasible after simulation verification by more than 40%. Because uniform sampling results in sparse samples near the safety threshold u0, the baseline model exhibits larger errors near u0, which can guide the search towards operating modes that violate transient overvoltage safety constraints. Active data augmentation concentrates samples within a narrow band region near u0, thereby reducing boundary errors and improving decision reliability.

[0129] (3) Changes in operating costs and control measures.

[0130] The impact of rule-guided hierarchical sequencing and the initialization of day-ahead standby heat collection on intraday rescheduling is analyzed from two perspectives: changes in operating costs and control measures. Table 9 compares the two methods presented with the proposed method. Figure 11 and Figure 12 The distribution of optimized trajectories and scheduling schemes are given separately, and the calculation time is also given in Table 9.

[0131] Figure 11 The intraday rescheduling results with and without rule-guided hierarchical sorting were compared under the same population size and number of iterations. Feasibility rate represents the proportion of candidate solutions that satisfy the transient overvoltage constraint, and minimum operating cost represents the lowest cost among feasible solutions.

[0132] Depend on Figure 11As can be seen, the feasibility rate of both methods reaches 100% at approximately 5% of the iteration progress. The rapid increase in feasibility rate is accompanied by a simultaneous increase in operating costs, indicating that feasibility is prioritized in the early stages of the search. Compared to Method 1, the proposed method reduces the peak cost offset by approximately 30%, and the final feasible cost is reduced by approximately 25%. This is because the auxiliary criteria in the rule-guided hierarchical ranking guide infeasible solutions towards lower total active power output from energy storage and higher output balance. Lower total active power output increases the maximum absorbable reactive power under apparent power constraints; higher balance prevents active power output from concentrating in a few energy storage units, thus maintaining stronger reactive power absorption capacity across the entire generation side. Therefore, with the same number of convergence iterations, the search can cross the safety boundary with a smaller economic deviation.

[0133] Figure 12 The distribution of scheduling schemes is visualized using t-distributed random neighborhood embedding (t-SNE). This method can preserve local neighborhood relationships in a two-dimensional embedding space, thus allowing for a direct comparison of the distribution characteristics of different solution sets.

[0134] like Figure 12 As shown, after hot-start initialization using the day-ahead backup scheme set, the solution set will be concentrated near the day-ahead scheduling scheme; while the solution set obtained from random initialization is more dispersed. This concentrated distribution shortens the distance to the day-ahead scheme, reduces changes in control measures, and thus improves the continuity between day-ahead and intraday scheduling.

[0135] Table 9. Calculation time for different methods

[0136] Table 9 shows the computation time for the three methods described above. The computation times for the three methods are basically the same, indicating that the proposed method can achieve the aforementioned performance improvement within the same computation time. In 10 independent experiments, the average computation time for the day-ahead phase was 336.67 s, and the average computation time for the intraday phase was 51.76 s. Considering that the scheduling interval is typically 15 minutes, the computation time for the intraday phase meets the requirements for online applications.

[0137] This embodiment proposes an optimized scheduling method based on boundary-aware active data augmentation to suppress transient overvoltages in high-proportion renewable energy power grids. Simulation results based on a real transmission system show that the proposed method can improve the feasibility of augmentation samples and effectively suppress transient overvoltages. The main conclusions are as follows: The proposed active data augmentation method improves the performance of the transient overvoltage assessment model near the safety boundary, increasing the proportion of scheduling schemes that are feasible after simulation verification by more than 40%; the proposed NSGA-III-based boundary exploration strategy can maintain 100% feasibility of augmented samples and improve the decision boundary coverage. The derived correlation between energy storage operating status and transient overvoltage indicates that reducing the total active power output of energy storage and improving the output balance of multiple energy storage units can enhance the transient overvoltage suppression capability. After embedding this mechanism into the metaheuristic search, the hierarchical sorting mechanism can reduce the operating cost by about 25% under the same computation time. The proposed two-stage optimized scheduling model, which combines day-ahead and intraday scheduling, can suppress transient overvoltages within safe boundaries in the test scenario. The day-ahead backup scheme set provides a hot start for intraday rescheduling and reduces changes in control measures. The computational efficiency of the proposed method meets the requirements for online applications.

[0138] Example 2 The transient overvoltage suppression and optimized dispatching system for high-proportion renewable energy power grids provided in this embodiment includes: The offline information acquisition module is configured to acquire power grid topology information, equipment parameters, and historical operating data.

[0139] The online information acquisition module is configured to acquire forecast data on base load, wind power and photovoltaic output, and available energy storage capacity for the day-ahead and intraday phases of the power system.

[0140] The time-domain simulation model calling module is configured to deploy a time-domain simulation model based on the acquired offline information.

[0141] The deep learning offline training module is configured to train a deep learning transient overvoltage assessment model based on the acquired offline information and time-domain simulation model using a boundary-aware active data augmentation method.

[0142] The transient overvoltage assessment module is configured to deploy a deep learning transient overvoltage assessment model to assess the transient overvoltage value of the system based on the input operating mode information.

[0143] The optimized scheduling module includes: day-ahead optimized scheduling module, intraday optimized scheduling trigger module, and intraday optimized scheduling module.

[0144] The scheduling module was recently optimized and configured as follows: Based on the day-ahead information acquired online, the two-stage NSGA-III algorithm is used to solve the day-ahead optimal scheduling model, and the day-ahead optimal scheduling scheme and the set of backup schemes are obtained. Transient overvoltage safety is verified by the transient overvoltage assessment module based on the pre-emptive optimized scheduling scheme.

[0145] The intraday optimization scheduling trigger module is configured as follows: Obtain the day-ahead optimized scheduling scheme based on the day-ahead optimized scheduling module; Intraday information is obtained based on the online information acquisition module; The transient overvoltage value of the day-ahead optimized scheduling scheme is calculated based on the transient overvoltage assessment module; If the transient overvoltage value of the optimized scheduling scheme is lower than the safety threshold, the intraday optimized scheduling module will not be triggered. If the transient overvoltage value of the optimized scheduling scheme is higher than the safety threshold, the intraday optimized scheduling module will be triggered.

[0146] The intraday optimized scheduling module is configured as follows: Based on intraday information acquired online, the NSGA-III algorithm with an embedded hierarchical sorting mechanism is used to solve the intraday optimal scheduling model and obtain the intraday optimal scheduling scheme. The transient overvoltage safety of the intraday optimized scheduling scheme is verified based on the transient overvoltage assessment module.

[0147] It should be noted that each module in this embodiment corresponds one-to-one with each step in Embodiment 1, and their specific implementation processes are the same, so they will not be repeated here.

[0148] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A transient overvoltage suppression and optimized scheduling method for high-proportion renewable energy power grids, characterized in that, include: Acquire power grid topology information, equipment parameters, and historical operating data to establish a time-domain simulation model of the power grid; Based on the time-domain simulation model and historical operating data of the power grid, a sample of the power grid's operating modes is generated, and the transient overvoltage assessment model is optimized using the boundary-aware active data augmentation method. The system acquires new energy power generation forecast data for the day-ahead and intraday phases, and obtains an optimized scheduling scheme through a two-stage day-ahead-intraday optimization scheduling model. The two-stage day-ahead-intraday optimization scheduling model includes: in the day-ahead phase, generating an optimal scheduling scheme and a set of backup scheduling schemes by minimizing operating costs and maximizing the minimum neighborhood distance; evaluating the optimal scheduling scheme using a transient overvoltage assessment model, and triggering intraday rescheduling if the assessment value exceeds the transient overvoltage safety constraint; and in the intraday phase, using the set of backup scheduling schemes as the initial solution set, generating an updated intraday optimized scheduling scheme by minimizing operating costs and control measure adjustments. The steps of the boundary-aware active data augmentation method include: Step 301: Based on power grid topology information, equipment parameters and historical operating data, generate a training set and train the transient overvoltage assessment model; Step 302: Based on the transient overvoltage assessment model, use the two-stage NSGA-III algorithm to generate the operating mode located near the transient overvoltage safety boundary; Step 303: Calculate the transient overvoltage of the generated operating mode using the time-domain simulation model, and include it as an enhanced sample in the training set to retrain the transient overvoltage evaluation model; Step 304: Repeat steps 302 and 303. If the evaluation error of the transient overvoltage assessment model for the enhanced sample is less than the threshold, then the transient overvoltage assessment model is the final transient overvoltage assessment model.

2. The transient overvoltage suppression and optimized scheduling method for high-proportion renewable energy power grids as described in claim 1, characterized in that, In the current stage, a two-stage NSGA-III algorithm is used to solve the problem: the first stage solves for the optimal scheduling scheme with the minimum running cost; The second stage generates a set of backup scheduling schemes by minimizing operating costs and maximizing the minimum neighborhood distance, spreading the optimal scheduling scheme around it.

3. The transient overvoltage suppression and optimized scheduling method for high-proportion renewable energy power grids as described in claim 1, characterized in that, During the intraday phase, the NSGA-III algorithm with a hierarchical sorting mechanism is used for solving the problem. The hierarchical sorting mechanism is as follows: for operating modes that violate transient overvoltage safety constraints, the merits of the operating mode are evaluated based on the minimum sum of squares of energy storage output; for operating modes that do not violate transient overvoltage safety constraints, the merits of the operating mode are evaluated according to auxiliary criteria.

4. The transient overvoltage suppression and optimized scheduling method for high-proportion renewable energy power grids as described in claim 1, characterized in that, When generating operating modes near the transient overvoltage safety boundary, the two-stage NSGA-III algorithm includes: the first stage adopts a single-objective optimization model, which generates operating modes near the transient overvoltage safety boundary by minimizing the distance from the operating mode to the decision boundary, and the generated set of operating modes is used as the initial solution set; based on the initial solution set, the second stage adopts a multi-objective optimization model, which maximizes the minimum neighborhood distance between samples by minimizing the distance from the operating mode to the decision boundary, and diffuses the operating modes.

5. The transient overvoltage suppression and optimized scheduling method for high-proportion renewable energy power grids as described in claim 4, characterized in that, When generating the operating mode located near the transient overvoltage safety boundary, the objective function of the second stage is: ； Where w represents the distance between the running point vector X and the decision boundary; This represents the transient overvoltage assessment model in the k-th iteration; u0 is the decision boundary; Operational Constraints are the operational constraints; For X i The normalized form of X; i F represents the operating mode of the i-th power system; spread (i) is The minimum neighborhood Euclidean distance; q is the slack variable; μ is the approximation factor.

6. The transient overvoltage suppression and optimized scheduling method for high-proportion renewable energy power grids as described in claim 5, characterized in that, The approximation factor is: ; among them, tanh( ) is the hyperbolic tangent function; n is the current iteration number of NSGA-III; N is the maximum iteration number; c is a constant that controls the growth rate of μ.

7. A transient overvoltage suppression and optimized dispatch system for high-proportion renewable energy power grids, characterized in that, include: The time-domain simulation model calling module is configured to: acquire power grid topology information, equipment parameters and historical operating data, and establish a time-domain simulation model of the power grid; The deep learning offline training module is configured to: generate power grid operation mode samples based on the power grid time-domain simulation model and historical operating data, and optimize the transient overvoltage assessment model using the boundary-aware active data augmentation method. The optimized scheduling module is configured to: acquire new energy power generation forecast data for the day-ahead and intraday periods, and obtain an optimized scheduling scheme through a two-stage optimized scheduling model; wherein, the two-stage optimized scheduling model includes: in the day-ahead period, generating an optimal scheduling scheme and a set of backup scheduling schemes by minimizing operating costs and maximizing the minimum neighborhood distance; evaluating the optimal scheduling scheme using a transient overvoltage assessment model, and triggering intraday rescheduling if the assessment value exceeds the transient overvoltage safety constraint; in the intraday period, using the set of backup scheduling schemes as the initial solution set, generating an intraday updated optimized scheduling scheme by minimizing operating costs and control measure adjustments; The steps of the boundary-aware active data augmentation method include: Step 301: Based on power grid topology information, equipment parameters and historical operating data, generate a training set and train the transient overvoltage assessment model; Step 302: Based on the transient overvoltage assessment model, use the two-stage NSGA-III algorithm to generate the operating mode located near the transient overvoltage safety boundary; Step 303: Calculate the transient overvoltage of the generated operating mode using the time-domain simulation model, and include it as an enhanced sample in the training set to retrain the transient overvoltage evaluation model; Step 304: Repeat steps 302 and 303. If the evaluation error of the transient overvoltage assessment model for the enhanced sample is less than the threshold, then the transient overvoltage assessment model is the final transient overvoltage assessment model.

8. The transient overvoltage suppression and optimized dispatching system for a high-proportion renewable energy power grid as described in claim 7, characterized in that, The optimized scheduling module includes a day-ahead optimized scheduling module, which is configured to: in the day-ahead phase, use a two-stage NSGA-III algorithm to solve the problem: the first stage solves for the optimal scheduling scheme with the minimum running cost; The second stage generates a set of backup scheduling schemes by minimizing operating costs and maximizing the minimum neighborhood distance, spreading the optimal scheduling scheme around it.

9. The transient overvoltage suppression and optimized dispatching system for a high-proportion renewable energy power grid as described in claim 7, characterized in that, The optimized scheduling module includes an intraday optimized scheduling module, which is configured to: during the intraday phase, use the NSGA-III algorithm with a hierarchical sorting mechanism for solving the problem. The hierarchical sorting mechanism is as follows: for operating modes that violate transient overvoltage safety constraints, the merits of the operating mode are evaluated based on the minimum sum of squares of energy storage output; for operating modes that do not violate transient overvoltage safety constraints, the merits of the operating mode are evaluated according to auxiliary criteria.

Images