Modeling Method and System for Primary Frequency Regulation of Photovoltaic Power Plants Considering Low Voltage Ride-Through Characteristics

Abstract

This invention discloses a method and system for primary frequency regulation modeling of photovoltaic (PV) power plants considering low-voltage ride-through characteristics, belonging to the field of primary frequency regulation modeling technology. The method includes acquiring active and reactive transient data of each power generation unit in a PV power plant during low-voltage ride-through and primary frequency regulation disturbances; extracting low-dimensional potential dynamic features through an adversarial feature extraction network; inputting these features into a deep embedding clustering network for clustering and selecting central representative units; establishing an aggregated equivalent transfer function and converting it into a state-space physical constraint equation; constructing a physical information neural network with time as input and state variables as output; setting the parameters to be identified as Bayesian random variables; training based on a joint optimization objective; and obtaining the parameter identification results for each cluster through posterior sampling; and aggregating the equivalent models of each cluster into an overall primary frequency regulation model for the PV power plant according to capacity multiples. This invention can construct a high-precision, physically consistent, and uncertainty-quantified equivalent model for primary frequency regulation of PV power plants.

Description

Technical Field

[0001] This invention relates to the field of primary frequency regulation modeling technology, and more specifically, to a method and system for primary frequency regulation modeling of photovoltaic power plants that takes into account low voltage ride-through characteristics. Background Technology

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

[0003] With the continuous expansion of photovoltaic (PV) power generation capacity, the PV penetration rate in the power system is constantly increasing. PV power generation units are connected to the grid via power electronic converters and lack the inherent rotational inertia and damping characteristics of synchronous machines. When the system experiences disturbances such as sudden load changes, unit tripping, or fluctuations in renewable energy output, a high proportion of PV integration leads to a reduction in the system's equivalent inertia, posing a severe challenge to the frequency stability of the power system. Therefore, establishing a model that can accurately describe the dynamic response characteristics of primary frequency regulation in PV power plants is of great significance.

[0004] The inventors discovered that existing photovoltaic power plant primary frequency regulation modeling methods can describe the overall effect of photovoltaic power plants participating in frequency regulation to a certain extent. However, in terms of cluster modeling, existing methods mostly rely on manually selected feature quantities or use shallow distance metrics for classification, making it difficult to fully extract high-dimensional and nonlinear time-series features during low-voltage ride-through. In terms of parameter identification, recursive least squares method relies on linearized models and is not applicable to nonlinear primary frequency regulation processes that include low-voltage ride-through control links. While particle swarm optimization algorithm can handle nonlinear problems, it has a large computational load and the stability of the results is easily affected by parameter settings. At the same time, traditional methods are difficult to explicitly introduce physical constraints and uncertainty assessments, resulting in insufficient physical consistency of identification results and a lack of generalization ability for complex operating conditions. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for primary frequency regulation modeling of photovoltaic power plants that considers low voltage ride-through characteristics.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solution: In a first aspect, the technical solution of the present invention provides a method for modeling the primary frequency regulation of a photovoltaic power station considering low voltage ride-through characteristics, comprising: The active power transient response data and reactive power transient response data of each photovoltaic power generation unit in the photovoltaic power station during low voltage ride-through and primary frequency regulation disturbance are obtained, and the low-dimensional potential dynamic features of each photovoltaic power generation unit are extracted based on the adversarial contrast feature extraction network. Low-dimensional potential dynamic features are input into a deep embedded clustering network. Feature reconstruction loss constrains the encoding of features to preserve time-series information, and cluster distribution optimization loss adjusts the sample soft assignment probability to obtain the cluster label corresponding to each photovoltaic power generation unit and determine the central representative unit of each cluster. Based on the primary frequency modulation perturbation input signal and active power response output signal of the central representative unit of each cluster, standardized input-output time series samples for parameter identification are generated. Based on the dynamic response mechanism of primary frequency regulation of photovoltaic power generation unit, an aggregated equivalent transfer function model from disturbance input to active power response output is established, and it is converted into a state-space model to construct physical constraint equations; A physical information neural network is constructed using time variables as input and predicted values ​​of state variables as output. The parameters to be identified and the implicit initial states are set as Bayesian random variables, and the network is trained based on a joint optimization objective. After training, the posterior distribution of the parameters to be identified is sampled multiple times to obtain the identification results of each cluster of parameters. The joint optimization objective includes: data fitting error, physical equation residual error, discrete simulation error of the equivalent transfer function, initial state error, and divergence between the Bayesian posterior distribution and the prior distribution. For each cluster, a primary frequency regulation equivalent transfer model is constructed based on the aggregated equivalent transfer function model and the parameter identification results of the cluster; the primary frequency regulation equivalent models of each cluster are aggregated to obtain the overall primary frequency regulation response model of the photovoltaic power station.

[0007] In at least one embodiment, active power transient response data and reactive power transient response data of each photovoltaic power generation unit in the photovoltaic power station during low voltage ride-through and primary frequency regulation disturbance are acquired, and low-dimensional potential dynamic features of each photovoltaic power generation unit are extracted based on an adversarial contrast feature extraction network, specifically including: For each photovoltaic power generation unit, its active power sequence and reactive power sequence during low voltage ride-through and primary frequency regulation disturbance are stacked according to the channel dimension to form a multi-channel time series sample, thus obtaining the transient response sample of each photovoltaic power generation unit. Gaussian noise and amplitude scaling are applied to each original transient response sample to construct the corresponding enhanced transient response sample; The original transient response samples and their corresponding enhanced transient response samples are input into an adversarial contrastive feature extraction network to extract low-dimensional potential dynamic features of each photovoltaic power generation unit. Contrastive learning is used to constrain the feature consistency of the same source disturbance view, and adversarial learning is used to weaken the influence of noise disturbance and operating condition changes on feature distribution.

[0008] In at least one embodiment, the adversarial contrastive feature extraction network includes a one-dimensional convolutional encoder, a contrastive learning projector head, and an adversarial discriminator. Among them, the one-dimensional convolutional encoder gradually compresses the time dimension and expands the feature channels through a multi-level one-dimensional convolutional structure, which is used to extract the rapid transient change features in the early stage of low voltage ride-through, the power recovery features in the voltage recovery stage, and the dynamic response features in the primary frequency modulation process. The contrastive learning projector maps the features output by the encoder to the contrastive learning space. Through contrastive learning loss, it narrows the feature distance between the original samples and the enhanced samples of the same photovoltaic power generation unit, and widens the feature distance between samples of different photovoltaic power generation units or different response modes. The adversarial discriminator is used to distinguish between the original sample features and the enhanced sample features. The encoder is trained adversarially to make the distribution of the enhanced sample features close to that of the original sample features, thereby reducing the impact of noise disturbances and changes in operating conditions on the feature distribution. The joint optimization objective of the adversarial contrastive feature extraction network can be specifically expressed as:

[0009] In the formula, Indicates the contrast learning loss. Indicates the loss due to resistance constraints. This represents the weighting coefficient.

[0010] In at least one embodiment, the deep embedding clustering network includes a feature-level autoencoder and a deep embedding clustering layer; The feature-level autoencoder maps latent dynamic features to the latent space and reconstructs the latent space features back to the original feature space. It uses feature reconstruction loss to constrain the encoding features and preserve temporal information. Specifically, it is represented as follows:

[0011] In the formula, Represents the feature reconstruction loss, This represents the potential dynamic characteristics after standardization. Represents reconstructed features; In the deep embedding clustering layer, the soft assignment probability is calculated based on the latent space features of photovoltaic power generation units and the distance between each cluster center. The feature representation and cluster center are adjusted by optimizing the cluster distribution loss. The joint loss of the deep embedding clustering network is constructed based on the feature reconstruction loss and the clustering distribution optimization loss, which can be specifically expressed as:

[0012] in, This represents the clustering distribution optimization loss; Represents the feature reconstruction loss; This indicates the weight of the reconfigured constraint.

[0013] In at least one embodiment, the aggregated equivalent transfer function model is specifically represented as follows:

[0014] In the formula, This represents the equivalent time constant of the output stage. This represents the equivalent time constant of the drooping channel. Represents the equivalent time constant of the differential channel. This indicates the droop channel aggregation gain. Indicates the differential channel aggregation gain. , , , and This constitutes the set of parameters to be identified; The aggregated equivalent transfer function model is transformed into a state-space model with physical meaning, and physical constraint equations are constructed using state variables, specifically as follows:

[0015] In the formula, Represents the state variables of the first-order inertial element in the drooping channel; This represents the state variable of the first-order inertial element in the differential channel; Indicates a disturbance input; This indicates the output deviation of the primary frequency regulation active power response of a photovoltaic power generation unit or photovoltaic power station; , , They represent , , The derivative with respect to time.

[0016] In at least one embodiment, on the network structure of the physical information neural network, the time variable is first normalized, then Fourier time features are constructed, and then a residual multilayer perceptron is used to perform nonlinear mapping on the Fourier features to output the predicted values ​​of the state variables.

[0017] In at least one embodiment, the joint optimization objective is specifically expressed as:

[0018] In the formula, This refers to the data error between the neural network's predicted output and the measured output. The physical residual error is obtained from the physical constraint equation of primary frequency modulation. The error between the simulated output and the measured output of the equivalent transfer function constructed from the current identified parameters. The initial state constraint error, Let be the divergence between the posterior and prior distributions. , , , and These are the weights of the corresponding loss terms.

[0019] In at least one embodiment, the training of the physical information neural network employs a phased training strategy, specifically including: In the first training phase, deterministic warm-up training is performed using the posterior mean of the parameters to be identified, so that the neural network learns the basic response relationship between the primary frequency modulation input and the active power output first. Then, physical equation residual constraints and equivalent transfer function simulation constraints are gradually introduced to constrain the network output state to gradually satisfy the primary frequency modulation dynamic equation while updating the parameters to be identified. In the second training phase, Bayesian variational inference is initiated, posterior sampling is performed on the parameters to be identified and the implicit initial state, and the weight of the Bayesian divergence term is gradually increased so that the identification results are constrained by both the physical equation and the posterior distribution while fitting the measured response. Through Bayesian variational inference, the model can obtain the mean, variance and confidence interval of the parameters while maintaining data fit and physical consistency.

[0020] In at least one embodiment, the overall primary frequency regulation response model of the photovoltaic power station is specifically represented as follows:

[0021] In the formula, Indicates the first The equivalent capacity multiple of each cluster; This represents a primary frequency modulation disturbance input; Indicates the first The first-order frequency modulation equivalent transfer model for a cluster is specifically represented as follows:

[0022] In the formula, Indicates the first The equivalent time constant of the output stage of each cluster Indicates the first Equivalent time constants for each cluster's drooping channel Indicates the first The equivalent time constant of the differential channels of each cluster, Indicates the first Cluster-specific down-channel aggregation gain Indicates the first Cluster differential channel aggregation gain.

[0023] Secondly, the technical solution of the present invention also provides a photovoltaic power station primary frequency regulation modeling system considering low voltage ride-through characteristics, comprising: The data acquisition and feature extraction module is configured to: acquire active power transient response data and reactive power transient response data of each photovoltaic power generation unit in the photovoltaic power station during low voltage ride-through and primary frequency regulation disturbance, and extract low-dimensional potential dynamic features of each photovoltaic power generation unit based on the adversarial comparison feature extraction network. The deep clustering module is configured to: input low-dimensional potential dynamic features into a deep embedded clustering network; use feature reconstruction loss to constrain the encoding of features to preserve time-series information; use cluster distribution optimization loss to adjust the sample soft assignment probability; obtain the cluster label corresponding to each photovoltaic power generation unit; and determine the central representative unit of each cluster. Based on the primary frequency modulation perturbation input signal and active power response output signal of the central representative unit of each cluster, generate standardized input-output time series samples for parameter identification. The physical constraint construction module is configured to: establish an aggregated equivalent transfer function model from disturbance input to active power response output based on the primary frequency regulation dynamic response mechanism of the photovoltaic power generation unit, and convert it into a state-space model to construct physical constraint equations; The Bayesian uncertainty quantification module is configured to: construct a physical information neural network with time variables as input and predicted state variables as output; set the parameters to be identified and the implicit initial state as Bayesian random variables; and train them based on a joint optimization objective; after training, the posterior distribution of the parameters to be identified is sampled multiple times to obtain the identification results of each cluster of parameters; the joint optimization objective includes: data fitting error, physical equation residual error, discrete simulation error of the equivalent transfer function, initial state error, and divergence between the Bayesian posterior distribution and the prior distribution; The equivalent model aggregation module is configured to: for each cluster, construct a primary frequency regulation equivalent transfer model based on the aggregated equivalent transfer function model and the parameter identification results of the cluster; aggregate the primary frequency regulation equivalent models of each cluster to obtain the overall primary frequency regulation response model of the photovoltaic power station.

[0024] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention designs the clustering and parameter identification stages as an organic whole. The clustering stage directly uses raw transient time-series data of active and reactive power as input, and automatically extracts low-dimensional potential dynamic features through an adversarial feature extraction network, avoiding the problem of relying on manual feature selection in traditional methods. Simultaneously, the contrastive learning and adversarial discrimination mechanisms synergistically enhance noise resistance and robustness. Based on this, the optimal number of clusters is automatically selected using silhouette coefficients, achieving an optimal trade-off between accuracy preservation and model simplification. The selection result is applied to all subsequent stages, including deep embedding clustering, selection of central representative units, parameter identification sample quantity, and final aggregation term count. In the identification stage, for the central representative unit of each cluster, the aggregation equivalent transfer function is converted into a state-space model and physical constraint equations are constructed, hard-constraining Bayesian objects in the form of physical residuals. The output of the information neural network ensures that the identification results meet the dynamic response law of photovoltaic primary frequency regulation, overcoming the shortcomings of traditional methods that lack physical constraints and thus have insufficient physical consistency in the identification results. At the same time, the parameters to be identified are modeled as Bayesian random variables, and the posterior distribution and 95% confidence interval of the parameters are obtained through variational inference, overcoming the problem that traditional deterministic identification methods cannot quantify uncertainty. A network structure combining Fourier time feature coding and residual multilayer perceptron is adopted to effectively improve the fitting ability of the rapid transient process after low voltage ride-through. Finally, the equivalent models of each cluster are aggregated into the overall primary frequency regulation model of the photovoltaic power station according to the capacity multiple, realizing the fully automated construction of the equivalent model that can be used for transient stability simulation from measured data, which has high precision, strong robustness, physical interpretability and uncertainty quantification capabilities.

[0025] The advantages of the present invention, as well as its additional advantages, will be described in detail in the following specific embodiments. Attached Figure Description

[0026] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute a limitation thereof.

[0027] Figure 1 This is a schematic diagram of the photovoltaic power station primary frequency regulation modeling method considering low voltage ride-through characteristics disclosed in Embodiment 1 of the present invention; Figure 2 This is a flowchart of the photovoltaic power station primary frequency regulation modeling method considering low voltage ride-through characteristics disclosed in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the algorithm for the photovoltaic power station primary frequency regulation modeling method considering low voltage ride-through characteristics disclosed in Embodiment 1 of the present invention; Figure 4 This is a line graph of the evaluation of the optimal number of clusters disclosed in Embodiment 1 of the present invention; Figure 5This is a visualization of the clustering results disclosed in Embodiment 1 of the present invention; Figure 6 This is the active transient response curve of each photovoltaic unit within a cluster as disclosed in Embodiment 1 of the present invention; Figure 7 This is a dynamic response error diagram of the aggregation results and measured values ​​disclosed in Embodiment 1 of the present invention; Figure 8 This is an identification result diagram of each central representative unit disclosed in Embodiment 1 of the present invention; Figure 9 This is a comparison diagram of the parameters to be identified and the true values ​​disclosed in Embodiment 1 of the present invention; Figure 10 This is a comparison chart of the total identification results and measured values ​​of the photovoltaic power station disclosed in Embodiment 1 of the present invention. Detailed Implementation

[0028] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0029] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0030] It should be noted that the terminology used herein is for describing particular embodiments only and is not intended to limit the exemplary embodiments of the present invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof. It should be noted that, without conflict, the various embodiments and features within those embodiments can be combined with each other. The embodiments will now be described in detail with reference to the accompanying drawings.

[0031] Example 1 With the continuous expansion of installed capacity of new energy power generation, the penetration rate of photovoltaic (PV) power generation in the power system is constantly increasing. PV power generation units are typically connected to the grid via power electronic converters and lack the inherent rotational inertia and damping characteristics of synchronous machines. When disturbances such as sudden load changes, unit tripping, or fluctuations in new energy output occur in the power system, a high proportion of PV integration leads to a reduction in the system's equivalent inertia, placing higher demands on the stable operation of the power system. To improve the frequency dynamic characteristics under high-proportion PV grid connection conditions, PV power plants need to possess a certain degree of active frequency support capability. Therefore, establishing a model that can accurately describe the primary frequency regulation dynamic response characteristics of PV power plants is of great significance for improving the frequency stability of the power system.

[0032] Existing modeling methods for primary frequency regulation of photovoltaic (PV) power plants mostly employ system-level equivalent models, average frequency response models, or single-unit equivalent models. While these methods can describe the overall effect of PV power plants participating in frequency regulation to a certain extent, they often struggle to balance the differences between PV power generation units within the power plant and the simulation efficiency of large-scale models. Furthermore, current research lacks consideration of low-voltage ride-through characteristics. Therefore, to reduce the simulation complexity of large-scale PV power plants while ensuring modeling accuracy, it is necessary to perform cluster-based equivalent modeling of PV power plants based on the dynamic operating characteristics of PV power generation units during primary frequency regulation and low-voltage ride-through processes.

[0033] From the perspective of clustering modeling, existing photovoltaic power plant clustering methods mostly rely on manually selected features or use shallow distance metrics for classification. These methods have certain applicability under steady-state equivalent or simple operating conditions, but when facing the primary frequency regulation process that considers the low voltage ride-through stage, they need to process high-dimensional and nonlinear time series data. Their feature representation ability is relatively limited, making it difficult to fully extract the nonlinear time series features in the primary frequency regulation process of photovoltaic power plants.

[0034] From the perspective of parameter identification, existing parameter identification methods mostly aim to minimize input-output errors. Recursive least squares methods rely on linearized models and are unsuitable for nonlinear primary frequency modulation processes involving low-voltage ride-through control. While particle swarm optimization can handle nonlinear problems, it involves large computational costs and the stability of the results is easily affected by parameter settings. Furthermore, traditional methods struggle to explicitly introduce physical constraints, parameter boundaries, and uncertainty assessments, leading to insufficient physical consistency of the identification results and inadequate generalization ability under complex operating conditions.

[0035] In summary, for photovoltaic power plant clustering primary frequency regulation modeling technology, traditional clustering methods rely on manual selection of feature quantities, making it difficult to fully extract the nonlinear time-series features in the photovoltaic primary frequency regulation process; traditional parameter identification methods suffer from insufficient identification accuracy, inadequate expression of physical constraints, and insufficient adaptability to complex operating conditions when facing nonlinear photovoltaic models containing low-voltage ride-through control links.

[0036] To overcome the shortcomings of the prior art, in a typical embodiment of the present invention, such as... Figures 1 to 10 As shown in the figure, this embodiment discloses a primary frequency regulation modeling method for photovoltaic power plants considering low-voltage ride-through characteristics. It uses active and reactive transient response sequences as clustering criteria and incorporates transfer function structure, state-space constraints, and Bayesian uncertainty quantification into the parameter identification process to generate an equivalent model of primary frequency regulation for photovoltaic power plants that considers low-voltage ride-through characteristics, has clear physical meaning, strong noise resistance, good generalization ability, and is suitable for engineering simulation. The specific steps of this primary frequency regulation modeling method for photovoltaic power plants considering low-voltage ride-through characteristics are as follows: S1. Obtain the active power transient response data and reactive power transient response data of each photovoltaic power generation unit in the photovoltaic power station during the low voltage ride-through and primary frequency regulation disturbance process, and extract the low-dimensional potential dynamic features of each photovoltaic power generation unit based on the adversarial comparison feature extraction network. S2. Low-dimensional potential dynamic features are input into a deep embedded clustering network. Feature reconstruction loss is used to constrain the encoding of features to preserve time-series information. Cluster distribution optimization loss is used to adjust the sample soft assignment probability to obtain the cluster label corresponding to each photovoltaic power generation unit and determine the central representative unit of each cluster. Based on the primary frequency modulation perturbation input signal and active power response output signal of the central representative unit of each cluster, standardized input-output time series samples for parameter identification are generated. S3. Based on the dynamic response mechanism of primary frequency regulation of photovoltaic power generation unit, establish an aggregated equivalent transfer function model from disturbance input to active power response output, and convert it into a state-space model to construct physical constraint equations; S4. Construct a physical information neural network with time variables as input and predicted state variables as output. Set the parameters to be identified and the implicit initial state as Bayesian random variables and train them based on the joint optimization objective. After training, sample the posterior distribution of the parameters to be identified multiple times to obtain the identification results of each cluster of parameters. The joint optimization objective includes: data fitting error, physical equation residual error, discrete simulation error of the equivalent transfer function, initial state error, and divergence between the Bayesian posterior distribution and the prior distribution. S5. For each cluster, construct a primary frequency regulation equivalent transfer model based on the aggregated equivalent transfer function model and the parameter identification results of the cluster; aggregate the primary frequency regulation equivalent models of each cluster to obtain the overall primary frequency regulation response model of the photovoltaic power station.

[0037] The following section provides a detailed explanation of the photovoltaic power station primary frequency regulation modeling method that considers low voltage ride-through characteristics, using specific implementation methods as examples.

[0038] S1. Obtain the active power transient response data and reactive power transient response data of each photovoltaic power generation unit in the photovoltaic power station during the low voltage ride-through and primary frequency regulation disturbance process, and extract the low-dimensional potential dynamic features of each photovoltaic power generation unit based on the adversarial contrast feature extraction network.

[0039] In this step, active power transient response data and reactive power transient response data of each photovoltaic power generation unit in the photovoltaic power station during low voltage ride-through and primary frequency regulation disturbance are first obtained. These data are used to characterize the differences in dynamic response of different photovoltaic power generation units after the disturbance and to construct transient response samples. For the first... Each photovoltaic power generation unit, its active power sequence during low voltage ride-through and primary frequency regulation disturbance processes. and reactive power sequence Stack them according to the channel dimension to form multi-channel time series samples. , represented as:

[0040] In the formula, Indicates the first Transient response samples of a photovoltaic power generation unit.

[0041] Then, for each original transient response sample Data augmentation processes such as applying small Gaussian noise and random amplitude scaling are performed to construct corresponding enhanced transient response samples. This is to simulate common measurement errors, lighting fluctuations, and slight deviations in operating conditions found in actual operating data.

[0042] Next, the original transient response samples and its corresponding enhanced transient response samples The input adversarial contrastive feature extraction network extracts low-dimensional potential dynamic features of each photovoltaic power generation unit, and uses contrastive learning to constrain the feature consistency of the same source perturbation view. It also uses adversarial learning to weaken the influence of noise perturbation and operating condition changes on feature distribution, so as to improve the feature's ability to distinguish different transient response modes.

[0043] Specifically, the adversarial contrastive feature extraction network in this step includes a one-dimensional convolutional encoder, a contrastive learning projector, and an adversarial discriminator. The one-dimensional convolutional encoder sequentially employs multi-level convolution, batch normalization, ReLU activation, and pooling operations to extract multi-scale features from the input active and reactive power dual-channel transient sequences. In this process, the multi-level convolution is implemented through a multi-level one-dimensional convolutional structure. The first-level convolution captures rapid transient change features, the second-level convolution extracts medium- to long-term recovery process features, and the third-level convolution, combined with adaptive average pooling, achieves high-dimensional temporal information fusion, ultimately outputting low-dimensional latent dynamic features. By progressively compressing the time dimension and expanding the feature channels through a multi-level one-dimensional convolutional structure, this one-dimensional convolutional encoder can extract rapid transient changes in the early stage of low voltage ride-through, power recovery features in the voltage recovery stage, and medium- to long-term dynamic response features in the primary frequency modulation process. It can directly process high-dimensional time-series data composed of active and reactive power, avoiding the problem of traditional clustering methods requiring manual extraction of features such as response peaks, recovery slopes, or steady-state deviations, thus making the dynamic feature extraction process more adaptive.

[0044] Original transient response sample and enhanced transient response samples The data are input into a one-dimensional convolutional encoder and mapped to a contrast space using a contrastive learning projector. The original transient response samples of the same photovoltaic power generation unit are then processed. and enhanced transient response samples As positive sample pairs, samples from different photovoltaic power generation units or different response modes are used as negative sample pairs. The contrastive learning (InfoNCE) loss is employed to narrow the feature distance between the original and enhanced samples of the same photovoltaic power generation unit, while widening the feature distance between samples from different photovoltaic power generation units or different response modes. Specifically, the contrastive learning (InfoNCE) loss can be expressed as:

[0045] In the formula, and These represent the characteristics of the original transient response sample and the enhanced transient response sample in the contrast space, respectively. Represents the similarity function. This represents the temperature coefficient. Through this constraint, the characteristic representation of the same photovoltaic power generation unit under different perturbation views is brought closer together, while photovoltaic power generation units with significantly different dynamic responses are separated in the feature space.

[0046] In this step, an adversarial discriminator is introduced to distinguish between the original transient response sample features and the enhanced transient response sample features. A one-dimensional convolutional encoder learns to generate robust features that more closely approximate the true distribution through adversarial loss. Specifically, the adversarial constraint loss can be expressed as:

[0047] in, This indicates the number of samples in a training batch. Indicates adversarial discriminator, This represents the transient response sample characteristics before enhancement. This represents the enhanced transient response sample characteristics.

[0048] This mechanism enables the one-dimensional convolutional encoder to proactively mitigate non-essential feature shifts caused by noise disturbances, amplitude scaling, and slight changes in operating conditions during training, thereby improving the robustness of latent dynamic features to measurement noise and changes in operating conditions. Compared to shallow clustering methods that rely solely on distance metrics, this feature extraction approach is more suitable for processing high-dimensional nonlinear transient data in photovoltaic power plants that include low-voltage ride-through control mechanisms.

[0049] In this step, the one-dimensional convolutional encoder, the contrastive learning projector, and the adversarial discriminator are jointly trained. The one-dimensional convolutional encoder and projector are updated based on the contrastive learning loss, while the adversarial discriminator is updated based on the discrimination error between the original and enhanced features. The one-dimensional convolutional encoder also undergoes adversarial constraints, causing the enhanced features to gradually approach the distribution of the original features. The joint optimization objective can be specifically expressed as:

[0050] In the formula, Indicates the contrast learning loss. Indicates the loss due to resistance constraints. This represents the weight coefficients. After training, the one-dimensional convolutional encoder is retained as a photovoltaic transient feature extractor to extract low-dimensional potential dynamic features of photovoltaic power generation units.

[0051] S2. Low-dimensional potential dynamic features are input into a deep embedded clustering network. Feature reconstruction loss constrains the encoding of features to preserve time-series information. Cluster distribution optimization loss adjusts the sample soft assignment probability to obtain the cluster label corresponding to each photovoltaic power generation unit and determines the central representative unit of each cluster. Based on the primary frequency modulation disturbance input signal and active power response output signal of the central representative unit of each cluster, standardized input-output time series samples for parameter identification are generated.

[0052] In this step, the extracted potential dynamic features of the photovoltaic power generation units are first assembled into a feature matrix and then standardized to eliminate differences in dimensions and scales between different feature dimensions. The standardized potential dynamic features are then used for subsequent cluster number selection and deep embedding clustering.

[0053] In the pre-defined candidate set of cluster numbers, initial clustering calculations are performed on the standardized potential dynamic features. The intra-cluster compactness and inter-cluster separation under different cluster numbers are evaluated based on the silhouette coefficient, and the optimal cluster number is selected. The contour coefficient can be expressed as:

[0054] In the formula, Indicates sample The average distance between the sample and the sample in the same cluster. Indicates sample The average distance between the nearest heterogeneous sample and the nearest heterogeneous sample.

[0055] The standardized latent dynamic features are input into an improved deep embedding clustering network, which includes a feature-level autoencoder and a deep embedding clustering layer. The feature-level autoencoder comprises an encoding part and a decoding part. The encoding part maps the latent dynamic features to a latent space more suitable for clustering, and the decoding part reconstructs the latent space features back to the original feature space. During this process, a feature reconstruction loss constrains the encoded features to retain temporal information. The feature reconstruction loss is specifically expressed as:

[0056] In the formula, Represents the feature reconstruction loss, This represents the potential dynamic characteristics after standardization. This indicates the reconstructed features.

[0057] Perform initial clustering on the latent space representation obtained from the feature-level autoencoder, and then select the optimal number of clusters. The initial cluster centers are written into a deep embedding clustering layer. In the deep embedding clustering layer, soft assignment probabilities are calculated based on the latent space characteristics of the photovoltaic power generation units and the distances between each cluster center. , indicating the first The photovoltaic power generation unit belongs to the first The probability of each cluster. Construct the target distribution based on the soft assignment probability. Specifically, it is expressed as:

[0058] In the formula, This indicates the number of photovoltaic power generation units. (Target distribution) With soft allocation distribution The Kullback–Leibler (KL) divergence between cluster centers and feature representations is used as the cluster distribution optimization loss to adjust the cluster center distribution. The cluster distribution optimization loss is specifically expressed as:

[0059] Based on feature reconstruction loss And clustering distribution optimization loss The joint loss of the deep embedding clustering network can be specifically expressed as:

[0060] in, This represents the clustering distribution optimization loss, i.e., the target distribution. With soft allocation distribution The divergence between them; Represents the feature reconstruction loss; This represents the weights of the reconstruction constraints. Through joint loss, the clustering layer can improve the clustering contribution of samples with high confidence while preserving the dynamic information in the original transient features.

[0061] During deep embedding clustering training, the target distribution is updated at preset intervals, and the cluster label change rate is calculated. When the label change rate is lower than a preset threshold or the training reaches a preset termination condition, clustering optimization is stopped, and the cluster labels and soft assignment probabilities of each photovoltaic power generation unit are obtained.

[0062] For each cluster, calculate the distance between the latent space features of each photovoltaic power generation unit within the cluster and the cluster center, and select the real photovoltaic power generation unit with the smallest distance as the central representative unit. Specifically, this can be expressed as:

[0063] in, Indicates the first A cluster, Indicates the first Hidden spatial characteristics of individual photovoltaic power generation units Indicates the first The central characteristics of each cluster This represents the central representative unit of the cluster.

[0064] Based on the final latent space features and cluster labels, cluster evaluation indicators are calculated. A dimensionality reduction visualization method is used to map the clustering results onto a two-dimensional plane and plot the active power curves of the representative units at the center of each cluster, which intuitively reflects the degree of separation and transient response differences of different clusters.

[0065] S3. Based on the dynamic response mechanism of primary frequency regulation of photovoltaic power generation unit, establish an aggregated equivalent transfer function model from disturbance input to active power response output, and convert it into a state-space model to construct physical constraint equations.

[0066] In this step, firstly, based on the primary frequency modulation input-output time series samples of the representative unit at the center of each cluster, the primary frequency modulation disturbance input signal and the active power response output signal are obtained. Then, the primary frequency modulation disturbance input signal and the active power response output signal are sequentially subjected to initial value debiasing, input polarity unification, integer sampling delay correction and robust scale normalization to generate standardized input-output time series samples for parameter identification.

[0067] Standardized input-output time series samples are divided into training and testing samples according to a preset ratio. During model training, only training samples are used to update neural network parameters and physical parameters to be identified, while testing samples are used only to evaluate the generalization ability of the identification model. The ratio of training to testing samples is 4:1.

[0068] Then, based on the primary frequency regulation dynamic response mechanism of photovoltaic power generation units or photovoltaic power plants, an aggregated equivalent transfer function for each cluster from disturbance input to active power response output is established, specifically expressed as:

[0069] In the formula, This represents the equivalent time constant of the output stage. This represents the equivalent time constant of the drooping channel. Represents the equivalent time constant of the differential channel. This indicates the droop channel aggregation gain. Indicates the differential channel aggregation gain. , , , and This constitutes the set of parameters to be identified.

[0070] The aggregated equivalent transfer function model is converted into a state-space model with physical meaning, and physical constraint equations are constructed using state variables. The state variables are represented as follows:

[0071] In the formula, This indicates the active power output deviation of a single frequency modulation. This represents the first-order inertial state of the drooping channel. This represents the first-order inertial state of the differential channel.

[0072] Based on the aggregate equivalent transfer function structure, the physical constraint equations are constructed, specifically as follows:

[0073] In the formula, Represents the state variables of the first-order inertial element in the drooping channel; This represents the state variable of the first-order inertial element in the differential channel; Indicates a disturbance input; This indicates the output deviation of the primary frequency regulation active power response of a photovoltaic power generation unit or photovoltaic power station; , , They represent , , The derivative with respect to time. This physical constraint equation constrains the subsequent neural network output state to meet the dynamic response law of primary frequency regulation of the photovoltaic power station.

[0074] S4. Construct a physical information neural network with time variables as input and predicted state variables as output. Set the parameters to be identified and the implicit initial state as Bayesian random variables and train them based on the joint optimization objective. After training, sample the posterior distribution of the parameters to be identified multiple times to obtain the identification results of each cluster of parameters. The joint optimization objective includes: data fitting error, physical equation residual error, discrete simulation error of the equivalent transfer function, initial state error, and divergence between the Bayesian posterior distribution and the prior distribution.

[0075] In this step, firstly, the time variable... As input, predicted values ​​of state variables A physical information neural network is constructed for the output. Among them, The active power response deviation predicted by the neural network. and These are the predicted values ​​for the hidden states.

[0076] Based on this physical information neural network structure, the time variable is first normalized, then Fourier time features are constructed. Subsequently, a residual multilayer perceptron is used to perform nonlinear mapping on the Fourier features, outputting the predicted values ​​of the state variables. The Fourier time features contain multiple sets of sine and cosine basis functions to improve the network's ability to fit the rapid recovery process after low-voltage ride-through. The residual multilayer perceptron improves the stability of the network during training through residual connections.

[0077] Then, the parameters to be identified , , , , And the implicit initial state is set as a Bayesian random variable, for any parameter to be identified. The uncertainty of the parameter identification results is represented by the posterior distribution, specifically as follows:

[0078] In the formula, and These are the learnable posterior mean and posterior standard deviation, respectively. Through Bayesian variational inference, the model can obtain the parameter mean, variance, and confidence interval while maintaining data fit and physical consistency.

[0079] Prior distribution and physical boundary constraints are set for the parameters to be identified to keep the time constant parameter positive, and sign constraints are applied to the droop channel aggregation gain and the differential channel aggregation gain according to the polarity of the primary frequency modulation disturbance input. At the same time, the parameters to be identified are restricted to a preset physical range through bounded transformation to avoid invalid solutions that do not meet the photovoltaic primary frequency modulation mechanism in the parameter identification results.

[0080] Next, a joint optimization objective is constructed, and the Bayesian physical information neural network is trained based on this objective. The joint optimization objective includes data fitting error, physical equation residual error, discrete simulation error of the equivalent transfer function, initial state error, and the divergence between the Bayesian posterior and prior distributions, specifically expressed as follows:

[0081] In the formula, This refers to the data error between the neural network's predicted output and the measured output. The physical residual error is obtained from the physical constraint equation of primary frequency modulation. The error between the simulated output and the measured output of the equivalent transfer function constructed from the current identified parameters. The initial state constraint error, Let be the divergence between the posterior and prior distributions. , , , and These are the weights of the corresponding loss terms.

[0082] In the training process of the Bayesian physical information neural network, a continuous state-space model is constructed based on the currently identified parameters, and the continuous state-space model is discretized using the zero-order preservation method to obtain the perturbation input. Simulation output of aggregate equivalent transfer function under action By comparing the deviation between the simulated output of the aggregated equivalent transfer function and the measured active power response output, dynamic consistency verification is performed on the parameters to be identified.

[0083] In this step, the training of the Bayesian physical information neural network adopts a phased training strategy, specifically: In the first training phase, deterministic warm-up training is performed using the posterior mean of the parameters to be identified, so that the neural network can learn the basic response relationship between the primary frequency modulation input and the active power output first. Subsequently, physical equation residual constraints and equivalent transfer function simulation constraints are gradually introduced to constrain the network output state to gradually satisfy the primary frequency modulation dynamic equation while updating the parameters to be identified.

[0084] In the second training phase, Bayesian variational inference is initiated. Posterior sampling is performed on the parameters to be identified and the implicit initial state, and the weight of the Bayesian divergence term is gradually increased. This ensures that the identification results, while fitting the measured response, are jointly constrained by the physical equations and the posterior distribution. Through Bayesian variational inference, the model can obtain the parameter mean, variance, and confidence interval while maintaining data fit and physical consistency.

[0085] After Bayesian training is completed, the posterior distribution of the parameters to be identified is sampled multiple times to obtain the mean, standard deviation and confidence interval of each parameter to be identified. Based on the posterior parameter samples, the posterior prediction interval of the first frequency modulation response output is generated, thus giving the uncertainty boundary at both the parameter domain and time domain response levels.

[0086] The goodness of fit, root mean square error, and mean absolute error of the direct output of the neural network and the simulated output of the equivalent transfer function are calculated on the training set, the test set, and the whole dataset, respectively. The coverage of the Bayesian posterior prediction interval, the average interval width, and the normalized average interval width are also calculated to evaluate the accuracy, generalization ability, and uncertainty reliability of the parameter identification results.

[0087] S5. For each cluster, construct a primary frequency regulation equivalent transfer model based on the aggregated equivalent transfer function model and the parameter identification results of the cluster; aggregate the primary frequency regulation equivalent models of each cluster to obtain the overall primary frequency regulation response model of the photovoltaic power station.

[0088] In this step, for each cluster, a primary frequency modulation equivalent transfer model is constructed based on the aggregated equivalent transfer function model and the cluster parameter identification results. For the , For a given cluster, its primary frequency modulation equivalent transfer model can be specifically represented as:

[0089] In the formula, Indicates the first The equivalent time constant of the output stage of each cluster Indicates the first Equivalent time constants for each cluster's drooping channel Indicates the first The equivalent time constant of the differential channels of each cluster, Indicates the first Cluster-specific droop channel aggregation gain Indicates the first Cluster differential channel aggregation gain.

[0090] Based on the number or capacity of photovoltaic power generation units within each cluster, an equivalent capacity multiple is generated for each cluster. The primary frequency regulation equivalent models of each cluster are then aggregated according to the equivalent capacity multiple to obtain the overall primary frequency regulation response model of the photovoltaic power station, specifically expressed as follows:

[0091] In the formula, Indicates the first The equivalent capacity multiple of each cluster, This represents a primary frequency modulation disturbance input.

[0092] Using data from a photovoltaic power station as an example, the primary frequency regulation modeling method for photovoltaic power stations that considers low voltage ride-through characteristics in this embodiment is verified.

[0093] S1. Obtain the active power transient response data and reactive power transient response data of each photovoltaic power generation unit in the photovoltaic power station during the low voltage ride-through and primary frequency regulation disturbance process, and extract the low-dimensional potential dynamic features of each photovoltaic power generation unit based on the adversarial contrast feature extraction network.

[0094] In this step, the operating data of 100 photovoltaic (PV) power generation units in the PV power plant are first acquired under conditions of a 0.1 pu voltage drop and frequency disturbance. Sampling is performed every 0.01 seconds to obtain the transient curves of the PV power generation units within 5 seconds. After data processing, the active power, reactive power, and primary frequency regulation response output sequences are obtained. The original active power and reactive power sequences are transposed according to the sample dimensions to construct a three-dimensional transient sample tensor of 100×2×500, where 100 represents the number of PV power generation units, 2 represents the two input channels (active and reactive power), and 500 represents the number of transient sampling points. In this example, the model uses the active power transient response curves and reactive power transient response curves of the PV power generation units as input to characterize the differences in the dynamic response of different PV power generation units after disturbance.

[0095] Next, a Deep Adversarial Contrastive Network (DACN) is constructed to extract low-dimensional latent dynamic features from the transient power curves of photovoltaic power generation units. This network consists of three parts: a one-dimensional convolutional encoder, a contrastive learning projector, and an adversarial discriminator. The one-dimensional convolutional encoder sequentially employs convolution, batch normalization, ReLU activation, and pooling operations to extract multi-scale features from the input dual-channel transient sequences of active power P and reactive power Q. The first convolutional layer captures rapid transient change features, the second convolutional layer extracts features from the medium- to long-term recovery process, and the third convolutional layer combines adaptive average pooling to achieve high-dimensional temporal information fusion, ultimately outputting a 16-dimensional latent dynamic feature vector. .

[0096] Data augmentation is performed on the original transient sequence, including adding small Gaussian noise and random amplitude scaling, to generate augmented samples. The original and augmented samples are simultaneously input into the DACN encoder. The projector head is mapped to the contrast space through contrastive learning. InfoNCE loss is used to narrow the feature distance between the original and augmented samples of the same photovoltaic power generation unit, while widening the feature representations between different samples. An adversarial discriminator is introduced to distinguish between real transient features and augmented perturbation features. The encoder learns to generate robust features that more closely approximate the true distribution through adversarial loss. After training, the encoder weights are saved for subsequent clustering feature extraction.

[0097] The P / Q transient sequences of all photovoltaic power generation units are input into a one-dimensional convolutional encoder to obtain 16-dimensional potential dynamic features corresponding to each photovoltaic power generation unit.

[0098] S2. Low-dimensional potential dynamic features are input into a deep embedded clustering network. Feature reconstruction loss constrains the encoding of features to preserve time-series information. Cluster distribution optimization loss adjusts the sample soft assignment probability to obtain the cluster label corresponding to each photovoltaic power generation unit and determines the central representative unit of each cluster. Based on the primary frequency modulation disturbance input signal and active power response output signal of the central representative unit of each cluster, standardized input-output time series samples for parameter identification are generated.

[0099] In this step, to eliminate the influence of differences in the dimensions of different feature dimensions on the clustering process, StandardScaler is used to standardize the latent dynamic features output by DACN to zero mean and unit variance. Subsequently, KMeans clustering is performed within the range of candidate cluster numbers K=2 to K=8, and the silhouette coefficient and DB index are calculated as evaluation indicators of cluster compactness and separability. The K corresponding to the optimal combination of these two indicators is taken as the optimal number of clusters. The cluster number selection curve is shown below. Figure 4 As shown.

[0100] The standardized latent dynamic features are input into an improved deep embedded clustering network (IDEC). This deep embedded clustering network has a feature-level autoencoder with an input dimension of 16, hidden layer dimensions of 32 and 16, and an embedding space dimension of 8. The autoencoder is pre-trained for reconstruction, and cluster centers are initialized in the autoencoder embedding space using KMeans. A student-defined clustering algorithm is then used. The distribution calculates the soft assignment probability of a sample to each cluster center. And according to the target distribution The clustering distribution is self-trained and optimized. The joint loss of the deep embedding clustering network consists of KL divergence clustering loss and feature reconstruction loss, which enhances intra-cluster compactness and inter-cluster discriminability while preserving the original transient feature structure. During training, the target distribution is updated every few rounds, and training stops early when the rate of change of cluster labels falls below a set threshold.

[0101] Based on soft allocation probability The maximum value determines the final category label for each photovoltaic power generation unit. The silhouette coefficient, Davies-Bouldin index, and Calinski-Harabasz index are calculated to evaluate the clustering results. t-SNE is used to reduce the IDEC embedding spatial features to a two-dimensional plane, and a cluster distribution map of the photovoltaic power generation units is plotted, as shown below. Figure 5 As shown; for each cluster, the real photovoltaic power generation unit closest to the cluster's latent space center is selected as the central representative unit. The active power transient response curve of each cluster's central representative unit and the dynamic response error diagram between the aggregated results and measured values ​​are plotted, as shown. Figure 6 and Figure 7 As shown. Export and save the response data of the representative unit at the center of each cluster.

[0102] S3. Based on the dynamic response mechanism of primary frequency regulation of photovoltaic power generation unit, establish an aggregated equivalent transfer function model from disturbance input to active power response output, and convert it into a state-space model to construct physical constraint equations.

[0103] In this step, transient response data of representative units at the centers of five clusters in the photovoltaic power station are acquired. The common input is the frequency perturbation sequence b(t), and the outputs are the active power responses of clusters 1 to 5, respectively. The sampling period is set to Δt = 0.01s, and an input-output identification sample is independently constructed for each cluster unit. The samples are randomly divided into training and test sets at a ratio of 4:1.

[0104] Each photovoltaic unit is equivalent to a third-order dynamic model containing a drooping channel and a differential channel, and state variables are introduced to establish physical constraint equations.

[0105] S4. Construct a physical information neural network with time variables as input and predicted state variables as output. Set the parameters to be identified and the implicit initial state as Bayesian random variables and train them based on the joint optimization objective. After training, sample the posterior distribution of the parameters to be identified multiple times to obtain the identification results of each cluster of parameters. The joint optimization objective includes: data fitting error, physical equation residual error, discrete simulation error of the equivalent transfer function, initial state error, and divergence between the Bayesian posterior distribution and the prior distribution.

[0106] In this step, a physical information neural network is constructed using the time variable t as input, and it is first mapped to the interval [ [1,1] Subsequently, a Fourier feature encoding consisting of 16 sets of sine and cosine basis functions with a highest frequency of 10 is employed, and transient features are extracted through a three-layer residual MLP with a width of 96, ultimately outputting three physical states. .

[0107] Parameters to be identified , , , , The unknown initial states are modeled as Bayesian random variables, using Gaussian variational posterior. The automatic differential calculation network outputs the derivative with respect to time, constructing a joint optimization objective that includes data fitting error, physical equation residual error, discrete simulation error of the equivalent transfer function, initial state error, and the divergence between the Bayesian posterior and prior distributions.

[0108] The model was trained for a total of 8000 rounds, employing a phased training strategy that gradually transitions from deterministic learning to Bayesian learning. In the first 3000 rounds, data error and initial condition constraints were implemented, allowing the neural network to learn the overall trend of active power response. From rounds 3001 to 6000, the weights of physical residuals and transfer function simulation losses were gradually increased, driving the network output to fit the photovoltaic dynamic equation. From round 6001 onwards, the variational Bayesian stage began, using MAP pre-training results to redetermine prior centers and implementing random sampling of physical parameters and a linear growth strategy for KL divergence. The neural network data error was calculated simultaneously on both the training and test sets in each training round. The test set error was used only to monitor generalization ability and did not participate in backpropagation or parameter updates.

[0109] After model training, 2000 parameter samples were independently extracted from the variational posterior. The posterior mean, standard deviation, and 95% confidence interval of each parameter to be identified were calculated, and the equivalent transfer function of each cluster unit was constructed using the posterior mean. Furthermore, 300 parameter samples were selected for zero-order preserve time-domain simulation to obtain the 95% posterior prediction interval of the output response. The identification results for each central representative unit are as follows: Figure 8 As shown; the comparison chart between the parameter to be identified and the true value is as follows. Figure 9 As shown; the comparison between the overall identification results and measured values ​​of the photovoltaic power station is as follows: Figure 10 As shown.

[0110] S5. For each cluster, construct a primary frequency regulation equivalent transfer model based on the aggregated equivalent transfer function model and the parameter identification results of the cluster; aggregate the primary frequency regulation equivalent models of each cluster to obtain the overall primary frequency regulation response model of the photovoltaic power station.

[0111] In this step, for each cluster, a primary frequency regulation equivalent transfer model is constructed based on the aggregated equivalent transfer function model and the cluster parameter identification results. The equivalent capacity multiple corresponding to each cluster is generated according to the number or capacity of photovoltaic power generation units in each cluster. The primary frequency regulation equivalent models of each cluster are aggregated according to the equivalent capacity multiple to finally obtain the overall primary frequency regulation response model of the photovoltaic power station.

[0112] Example 2 In a typical embodiment of the present invention, this embodiment discloses a photovoltaic power station primary frequency regulation modeling system considering low voltage ride-through characteristics, comprising: The data acquisition and feature extraction module is configured to: acquire active power transient response data and reactive power transient response data of each photovoltaic power generation unit in the photovoltaic power station during low voltage ride-through and primary frequency regulation disturbance, and extract low-dimensional potential dynamic features of each photovoltaic power generation unit based on the adversarial comparison feature extraction network. The deep clustering module is configured to: input low-dimensional potential dynamic features into a deep embedded clustering network; use feature reconstruction loss to constrain the encoding of features to preserve time-series information; use cluster distribution optimization loss to adjust the sample soft assignment probability; obtain the cluster label corresponding to each photovoltaic power generation unit; and determine the central representative unit of each cluster; based on the primary frequency modulation perturbation input signal and active power response output signal of the central representative unit of each cluster, generate standardized input-output time series samples for parameter identification. The physical constraint construction module is configured to: establish an aggregated equivalent transfer function model from disturbance input to active power response output based on the primary frequency regulation dynamic response mechanism of the photovoltaic power generation unit, and convert it into a state-space model to construct physical constraint equations; The Bayesian uncertainty quantification module is configured to: construct a physical information neural network with time variables as input and predicted state variables as output; set the parameters to be identified and the implicit initial state as Bayesian random variables; and train them based on a joint optimization objective; after training, the posterior distribution of the parameters to be identified is sampled multiple times to obtain the identification results of each cluster of parameters; the joint optimization objective includes: data fitting error, physical equation residual error, discrete simulation error of the equivalent transfer function, initial state error, and divergence between the Bayesian posterior distribution and the prior distribution; The equivalent model aggregation module is configured to: for each cluster, construct a primary frequency regulation equivalent transfer model based on the aggregated equivalent transfer function model and the parameter identification results of the cluster; aggregate the primary frequency regulation equivalent models of each cluster to obtain the overall primary frequency regulation response model of the photovoltaic power station.

[0113] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for modeling the primary frequency regulation of photovoltaic power plants considering low voltage ride-through characteristics, characterized in that, include: The active power transient response data and reactive power transient response data of each photovoltaic power generation unit in the photovoltaic power station during low voltage ride-through and primary frequency regulation disturbance are obtained, and low-dimensional potential dynamic features of each photovoltaic power generation unit are extracted based on the adversarial contrast feature extraction network. Low-dimensional potential dynamic features are input into a deep embedded clustering network. Feature reconstruction loss constrains the encoding of features to preserve time-series information, and cluster distribution optimization loss adjusts the sample soft assignment probability to obtain the cluster label corresponding to each photovoltaic power generation unit and determine the central representative unit of each cluster. Based on the primary frequency modulation perturbation input signal and active power response output signal of the central representative unit of each cluster, standardized input-output time series samples for parameter identification are generated. Based on the dynamic response mechanism of primary frequency regulation of photovoltaic power generation unit, an aggregated equivalent transfer function model from disturbance input to active power response output is established, and it is converted into a state-space model to construct physical constraint equations; A physical information neural network is constructed with time variables as input and predicted values ​​of state variables as output. The parameters to be identified and the implicit initial states are set as Bayesian random variables, and the network is trained based on a joint optimization objective. After training, the posterior distribution of the parameters to be identified is sampled multiple times to obtain the identification results of each cluster of parameters. The joint optimization objectives include: data fitting error, physical equation residual error, discrete simulation error of equivalent transfer function, initial state error, and divergence between Bayesian posterior and prior distributions. For each cluster, a primary frequency regulation equivalent transfer model is constructed based on the aggregated equivalent transfer function model and the parameter identification results of the cluster; the primary frequency regulation equivalent models of each cluster are aggregated to obtain the overall primary frequency regulation response model of the photovoltaic power station.

2. The photovoltaic power station primary frequency regulation modeling method considering low voltage ride-through characteristics as described in claim 1, characterized in that, The active power transient response data and reactive power transient response data of each photovoltaic power generation unit in the photovoltaic power station during low voltage ride-through and primary frequency regulation disturbance are obtained. Low-dimensional potential dynamic features of each photovoltaic power generation unit are extracted based on an adversarial contrast feature extraction network, specifically including: For each photovoltaic power generation unit, its active power sequence and reactive power sequence during low voltage ride-through and primary frequency regulation disturbance are stacked according to the channel dimension to form a multi-channel time series sample, thus obtaining the transient response sample of each photovoltaic power generation unit. Gaussian noise and amplitude scaling are applied to each original transient response sample to construct the corresponding enhanced transient response sample; The original transient response samples and their corresponding enhanced transient response samples are input into an adversarial contrastive feature extraction network to extract low-dimensional potential dynamic features of each photovoltaic power generation unit. Contrastive learning is used to constrain the feature consistency of the same source disturbance view, and adversarial learning is used to weaken the influence of noise disturbance and operating condition changes on feature distribution.

3. The photovoltaic power station primary frequency regulation modeling method considering low voltage ride-through characteristics as described in claim 1, characterized in that, The adversarial contrastive feature extraction network consists of a one-dimensional convolutional encoder, a contrastive learning projector head, and an adversarial discriminator. Among them, the one-dimensional convolutional encoder gradually compresses the time dimension and expands the feature channels through a multi-level one-dimensional convolutional structure, which is used to extract the rapid transient change features in the early stage of low voltage ride-through, the power recovery features in the voltage recovery stage, and the dynamic response features in the primary frequency modulation process. The contrastive learning projector maps the features output by the encoder to the contrastive learning space. Through contrastive learning loss, it narrows the feature distance between the original samples and the enhanced samples of the same photovoltaic power generation unit, and widens the feature distance between samples of different photovoltaic power generation units or different response modes. The adversarial discriminator is used to distinguish between the original sample features and the enhanced sample features. The encoder is trained adversarially to make the distribution of the enhanced sample features close to that of the original sample features, thereby reducing the impact of noise disturbances and changes in operating conditions on the feature distribution. The joint optimization objective of the adversarial contrastive feature extraction network can be specifically expressed as: In the formula, Indicates the contrast learning loss. Indicates the loss due to resistance constraints. This represents the weighting coefficient.

4. The photovoltaic power station primary frequency regulation modeling method considering low voltage ride-through characteristics as described in claim 1, characterized in that, Deep embedding clustering networks consist of feature-level autoencoders and deep embedding clustering layers; In this process, the feature-level autoencoder maps latent dynamic features to the latent space and reconstructs the latent space features back to the original feature space. The feature reconstruction loss constrains the encoding features to retain temporal information, specifically as follows: In the formula, Represents the feature reconstruction loss, This represents the potential dynamic characteristics after standardization. Represents reconstructed features; In the deep embedding clustering layer, the soft assignment probability is calculated based on the latent space features of photovoltaic power generation units and the distance between each cluster center. The feature representation and cluster center are adjusted by optimizing the cluster distribution loss. The joint loss of the deep embedding clustering network is constructed based on the feature reconstruction loss and the clustering distribution optimization loss, which can be specifically expressed as: in, This represents the clustering distribution optimization loss; Represents the feature reconstruction loss; This indicates the weight of the reconfigured constraint.

5. The photovoltaic power station primary frequency regulation modeling method considering low voltage ride-through characteristics as described in claim 1, characterized in that, The aggregate equivalence transfer function model is specifically represented as follows: In the formula, This represents the equivalent time constant of the output stage. This represents the equivalent time constant of the drooping channel. Represents the equivalent time constant of the differential channel. This indicates the droop channel aggregation gain. Indicates the differential channel aggregation gain. , , , and This constitutes the set of parameters to be identified; The aggregated equivalent transfer function model is transformed into a state-space model with physical meaning, and physical constraint equations are constructed using state variables, specifically as follows: In the formula, Represents the state variables of the first-order inertial element in the drooping channel; This represents the state variable of the first-order inertial element in the differential channel; Indicates a disturbance input; This indicates the output deviation of the primary frequency regulation active power response of a photovoltaic power generation unit or photovoltaic power station; , , They represent , , The derivative with respect to time.

6. The photovoltaic power station primary frequency regulation modeling method considering low voltage ride-through characteristics as described in claim 1, characterized in that, In the network structure of the physical information neural network, the time variable is first normalized, then Fourier time features are constructed, and then a residual multilayer perceptron is used to perform nonlinear mapping on the Fourier features to output the predicted values ​​of the state variables.

7. The photovoltaic power station primary frequency regulation modeling method considering low voltage ride-through characteristics as described in claim 1, characterized in that, The joint optimization objective is specifically expressed as follows: In the formula, This refers to the data error between the neural network's predicted output and the measured output. The physical residual error is obtained from the physical constraint equation of primary frequency modulation. The error between the simulated output and the measured output of the equivalent transfer function constructed from the current identified parameters. The initial state constraint error, Let be the divergence between the posterior and prior distributions. , , , and These are the weights of the corresponding loss terms.

8. The photovoltaic power station primary frequency regulation modeling method considering low voltage ride-through characteristics as described in claim 1, characterized in that, The training of the physical information neural network employs a phased training strategy, specifically including: In the first training phase, deterministic warm-up training is performed using the posterior mean of the parameters to be identified, so that the neural network learns the basic response relationship between the primary frequency modulation input and the active power output first. Then, physical equation residual constraints and equivalent transfer function simulation constraints are gradually introduced to constrain the network output state to gradually satisfy the primary frequency modulation dynamic equation while updating the parameters to be identified. In the second training phase, Bayesian variational inference is initiated, posterior sampling is performed on the parameters to be identified and the implicit initial state, and the weight of the Bayesian divergence term is gradually increased so that the identification results are constrained by both the physical equation and the posterior distribution while fitting the measured response. Through Bayesian variational inference, the model can obtain the mean, variance and confidence interval of the parameters while maintaining data fit and physical consistency.

9. The photovoltaic power station primary frequency regulation modeling method considering low voltage ride-through characteristics as described in claim 1, characterized in that, The overall primary frequency regulation response model of a photovoltaic power station is specifically represented as follows: In the formula, Indicates the first The equivalent capacity multiple of each cluster; This represents a primary frequency modulation disturbance input; Indicates the first The first-order frequency modulation equivalent transfer model for a cluster is specifically represented as follows: In the formula, Indicates the first The equivalent time constant of the output stage of each cluster Indicates the first Equivalent time constants for each cluster's drooping channel Indicates the first The equivalent time constant of the differential channels of each cluster, Indicates the first Cluster-specific droop channel aggregation gain Indicates the first Cluster differential channel aggregation gain.

10. A photovoltaic power station primary frequency regulation modeling system considering low voltage ride-through characteristics, characterized in that, include: The data acquisition and feature extraction module is configured to: acquire active power transient response data and reactive power transient response data of each photovoltaic power generation unit in the photovoltaic power station during low voltage ride-through and primary frequency regulation disturbance, and extract low-dimensional potential dynamic features of each photovoltaic power generation unit based on the adversarial comparison feature extraction network. The deep clustering module is configured to: input low-dimensional potential dynamic features into a deep embedded clustering network; use feature reconstruction loss to constrain the encoding of features to preserve time-series information; use cluster distribution optimization loss to adjust the sample soft assignment probability; obtain the cluster label corresponding to each photovoltaic power generation unit; and determine the central representative unit of each cluster; based on the primary frequency modulation perturbation input signal and active power response output signal of the central representative unit of each cluster, generate standardized input-output time series samples for parameter identification. The physical constraint construction module is configured to: establish an aggregated equivalent transfer function model from disturbance input to active power response output based on the primary frequency regulation dynamic response mechanism of the photovoltaic power generation unit, and convert it into a state-space model to construct physical constraint equations; The Bayesian uncertainty quantification module is configured to: construct a physical information neural network with time variables as input and predicted values ​​of state variables as output; set the parameters to be identified and the implicit initial states as Bayesian random variables; and train them based on the joint optimization objective. After training, the posterior distribution of the parameters to be identified is sampled multiple times to obtain the identification results of each cluster of parameters. The joint optimization objectives include: data fitting error, physical equation residual error, discrete simulation error of equivalent transfer function, initial state error, and divergence between Bayesian posterior and prior distributions. The equivalent model aggregation module is configured to: for each cluster, construct a primary frequency regulation equivalent transfer model based on the aggregated equivalent transfer function model and the parameter identification results of the cluster; aggregate the primary frequency regulation equivalent models of each cluster to obtain the overall primary frequency regulation response model of the photovoltaic power station.

Images