A power system transient frequency stability index safety evaluation method and system

Abstract

This invention discloses a method and system for security assessment of transient frequency stability indicators in power systems, belonging to the field of power system security analysis technology. It aims to solve the technical problem of existing assessment methods struggling to balance computational speed, accuracy, and reliability. The invention includes constructing a deep neural network model guided by physical mechanisms, comprising linear mapping branches, nonlinear mapping branches, and feature fusion and output layers. First, a training dataset generated through time-domain simulation is acquired and used to train the model to obtain a surrogate model. Then, the input features of the system to be assessed are input into the surrogate model to quickly obtain the predicted value of the transient frequency stability indicator. This technical solution enables rapid and accurate calculation of transient frequency stability indicators in power systems, facilitating online security assessment of the power grid, dispatch decision support, and frequency stability early warning.

Description

Technical Field

[0001] This invention relates to the field of power system security analysis technology, and in particular to a method for security assessment of power system transient frequency stability indicators. Background Technology

[0002] With the high proportion of new energy sources such as wind power and photovoltaics being integrated into the power grid, the frequency stability problem of the power system is becoming increasingly prominent. When conducting security analysis, optimized scheduling, and real-time control of the power grid, it is crucial to quickly and accurately calculate key transient frequency stability indicators (such as minimum frequency deviation and maximum frequency change rate). However, existing technologies suffer from an irreconcilable contradiction between computational accuracy, computational speed, and model reliability.

[0003] Currently, the main frequency stability assessment methods can be divided into three categories:

[0004] The first category is based on simplified physical models, such as the System Frequency Response (SFR) model. A typical structure of this model is as follows: Figure 1 As shown, the main components include system inertia, speed governor, and load damping. This method simplifies the system to a low-order linear model through equivalent aggregation, resulting in extremely fast computation speed, which meets the speed requirements for online analysis. However, its drawback is that the model is oversimplified and cannot accurately characterize the high-order, nonlinear dynamic characteristics of actual power systems containing diverse components such as steam turbines, water turbines, wind turbines, and photovoltaic inverters. In particular, it cannot effectively handle key aspects such as amplitude limiting and dead zones in various speed governor and inverter control strategies. In complex scenarios with high proportions of renewable energy integration and varied operating modes, its calculation error is relatively large, and the reliability of the evaluation results is insufficient, which may lead to overly conservative or risky scheduling decisions.

[0005] The second category is based on detailed time-domain simulation. This method utilizes specialized tools such as Simulink and PSASP to construct sophisticated simulation models that include detailed equipment models and high-order differential equations, enabling the acquisition of high-precision frequency response curves and indicators. However, its modeling process is complex, and each simulation is extremely time-consuming, requiring massive simulations to cover diverse operating scenarios and build a data sample library for subsequent analysis. The enormous computational overhead makes it difficult to directly apply to scenarios requiring rapid response, such as online safety verification, real-time early warning, or unit combination optimization with frequency constraints.

[0006] The third category comprises methods that have emerged in recent years based on general data-driven models, such as using traditional deep neural networks like multilayer perceptrons as surrogate models. The structural components of this model include, for example... Figure 2Typically, these models consist of an input layer (receiving features such as unit status, output, and power disturbances), multiple fully connected hidden layers (performing feature transformations through nonlinear activation functions), and an output layer (outputting the minimum frequency deviation). This method attempts to bypass complex physical equation solving and achieve rapid prediction by learning the input-output mapping relationships from a large amount of simulation data. While such models possess strong nonlinear fitting capabilities, they are essentially "data black boxes" unrelated to physical mechanisms. Their training depends entirely on the scale and quality of the data, and the internal logic of the model lacks interpretability. More importantly, in new operating modes or boundary scenarios where the training data is insufficiently covered, the model's predictions may severely violate basic physical laws, raising questions about its generalization ability and engineering reliability. This makes it difficult to gain the trust of operators and apply to the analysis of power systems with high reliability requirements.

[0007] Therefore, existing technologies have not yet provided a method for assessing the safety of power system transient frequency stability indicators that can balance computational speed, high result accuracy, and strong model reliability in practical engineering. This is precisely the technical problem that urgently needs to be solved in this field. Summary of the Invention

[0008] The technical problem to be solved and the technical task proposed by this invention is to improve and refine existing technical solutions, and to provide a method and system for safety assessment of transient frequency stability indicators of power systems, aiming to balance speed, accuracy and reliability. To this end, this invention adopts the following technical solution.

[0009] In a first aspect, the present invention provides a method for security assessment of transient frequency stability indicators of a power system, comprising the following steps: 1) Construct a deep neural network model guided by physical mechanisms; The deep neural network model includes: The linear mapping branch receives input features representing the start-up and shutdown states of each unit in the system, which are used to learn and extract equivalent system-level frequency response lumped parameters. The nonlinear mapping branch receives input characteristics representing the active power output of each unit and the power disturbance of the system, and is used to simulate the frequency response process of the system under the influence of the nonlinear dynamic characteristics of the equipment under power disturbance. The feature fusion and output layer is used to fuse the output of the linear mapping branch and the output of the nonlinear mapping branch, and output at least one transient frequency stability index. 2) Obtain the training dataset, which includes the input features obtained through time-domain simulation under different operating scenarios and the corresponding true values ​​of transient frequency stability indicators; 3) The deep neural network model is trained using the training dataset to obtain a trained surrogate model; 4) Input the input features of the power system operation scenario to be evaluated into the trained surrogate model to obtain the corresponding predicted value of transient frequency stability index and perform a security assessment.

[0010] Step 1) Model building and Step 2) Training data acquisition are both preparatory steps and can be interchanged; the order is not fixed or mandatory. This technical solution processes input features such as unit start-up and shutdown status, active power output, and power disturbances through dedicated linear and nonlinear mapping branches. This is not a simple data black box, but rather an explicit embedding of the physical understanding that "system-level lumped parameters are mainly linearly related to unit combination status" and "dynamic process details are affected by nonlinear links triggered by power levels and disturbances" in the power system frequency response into the model architecture. This provides physical prior guidance for the neural network learning process, allowing the model to learn mapping relationships that are more in line with physical essence with higher data efficiency and faster convergence speed, and enhancing generalization ability and prediction reliability in scenarios where data is not fully covered. This technical solution transforms the complex dynamic calculation of high-order system frequency response into a forward propagation calculation using a pre-trained neural network model through an "offline training, online application" model. In the application phase (step 4), only the operating features need to be input into the model to obtain high-precision index predictions in a very short time. This fundamentally overcomes the bottleneck of excessively long computation times in traditional time-domain simulation methods, making it possible to perform rapid batch scanning and online evaluation of frequency safety for a large number of operating modes or anticipated faults. This provides a crucial technical tool for real-time power grid dispatching and safety early warning. Unlike general black-box neural networks with obscure input-output relationships, the model used in this solution has a clear physical correspondence: linear mapping branches are associated with system-level equivalent parameters, and nonlinear mapping branches are associated with device-level dynamic details. This structural interpretability allows operators and R&D personnel to understand the model's decision-making basis to a certain extent. When the model makes a prediction, its physical logic can be traced, increasing the credibility of the analysis results and reducing the trust barrier for applying intelligent models in highly reliable power system engineering, which is conducive to the practical deployment and promotion of the technology. The predicted value can be compared with a preset safety threshold to perform a frequency stability safety assessment and generate early warning information.

[0011] As a preferred technical means: the linear mapping branch constructed in step 1) is specifically used to combine the input unit start-up and shutdown states, and through at least one linear transformation layer that only contains linear weight operations and does not contain nonlinear activation functions, map it into lumped parameter features that characterize the overall frequency response of the system. The lumped parameter features are used to reflect the system's equivalent frequency regulation capability determined by the unit combination.

[0012] This technical solution maintains strong learning capabilities while ensuring that the core of the model strictly adheres to physical laws, thus achieving significant advantages in interpretability, learning efficiency, generalization ability, and reliability that traditional black-box models lack. Specifically, compared to traditional networks that use nonlinear layers to process all features, entrusting the discrete combination feature of unit status to a purely linear branch effectively reduces the dimensionality of complex nonlinear relationships that the model needs to learn, lowering the overall complexity of the model. During training, this branch can converge to a solution that conforms to physical laws faster and more stably, reducing the need for massive amounts of training data. Especially when facing entirely new unit combination methods that have not appeared in the training data, the linear branch performs linear extrapolation based on the learned weights, and its prediction results still maintain physical rationality, thereby enhancing the model's generalization ability and reliability for new operating modes. By "hard-coding" physical constraints (the linear relationship between unit state and system-level parameters) through network structure design, the model is prevented from learning or outputting absurd mapping relationships that violate basic physical common sense, such as "the relationship between unit state and system regulation capability is complex, nonlinear, or even inversely proportional." This ensures that even with poor data quality or noise, the model's core judgments about system-level characteristics remain within a reasonable physical framework, thus improving the overall robustness of the method.

[0013] As a preferred technical means: the nonlinear mapping branch constructed in step 1) includes an input layer, at least two fully connected hidden layers and an interface for output to the feature fusion layer connected in sequence; wherein, each of the fully connected hidden layers contains trainable weight parameters, bias parameters and nonlinear activation functions, which are used to extract and combine the high-order nonlinear interaction features in the active power output of the unit and the power disturbance input of the system layer by layer, so as to simulate the nonlinear frequency dynamic process caused by the speed regulation system limiting, dead zone and control saturation of the new energy inverter.

[0014] This technical solution constructs a deep structure containing at least two fully connected hidden layers, each equipped with a nonlinear activation function. This endows the network with powerful nonlinear fitting capabilities, enabling the model to accurately characterize and quantify the impact of strongly nonlinear elements commonly found in real power equipment (such as traditional generator speed governors and new energy inverters) on the overall system frequency dynamics. These impacts are completely ignored or severely distorted by traditional linearized models (such as SFR models). This technical solution models these impacts specifically through a dedicated nonlinear branch, thereby significantly improving the accuracy and realism of frequency stability index predictions under complex operating conditions (especially when there are large disturbances or equipment is in the nonlinear operating region). This branch takes the generator's active power output and system power disturbance as core inputs. Through the step-by-step transformation and combination of multiple nonlinear hidden layers, it can automatically learn and extract the complex, high-order nonlinear interaction relationships between these continuous variables, enabling the model to achieve a much more refined dynamic representation than traditional simplified models. This technical solution specifically assigns the nonlinear dynamic modeling task to this branch, forming a clear and complementary division of responsibilities with the linear mapping branch responsible for handling unit state and system-level linear parameters. This allows the model to learn without confusing the mapping relationships between two different properties, linear and nonlinear. The nonlinear branch can focus on learning complex dynamic patterns in the power and disturbance domains, avoiding the feature confusion and training difficulties that may occur when using a general network to learn all relationships at the same time. This guides the network to learn the complete physical laws more efficiently and accurately, improving the convergence speed and final performance of the overall model.

[0015] As a preferred technical means: In step 1), the feature fusion and output layer receives at least one system-level artificial feature in addition to receiving the output features of the linear mapping branch and the nonlinear mapping branch; the system-level artificial feature is a scalar that reflects the macroscopic state of the system and is obtained by statistical calculation of the system operation data through a predefined aggregation formula.

[0016] This technical solution introduces system-level artificial features during the model fusion stage, combining human experts' physical knowledge with data-driven models. This not only directly improves the model's accuracy and reliability by providing "physical priors," but also addresses a core obstacle to deploying intelligent models in critical industrial scenarios by enhancing interpretability and generalization capabilities. Specifically, this solution directly inputs predefined and calculated system-level artificial features, based on domain knowledge, into the feature fusion layer. This injects key, explicit macroscopic system state information into the model. These artificial features are a condensed summary of the system's operating state by human experts, containing core physical elements affecting frequency response. The model no longer needs to laboriously infer this global information implicitly from raw, fine-grained unit data; instead, it directly obtains high-value "physical cues," greatly supplementing the data-driven learning process. This guides the model to grasp the overall system situation more quickly and accurately, thereby making predictions more consistent with physical laws and significantly improving the accuracy and reliability of the output. The introduced system-level artificial features are scalars calculated based on explicit physical formulas (such as aggregation formulas), with clear and unambiguous physical meanings. The model's output establishes a direct data correlation with these understandable macroscopic indicators, enabling analysts to partially trace and understand the logic behind the model's predictions. This breaks the limitations of a purely data-driven model's "black box" and enhances the model's credibility and acceptability in power system engineering applications requiring high reliability. When facing new operating modes or extreme boundary scenarios not fully covered by training samples, models relying solely on data fitting may become inaccurate. However, the artificial features introduced in this solution are calculated based on universal physical definitions. Even for entirely new operating points, as long as data is substituted, the calculation of these features remains valid and the physical meaning remains unchanged. This helps the model maintain basic physical logic judgments in unfamiliar data environments, thereby improving its generalization ability and robustness under unknown or rare operating conditions. Since key macroscopic state information is provided directly through artificial features, the model does not need to learn and mine these complex high-level patterns from scratch from massive amounts of data. This simplifies the complexity of the mapping relationships that the model needs to learn, provides a better starting point for optimization, effectively accelerates the training convergence process of neural networks, and may reduce the requirements for the scale and coverage of training data, making it possible to build high-performance models under limited data conditions.

[0017] As a preferred technical means, the system-level artificial features include: Total generator standby capacity percentage: This is the difference between the sum of the maximum available capacity of all generating equipment in the system and the sum of the current active power output, and is the proportion of the sum of the maximum available capacity. And / or, The proportion of power output from new energy power generation equipment: This is the ratio of the total current active power output of all new energy power generation equipment in the system to the total current active power output of all power generation equipment in the system.

[0018] This technical solution injects the model with the most valuable prior physical knowledge, fundamentally guiding the model to focus on and understand the macroscopic physical essence that determines the frequency response, thereby improving accuracy, enhancing generalization, accelerating convergence, and ensuring physical rationality.

[0019] As a preferred technical approach, the process of obtaining the training dataset in step 2) includes: Based on the equipment parameters of the power system, a high-order system frequency response simulation model is constructed. Determine a multi-dimensional sampling space that includes the total power output of generators, the power output of new energy equipment, the number of operating units, and the magnitude of power disturbances; A stratified sampling method is used to sample various operating conditions within the sampling space; Time-domain simulations are performed for each operating condition to obtain its input features and the true values ​​of the corresponding transient frequency stability index, which constitute the training dataset.

[0020] The training dataset acquisition scheme of this technical solution combines high-fidelity simulation model construction, key dimension definition for the problem, and efficient stratified sampling strategy to form a systematic solution. It can generate high-quality training data with comprehensive coverage and accurate annotation with high efficiency, which lays a crucial data foundation for the subsequent training of high-performance and highly reliable intelligent analysis models. The scientific nature and efficiency of this data preparation method constitute an important premise and independent contribution to the realization of the overall technical effect of this invention.

[0021] As a preferred technical means, the transient frequency stability index includes the minimum frequency deviation, the maximum frequency change rate, and the quasi-steady-state frequency deviation.

[0022] The maximum rate of frequency change quantifies the drastic drop in frequency at the initial moment of the disturbance, reflecting the system's inertia support capability; the minimum frequency deviation captures the deepest point of frequency drop during the entire dynamic process, serving as the core basis for assessing whether the system has reached the low-frequency protection threshold and faces the risk of instability; the quasi-steady-state frequency deviation characterizes the steady-state shift of the frequency after the dynamic process subsides, reflecting the adjustment effect of the primary frequency regulation and the new steady-state operating point. The combination of these three metrics provides operators with a comprehensive, multi-dimensional safety assessment of the entire process from instantaneous impact and transient minimum point to long-term steady state, achieving a refined and panoramic depiction of the frequency stability state.

[0023] Secondly, the present invention provides a power system transient frequency stability index analysis system, comprising: A model building module is used to build a deep neural network model guided by physical mechanisms, wherein the deep neural network model includes: The linear mapping branch is configured to receive input features representing the start-up and shutdown states of each unit in the system, which are used to learn and extract equivalent system-level frequency response lumped parameters. The nonlinear mapping branch is configured to receive input characteristics representing the active power output of each unit and the power disturbance of the system, which is used to simulate the frequency response process of the system under the influence of the nonlinear dynamic characteristics of the equipment under power disturbance. The feature fusion and output layer is configured to fuse the output of the linear mapping branch and the output of the nonlinear mapping branch, and output at least one transient frequency stability index. The data acquisition module is used to acquire a training dataset, which includes input features obtained through time-domain simulation under different operating scenarios and the corresponding true values ​​of transient frequency stability indicators. The model training module is used to train the deep neural network model using the training dataset to obtain a trained proxy model. The index calculation and analysis module is used to input the input features of the power system operation scenario to be evaluated into the trained surrogate model, obtain the corresponding transient frequency stability index prediction value, and perform a security assessment.

[0024] The system employs a dual-branch deep neural network designed through its model building module to accurately characterize the frequency response characteristics of power systems. The linear mapping branch specifically targets the start-up and shutdown characteristics of generating units, learning and extracting equivalent system-level frequency response lumped parameters to directly address the core physical laws governing power system frequency stability. The nonlinear mapping branch focuses on the active power output of generating units and system power disturbances, specifically simulating the frequency response impact of nonlinear dynamic characteristics of equipment. This clear division of labor between the two branches closely aligns with actual physical processes, avoiding the shortcomings of traditional "black box" models that lack physical basis. This enhances the model's interpretability and ensures accurate fitting of complex frequency dynamic processes. The data acquisition module obtains training data under different operating scenarios through time-domain simulation. Time-domain simulation is a recognized high-precision method for acquiring frequency response data in the power system field. Its output input features and corresponding true values ​​of transient frequency stability indicators provide reliable benchmark labels for model training. Simultaneously, the data covers "different operating scenarios," ensuring the training dataset has broad representativeness. This allows the subsequently trained surrogate model to adapt to diverse operating conditions of the power system, reducing prediction bias in special scenarios. After the system generates a trained proxy model through the model training module, the index calculation and analysis module only needs to input the input features of the scenario to be evaluated into the proxy model to quickly output the predicted value of the transient frequency stability index. This process does not rely on complex high-order simulation modeling and massive iterative calculations, effectively solving the bottleneck of slow calculation speed of traditional fine models. It can meet the application scenarios that require response speed, such as online safety verification and real-time control of power systems.

[0025] As a preferred technical means: the linear mapping branch constructed in the model construction module specifically includes at least one linear transformation layer that contains only linear weight operations and no nonlinear activation function, used to map the input unit start-up and shutdown state combination into lumped parameter features characterizing the overall frequency response characteristics of the system; The nonlinear mapping branch constructed in the model building module includes an input layer, at least two fully connected hidden layers, and an interface for output to the feature fusion layer, which are connected in sequence. Each fully connected hidden layer contains trainable weight parameters, bias parameters, and a nonlinear activation function to simulate the nonlinear frequency dynamic process caused by speed regulation system limiting, dead zone, and control saturation of new energy inverters.

[0026] The linear mapping branch of this technical solution only sets up a linear transformation layer containing linear weight operations and no nonlinear activation functions. This aligns with the physical law of power systems—the linear aggregation relationship between the combination of unit start-up and shutdown states and the lumped parameters of the system-level frequency response—avoiding the interference of nonlinear activation functions on the linear characteristics of physical parameters. It can directly and accurately extract lumped parameter features characterizing the overall frequency response of the system from the unit start-up and shutdown states, ensuring that the extracted features are highly consistent with the physical mechanism and improving the reliability of the model in depicting the macroscopic frequency response characteristics of the system. The nonlinear mapping branch, through a structure of "at least two fully connected hidden layers + nonlinear activation functions," is specifically adapted to the modeling needs of nonlinear dynamic processes such as speed control system limiting, dead zones, and control saturation of renewable energy inverters. Multiple fully connected hidden layers can extract high-order nonlinear interaction features in the active power output of the units and system power disturbances layer by layer, while the nonlinear activation functions can effectively characterize the frequency dynamic distortion caused by the aforementioned nonlinear effects. This solves the problem that traditional linear models cannot accurately describe nonlinear frequency responses, significantly improving the simulation accuracy of complex nonlinear dynamic processes. The linear mapping branch focuses on capturing the linear physical essence (lumped parameter characteristics) of the system's frequency response, while the nonlinear mapping branch focuses on fitting nonlinear dynamic effects. The functional boundaries of the two are clear and their division of labor is precise, avoiding the confusion of learning objectives caused by a single network structure learning both linear laws and nonlinear effects at the same time. This allows the model to learn efficiently for linear relationships and nonlinear processes respectively, reducing mutual interference between learning different characteristics and improving the convergence speed and learning efficiency of model training.

[0027] As a preferred technical approach, the feature fusion and output layer is further configured to receive system-level artificial features, which include: Total generator standby capacity percentage: This is the difference between the sum of the maximum available capacity of all generating equipment in the system and the sum of the current active power output, and is the proportion of the sum of the maximum available capacity. And / or, The proportion of power output from new energy power generation equipment: This is the ratio of the total current active power output of all new energy power generation equipment in the system to the total current active power output of all power generation equipment in the system.

[0028] In this technical solution, the proportion of total generator reserve capacity and the proportion of output from renewable energy generation equipment are both system-level artificial features with clear physical significance. They are directly related to the core influencing mechanisms of power system frequency stability. The former reflects the system's active power regulation reserve capacity to cope with power shortages, while the latter characterizes the structural changes in the system's inertia level and frequency response characteristics. The feature fusion and output layer reception of these two types of features allow the model to directly utilize this macroscopic physical information during the learning process, avoiding the blindness of a purely data-driven "black box" learning model. This makes the model's prediction logic more aligned with the physical laws of the power system, significantly improving the model's interpretability. The proportion of total generator reserve capacity and the proportion of output from renewable energy generation equipment are core factors determining transient frequency stability indicators. Insufficient reserve capacity leads to weak frequency recovery capability, while an excessively high proportion of renewable energy causes a decrease in system inertia and increased frequency fluctuations. Integrating these two types of features into the fusion layer allows the model to directly focus on these key influencing variables without implicitly mining from massive amounts of raw features. This effectively reduces the interference of irrelevant features, enabling the model to more accurately capture the correlation between frequency stability indicators and core influencing factors, significantly improving prediction accuracy. System-level artificial features are highly aggregated and refined from raw operational data, pre-condensing key macroscopic information affecting frequency stability. These features eliminate the need for complex, layer-by-layer derivation by the model; the feature fusion and output layers directly receive these features, significantly reducing the difficulty of extracting core information from raw features, minimizing redundant calculations and iterations during model training, accelerating convergence, and improving overall training efficiency. The renewable energy power generation equipment output ratio feature is specifically designed for high-proportion renewable energy integration scenarios, accurately quantifying the impact of renewable energy penetration on system frequency response characteristics. The total generator reserve capacity ratio adapts to the active power balance uncertainty caused by renewable energy output fluctuations. The introduction of these two types of features enables the model to quickly perceive system structural changes and adjustment capabilities in this scenario, effectively improving the model's adaptability in scenarios with high renewable energy integration and variable operating modes, and avoiding prediction biases caused by inaccurate capture of scenario characteristics.

[0029] Beneficial effects: By deeply integrating the physical mechanism of power system frequency response into neural network design, this invention achieves performance improvements in three dimensions: accuracy, efficiency, and reliability.

[0030] This invention designs a network structure with a clear physical division of labor. The upper part uses linear mapping to extract system-level lumped parameters, while the lower part uses a nonlinear network combined with artificial features to simulate nonlinear effects such as amplitude limiting, providing powerful physical prior guidance for model learning. This overcomes the inefficiency of general "black box" neural networks that learn physical laws implicitly from scratch, enabling them to converge to more accurate solutions with less data and higher efficiency. It also significantly improves training speed and computational accuracy with the same number of parameters.

[0031] Meanwhile, since the model structure design itself is based on physical principles, it also significantly enhances the generalization ability and reliability under new operating modes where the training data is not fully covered. Furthermore, the correspondence between model parameters and physical concepts also gives the model a certain degree of interpretability, achieving the invention goal of balancing speed and accuracy. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of a typical system frequency response (SFR) model structure in existing technologies.

[0033] Figure 2 This is a schematic diagram of a multilayer perceptron structure commonly used in existing technologies.

[0034] Figure 3 This is a flowchart of the present invention.

[0035] Figure 4 This is a diagram of the neural network structure of the present invention. Detailed Implementation

[0036] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings.

[0037] Example 1 This embodiment provides a method for security assessment of transient frequency stability indicators in power systems, the process of which is as follows: Figure 3 As shown, it includes the following steps: S1: Based on the equipment parameters and network topology of the power system, a high-order system frequency response model is constructed. Different operating scenarios are obtained based on hierarchical sampling method and time-domain simulation. Initial features and transient frequency stability indicators of different scenarios are collected to obtain an initial feature sample set.

[0038] S2: Calculate the total generator standby capacity percentage across all operating scenarios in the initial feature sample set. and the proportion of power output from new energy power generation equipment As a system-level artificial feature, it is merged with the initial feature to obtain a complete feature sample set for calculating the transient frequency stability index.

[0039] S3: Construct a deep neural network model guided by physical mechanisms. The deep neural network model includes: The linear mapping branch receives input features representing the start-up and shutdown states of each unit in the system, which are used to learn and extract equivalent system-level frequency response lumped parameters. The nonlinear mapping branch receives input characteristics representing the active power output of each unit and the power disturbance of the system, and is used to simulate the frequency response process of the system under the influence of the nonlinear dynamic characteristics of the equipment under power disturbance. The feature fusion and output layer is used to fuse the outputs of the linear mapping branch and the nonlinear mapping branch, and output at least one transient frequency stability index.

[0040] In this embodiment, the input to the network is the unit's start-up and shutdown status. Output of each unit System power disturbance In addition to the artificial features introduced by S2, the output is the minimum predicted frequency deviation. Maximum frequency change rate and quasi-steady-state deviation .

[0041] S4: Obtain the training dataset and use the training dataset to train the deep neural network model.

[0042] In this embodiment, the complete feature sample set obtained in step S2 is used to train the deep neural network constructed in step S3 using the gradient descent method based on adaptive momentum estimation (Adam optimizer) to obtain the surrogate model used to calculate the above transient frequency stability index.

[0043] S5: Obtain and calculate the initial and artificial features of the power system operation scenario to be analyzed. Input the initial and artificial features into the surrogate model obtained in step S4 to obtain the transient frequency stability index to be calculated, thereby realizing the rapid calculation and analysis of the transient frequency stability index under different power system operation scenarios.

[0044] The following provides further explanation of some of the steps.

[0045] I. Step S1 specifically includes: S101: For a given... generator Taiwan centralized wind power equipment, For a centralized photovoltaic system, acquire system equipment and network parameters, including the frequency coefficients of each generator speed controller under power step disturbances (such as the static gain of each generator speed controller). , , , Inertial time constant Frequency regulation coefficient of new energy units (Wind power) (Photovoltaics), construct a refined high-order system frequency response model; S102: Based on historical power system operation data and expert experience, set the total generator output level. RoP total Power output level of new energy equipment R renew Number of machines turned on W total and the magnitude of power disturbance ΔP L The upper and lower limits of the range are expressed as follows:

[0046] in, and The first The upper and lower limits of the output of a generator (steam turbine and water turbine), For the first The active power of each generator.

[0047] This represents the set of steam turbines and water turbines that are currently in operation. The range of values ​​is .

[0048]

[0049] in This represents a collection of new energy power generation equipment. Its value range is determined by the upper and lower limits of generator output and the upper and lower limits of new energy power generation equipment output.

[0050]

[0051] in Indicates the first The generator's on / off status is represented by 1 if on and 0 if off. For The system of generators has a value range of 1. In one embodiment of the present invention, The range of values ​​is .

[0052] The specific power disturbance is expressed in MW, and its value range is manually determined. In one embodiment of the present invention, the selected... The range of values ​​is .

[0053] Will RoP total , R renew , W total as well as ΔP L Divide into four dimensions evenly The sampling space is divided into several equal-width intervals. Subspace; in one embodiment of the invention, the number of partition intervals is selected. A total of 81 sampling subspaces were identified.

[0054] S103: The result obtained in step S102 Random sampling is performed on each subspace to extract... Various operating conditions, collecting all data under all operating conditions. generator Taiwan centralized wind power equipment, The power of the centralized photovoltaic equipment in Taiwan, all The start-up and shutdown status of the generators, and the magnitude of system power disturbances. As Based on the initial characteristics, a time-domain simulation algorithm is used to analyze the power system in this... Simulations were performed to determine the minimum frequency deviation after a set power disturbance occurred under certain operating conditions. Maximum frequency change rate and quasi-steady-state deviation There are three transient frequency stability indices. The initial features and the three transient frequency stability indices are combined to form the initial frequency stability feature sample set.

[0055] II. Step S2 specifically includes: S201: Total generator standby capacity percentage ,in For the first Maximum available capacity of each power generation unit At this moment The active power of each generating unit. This feature characterizes the overall active power regulation capability of the system.

[0056] S202: The ratio of output from new energy power generation equipment to total power generation ,in This represents a collection of new energy power generation equipment. This feature is used to characterize structural changes in the system's inertia and frequency response characteristics.

[0057] The introduced artificial features are combined with the initial frequency-stabilized features in step S1 to obtain a complete frequency-stabilized feature sample set.

[0058] III. The deep neural network structure constructed in step S3 is as follows: Figure 4 Its core consists of two parallel processing branches with clear physical meaning, and its core features include: 1. The structure of a neural network mainly consists of linear and nonlinear mapping branches. Each branch contains an input layer, a feature cross layer, and an output layer. Multiple hidden layers are located between the input layer and the feature cross layer, and between the feature cross layer and the output layer.

[0059] 2. The upper part of the neural network is a linear mapping branch. The input to this branch is the unit's start-up and shutdown status. It is mapped to the feature cross layer through a linear mapping layer (without a non-linear activation function). The lumped parameters of the equivalent system-level frequency response module are used to characterize the system-level frequency response module. These parameters are essentially a linear weighted sum of unit states. Therefore, this linear layer is designed to directly learn and characterize these equivalent system-level lumped parameter features from unit start-up and shutdown states.

[0060] 3. The lower part is the nonlinear mapping branch. The input to this branch is the active power output of each unit. With power disturbance This branch contains m fully connected hidden layers, each with dimension d, where the values ​​of m and d are manually specified. In one embodiment of the invention, m=2 and d=64 are selected. Each hidden layer is activated using ReLU (Rectified Linear Function) to progressively extract high-dimensional features, which are ultimately mapped to a feature cross layer. This branch aims to simulate complex nonlinear dynamics that traditional simplified models cannot describe. The specific output level and disturbance magnitude of the unit directly affect whether individual devices (especially their governors and inverter controllers) enter the limiting region or control dead zone. Stacking multiple nonlinear layers allows us to learn how these nonlinear dynamics ultimately affect the overall frequency response of the system.

[0061] 4. Feature Crossover and Output. The features output from the upper branch... Features of the lower half branch output and the artificial features introduced in step S2. and The features are then fused and processed together through n r-dimensional hidden layers. The values ​​of n and r are manually specified; in one embodiment of this invention, n=1 and r=32 are selected. Each hidden layer is activated using the ReLU activation function, and the results are finally mapped to the output layer, outputting the minimum frequency deviation. Maximum frequency change rate and quasi-steady-state deviation .

[0062] IV. The training process in step S4 specifically includes: 1. The mean squared error (MSE) between the predicted values ​​and the simulated true values ​​in the dataset is used as the loss function. 2. The Adam optimizer (adaptive momentum estimation) is used, and the network is trained via backpropagation. The training process essentially involves the two branches of the network learning collaboratively; the upper branch learns how to... Extracting effective system-level features, the second branch learns how to extract from... and The model captures nonlinear effects. After training, the model's accuracy is verified using a test set. If the accuracy requirements are met, it can proceed to the application stage; otherwise, training needs to continue.

[0063] V. Step S5 specifically includes: Obtain new operating characteristics Calculate the artificial features according to the formula in step S2. and The original features and artificial features are input into the trained model to obtain the transient frequency stability index to be calculated. , and .

[0064] This embodiment has the following characteristics: 1. Specific neural network structures that integrate physical mechanisms A deep neural network structure with explicit physical meaning was constructed, with the input being a combination of start and stop states. Output level of each unit and the magnitude of system power disturbance The output is the transient frequency stability index of the system to be calculated. The parameters of the upper half of the network directly correspond to the lumped parameters of the system-level frequency response, while the lower half is specifically used to characterize the comprehensive impact of nonlinear elements such as limiting and dead zones in equipment such as speed governors. The two parts are coupled through the output layer and output the corresponding transient frequency stability index. This structure deeply integrates power system expertise with deep learning models, unlike general "black box" neural networks.

[0065] 2. A rapid calculation method for end-to-end transient frequency stability indices This method provides a fast end-to-end calculation approach that directly calculates the minimum frequency deviation from operating variables such as "unit combination, output, and power disturbance". Using a specific neural network as its core, this method transforms the problem of solving complex high-order nonlinear differential equations into a single, efficient feedforward neural network inference, reducing the computation time from seconds to milliseconds while maintaining engineering accuracy.

[0066] 3. Adaptive sample construction method combining expert experience In constructing the training dataset, this method employs expert-guided initial sampling: using a stratified sampling method, and based on the predefined ground-state operating scenarios, power fluctuations, and new energy output distributions from expert experience, the high-dimensional sampling space is divided into multiple physically meaningful subspaces for initial sampling, ensuring that the samples can broadly cover various possible key operating conditions.

[0067] 4. Introduce system-level physical artificial features In the lower part of the neural network, in addition to the original unit output and power disturbance data, aggregated features with clear physical meaning that reflect the overall operating state of the system are constructed and introduced as input, specifically including: Total generator standby capacity percentage This characterizes the active power regulation capability of the system.

[0068] proportion of new energy power generation , characterizing the structural changes in the system's inertia and frequency response characteristics.

[0069] This artificial feature provides the model with system-level macroscopic information, significantly improving the model's convergence, accuracy, and interpretability.

[0070] Example 2 This embodiment provides a power system transient frequency stability index analysis system, which includes: a model building module, a data acquisition module, a model training module, and an index calculation and analysis module.

[0071] I. Model building module, used to build a deep neural network model guided by physical mechanisms, wherein the deep neural network model includes: The linear mapping branch is configured to receive input features representing the start-up and shutdown states of each unit in the system, which are used to learn and extract equivalent system-level frequency response lumped parameters. The nonlinear mapping branch is configured to receive input characteristics representing the active power output of each unit and the power disturbance of the system, which is used to simulate the frequency response process of the system under the influence of the nonlinear dynamic characteristics of the equipment under power disturbance. The feature fusion and output layer is configured to fuse the output of the linear mapping branch and the output of the nonlinear mapping branch, and output at least one transient frequency stability index.

[0072] II. Data Acquisition Module, used to acquire training dataset, which includes input features obtained through time-domain simulation under different operating scenarios and the corresponding true values ​​of transient frequency stability indicators; including system equipment parameters (such as generator inertia time constant, governor parameters, new energy controller parameters), network parameters, load parameters, as well as real-time or historical unit operating status, output information and disturbance data.

[0073] III. Model training module, used to train the deep neural network model using the training dataset, and to execute the model training process, including parameter initialization, forward propagation calculation, loss function calculation, back propagation and parameter update, to obtain a trained surrogate model.

[0074] IV. The index calculation and analysis module is used to input the input features of the power system operation scenario to be evaluated into the trained surrogate model to obtain the corresponding predicted values ​​of transient frequency stability indicators. The index calculation and analysis module can perform online inference and receive real-time or predicted system operation data. ), using the trained model for forward computation, to quickly output , and The model's predicted results are compared with preset safety thresholds to determine the system's frequency safety status. When the index exceeds the limit or a risk is detected, corresponding early warning information is generated and pushed to dispatching and operation personnel.

[0075] To effectively accelerate the training and convergence process of the neural network and reduce the requirements for the scale and coverage of training data, this embodiment also includes an artificial feature calculation module, which is used to calculate system-level artificial features from the initial frequency-stabilized data obtained by the data acquisition module. and .

[0076] It is understood that the detailed functional implementation of each of the above modules can be found in the description of the aforementioned method embodiments, and will not be elaborated further here.

[0077] The above are specific embodiments of the present invention, which demonstrate the substantial features and progress of the present invention. Based on the actual needs of use, equivalent modifications in shape, structure, etc., can be made to it according to the teachings of the present invention, and all such modifications are within the scope of protection of this solution.

Claims

1. A method for security assessment of transient frequency stability indicators in power systems, characterized in that, Includes the following steps: 1) Construct a deep neural network model guided by physical mechanisms; The deep neural network model includes: The linear mapping branch receives input features representing the start-up and shutdown states of each unit in the system, which are used to learn and extract equivalent system-level frequency response lumped parameters. The nonlinear mapping branch receives input characteristics representing the active power output of each unit and the power disturbance of the system, and is used to simulate the frequency response process of the system under the influence of the nonlinear dynamic characteristics of the equipment under power disturbance. The feature fusion and output layer is used to fuse the output of the linear mapping branch and the output of the nonlinear mapping branch, and output at least one transient frequency stability index. 2) Obtain the training dataset, which includes the input features obtained through time-domain simulation under different operating scenarios and the corresponding true values ​​of transient frequency stability indicators; 3) The deep neural network model is trained using the training dataset to obtain a trained surrogate model; 4) Input the input features of the power system operation scenario to be evaluated into the trained surrogate model to obtain the corresponding predicted value of transient frequency stability index and perform a security assessment.

2. The method for security assessment of transient frequency stability indicators of a power system according to claim 1, characterized in that: The linear mapping branch constructed in step 1) is specifically used to combine the input unit start-up and shutdown states and map them into lumped parameter features that characterize the overall frequency response of the system through at least one linear transformation layer that contains only linear weight operations and no nonlinear activation functions. The lumped parameter features are used to reflect the system's equivalent frequency regulation capability determined by the unit combination.

3. The method for security assessment of transient frequency stability indicators of a power system according to claim 1, characterized in that: The nonlinear mapping branch constructed in step 1) includes an input layer, at least two fully connected hidden layers, and an interface for output to the feature fusion layer, which are connected in sequence. Each of the fully connected hidden layers contains trainable weight parameters, bias parameters, and a nonlinear activation function, which are used to extract and combine the high-order nonlinear interaction features in the active power output of the unit and the power disturbance input of the system layer by layer, so as to simulate the nonlinear frequency dynamic process caused by the speed regulation system limiting, dead zone, and control saturation of the new energy inverter.

4. The method for security assessment of transient frequency stability indicators of a power system according to claim 1, characterized in that: In step 1), the feature fusion and output layer receives at least one system-level artificial feature in addition to the output features of the linear mapping branch and the nonlinear mapping branch. The system-level artificial feature is a scalar that reflects the macroscopic state of the system and is obtained by statistical calculation of the system operation data through a predefined aggregation formula.

5. The method for security assessment of transient frequency stability indicators of a power system according to claim 4, characterized in that: The system-level artificial features include: Total generator standby capacity percentage: This is the difference between the sum of the maximum available capacity of all generating equipment in the system and the sum of the current active power output, and is the proportion of the sum of the maximum available capacity. And / or, The proportion of power output from new energy power generation equipment: This is the ratio of the total current active power output of all new energy power generation equipment in the system to the total current active power output of all power generation equipment in the system.

6. The method for security assessment of transient frequency stability indicators of a power system according to claim 1, characterized in that: Step 2) involves obtaining the training dataset, which includes: Based on the equipment parameters of the power system, a high-order system frequency response simulation model is constructed. Determine a multi-dimensional sampling space that includes the total power output of generators, the power output of new energy equipment, the number of operating units, and the magnitude of power disturbances; A stratified sampling method is used to sample various operating conditions within the sampling space; Time-domain simulations are performed for each operating condition to obtain its input features and the true values ​​of the corresponding transient frequency stability index, which constitute the training dataset.

7. The method for security assessment of transient frequency stability indicators of a power system according to claim 1, characterized in that: The transient frequency stability indicators include the minimum frequency deviation, the maximum rate of frequency change, and the quasi-steady-state frequency deviation.

8. A power system transient frequency stability index analysis system, characterized in that, include: A model building module is used to build a deep neural network model guided by physical mechanisms, wherein the deep neural network model includes: The linear mapping branch is configured to receive input features representing the start-up and shutdown states of each unit in the system, which are used to learn and extract equivalent system-level frequency response lumped parameters. The nonlinear mapping branch is configured to receive input characteristics representing the active power output of each unit and the power disturbance of the system, which is used to simulate the frequency response process of the system under the influence of the nonlinear dynamic characteristics of the equipment under power disturbance. The feature fusion and output layer is configured to fuse the output of the linear mapping branch and the output of the nonlinear mapping branch, and output at least one transient frequency stability index. The data acquisition module is used to acquire the training dataset, which includes the input features obtained through time-domain simulation under different operating scenarios and the corresponding true values ​​of transient frequency stability indicators. The model training module is used to train the deep neural network model using the training dataset to obtain a trained proxy model. The index calculation and analysis module is used to input the input features of the power system operation scenario to be evaluated into the trained surrogate model, obtain the corresponding transient frequency stability index prediction value, and perform a security assessment.

9. A power system transient frequency stability index analysis system according to claim 8, characterized in that: The linear mapping branch constructed in the model building module specifically includes at least one linear transformation layer that contains only linear weight operations and no nonlinear activation function, used to map the input unit start-up and shutdown state combination into lumped parameter features characterizing the overall frequency response characteristics of the system. The nonlinear mapping branch constructed in the model building module includes an input layer, at least two fully connected hidden layers, and an interface for output to the feature fusion layer, which are connected in sequence. Each fully connected hidden layer contains trainable weight parameters, bias parameters, and a nonlinear activation function to simulate the nonlinear frequency dynamic process caused by speed regulation system limiting, dead zone, and control saturation of new energy inverters.

10. A power system transient frequency stability index analysis system according to claim 8, characterized in that: The feature fusion and output layer is also configured to receive system-level artificial features, which include: Total generator standby capacity percentage: This is the difference between the sum of the maximum available capacity of all generating equipment in the system and the sum of the current active power output, and is the proportion of the sum of the maximum available capacity. And / or, The proportion of power output from new energy power generation equipment: This is the ratio of the total current active power output of all new energy power generation equipment in the system to the total current active power output of all power generation equipment in the system.

Images