A method for estimating inertia and primary frequency modulation coefficient of power electronic devices
By combining the proxy network and particle transformation network of the cooperative adaptive adversarial particle filter module with the dual discriminator diffusion generative adversarial network, the problem of real-time and accurate estimation of the inertia and primary frequency regulation coefficient of power electronic equipment is solved, realizing long-term adaptive parameter evaluation and improving the reliability of frequency support capability assessment and operation scheduling.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HUNAN UNIV
- Filing Date
- 2026-03-27
- Publication Date
- 2026-06-09
AI Technical Summary
Existing technologies struggle to estimate the inertia and primary frequency regulation coefficient of power electronic equipment in real time, accurately, and robustly. This is especially problematic when the penetration rate of new energy sources is high and operating conditions change, leading to increased frequency fluctuations and slower recovery.
A collaborative adaptive adversarial particle filter module, including a proxy network and a particle transformation network, is adopted. By training the generator offline and updating it online, a posterior particle set is generated. Combined with a dual discriminator diffusion generative adversarial network, the inertia and primary frequency modulation coefficient of power electronic equipment are estimated.
It enables long-term, adaptive, point, and interval estimation of the inertia and primary frequency regulation coefficient of power electronic equipment, improving the robustness and engineering applicability of dynamic parameter evaluation, and maintaining estimation accuracy and stability when operating conditions change.
Smart Images
Figure CN121939440B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of power systems and their automation, and in particular to a method for estimating the inertia and primary frequency regulation coefficient of power electronic equipment. Background Technology
[0002] The equivalent inertia parameter and primary frequency regulation coefficient of power electronic equipment are key indicators characterizing its frequency support strength and frequency recovery capability, directly affecting the tuning of grid-connected / grid-linked control parameters, the configuration of primary frequency regulation support, and the assessment of grid frequency safety margin. With the increasing penetration rate of inverter interface power sources such as wind power and photovoltaics, the proportion of traditional synchronous machines is declining, making system frequency dynamics more sensitive to the inertia support and primary frequency regulation support of power electronic equipment. Simultaneously, influenced by factors such as control strategy switching, amplitude constraints, changes in resource availability, and external operating condition fluctuations, the equivalent inertia and primary frequency regulation coefficient of power electronic equipment exhibit significant time-varying and non-stationary characteristics, easily leading to problems such as increased frequency fluctuations and slower recovery. Therefore, real-time, accurate, and robust online estimation of the equivalent inertia parameter and primary frequency regulation coefficient of power electronic equipment has become an important technical foundation for frequency support capability assessment and operation scheduling.
[0003] In recent years, artificial intelligence technology has been used for parameter estimation. Methods such as artificial neural networks, deep neural networks, and long short-term memory networks are used to directly learn the nonlinear mapping between grid-connected power and frequency, enabling online identification of equivalent inertia and frequency modulation parameters. While these methods reduce the difficulty of mechanistic modeling to some extent, they typically rely on fixed offline data for training, making them prone to performance degradation when faced with changes in operating conditions and concept drift. Furthermore, most methods are primarily point estimations, lacking quantitative representation of parameter uncertainties, making it difficult to support the requirements of credibility intervals and risk boundaries in frequency safety assessments. On the other hand, particle filtering, as a typical Bayesian recursive method, can approximate the posterior distribution through Monte Carlo sampling, naturally possessing point and interval estimation capabilities and being suitable for nonlinear and non-Gaussian scenarios. However, when parameter dimensions are high or observation information is insufficient, weight degradation easily occurs, requiring frequent resampling, leading to sample depletion and decreased diversity. Moreover, its performance is still affected by the mismatch between the state model and noise statistical assumptions, making it difficult to guarantee long-term stability and robust effectiveness under conditions of high new energy penetration, time-varying noise, and rapid switching of operating conditions. Summary of the Invention
[0004] This application provides a method for estimating the inertia and primary frequency regulation coefficient of power electronic equipment. To solve the above-mentioned technical problems, this application adopts the following technical method:
[0005] In a first aspect, this application provides a method for estimating the inertia and primary frequency regulation coefficient of power electronic equipment, including:
[0006] Collect real-time data from units, areas, and systems;
[0007] The real-time data of the unit, region, and system are input into the trained cooperative adaptive adversarial particle filter module, which outputs a posterior particle set; the cooperative adaptive adversarial particle filter module includes a proxy network and a particle transformation network.
[0008] Statistical operations are performed on the posterior particle set to obtain estimated values of the inertia and primary frequency regulation coefficient of the power electronic equipment.
[0009] The training process of the trained cooperative adaptive adversarial particle filter module includes the following steps:
[0010] During the offline training phase, the generator is trained based on the training sample data, and the network parameters of the generator are saved after training is completed.
[0011] During the online training phase, the generator network parameters saved during the offline training phase are loaded. The generator obtained from the offline training is used as a proxy network to generate a prior particle set. A particle transformation network is constructed to correct the distribution of the prior particles so that the generated posterior particles are consistent with the posterior features of the parameters under the current measurement conditions.
[0012] The particle transformation network and the proxy network are updated to obtain the trained proxy network and particle transformation network.
[0013] Optionally, the step of training the generator based on training sample data and saving the network parameters of the generator after training includes:
[0014] Collect historical data on grid-connected operation of power electronic equipment, and construct training sample data related to inertia and primary frequency regulation coefficient;
[0015] The training sample data is input into the dual-discriminator diffusion generative adversarial network for training, and the network parameters of the generator are saved after training is completed; the dual-discriminator diffusion generative adversarial network includes a generator, a first discriminator and a second discriminator.
[0016] Optionally, the generator takes training sample data as input to generate a parameter sample set of inertia and primary frequency modulation coefficients, and uses this parameter sample set as a priori particle set. The generator internally embeds a forward noise addition mechanism and a reverse noise reduction mechanism of the diffusion model. The reverse noise reduction mechanism of the diffusion model is configured to: take training sample data as input, first perform progressive noise addition processing on the training sample data, then iteratively optimize the generator's model parameters by learning the reverse noise reduction generation process of the noise-added samples, and finally, based on the iteratively optimized model parameters, gradually restore the initial pure noise to a sample set of power electronic equipment inertia and primary frequency modulation coefficients that approximates the parameter distribution of the real power system.
[0017] Optionally, the first discriminator works in conjunction with the generator; the first discriminator is used to determine the consistency of the prior particles and training samples in statistical distribution, specifically by calculating the distribution distance or the similarity of the fitted distribution curve between the prior particles and training samples, and feeds the discrimination result back to the generator as a constraint basis for optimizing the generator parameters.
[0018] Optionally, the second discriminator works in conjunction with the generator; the second discriminator is used to determine the physical feasibility of the prior particle under the constraints of frequency dynamic mechanism and power electronic equipment; wherein, the discrimination criteria corresponding to the frequency dynamic mechanism include: whether the frequency deviation of the power system corresponding to the prior particle is within the allowable range of the preset frequency, and whether the frequency oscillation damping torque contribution meets the requirements for stable system operation; the discrimination criteria corresponding to the constraints of power electronic equipment include: the voltage amplitude deviation of the equipment corresponding to the prior particle does not exceed the preset error range, the operating temperature of the equipment is not higher than the preset temperature, and the output power is within the rated power range of the equipment; and the second discriminator feeds back the physical feasibility judgment result to the generator, which together with the judgment result of the first discriminator constitutes the dual constraint conditions for the iterative optimization of generator parameters.
[0019] Optionally, updating the particle transformation network and the proxy network to obtain the trained proxy network and particle transformation network includes:
[0020] Collect real-time data on the grid-connected operation of power electronic equipment;
[0021] The real-time data is input into the agent network to generate a priori particle set;
[0022] The prior particle set is input into the particle transformation network to generate the posterior particle set.
[0023] The posterior particle set is stored in the experience pool as a historical posterior sample.
[0024] The real-time data is input into a sliding window to calculate the maximum mean difference between the current real-time data distribution and the historical sample distribution in the experience pool.
[0025] Determine whether the maximum mean difference is greater than the adaptive drift threshold;
[0026] If so, representative samples are extracted from the experience pool to form a small batch of training data;
[0027] Based on the small batch training data, the particle transformation network and the proxy network are updated to obtain the trained proxy network and particle transformation network.
[0028] Optionally, the particle transformation network loss function includes a kernel maximum mean difference loss function and an optimal regularization function.
[0029] This application has the following beneficial effects:
[0030] The method proposed in this application can achieve long-term, adaptive, point, and interval estimation of inertia and primary frequency regulation coefficients at multiple levels, including unit, region, and system. It solves the problems of model dependence and particle degradation in existing methods and can improve the robustness and engineering applicability of dynamic parameter evaluation of power electronic equipment. Attached Figure Description
[0031] Figure 1 A flowchart illustrating a method for estimating the inertia and primary frequency regulation coefficient of power electronic equipment provided in this application embodiment;
[0032] Figure 2 A flowchart illustrating the offline training phase provided in an embodiment of this application;
[0033] Figure 3 A topology diagram of the New England 39-node system provided for embodiments of this application;
[0034] Figure 4 The inertia and primary frequency modulation coefficient evaluation results are shown in the figure provided for the embodiments of this application; Figure 4 (a) is a graph showing the inertia assessment results; Figure 4 (b) is a graph showing the evaluation results of the primary frequency modulation coefficient. Detailed Implementation
[0035] To facilitate understanding by those skilled in the art, the present application will be further described below in conjunction with embodiments and accompanying drawings. The content mentioned in the embodiments is not intended to limit the present application.
[0036] To solve the above technical problems, such as Figure 1 As shown, this application proposes a method for estimating the inertia and primary frequency regulation coefficient of power electronic equipment, including:
[0037] Step S101: Collect real-time data from the unit, area, and system;
[0038] Real-time data for generating units typically includes real-time operating data for individual power electronic devices (such as wind turbine converters, photovoltaic inverters, and energy storage converters). Key data includes: equipment output active power, output voltage / frequency, converter switching status, control commands (such as primary frequency regulation trigger signals), and DC-side power. Regional real-time data typically includes the aggregated active / reactive power of all power electronic devices within the region, total load power within the region, regional bus frequency, and regional tie-line power. System real-time data typically includes system frequency, total system active power deficit, total power output on both the generation and load sides of the entire grid, tie-line power in each region, and the total primary frequency regulation response of the system.
[0039] When the system is disturbed, the frequency recovery process and the speed of the active power output response of the equipment will change with the values of the equivalent inertia and the primary frequency regulation coefficient. Therefore, the above real-time data can be used as the observation basis for estimating parameters.
[0040] Step S102: Input the real-time data of the unit, region and system into the trained cooperative adaptive adversarial particle filter module, and output the posterior particle set; the cooperative adaptive adversarial particle filter module includes a proxy network and a particle transformation network;
[0041] By inputting the real-time data from the aforementioned units, regions, and systems into the trained cooperative adaptive adversarial particle filter module, a posterior particle set can be generated. The specific process is as follows:
[0042] The aforementioned particle filtering module includes a proxy network and a particle transformation network. Real-time data from the unit, region, and system are input into the proxy network to generate a priori particle set, and the priori particle set is input into the particle transformation network to generate a posterior particle set.
[0043] The training process of the aforementioned cooperative adaptive adversarial particle filter module is generally divided into an offline training phase and an online training phase. The training process is described below:
[0044] First, during the offline training phase, the generator is trained based on the training sample data. After training, the generator's network parameters are saved, such as... Figure 2 As shown, historical data of grid-connected operation of power electronic equipment is first collected, and training sample data related to inertia and primary frequency regulation coefficient are constructed. Then, the training sample data is input into a dual-discriminator diffusion generative adversarial network for training. The dual-discriminator diffusion generative adversarial network includes a generator, a first discriminator 1, and a second discriminator 2. Then, the generator and discriminator are optimized through adversarial training, and the generator parameters are saved after training is completed.
[0045] The generator described above takes training sample data as input to generate a parameter sample set of inertia and primary frequency modulation coefficients, and uses this parameter sample set as a priori particle set. The generator embeds a forward noise addition mechanism and a reverse noise reduction mechanism of the diffusion model. The reverse noise reduction mechanism of the diffusion model is configured to: take training sample data as input, first perform progressive noise addition on the training sample data, then learn the reverse noise reduction generation process of the noise-added samples, iteratively optimize the generator's model parameters, and finally, based on the iteratively optimized model parameters, gradually restore the initial pure noise to a sample set of power electronic equipment inertia and primary frequency modulation coefficients that approximate the parameter distribution of the real power system.
[0046] The first discriminator 1 works in conjunction with the generator. The first discriminator is used to determine the consistency of the prior particles and training samples in statistical distribution. Specifically, it is achieved by calculating the distribution distance or the similarity of the fitted distribution curve between the prior particles and training samples. The discriminator results are fed back to the generator as a constraint for optimizing the generator parameters.
[0047] The second discriminator 2 works in conjunction with the generator. The second discriminator is used to determine the physical feasibility of the prior particle under the constraints of frequency dynamic mechanism and power electronic equipment. The discrimination criteria corresponding to the frequency dynamic mechanism include: whether the frequency deviation of the power system corresponding to the prior particle is within the allowable range of the preset frequency, and whether the frequency oscillation damping torque contribution meets the requirements for stable system operation. The discrimination criteria corresponding to the constraints of power electronic equipment include: the voltage amplitude deviation of the equipment corresponding to the prior particle does not exceed the preset error range, the operating temperature of the equipment is not higher than the preset temperature, and the output power is within the rated power range of the equipment. The second discriminator feeds back the physical feasibility judgment result to the generator, which together with the judgment result of the first discriminator constitutes the dual constraint conditions for the iterative optimization of the generator parameters.
[0048] The loss functions of the first discriminator 1 and the second discriminator 2 are as follows:
[0049] (1)
[0050] (2)
[0051] in, This represents the loss function of the first discriminator (1). To obtain noise distribution The noise vector obtained from sampling; The prior particles output by the generator; This represents the output probability of the first discriminator (1). For the generator, an adversarial loss term; To obtain from the training data distribution The real sample obtained through sampling; This represents the physical loss function of the second discriminator 2; For system inertia; This is the primary frequency modulation coefficient; This is for frequency deviation; The rate of change of frequency deviation; For system power deviation, This is the physical residual mapping function, used to map the frequency dynamic mechanism imbalance within the brackets to the physical constraint characteristics of the second discriminator 2.
[0052] The diffusion mechanism gradually adds Gaussian noise to the training samples and gradually denoises them during the reverse process, enabling the generator to learn a smooth mapping from the noise distribution to the true prior distribution, thereby significantly improving the stability of training convergence and the diversity of generated samples.
[0053] (3)
[0054] in, Represents the diffusion loss function; This represents a noise estimate derived from real samples. Represents the square of the L2 norm; This represents the expected value.
[0055] Based on this, the overall loss function of the generator is defined as follows:
[0056] (4)
[0057] in, This represents the overall loss function of the generator.
[0058] Historical data is used to train the generator. In the offline phase, sufficient historical samples are used to learn the prior structure of the parameter distribution. A dual discriminator simultaneously constrains statistical consistency and physical feasibility, enabling the rapid generation of prior particles that closely approximate the true parameter range and exhibit good diversity under real-time measurement conditions in the online phase. This significantly reduces the proportion of invalid particles and prior bias, thereby mitigating sources of estimation error.
[0059] During the online training phase, the generator parameters saved during the offline training phase are loaded. The generator obtained from the offline training is used as a surrogate network, and a particle transformation network is constructed to correct the distribution of prior particles, so that the generated posterior particles are consistent with the posterior features of the parameters under the current measurement conditions. This avoids the sample degradation problem caused by repeated resampling in traditional particle filtering, thereby improving the approximation accuracy and stability of the true parameters while maintaining particle diversity. The particle transformation network and the surrogate network are then updated to obtain the trained surrogate network and particle transformation network.
[0060] The online training process is as follows:
[0061] First, real-time data on the grid-connected operation of power electronic equipment is collected. Then, the real-time data is input into the agent network to generate a set of prior particles that combines diversity and physical consistency.
[0062] (5)
[0063] in, It is a moment Real-time data; For proxy networks; To indicate time The generated first A priori particle; For the first One noise vector.
[0064] The prior particle set is then input into the particle transformation network to generate the posterior particle set:
[0065] (6)
[0066] in, Let i be the i-th posterior particle generated at time t. It is a particle transformation network.
[0067] The posterior particle set is stored in the experience pool to provide a stable reference and suppress estimation jitter, serving as historical posterior samples to also provide a stable reference and suppress estimation jitter. Furthermore, real-time data is input into a sliding window to calculate the maximum mean difference between the current real-time data distribution and the historical distributions in the experience pool in real time.
[0068] (7)
[0069] Where Nw represents the number of samples in the sliding window; Nb represents the number of samples in the experience pool. To represent the number of the current sliding window One real-time sample; Represents the first in the experience pool One reference sample, This is the Gaussian kernel function.
[0070] Then determine whether the maximum mean difference is greater than the adaptive drift threshold;
[0071] The aforementioned adaptive drift threshold is set using Bernstein's inequality, and parameter updates are only triggered when a significant drift is detected, thus avoiding redundant calculations.
[0072] (8)
[0073] in, This is the drift detection sensitivity coefficient; This is the drift detection threshold.
[0074] If the above judgment result is yes, then representative samples are extracted from the experience pool to form a small batch of training data. Then, based on the small batch of training data, the particle transformation network and the surrogate network are updated to obtain the trained surrogate network and particle transformation network.
[0075] Particle transformation networks aim to minimize the maximum mean difference, while the supremum operator in the maximum mean difference simultaneously maximizes the distribution difference in the function space, thus naturally forming a minimax adversarial learning structure.
[0076] (9)
[0077] in, This is the maximum mean difference loss function; For supremum operators, it is indicated that in the regenerating kernel Hilbert space... Find the optimal discriminant function that maximizes the difference between the two distributions. ; For the transformed particle distribution and This represents the true posterior distribution; These are real posterior samples.
[0078] To simplify the minimax optimization problem, a kernel method is adopted. By restricting the function space to a unit sphere in the kernel functional space, the two-layer adversarial optimization is transformed into a single-layer optimization, reducing computational overhead.
[0079] (10)
[0080] in, is the kernel maximum mean difference loss function; N is the number of particles; m represents all particles of different i in the same batch; This represents the m-th prior particle generated at time t; This represents the weight of the m-th prior particle; Gaussian kernel function.
[0081] To enhance distribution consistency and prevent overfitting, optimal transfer regularization is introduced, which penalizes the deviation between the prior distribution and the transformed distribution. The optimal transfer regularization function is:
[0082] (11)
[0083] The final loss function of the particle transformation network is:
[0084] (12)
[0085] For a proxy network, the goal is to make the prior distribution approximate the historical posterior distribution as closely as possible, preventing particles from deviating from their true state. Its loss function is:
[0086] (13)
[0087] in, Indicates the first A historical posterior particle.
[0088] Based on small batch training data, the loss functions of equations (12) and (13) are updated to update the surrogate network and the particle transformation network, so that both the prior and posterior ends are close to the new working conditions, thereby maintaining the estimation accuracy and robustness for a long time. After the update is completed, the trained surrogate network and particle transformation network can be obtained and used as the trained cooperative adaptive adversarial particle filter module for parameter estimation in the final stage.
[0089] The advantages of this method are as follows: First, the proxy network utilizes the prior parameter structure learned offline to quickly generate prior particles within a reasonable range and with sufficient diversity under real-time measurement conditions, reducing invalid samples and prior bias. Second, the particle transformation network corrects the distribution of prior particles with reference to historical posterior samples in the experience pool, avoiding sample degradation caused by traditional resampling, and generating posterior parameter particles that are closer to the current measurement conditions while maintaining diversity. Finally, a sliding window compares the current and historical measurement distributions, and triggers small-batch updates only when significant drift is detected through an adaptive threshold, simultaneously correcting both the prior and posterior generation parts. This allows the model to track changes in operating conditions in a timely manner while avoiding error accumulation and overfitting caused by frequent updates, thereby providing a more accurate inertia and first-order frequency modulation coefficient point estimate and a more reliable confidence interval in the long term.
[0090] Step S103: Perform statistical calculations on the posterior particles to obtain estimated values of the inertia and primary frequency regulation coefficient of the power electronic equipment.
[0091] Posterior particle It is a two-dimensional vector containing inertia and primary frequency modulation coefficients. The estimated values of inertia and primary frequency modulation coefficients are obtained by averaging the values of posterior particles.
[0092] (14)
[0093] in, yes A two-dimensional point estimate vector of two types of values: moment of inertia and primary frequency modulation coefficient.
[0094] The interval estimate is the empirical quantile of the posterior particle:
[0095] (15)
[0096] in, It is the confidence level; and posterior particles in and Empirical quantiles.
[0097] Simulation analysis:
[0098] Build such a power system simulation platform Figure 3 The New England 39-node system is illustrated, and an operating scenario dominated by power electronic equipment is constructed: synchronous units include 800MW, 800MW, and 700MW grid-connected converter-type wind turbines (denoted as units W1, W2, and W3), whose grid connection points are connected to the corresponding buses via power electronic interfaces. The remaining synchronous units in the system can provide 24,470 MW·s of inertia resources; simultaneously, adjustable virtual inertia resources are deployed at nodes 30, 31, 32, 33, 34, 35, 36, 37, and 39, with a total adjustable virtual inertia resource of 42,000 MW·s within the system. Furthermore, to support multi-level evaluation and control at the unit, regional, and system levels, the 39-node system is divided into three regions based on electrical coupling and topological connectivity. Measurements at each region level are obtained by aggregating measurements such as frequency and active power from multiple power electronic devices within the region, while system-level measurements are formed by aggregation from the entire network.
[0099] Figure 4 This is a graph showing the inertia and primary frequency regulation coefficient evaluation results of the method proposed in this application. By applying a bounded random active power disturbance to the New England 39-node system, this application presents the evaluation results for unit W1, region 1, and multiple system levels over a continuous operating period. Figure 4 (a) Show the point estimation curve of inertia and its interval estimation results with 80%, 90%, and 95% confidence intervals. Figure 4 (b) The point estimation curve of the primary frequency regulation coefficient and the interval estimation results of its corresponding confidence interval are shown. As can be seen from the figure, the method of this application can continuously output stable point and interval estimation results during the operating cycle, and can track parameter changes when operating conditions change; the interval estimation can quantitatively characterize the uncertainty of the frequency support parameters of power electronic equipment, and when the data distribution changes, the collaborative adaptive update mechanism can trigger network updates as needed, suppress the degradation of estimation performance and maintain long-term effectiveness. Therefore, this application can realize real-time evaluation of the inertia and primary frequency regulation coefficient of power electronic equipment at the unit, region and system levels, providing a quantifiable basis for frequency support capability analysis and operation decision-making.
[0100] In summary, the method proposed in this application can achieve long-term, adaptive, point, and interval estimation of inertia and primary frequency regulation coefficients at multiple levels, including unit, region, and system. It solves the problems of model dependence and particle degradation in existing methods, and can improve the robustness and engineering applicability of dynamic parameter evaluation of power electronic equipment.
[0101] In some embodiments, this application also provides a computer system including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.
[0102] This application also provides a computer-readable storage medium for storing a computer program. This computer-readable storage medium can be applied to a computer device, and the computer program causes the computer device to execute the corresponding processes in the methods described above in the embodiments of this application; for brevity, further details are omitted here.
[0103] The above embodiments are preferred implementations of this application. In addition, this application can be implemented in other ways. Any obvious substitutions without departing from the concept of this technical solution are within the protection scope of this application.
[0104] To facilitate understanding by those skilled in the art of the improvements made by this application compared to the prior art, some of the accompanying drawings and descriptions have been simplified, and for clarity, some other elements have been omitted from this application. Those skilled in the art should realize that these omitted elements may also constitute the content of this application.
Claims
1. A method for estimating the inertia and primary frequency regulation coefficient of power electronic equipment, characterized in that, include: Collect real-time data from units, areas, and systems; The real-time data of the unit, region, and system are input into the trained cooperative adaptive adversarial particle filter module, which outputs a posterior particle set; the cooperative adaptive adversarial particle filter module includes a proxy network and a particle transformation network. Statistical operations are performed on the posterior particle set to obtain estimated values of the inertia and primary frequency regulation coefficient of the power electronic equipment. The training process of the trained cooperative adaptive adversarial particle filter module includes the following steps: During the offline training phase, historical data of grid-connected operation of power electronic equipment is collected to construct training sample data related to inertia and primary frequency regulation coefficient. The training sample data is input into a dual-discriminator diffusion generative adversarial network for training, and the network parameters of the generator are saved after training is completed; the dual-discriminator diffusion generative adversarial network includes a generator, a first discriminator, and a second discriminator. During the online training phase, the generator network parameters saved during the offline training phase are loaded. The generator obtained from the offline training is used as a proxy network to generate a prior particle set. A particle transformation network is constructed to correct the distribution of the prior particles so that the generated posterior particles are consistent with the posterior features of the parameters under the current measurement conditions. Collect real-time data on the grid-connected operation of power electronic equipment; The real-time data is input into the agent network to generate a priori particle set; The prior particle set is input into the particle transformation network to generate the posterior particle set. The posterior particle set is stored in the experience pool as a historical posterior sample. The real-time data is input into a sliding window to calculate the maximum mean difference between the current real-time data distribution and the historical sample distribution in the experience pool. Determine whether the maximum mean difference is greater than the adaptive drift threshold; If so, representative samples are extracted from the experience pool to form a small batch of training data; Based on the small batch training data, the particle transformation network and the proxy network are updated to obtain the trained proxy network and particle transformation network. The second discriminator works in conjunction with the generator. The second discriminator is used to determine the physical feasibility of the prior particle under the constraints of frequency dynamic mechanism and power electronic equipment. The discrimination criteria corresponding to the frequency dynamic mechanism include: whether the frequency deviation of the power system corresponding to the prior particle is within the allowable range of the preset frequency, and whether the frequency oscillation damping torque contribution meets the requirements for stable system operation. The discrimination criteria corresponding to the constraints of power electronic equipment include: the voltage amplitude deviation of the equipment corresponding to the prior particle does not exceed the preset error range, the operating temperature of the equipment is not higher than the preset temperature, and the output power is within the rated power range of the equipment. The second discriminator feeds back the physical feasibility determination result to the generator, which together with the determination result of the first discriminator constitutes the dual constraint conditions for the iterative optimization of the generator parameters.
2. The method according to claim 1, characterized in that, The generator takes training sample data as input to generate a parameter sample set of inertia and primary frequency modulation coefficients, and uses this parameter sample set as a priori particle set. The generator internally embeds a forward noise addition mechanism and a reverse noise reduction mechanism of the diffusion model. The reverse noise reduction mechanism of the diffusion model is configured to: take training sample data as input, first perform progressive noise addition processing on the training sample data, then learn the reverse noise reduction generation process of the noise-added samples, iteratively optimize the generator's model parameters, and finally, based on the iteratively optimized model parameters, gradually restore the initial pure noise to a sample set of power electronic equipment inertia and primary frequency modulation coefficients that approximates the parameter distribution of the real power system.
3. The method according to claim 2, characterized in that, The first discriminator works in conjunction with the generator. The first discriminator is used to determine the consistency of the prior particles and training samples in statistical distribution. Specifically, it is achieved by calculating the distribution distance or the similarity of the fitted distribution curve between the prior particles and training samples. The discriminator also feeds back the discrimination result to the generator as a constraint for optimizing the generator parameters.
4. The method according to claim 1, characterized in that, The loss function of the particle transformation network includes the kernel maximum mean difference loss function and the optimal regularization function.