Phase modifier reactive transient stability quantitative analysis system based on time constant
By generating an expanded sample set using an adversarial network model and optimizing the parameters of the synchronous condenser simulation model, the problem of inaccurate identification of damping coefficients under sparse measured data is solved, and high-precision reactive transient stability performance evaluation of the synchronous condenser is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA POWER INVESTMENT XINJIANG ENERGY & CHEMICAL GROUP TOLI CO LTD
- Filing Date
- 2026-01-16
- Publication Date
- 2026-05-08
AI Technical Summary
Existing technologies struggle to accurately identify the damping coefficient of a synchronous condenser under sparse measured disturbance data, leading to inaccurate time constant calculations and impacting the assessment of reactive transient stability.
A data augmentation method based on an adversarial network model is adopted to generate an expanded sample set. Multi-parameter collaborative optimization is performed in conjunction with the simulation model. The sparse measured disturbance data is intelligently amplified by the adversarial network model to generate an expanded sample set covering multiple operating conditions. Based on this, the simulation model is driven to perform closed-loop optimization and the parameters of the synchronous condenser group simulation model are adjusted to extract the damping coefficient.
It significantly improves the identification accuracy of the damping coefficient, ensures the reliability of the time constant calculation, and provides a high-precision assessment of the reactive transient stability performance of the synchronous condenser.
Smart Images
Figure CN122000951A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of synchronous condenser technology, and more specifically, to a reactive transient stability quantitative analysis system for synchronous condensers based on time constant. Background Technology
[0002] The time constant τ is a core comprehensive parameter characterizing the dynamic response of a synchronous condenser rotor. It directly determines the speed at which the system recovers synchronization after a disturbance and the decay rate of power angle oscillations. A smaller τ value indicates a stronger synergistic effect between rotor kinetic energy and damping torque, a more prominent ability of the unit to maintain synchronous operation under grid voltage fluctuations or fault disturbances, and better reactive power support performance during transient processes (such as voltage recovery speed and oscillation suppression efficiency). Therefore, τ is a key indicator for evaluating its transient stability margin.
[0003] The physical basis of the time constant τ relies on two core parameters: the rotor inertia constant (H) and the damping coefficient (D). H characterizes the inherent ability of the unit's rotor to store kinetic energy, determined by the equipment's mechanical structure, and is usually obtained directly from the factory calibration value. D, on the other hand, reflects the system's damping effectiveness in suppressing oscillations, essentially a combination of electrical and mechanical damping. Because D is affected by actual operating conditions, it needs to be identified and calculated on-site through dynamic testing.
[0004] Current methods for dynamically identifying the damping coefficient D (such as frequency domain response method, Prony analysis, or least squares fitting) rely on high-precision synchronously acquired multi-dimensional operational data, including but not limited to rotor power angle, terminal voltage / current, active / reactive power, and excitation system variables. However, such high-dimensional data available in actual power grids is often sparse and discrete: limited by the deployment of measurement devices, communication costs, and data storage capabilities, continuous and complete disturbance process data are difficult to fully cover. Summary of the Invention
[0005] The summary section of this application is intended to provide a brief overview of the concepts, which will be described in detail in the detailed description section below. This summary section is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.
[0006] Some embodiments of this application propose a time constant-based reactive power transient stability quantitative analysis system to solve the technical problems mentioned in the background section above.
[0007] As a first aspect of this application, some embodiments of this application provide a reactive power transient stability quantitative analysis system based on a time constant, including: The data acquisition device is used to collect the power angle deviation curve and rotor speed deviation curve of the synchronous condenser group under different disturbance information, and generate an actual sample set; The data augmentation device uses an adversarial network model to augment the actual sample set and generate an expanded sample set. The simulation test device has a built-in simulation model of the synchronous condenser group. Based on the disturbance information in the expanded sample set, it generates simulation results, including the simulated power angle deviation curve and the simulated rotor speed deviation curve. The model parameter correction device aims to achieve the best fit between the simulation results and the actual power angle deviation curve and the actual rotor speed deviation curve in the expanded sample set. It adjusts the simulation parameters of the synchronous condenser group simulation model and extracts the damping coefficient of the synchronous condenser group from the simulation parameters. The time constant calculation device calculates the time constant based on the damping coefficient and the inertia constant, and generates a performance report of the synchronous condenser based on the time constant.
[0008] This solution uses an adversarial network model to intelligently amplify sparse measured disturbance data (power angle / speed deviation curves), generating an expanded sample set covering multiple operating conditions. Based on this, the simulation model is optimized in a closed loop, significantly improving the identification accuracy of the damping coefficient (D), thereby ensuring the reliability of the time constant calculation. Ultimately, it provides a high-precision, low-data-dependency authoritative analysis report for the quantitative evaluation of the reactive transient stability performance of synchronous condensers.
[0009] Furthermore, the simulation parameters include damping coefficient, synchronous torque coefficient, AVR gain, AVR time constant, PSS gain, PSS lead time constant, and PSS lag time constant.
[0010] This scheme uses an adversarial network model to amplify sparse measured disturbance data (power angle / speed deviation curves) under multiple operating conditions. Based on the expanded samples, it drives multi-parameter collaborative optimization of the simulation model (dynamically adjusting key parameters such as damping coefficient, synchronous torque coefficient, AVR gain / time constant, PSS gain / lead / lag time constant, etc.), effectively avoiding overfitting problems caused by adjusting a single damping coefficient, and significantly improving the global fitting accuracy between the simulation curve and the actual curve.
[0011] Furthermore, the disturbance information is the change information of the reference voltage of the automatic voltage regulator of the synchronous condenser group; A single disturbance can be any one of a step disturbance, an impulse disturbance, or a pseudo-random disturbance; ; Where t represents time, and t0 represents the injection time of the disturbance information. This represents the reference voltage before the disturbance. Indicates the step amplitude; When the disturbance information Vref (t) represents a step disturbance: ; Indicates the step amplitude; When the disturbance information V ref (t) represents the pulse disturbance: ; Indicates the pulse amplitude. Indicates the pulse width; When the disturbance information V ref When (t) is a pseudo-random perturbation: , This represents the amplitude of the k-th frequency, where N represents the number of frequency components and k represents the index of the frequency component. This represents the k-th frequency phase.
[0012] This scheme injects three typical disturbances (step / pulse / pseudo-random) into the reference voltage (V) of the automatic voltage regulator by defining them in a structured manner. ref While ensuring the interpretability of disturbances, it avoids simulation distortion caused by complex disturbances and breaks through the limitations of single disturbance conditions. By combining the sample amplification and multi-parameter collaborative optimization mechanism of the adversarial network model, it significantly improves the cross-condition fitting accuracy of the simulation curve and the actual power angle / speed deviation curve.
[0013] Furthermore, the simulation model of the synchronous condenser group is as follows: ; ; Where D represents the damping coefficient, K represents the synchronous torque coefficient, and K A T represents the AVR gain. A K represents the AVR time constant. PSS T1 represents the PSS gain (signal amplification factor of the stabilizer), T2 represents the PSS lead time constant, and T2 represents the PSS lag time constant. M represents the synchronous angular velocity, H represents the inertial time constant, and H represents the inertial constant. Indicates electromagnetic torque deviation. Indicates the terminal voltage. This represents the PSS washing and filtration time constant. Represents state variables, Indicates the generator power angle. Indicates the actual rotor speed. Indicates the excitation voltage. This indicates the PSS output signal. This represents the disturbance signal, where f1, f2, f3, and f4 are the reciprocals of the state variables. This represents the initial steady-state work angle.
[0014] As a second aspect of this application, the parameter identification method relies on the matching verification between the simulation curve and the actual operating curve (power angle / speed deviation): if the two are consistent, the simulation model parameters can be equivalent to the actual unit parameters. However, this method requires massive training data to drive the simulation results to approximate the actual response. In scenarios where measured samples are sparse, the accuracy of parameter identification is severely limited due to insufficient data support, necessitating enhancement techniques to overcome the sample bottleneck. Based on this, this application provides the following technical solution: Furthermore, the data enhancement device includes: The data classification module acquires the actual sample set and divides the disturbance information in the actual sample set into step disturbance information sample set, impulse disturbance sample set and pseudo-random disturbance sample set according to the type of disturbance information. The sample expansion module has a built-in adversarial network model; The training control module uses a step disturbance information sample set to train the adversarial network model and obtain the step model parameters. The training control module uses a set of impulse perturbation information samples to train the adversarial network model and obtain the impulse model parameters; The training control module uses a pseudo-random perturbation information sample set to train the adversarial network model and obtain pseudo-random model parameters. Among them, the adversarial network model loads step model parameters to expand the step perturbation information sample set; The adversarial network model is loaded with pulse model parameters to expand the pulse perturbation information sample set; The adversarial network model is loaded with pseudo-random model parameters to expand the sample set of pseudo-random perturbation information.
[0015] This scheme trains adversarial network sub-models (with independent parameters) according to perturbation type (step / pulse / pseudo-random) to accurately learn the dynamic physical mapping relationship of "reference voltage → power angle / speed deviation" under different perturbations, generating high-fidelity augmented samples. This model-splitting enhancement mechanism significantly suppresses the distribution deviation between the generated data and the actual samples, providing a high-confidence data foundation with strong physical constraints for multi-parameter collaborative optimization of the simulation model, and ultimately improving the identification accuracy of damping coefficient and time constant.
[0016] Furthermore, adversarial network models include: The generator produces augmented samples based on the input random information; The discriminator is used to randomly input real samples and augmented samples, and outputs the probability that the input information is a real sample; Both the generator and the discriminator are deep neural network models, and the generator and the discriminator are trained simultaneously. The loss function of the discriminator is: : ; The loss function of the generator is : ; This represents the i-th real sample. This represents the probability that the discriminator classifies a real sample. This represents the i-th augmented sample. This represents the probability that the discriminator will distinguish the expanded samples, and m represents the number of samples.
[0017] Furthermore, the random information is set to d-dimensional random information, with each dimension following a standard normal distribution; Among them, the dimensionality of random information in the generation of step disturbance information sample sets, pulse disturbance information sample sets, and pseudo-random disturbance information sample sets is getting higher and higher.
[0018] This scheme precisely controls the random intensity of generated samples by designing an adaptive random information dimension for perturbation types (step → low dimension, impulse → medium dimension, pseudo-random → high dimension d-dimensional normal distribution): while ensuring the determinism of step / impulse perturbations, it enhances the spectral diversity of pseudo-random perturbations; combined with the physical constraints of the sub-models in classification training, it ensures that the generated samples not only conform to the actual dynamic response laws, but also cover multi-scale random conditions, providing high-fidelity and wide-coverage augmented data for simulation models.
[0019] Furthermore, the generator includes: The input layer is used to input a noise vector, and to reshape the noise vector to obtain reshaped features; The upsampling layer upsamples the reconstructed features to obtain upsampled features. The normalization layer normalizes the upsampled features to obtain normalized features; The activation function layer performs RelU activation on the normalized features to generate activated features; Convolutional layers, which have multiple layers, repeatedly perform convolution operations on activation features to obtain convolutional features; In the output layer, the convolutional features are activated using the tanh function to obtain the output features, and the output features are then denormalized to generate augmented samples. The discriminator includes: The information input layer is used to input augmented or real samples to generate initial features; Information convolutional layers perform convolution operations on initial features to generate convolutional features; The information pooling layer performs global average pooling on the convolutional features to generate pooled features. Fully connected layers map pooled features to a low-dimensional space to generate low-dimensional features. The information output layer uses the Sigmoid function on low-dimensional features to generate the true probability of the samples.
[0020] Furthermore, the loss function of the adversarial network during joint training is: ; Where G represents the generator and D represents the discriminator. Represents the true data distribution. Represents a probability distribution. express The average value on the real data distribution. express The average value over the probability distribution, where x represents the input to the discriminator and z represents random information; Represents the core value function; When training the adversarial network model, the following training loop is executed: Fixed generator G, updated discriminator D: maximize ; With a fixed discriminator D, update the generator G: minimize ; When fixing the generator G and updating the discriminator D, several noise information is used to generate augmented samples. The augmented samples are mixed with several real samples and sent to the discriminator D. For each sample input to the discriminator D, calculate the loss function. According to the loss function Backpropagation updates the weight parameters inside the discriminator D; With a fixed discriminator D, update the generator G: minimize hour: Several new random information samples are sampled and used to generate several expanded samples using generator G. The expanded samples are then mixed with several real samples and sent to discriminator D. For each sample, calculate the generator's loss function. According to the loss function Backpropagation updates the weight parameters inside the generator.
[0021] This scheme utilizes a min-max game training framework of Generative Adversarial Networks (GANs) (where the generator G and discriminator D alternately optimize the value function V(D,G)) to progressively improve the physical fidelity of generated samples in a dynamic game: the generator G learns the real perturbation-response mapping, while the discriminator D enhances the ability to distinguish sample distributions. The two work in sync and iteratively drive the generated samples to infinitely approximate the actual power angle / speed deviation curve. Combined with perturbation classification and dimensionality adaptation mechanisms, this ensures that the expanded samples have both high fidelity and operating condition coverage, providing strong data support for the accurate identification of the damping coefficient (D) and time constant (τ).
[0022] Furthermore, at the beginning of each loop, for each training of the generator G, based on the pre-set random probability h, it is determined whether each neuron in the hidden layer is hidden, the hidden neurons are frozen, a new hidden layer structure is generated, the generator G is then trained, and the weight parameters of the unfrozen hidden layers are updated. When the current loop ends, the frozen neurons are restored.
[0023] This scheme introduces a random hiding regularization mechanism for neurons in adversarial network training (dynamically freezing the hidden layer neurons of generator G with probability h). By forcing the network to learn in a fragmented structure, it effectively suppresses the generator's tendency to overfit training samples. Combined with a min-max game training framework and perturbation classification-dimensional adaptive design, the generated power angle / speed deviation samples have high physical fidelity, strong generalization and wide operating condition coverage, providing an enhanced data foundation with anti-overfitting and high robustness for the accurate identification of damping coefficient and time constant. Attached Figure Description
[0024] Figure 1 This is a schematic diagram of a reactive transient stability quantification analysis system based on time constant of a synchronous condenser.
[0025] Figure 2 This is a schematic diagram of the adversarial network model.
[0026] Figure 3 This is a partial schematic diagram of the convolutional network in the generator.
[0027] Figure 4 This is a schematic diagram showing the hidden structure of the neurons in the convolutional network within the generator. Detailed Implementation
[0028] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments. The same reference numerals in the accompanying drawings represent the same components. It should be noted that the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the described embodiments of this application without creative effort are within the scope of protection of this application.
[0029] refer to Figure 1 Example 1: A time-constant-based reactive power transient stability quantitative analysis system for synchronous condensers includes a data acquisition device, a data augmentation device, a simulation testing device, a model parameter correction device, and a time constant calculation device. The data acquisition device, data augmentation device, simulation testing device, model parameter correction device, and time constant calculation device are connected sequentially.
[0030] The calculation of the time constant τ requires the rotor inertia constant and damping coefficient. The rotor inertia constant is calibrated after the synchronous condenser unit leaves the factory. Therefore, it is necessary to calculate the damping coefficient D of the synchronous condenser unit. The damping coefficient D is affected by various internal information of the synchronous condenser unit and is a dynamic characteristic. Therefore, the damping coefficient D is difficult to calculate directly through a model or by measuring and calculating using relevant data. To this end, this application provides the following solution: The data acquisition device is used to collect the power angle deviation curves and rotor speed deviation curves of the synchronous condenser unit under different disturbance information, generating a real sample set. The real sample set includes multiple samples, for example, 1000 samples. Each sample includes features and labels: the features are the disturbance information, and the labels are the power angle deviation curve and the rotor speed deviation curve. In other words, the synchronous condenser unit is deployed in a power grid operating environment, and then provided with disturbance information (features). When responding to this disturbance information, the synchronous condenser unit records the power angle deviation curve and the rotor speed deviation curve (label). In this way, a sample is obtained. Through multiple experiments, the synchronous condenser unit can obtain multiple samples. All these samples are collected to form the real dataset.
[0031] The data augmentation device uses an adversarial network model to augment the actual sample set, generating an expanded sample set.
[0032] Data augmentation devices primarily expand the sample set of an actual dataset. For example, to obtain the most accurate damping coefficient, 1000 samples are needed. However, conducting 1000 experiments with a camera shifter group is difficult and time-consuming. Therefore, it is necessary to expand the sample set based on the original sample set to obtain an expanded sample set.
[0033] Thus, the expanded sample set includes all samples in the actual sample set, as well as some samples that have been diffused out of the set itself.
[0034] In practice, whether or not it is necessary to expand the actual sample is mainly determined by the accuracy requirements of the damping coefficient and the number of actual samples. If the number of actual samples is large and the accuracy requirements of the damping coefficient are low, then it is not necessary to expand the sample.
[0035] The simulation test device has a built-in simulation model of the synchronous condenser group. Based on the disturbance information in the expanded sample set, it generates simulation results, including the simulated power angle deviation curve and the simulated rotor speed deviation curve.
[0036] The simulation model of the synchronous condenser group is as follows: ; ; Where D represents the damping coefficient, K represents the synchronous torque coefficient, and K A T represents the AVR gain.A K represents the AVR time constant. PSS T1 represents the PSS gain (signal amplification factor of the stabilizer), T2 represents the PSS lead time constant, and T2 represents the PSS lag time constant. M represents the synchronous angular velocity, H represents the inertial time constant, and H represents the inertial constant. Indicates electromagnetic torque deviation. Indicates the terminal voltage. This represents the PSS washing and filtration time constant. Represents state variables, Indicates the generator power angle. Indicates the actual rotor speed. Indicates the excitation voltage. This indicates the PSS output signal. This represents the disturbance signal, where f1, f2, f3, and f4 are the reciprocals of the state variables. This represents the initial steady-state work angle.
[0037] The synchronous condenser simulation model is essentially the theoretical calculation model of the synchronous condenser. After defining the simulation parameters and the initial state, and inputting disturbance information to solve the synchronous condenser simulation model, the complete power angle deviation curve and rotor speed deviation curve can be obtained.
[0038] The simulation parameters include damping coefficient, synchronous torque coefficient, AVR gain, AVR time constant, PSS gain, PSS lead time constant, and PSS lag time constant. Numerical integration methods (such as the fourth-order Runge-Kutta method) are generally used to solve the simulation model of the synchronous condenser group; how to solve this model is currently unavailable. The key to this application lies in the targeted setting of disturbance information.
[0039] The disturbance information is the change information of the reference voltage of the automatic voltage regulator of the synchronous condenser group; A single disturbance can be any one of a step disturbance, an impulse disturbance, or a pseudo-random disturbance; ; Where t represents time, and t0 represents the injection time of the disturbance information. This represents the reference voltage before the disturbance. Indicates the step amplitude; When the disturbance information V ref (t) represents a step disturbance: ; Indicates the step amplitude; When the disturbance information V ref (t) represents the pulse disturbance: ; Indicates the pulse amplitude. Indicates the pulse width; When the disturbance information V ref When (t) is a pseudo-random perturbation: , This represents the amplitude of the k-th frequency, where N represents the number of frequency components and k represents the index of the frequency component. This represents the k-th frequency phase.
[0040] The model parameter correction device aims to achieve the best fit between the simulation results and the actual power angle deviation curve and the actual rotor speed deviation curve in the expanded sample set. It adjusts the simulation parameters of the synchronous condenser group simulation model and extracts the damping coefficient of the synchronous condenser group from the simulation parameters.
[0041] Specifically, the expanded sample set contains 1000 samples. The disturbance information in these 1000 samples is substituted into the simulation model of the synchronous condenser group to obtain the simulated power angle deviation curve and the simulated rotor speed deviation curve. This simulated power angle deviation curve and the simulated rotor speed deviation curve are generally different from the actual power angle deviation curve and the actual rotor speed deviation curve in the sample. In this way, the model parameter correction device corrects the simulation parameters, so that the simulated power angle deviation curve and the simulated rotor speed deviation curve can be closer to the actual power angle deviation curve and the actual rotor speed deviation curve.
[0042] The model parameter correction device uses the acquired expanded sample set and the least squares method to continuously correct the simulation parameters, so that the simulated power angle deviation curve and the simulated rotor speed deviation curve gradually approach the actual power angle deviation curve and the actual rotor speed deviation curve.
[0043] When a set of simulation parameters is adjusted so that the simulated power angle deviation curve and the simulated rotor speed deviation curve are closest to the actual power angle deviation curve and the actual rotor speed deviation curve, the damping coefficient in the simulation parameters is closest to the actual damping coefficient.
[0044] The time constant calculation device calculates the time constant based on the damping coefficient and the inertia constant, and generates a performance report for the synchronous condenser based on the time constant. The performance report is simply the magnitude of the time constant.
[0045] ; ; Let M represent the time constant, D represent the damping coefficient, M represent the inertial time constant, and H represent the inertial constant. Indicates synchronous angular velocity; This solution constructs a simulation model of the synchronous condenser group, which incorporates sufficient simulation parameters. Therefore, by simply adjusting the model parameters of the synchronous condenser group simulation model using the obtained samples, and ensuring that the simulation results coincide with the actual results, the required simulation parameters can be obtained.
[0046] After constructing the simulation model of the synchronous condenser group, the model can be solved using the fourth-order Runge-Kutta method. The simulation parameters can be adjusted using a greedy algorithm, continuously adjusting the model parameters step-by-step until the optimal simulation parameters are obtained.
[0047] refer to Figure 2 Example 2: Example 1 provides how to calculate the time constant. In fact, the key to the technical solution of Example 1 lies in how to obtain sufficient augmented samples relatively inexpensively. Based on this, this application provides the following technical solution: The data augmentation device includes a data classification module, a sample augmentation module, and a training control module. The data classification module is connected to the data acquisition device and classifies the samples in the acquired actual sample set. The sample augmentation module is connected to the training control module, which in turn is connected to the data classification module. The training control module uses the actual sample set to train the sample augmentation module, which then generates augmented samples. These augmented samples are then combined with the actual samples to form an augmented sample set.
[0048] Specifically, the data classification module acquires the actual sample set and classifies the perturbation information within it into step perturbation sample sets, impulse perturbation sample sets, and pseudo-random perturbation sample sets based on the type of perturbation information. The sample expansion module has a built-in adversarial network model. The training control module uses the step perturbation sample set to train the adversarial network model and obtain the step model parameters; the training control module uses the impulse perturbation sample set to train the adversarial network model and obtain the impulse model parameters; the training control module uses the pseudo-random perturbation sample set to train the adversarial network model and obtain the pseudo-random model parameters. The adversarial network model is loaded with step model parameters to expand the step perturbation information sample set; the adversarial network model is loaded with pulse model parameters to expand the pulse perturbation information sample set; the adversarial network model is loaded with pseudo-random model parameters to expand the pseudo-random perturbation information sample set.
[0049] For example, a real sample set might contain 100 step perturbation samples, 100 impulse perturbation samples, and 100 pseudo-random perturbation samples. The adversarial network model can then be trained in parallel using these samples to obtain the step model parameters, impulse model parameters, and pseudo-random model parameters, respectively. These corresponding model parameters can then be loaded into the adversarial network model to generate the required augmented samples.
[0050] Therefore, the key to obtaining augmented samples lies in the design of the adversarial network model and the training process of the adversarial network model.
[0051] An adversarial network model is essentially a simple convolutional network where a generator and a discriminator are trained adversarially. Specifically: Adversarial network models include: The generator produces augmented samples based on the input random information; The discriminator is used to randomly input real samples and augmented samples, and outputs the probability that the input information is a real sample; Both the generator and the discriminator are deep neural network models, and the generator and the discriminator are trained simultaneously. The generator includes: The input layer is used to input a noise vector, and to reshape the noise vector to obtain reshaped features; The upsampling layer upsamples the reconstructed features to obtain upsampled features. The normalization layer normalizes the upsampled features to obtain normalized features; The activation function layer performs RelU activation on the normalized features to generate activated features; Convolutional layers, which have multiple layers, repeatedly perform convolution operations on activation features to obtain convolutional features; In the output layer, the convolutional features are activated using the tanh function to obtain the output features, and the output features are then denormalized to generate augmented samples. The discriminator includes: The information input layer is used to input augmented or real samples to generate initial features; Information convolutional layers perform convolution operations on initial features to generate convolutional features; The information pooling layer performs global average pooling on the convolutional features to generate pooled features. Fully connected layers map pooled features to a low-dimensional space to generate low-dimensional features. The information output layer uses the Sigmoid function on low-dimensional features to generate the true probability of the samples.
[0052] The loss function of the discriminator is: : ; The loss function of the generator is : ; This represents the i-th real sample. This represents the probability that the discriminator classifies a real sample. This represents the i-th augmented sample. This represents the probability that the discriminator will distinguish the expanded samples, and m represents the number of samples.
[0053] The random information is set to d-dimensional random information, with each dimension following a standard normal distribution; among them, the dimensions of the random information used to generate step disturbance information sample sets, pulse disturbance information sample sets, and pseudo-random disturbance information sample sets increase with each other.
[0054] Random information serves as the input to the generator, meaning the generator uses perturbation information and built-in model parameters to generate augmented samples. Different types of augmented samples require different model parameters, and correspondingly, different types of random information need to be input to ensure a proper match between the random information and the model parameters, thus preventing model overfitting.
[0055] The following is the training process of the adversarial network model: The loss function of the adversarial network during joint training is: ; Where G represents the generator and D represents the discriminator. Represents the true data distribution. Represents a probability distribution. express The average value on the real data distribution. express The average value over the probability distribution, where x represents the input to the discriminator and z represents random information; Represents the core value function; When training the adversarial network model, the following training loop is executed: Fixed generator G, updated discriminator D: maximize ; With a fixed discriminator D, update the generator G: minimize ; When fixing the generator G and updating the discriminator D, several noise information is used to generate augmented samples. The augmented samples are mixed with several real samples and sent to the discriminator D. For each sample input to the discriminator D, calculate the loss function. According to the loss function Backpropagation updates the weight parameters inside the discriminator D; With a fixed discriminator D, update the generator G: minimize hour: Several new random information samples are sampled and used to generate several expanded samples using generator G. The expanded samples are then mixed with several real samples and sent to discriminator D. For each sample, calculate the generator's loss function. According to the loss function Backpropagation updates the weight parameters inside the generator.
[0056] Specifically, taking the acquisition of pulse model parameters as an example, the following details how to perform iterative training: S1: Obtain 1000 augmented samples (all from the impulse perturbation information sample set), and divide the 1000 augmented samples into 10 smaller sample sets, each with 100 samples.
[0057] S2: Fixed generator G, updated discriminator D: Maximization ; 100 noise information points are used to generate 100 augmented samples, which are then mixed with 100 real samples (all from the first small sample set) and fed into the discriminator D; For each sample input to the discriminator D, calculate the loss function. According to the loss function Backpropagation updates the weight parameters inside the discriminator D; S3: Fixed discriminant D, updated generator G: Minimization : 100 new random information samples are sampled and used to generate 100 augmented samples using generator G. These augmented samples are then mixed with 100 real samples (all from the second small sample set) and fed into discriminator D. For each sample, the loss function of the generator is calculated. According to the loss function Backpropagation updates the weight parameters inside the generator.
[0058] S4: Repeat S2 and S3 continuously until all small sample sets have been trained.
[0059] The above describes the joint training process of the adversarial network model.
[0060] In adversarial network models, the generator G plays a decisive role. Therefore, conventional approaches continuously increase the complexity of the generator G. However, excessive complexity and a large number of neurons in the generator G can lead to overfitting and path dependence during training. To address this, a more advanced implementation involves, at the beginning of each iteration, determining whether each neuron in the hidden layer is hidden based on a pre-set random probability h. Hidden neurons are frozen, a new hidden layer structure is generated, and the generator G is then trained, updating the unfrozen weights in the hidden layer. At the end of the iteration, the frozen neurons are restored.
[0061] For example, if a generator G has 10,000 neurons with a random probability h = 0.5, approximately 5,000 neurons will be hidden during each training iteration. Thus, the internal structure of the generator's convolutional network is not fixed during each iteration, thereby greatly avoiding path dependence during training. Figure 3 As shown, Figure 3 For a complete neural network, Figure 4 This is a network structure that hides some neurons.
[0062] The above are merely preferred embodiments of this application and are not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.
Claims
1. A reactive transient stability quantitative analysis system based on time constant for synchronous condensers, characterized in that, include: The data acquisition device is used to collect the power angle deviation curve and rotor speed deviation curve of the synchronous condenser group under different disturbance information, and generate an actual sample set; The data augmentation device uses an adversarial network model to augment the actual sample set and generate an expanded sample set. The simulation test device has a built-in simulation model of the synchronous condenser group. Based on the disturbance information in the expanded sample set, it generates simulation results, including the simulated power angle deviation curve and the simulated rotor speed deviation curve. The model parameter correction device aims to achieve the best fit between the simulation results and the actual power angle deviation curve and the actual rotor speed deviation curve in the expanded sample set. It adjusts the simulation parameters of the synchronous condenser group simulation model and extracts the damping coefficient of the synchronous condenser group from the simulation parameters. The time constant calculation device calculates the time constant based on the damping coefficient and the inertia constant, and generates a performance report of the synchronous condenser based on the time constant.
2. The reactive transient stability quantitative analysis system for synchronous condensers based on time constant as described in claim 1, characterized in that: The simulation parameters include damping coefficient, synchronous torque coefficient, AVR gain, AVR time constant, PSS gain, PSS lead time constant, and PSS lag time constant.
3. The reactive transient stability quantitative analysis system for synchronous condensers based on time constant as described in claim 1, characterized in that: The disturbance information is the change information of the reference voltage of the automatic voltage regulator of the synchronous condenser group; A single disturbance can be any one of a step disturbance, an impulse disturbance, or a pseudo-random disturbance; ； Where t represents time, and t0 represents the injection time of the disturbance information. This represents the reference voltage before the disturbance. Indicates the step amplitude; When the disturbance information V ref (t) represents a step disturbance: ; Indicates the step amplitude; When the disturbance information V ref (t) represents the pulse disturbance: ; Indicates the pulse amplitude. Indicates the pulse width; When the disturbance information V ref When (t) is a pseudo-random perturbation: , This represents the amplitude of the k-th frequency, where N represents the number of frequency components and k represents the index of the frequency component. This represents the k-th frequency phase.
4. The reactive transient stability quantitative analysis system for synchronous condensers based on time constant as described in claim 1, characterized in that: The simulation model of the synchronous condenser group is as follows: ； ； Where D represents the damping coefficient, K represents the synchronous torque coefficient, and K A T represents the AVR gain. A K represents the AVR time constant. PSS T1 represents the PSS gain, T2 represents the PSS lead time constant, and T2 represents the PSS lag time constant. M represents the synchronous angular velocity, H represents the inertial time constant, and H represents the inertial constant. Indicates electromagnetic torque deviation. Indicates the terminal voltage. This represents the PSS washing and filtration time constant. Represents state variables, Indicates the generator power angle. Indicates the actual rotor speed. Indicates the excitation voltage. This indicates the PSS output signal. This represents the disturbance signal, where f1, f2, f3, and f4 are the reciprocals of the state variables. This represents the initial steady-state work angle.
5. The reactive transient stability quantitative analysis system for synchronous condensers based on time constant according to any one of claims 1 to 4, characterized in that: The data enhancement device includes: The data classification module acquires the actual sample set and divides the disturbance information in the actual sample set into step disturbance information sample set, impulse disturbance sample set and pseudo-random disturbance sample set according to the type of disturbance information. The sample expansion module has a built-in adversarial network model; The training control module uses a step disturbance information sample set to train the adversarial network model and obtain the step model parameters. The training control module uses a set of impulse perturbation information samples to train the adversarial network model and obtain the impulse model parameters; The training control module uses a pseudo-random perturbation information sample set to train the adversarial network model and obtain pseudo-random model parameters. Among them, the adversarial network model loads step model parameters to expand the step perturbation information sample set; The adversarial network model is loaded with pulse model parameters to expand the pulse perturbation information sample set; The adversarial network model is loaded with pseudo-random model parameters to expand the sample set of pseudo-random perturbation information.
6. The reactive transient stability quantitative analysis system for synchronous condensers based on time constant as described in claim 5, characterized in that: Adversarial network models include: The generator produces augmented samples based on the input random information; The discriminator is used to randomly input real samples and augmented samples, and outputs the probability that the input information is a real sample; Both the generator and the discriminator are deep neural network models, and the generator and the discriminator are trained simultaneously. The loss function of the discriminator is: : ； The loss function of the generator is : ； This represents the i-th real sample. This represents the probability that the discriminator classifies a real sample. This represents the i-th augmented sample. This represents the probability that the discriminator will distinguish the expanded samples, and m represents the number of samples.
7. The reactive transient stability quantitative analysis system for synchronous condensers based on time constant as described in claim 5, characterized in that: The random information is set to d-dimensional random information, with each dimension following a standard normal distribution; Among them, the dimensionality of random information in the generation of step disturbance information sample sets, pulse disturbance information sample sets, and pseudo-random disturbance information sample sets is getting higher and higher.
8. The reactive transient stability quantitative analysis system for synchronous condensers based on time constant as described in claim 5, characterized in that: The generator includes: The input layer is used to input a noise vector, and to reshape the noise vector to obtain reshaped features; The upsampling layer upsamples the reconstructed features to obtain upsampled features. The normalization layer normalizes the upsampled features to obtain normalized features; The activation function layer performs RelU activation on the normalized features to generate activated features; Convolutional layers, which have multiple layers, repeatedly perform convolution operations on activation features to obtain convolutional features; In the output layer, the convolutional features are activated using the tanh function to obtain the output features, and the output features are then denormalized to generate augmented samples. The discriminator includes: The information input layer is used to input augmented or real samples to generate initial features; Information convolutional layers perform convolution operations on initial features to generate convolutional features; The information pooling layer performs global average pooling on the convolutional features to generate pooled features. Fully connected layers map pooled features to a low-dimensional space to generate low-dimensional features. The information output layer uses the Sigmoid function on low-dimensional features to generate the true probability of the samples.
9. The reactive transient stability quantitative analysis system for synchronous condensers based on time constant as described in claim 8, characterized in that: The loss function of the adversarial network during joint training is: ； Where G represents the generator and D represents the discriminator. Represents the true data distribution. Represents a probability distribution. express The average value on the real data distribution. express The average value over the probability distribution, where x represents the input to the discriminator and z represents random information; Represents the core value function; When training the adversarial network model, the following training loop is executed: Fixed generator G, updated discriminator D: maximize ; With a fixed discriminator D, update the generator G: minimize ; When fixing the generator G and updating the discriminator D, several noise information is used to generate augmented samples. The augmented samples are mixed with several real samples and sent to the discriminator D. For each sample input to the discriminator D, calculate the loss function. According to the loss function Backpropagation updates the weight parameters inside the discriminator D; With a fixed discriminator D, update the generator G: minimize hour: Several new random information samples are sampled and used to generate several expanded samples using generator G. The expanded samples are then mixed with several real samples and sent to discriminator D. For each sample, calculate the generator's loss function. According to the loss function Backpropagation updates the weight parameters inside the generator.
10. The reactive transient stability quantitative analysis system for synchronous condensers based on time constant according to claim 9, characterized in that: At the beginning of each loop, for each training iteration of the generator G, based on a pre-set random probability h, it is determined whether each neuron in the hidden layer is hidden. Hidden neurons are frozen, a new hidden layer structure is generated, and then the generator G is trained, updating the unfrozen weight parameters in the hidden layer. When the current loop ends, the frozen neurons are restored.