Method for judging virtual inertia of frequency modulation of thermal power plant and related device

Abstract

The application discloses a kind of thermal power plant frequency modulation virtual inertia discrimination method and related device, comprising: 1) construct disturbance input sequence;2) using historical disturbance input sequence and corresponding historical virtual inertia support quantity and historical support duration training to obtain joint discrimination model;3) the disturbance input sequence is input into the Causal-TCN of joint discrimination model, and the disturbance representation vector is obtained;4) form support quantity-duration joint distribution;5) based on support quantity-duration joint distribution, obtain initial virtual inertia support quantity and initial support duration;6) under the constraint of maximum exchange power of energy storage system and the constraint of available energy of energy storage system, initial virtual inertia support quantity and initial support duration are converted into virtual inertia support instruction, the method and related device can reliably discriminate virtual inertia support quantity-support duration, and closed loop generates executable support instruction.

Description

Technical Field

[0001] This invention belongs to the field of power system frequency control technology, and relates to a method and related device for determining the virtual inertia of frequency regulation in thermal power plants. Background Technology

[0002] With a high proportion of renewable energy connected to the grid, fluctuations and sudden changes in wind power output are more likely to cause active power imbalances, leading to a drop or rise in system frequency in the early stages of disturbances, accompanied by a short-term sharp change in the rate of frequency change (RoCoF). As the rotational inertia of traditional synchronous machines decreases, the frequency inertial response weakens, shortening the processing time available for frequency regulation devices in the early stages of disturbances. Thermal power plants, during grid-connected frequency regulation operation, typically need to coordinate with fast-power devices such as energy storage to provide rapid active power support to the system in the early stages of disturbances through virtual inertia. This aims to suppress RoCoF, mitigate frequency drift, and buy time for subsequent primary / secondary frequency regulation. This type of support depends not only on the strength of the disturbance but also on its duration. In actual control, it is usually necessary to simultaneously determine the amount and duration of virtual inertia support and translate it into executable commands under the constraints of maximum energy storage exchange power and available energy, thus forming a closed-loop chain from disturbance observation, trigger determination, support requirement determination to control issuance.

[0003] In existing technologies, virtual inertia support often adopts rule-based or parameterized control based on frequency change rate threshold triggering: after triggering, the active power of energy storage is given proportionally to RoCoF according to a preset coefficient, and a fixed or empirically set duration is used; some schemes estimate the equivalent inertia gap or power imbalance based on a simplified frequency response model, and then calculate the support amount and duration; some data-driven methods use historical operating data to train regression models or neural networks, and output the support amount and duration respectively for online setting or decision support.

[0004] The above-mentioned schemes generally treat the support quantity and duration as independent quantities or deterministic results, which makes it difficult to express the source of uncertainty caused by the differences in the mechanism of the disturbance: when the observations are similar in the early stage of the disturbance, different mechanism events may correspond to different paired support requirements. Independent regression or fixed mapping is prone to produce inconsistent quantity-time combinations, and lacks a characterization of the linkage between the two, which makes it easier for undeliverable instructions that need to be manually corrected to appear under the constraints of energy storage power / energy. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and related device for determining the virtual inertia of frequency regulation in thermal power plants. This method and related device can reliably determine the virtual inertia support quantity and support duration, and generate executable support commands in a closed loop.

[0006] To achieve the above objectives, this invention discloses a method for determining the virtual inertia of frequency regulation in thermal power plants, comprising: 1) During the grid-connected frequency regulation operation of thermal power plants, real-time wind power, system frequency, and frequency change rate are obtained, and the input sequence is perturbed accordingly; 2) A joint discrimination model is trained using the historical perturbation input sequence and the corresponding historical virtual inertia support quantity and historical support duration. The joint discrimination model includes the Causal-TCN and the perturbation latent variable conditional Copula joint distribution head. 3) Input the perturbation input sequence into the Causal-TCN of the joint discriminant model to obtain the perturbation representation vector; 4) Input the disturbance characterization vector into the disturbance latent variable conditional Copula joint distribution head. The disturbance latent variable conditional Copula joint distribution head generates the disturbance mechanism latent variable, the virtual inertia support quantity probability parameter, the support duration probability parameter, and the Copula correlation parameter in sequence. Based on the Copula correlation parameter, the virtual inertia support quantity probability parameter and the support duration probability parameter are combined to form the support quantity-duration joint distribution. 5) Determine the trigger state based on the frequency change rate and the preset frequency change rate threshold. In the trigger state, obtain the initial virtual inertia support amount and the initial support duration based on the joint distribution of support amount and duration. 6) Under the constraints of the maximum exchange power of the energy storage system and the available energy of the energy storage system, the initial virtual inertia support quantity and the initial support duration are converted into virtual inertia support commands.

[0007] Furthermore, the specific operation of step 1) is as follows: Set the sampling period and preset sequence length of the perturbation input sequence; During grid-connected frequency regulation operation of thermal power plants, real-time wind power and system frequency are obtained at the same moment according to the sampling period. The frequency change rate is calculated based on the system frequency of adjacent sampling periods, and the frequency change rate is associated with the system frequency samples involved in the calculation at the same time. The real-time wind power, system frequency, and frequency change rate at the same moment are used as sampling points. A series of sampling points of a preset length are arranged in chronological order to form a disturbance input sequence. The disturbance input sequence is then organized into a three-channel time-series signal within the model. The three channels are, in order, real-time wind power, system frequency, and frequency change rate.

[0008] Furthermore, using the historical perturbation input sequence as input, forward computation is performed to output the joint distribution of support quantity and duration. The historical virtual inertia support quantity and historical support duration are used as support demand pairs for joint distribution matching, and the network parameters of the joint discriminant model are updated in reverse accordingly. The matching results synchronously constrain the probability parameters of virtual inertia support quantity, support duration, and Copula correlation parameters. The network parameters of Causal-TCN are then updated in reverse through the perturbation latent variable conditional Copula joint distribution head to obtain the trained joint discriminant model.

[0009] Furthermore, the loss function of the joint discriminant model during training is:

[0010] in, This is due to batch training losses; This refers to the batch sample size. The sample number; This is the historical virtual inertia support quantity; The duration of historical support; For virtual inertia support probability parameters; To support the duration probability parameter; For Copula-related parameters; For the reason The normal distribution probability density corresponding to the marginal probability expression of the determined virtual inertia support quantity; For the reason The marginal probability expression of the determined support duration corresponds to the normal distribution probability density; To and A consistent normal cumulative distribution function; To and A consistent normal cumulative distribution function; This represents the cumulative probability variable of historical virtual inertia support quantity under the marginal probability expression of virtual inertia support quantity; The cumulative probability variable representing the historical support duration under the marginal probability expression of the support duration; For the reason Determined Gaussian Copula density.

[0011] Furthermore, the perturbation input sequence is:

[0012] in, The preset sequence length; The sampling sequence number; This represents the real-time wind power output at the corresponding sampling number; This represents the system frequency at the corresponding sampling number; This represents the frequency change rate at the corresponding sampling number.

[0013] Furthermore, the specific operation of step 4) is as follows: The perturbation representation vector is input into the perturbation latent variable conditional Copula joint distribution head, which is composed of a latent variable generation layer, a marginal parameter generation layer and a correlation parameter generation layer connected in sequence. A latent variable generation layer is used to perform a fully connected mapping on the perturbation representation vector. The number of neurons in the fully connected mapping is sixteen, and the output perturbation mechanism latent variable has an eight-dimensional dimension. The hidden variables of the disturbance mechanism are input into the first set of fully connected networks of the edge parameter generation layer to generate virtual inertia support probability parameters, wherein the virtual inertia support probability parameters have a dimension of two. The hidden variables of the disturbance mechanism are input into the second set of fully connected networks of the edge parameter generation layer to generate the support duration probability parameter, which has a two-dimensional dimension. The hidden variables of the disturbance mechanism are input into the fully connected network of the association parameter generation layer to generate Copula association parameters, wherein the Copula association parameters have a dimension of one. The marginal probability expression of virtual inertia support quantity is determined based on the probability parameter of virtual inertia support quantity, and the marginal probability expression of support duration is determined based on the probability parameter of support duration. The marginal probability expression of virtual inertia support quantity and the marginal probability expression of support duration are associated and combined according to the Copula correlation parameter to obtain the joint distribution of support quantity-duration.

[0014] Furthermore, the specific operation of step 5) is as follows: The frequency change rate of the last sampling point in the perturbation input sequence is selected as the determination frequency change rate. The amplitude of the determined frequency change rate is taken and compared with a preset frequency change rate threshold. If the amplitude is not less than the preset frequency change rate threshold, it is determined to be a triggered state; if the amplitude is less than the preset frequency change rate threshold, it is determined to be a non-triggered state. In the triggered state, the joint distribution of support quantity and duration is used as the criterion for determining support demand pairs. The support demand pair with the highest joint probability is determined, and the support quantity in the support demand pair is determined as the initial virtual inertia support quantity, and the duration in the support demand pair is determined as the initial support duration.

[0015] This invention discloses a virtual inertia discrimination system for frequency regulation in thermal power plants, comprising: The acquisition module is used to acquire real-time wind power, system frequency and frequency change rate during grid-connected frequency regulation operation of thermal power plants, and to use these to perturb the input sequence. The judgment module is used to train a joint discrimination model using historical perturbation input sequences and corresponding historical virtual inertia support quantities and historical support durations. The joint discrimination model includes the Causal-TCN and the perturbation latent variable conditional Copula joint distribution head. The perturbation module is used to input the perturbation input sequence into the Causal-TCN of the joint discriminant model to obtain the perturbation representation vector. The generation module is used to input the disturbance characterization vector into the disturbance latent variable conditional Copula joint distribution head. The disturbance latent variable conditional Copula joint distribution head generates the disturbance mechanism latent variable, the virtual inertia support quantity probability parameter, the support duration probability parameter, and the Copula correlation parameter in sequence. Based on the Copula correlation parameter, the virtual inertia support quantity probability parameter and the support duration probability parameter are combined to form the support quantity-duration joint distribution. The triggering module is used to determine the triggering state based on the frequency change rate and the preset frequency change rate threshold. In the triggering state, the initial virtual inertia support amount and the initial support duration are obtained based on the joint distribution of support amount and duration. The conversion module is used to convert the initial virtual inertia support quantity and initial support duration into virtual inertia support commands under the constraints of the maximum exchange power of the energy storage system and the available energy of the energy storage system.

[0016] The present invention discloses a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method for generating and sending trusted status alarm information.

[0017] The present invention discloses a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the steps of the method for generating and uploading trusted status alarm information are implemented.

[0018] The present invention has the following beneficial effects: The virtual inertia discrimination method and related device for frequency regulation in thermal power plants described in this invention, in specific operation, uses real-time wind power, system frequency, and frequency change rate as causal time-series inputs. It extracts disturbance characteristics using an extended causal TCN and generates marginal probability parameters and Copula correlation parameters for support quantity and duration through latent variables of disturbance mechanisms, constructing a joint distribution of support quantity and duration and training it using joint likelihood. During operation, triggered by a frequency change rate threshold, the maximum probability support demand pair is determined from the joint distribution and converted into executable virtual inertia support commands under the constraints of maximum energy storage exchange power and available energy. This method can characterize the uncertainty and linkage relationship of support demand, improving the executability of initial disturbance discrimination and closed-loop control. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart of the method of the present invention; Figure 2 The flowchart for step 1); Figure 3 This is a flowchart for step 2); Figure 4 The flowchart for step 3); Figure 5 This is a flowchart for step 4); Figure 6 The flowchart for step 5); Figure 7 This is the flowchart for step 6). Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] In the description of this invention, it should be understood that the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0023] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0024] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Additionally, the character " / " in this invention generally indicates that the preceding and following objects have an "or" relationship.

[0025] It should be understood that although terms such as first, second, third, etc., may be used in the embodiments of the present invention to describe the preset range, these preset ranges should not be limited to these terms. These terms are only used to distinguish the preset ranges from one another. For example, without departing from the scope of the embodiments of the present invention, the first preset range may also be referred to as the second preset range, and similarly, the second preset range may also be referred to as the first preset range.

[0026] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0028] The accompanying drawings illustrate various structural schematic diagrams according to embodiments disclosed in this invention. These drawings are not to scale, and some details have been enlarged for clarity, and some details may have been omitted. The shapes of the various regions and layers shown in the drawings, as well as their relative sizes and positional relationships, are merely exemplary and may deviate from reality due to manufacturing tolerances or technical limitations. Furthermore, those skilled in the art can design regions / layers with different shapes, sizes, and relative positions as needed.

[0029] Example 1 refer to Figure 1 The method for determining the virtual inertia of frequency regulation in thermal power plants according to the present invention includes the following steps: 1) During the grid-connected frequency regulation operation of thermal power plants, real-time wind power, system frequency and frequency change rate are acquired, and the real-time wind power, system frequency and frequency change rate are arranged in chronological order to form a disturbance input sequence; It should be noted that the reference Figure 2 The sampling period and preset sequence length of the disturbance input sequence are set. During the grid-connected frequency regulation operation of the thermal power plant, the real-time wind power and system frequency at the same moment are obtained according to the sampling period. The frequency change rate is calculated based on the system frequency of adjacent sampling periods, and the frequency change rate is associated with the system frequency samples involved in the calculation at the same moment. The real-time wind power, system frequency and frequency change rate at the same moment are used as sampling points. The sampling points of the preset sequence length are arranged in chronological order to form the disturbance input sequence. The disturbance input sequence is organized in the model as a three-channel time-series signal, wherein the three channels are real-time wind power, system frequency and frequency change rate in sequence.

[0030] The specific operation of step 1) is as follows: To enable the joint discriminant model to use a three-channel time-series signal as input and to match the causal convolutional structure of Causal-TCN, the sampling period of the perturbed input sequence is set to be... Set the preset sequence length to .in, The sampling time interval used to constrain real-time wind power and system frequency. Used to constrain the time dimension length of the perturbation input sequence. The sampling sequence number uses characters. express, It is a positive integer; the first The sampling time corresponding to the first sampling is denoted as . The time interval between adjacent sampling times is Preset sequence length The time span coverage requirement of the dilated convolutional layer group is set so that the perturbation input sequence can simultaneously contain short-term change information of the rate of frequency change and continuous offset information of the system frequency within the same time window.

[0031] At each sampling time Obtain the real-time wind power corresponding to the sampling time. and system frequency ,in, For the wind power measurement channel at the sampling time The corresponding active power value, For the power grid frequency measurement channel at the sampling time The corresponding frequency value. When the original sampling time of real-time wind power is inconsistent with the sampling time of the system frequency, the sampling time of the system frequency shall be used. As an alignment benchmark, real-time wind power is generated using a forward hold method. Select a time marker no later than The latest raw wind power sampling value as ,make and Features of the same sampling point are formed at the same sampling time and can be directly spliced ​​together.

[0032] After obtaining the system frequencies at adjacent sampling times, calculate the rate of frequency change. Rate of change of frequency The input only uses , With sampling period The method is to divide the difference between two adjacent system frequencies by the sampling period to obtain the result. and will The system frequency samples used in the calculation are compared according to the sampling time. To associate, make Corresponding description from arrive The system frequency change rate, and without introducing any later than The sampling points participate in the calculation, which satisfies the causal connection requirement of Causal-TCN from the input structure, which only uses data from the current and historical moments.

[0033] Based on the same sampling time Below , and Construct sampling point vector ,in, This is a three-dimensional feature vector, with the three components in the following order: real-time wind power, system frequency, and rate of frequency change. (The last part, "continuous," appears to be a typo and can be omitted.) The sampling point vectors are arranged chronologically to form a perturbation input sequence. Perturbation input sequence Within the joint discriminant model, the three-channel time-series signals are organized into a three-row array. The data matrix consists of columns, with the first row representing the real-time wind power channel sequence. The second line is the system frequency channel sequence. The third row is the frequency change rate channel sequence. Thus, the channel-separated convolutional layers of Causal-TCN can perform causal convolution calculations on the three channels separately and output the corresponding convolutional channels, and the dilated convolutional layer group can form response representations with different time spans on the same perturbed input sequence.

[0034] Perturbation input sequence According to the sampling sequence number Rolling update: When the sampling time starts from... Updated to At that time, based on the newly added Generate a new perturbation input sequence And ensure that the perturbation input sequence always remains a continuous sequence. The data consists of a vector of sampling points, which continuously provides real-time input for the joint discrimination model in the early stages of perturbation. The information of the perturbation input sequence data items is shown in Table 1: Table 1

[0035] 2) A joint discrimination model is trained using the historical perturbation input sequence and the corresponding historical virtual inertia support quantity and historical support duration. The joint discrimination model includes the Causal-TCN and the perturbation latent variable conditional Copula joint distribution head. It should be noted that the reference Figure 3Multiple historical disturbance input sequences are acquired, and the corresponding historical virtual inertia support and historical support duration are associated with each historical disturbance input sequence to form training samples. A Causal-TCN joint discriminant model is constructed, using channel-separated convolutional layers to perform causal convolution on the real-time wind power, system frequency, and frequency change rate channels, outputting sixteen convolutional channels. Then, a channel fusion convolutional layer is used to fuse the channels, outputting sixty-four convolutional channels. After the channel fusion convolutional layer, four dilated convolutional layers are sequentially connected to form a dilated convolutional layer group. Each dilated convolutional layer has sixty-four convolutional channels, with the dilation factor increasing from smallest to largest. Bypass connections are set between the layers in the dilated convolutional layer group. The output of the dilated convolutional layer group is compressed in time dimension using a time convergence layer to obtain a time-series feature vector of length sixty-four, and then a disturbance representation vector of dimension thirty-two is obtained through a fully connected mapping layer. A disturbance latent variable conditional Copula joint distribution head is constructed for the joint discriminant model, using a latent variable generation layer with the disturbance representation vector as input and passing through ten neurons. The six-dimensional fully connected network outputs a perturbation mechanism latent variable with an eight-dimensional dimension. An edge parameter generation layer takes the perturbation mechanism latent variable as input and generates two-dimensional virtual inertia support quantity probability parameters and two-dimensional support duration probability parameters through two sets of fully connected networks. An association parameter generation layer takes the perturbation mechanism latent variable as input and generates a one-dimensional Copula association parameter. A support quantity-duration joint distribution is formed based on the virtual inertia support quantity probability parameter, support duration probability parameter, and Copula association parameter. A forward calculation is performed using the historical perturbation input sequence as input to output the support quantity-duration joint distribution. The historical virtual inertia support quantity and historical support duration are used as support demand pairs for joint distribution matching, and the joint discriminant model network parameters are updated accordingly. The matching result synchronously constrains the virtual inertia support quantity probability parameter, support duration probability parameter, and Copula association parameter. The Causal-TCN network parameters are updated in reverse through the perturbation latent variable conditional Copula joint distribution head, resulting in the trained joint discriminant model.

[0036] The specific operation of step 2) is as follows: The joint discriminant model is trained using historical perturbation input sequences as input and support demand pairs consisting of historical virtual inertia support quantities and historical support durations as supervision information. Let the number of historical samples be... The preset sequence length is For each sample number The historical perturbation input sequence is denoted as ,in, For three lines The data matrix consists of columns, with the three rows representing the real-time wind power channel, system frequency channel, and frequency change rate channel, respectively. (The rest of the text appears to be a list of data points and instructions, possibly related to a data matrix or dataset.) The historical perturbation input sequence tensor is formed by stacking samples according to their sequence numbers. , The sample dimension is The channel dimension is three, and the time dimension is... The original historical virtual inertia support quantity and the original historical support duration corresponding to each sample are normalized by the ratio of the preset benchmark virtual inertia support quantity and the preset benchmark support duration, respectively, to obtain the historical virtual inertia support quantity. Duration of historical support And constitutes a supporting demand for .

[0037] The first part of the joint discriminant model is Causal-TCN. The input channel is separated into convolutional layers, which perform one-dimensional causal convolution calculations on the three channels of real-time wind power, system frequency, and frequency change rate. At any given time, the causal convolution uses only the current and historical sampling points for kernel weighting and bias superposition. Each channel-separated convolutional layer outputs sixteen convolutional channels, resulting in three sets of channel feature matrices. , , Each channel feature matrix consists of sixteen rows. Column, will , , After concatenation along the channel dimension, the input is processed by the channel fusion convolutional layer to perform channel fusion convolution calculations. The channel fusion convolutional layer outputs sixty-four convolutional channels, resulting in a fused feature matrix. , Sixty-four lines List.

[0038] exist This is followed by a group of dilated convolutional layers, consisting of four dilated convolutional layers sequentially. Each dilated convolutional layer has 64 convolutional channels, and the dilation factor is set to increase from smallest to largest, so that the same... Different dilation factors generate response representations for different time spans. Bypass connections are established between adjacent layers in the dilated convolutional layer group, adding the input features of the previous layer to the output features of the current layer channel-wise to obtain inter-layer fusion features. This preserves the traceability of the initial perturbation patterns of system frequency and rate of change in deep convolutional features. The output of the last layer of the dilated convolutional layer group is input to a temporal convergence layer to compress the time dimension, resulting in a temporal feature vector. ,in, It is formed by splicing together 64 scalars obtained by time convergence of 64 convolution channels in channel order. Inputting the fully connected mapping layer yields the perturbation representation vector. ,in, It is formed by splicing together the 32 scalars output by 32 fully connected neurons in the order of the neurons. It serves as the sole input to the joint distribution head of the conditional Copula for the perturbation of latent variables.

[0039] The latter part of the joint discriminant model is the joint distribution head of the perturbation latent variable conditional Copula. The input is a latent variable generation layer, which uses a fully connected network with sixteen neurons to output perturbation mechanism latent variables. , The dimension is eight. The first set of fully connected networks in the input edge parameter generation layer generates virtual inertia support probability parameters. , The two dimensions are the mean parameter and the scale parameter, respectively, used to characterize the marginal probability expression of the virtual inertia support quantity; The second set of fully connected networks in the input edge parameter generation layer generates the support duration probability parameters. , The dimension is two, with the two dimensions representing the mean parameter and scale parameter, respectively, of the marginal probability expression supporting the duration; Input association parameter generation layer generates Copula association parameters , The dimension is one. (Regarding...) and The scale parameter in the model is obtained by a monotonically positive value mapping to obtain the positive scale value. Bounded mappings are used to obtain constrained association parameters. Based on Calculate the historical virtual inertia support using the normal cumulative distribution function. Corresponding cumulative probability variable ,based on Calculate the historical support duration using the normal cumulative distribution function. Corresponding cumulative probability variable And Gaussian Copula density is used as Calculation method: and Input the inverse cumulative distribution function of the standard normal distribution to obtain the standard normal quantiles, and then use the correlation coefficient as... The ratio of the two-dimensional standard normal distribution density to the two one-dimensional standard normal distribution densities is determined. This results in a joint distribution of support quantity and duration. .

[0040] Training is conducted in batches from Select batches of samples, and the number of samples in each batch is denoted as . For each batch of samples, a forward computation is performed to obtain the corresponding... , , , , and and will and As a supporting requirement for the input joint distribution matching process, the batch negative log-likelihood based on Copula decomposition is used as the training loss, calculated as follows:

[0041] in, This is due to batch training losses; This refers to the batch sample size. The sample number; This is the historical virtual inertia support quantity; The duration of historical support; For virtual inertia support probability parameters; To support the duration probability parameter; For Copula-related parameters; For the reason The normal distribution probability density corresponding to the marginal probability expression of the determined virtual inertia support quantity; For the reason The marginal probability expression of the determined support duration corresponds to the normal distribution probability density; To and A consistent normal cumulative distribution function; To and A consistent normal cumulative distribution function; This represents the cumulative probability variable of historical virtual inertia support quantity under the marginal probability expression of virtual inertia support quantity; The cumulative probability variable representing the historical support duration under the marginal probability expression of the support duration; For the reason Determined Gaussian Copula density.

[0042] Training loss Perform backpropagation to update the network parameters of the joint distribution head of the perturbation latent variable conditional Copula, ensuring that the virtual inertia support probability parameter, support duration probability parameter, and Copula association parameter are simultaneously subject to the joint distribution matching constraint of the support demand pair; then, propagate the backpropagation gradient along the perturbation representation vector. The data is passed to Causal-TCN and the network parameters of Causal-TCN are updated to obtain the trained joint discriminant model.

[0043] 3) Input the perturbation input sequence into the Causal-TCN of the joint discriminant model. The Causal-TCN performs convolution calculations sequentially using only the causal connections of the current and historical moments. Dilated convolutions are introduced in the convolution calculations to form response representations of different time spans, and the perturbation representation vector is output. The perturbation representation vector is used to represent the strength and persistence trend of the perturbation and serves as the sole input to the joint distribution head of the conditional Copula of the perturbation latent variables. It should be noted that the reference Figure 4 The perturbation input sequence is input into the channel-separated convolutional layer of the Causal-TCN as a three-channel time-series signal. The three channels are, in order, real-time wind power, system frequency, and frequency change rate. Causal convolution calculation is performed on each channel. This causal convolution calculation uses only the sampling points at any given time and those prior to that time for kernel weighting and bias superposition, resulting in three sets of channel-separated convolutional outputs, each with sixteen convolutional channels. These three sets of outputs are then input into a channel-fusion convolutional layer to perform channel-fusion convolution calculation, resulting in a fused feature with sixty-four convolutional channels. The fused feature is then sequentially input into a group of dilated convolutional layers according to the inter-layer causal connection method. This group consists of four dilated convolutional layers, each with sixty-four convolutional channels, and the dilation factor increases sequentially from... The dilation factor is incremented from small to large, allowing the same perturbation input sequence to exhibit response representations over different time spans at different dilation factors. Smaller dilation factors correspond to short-term changes in the rate of frequency change, while larger dilation factors correspond to long-term continuous shifts in the system frequency. Bypass connections are established between adjacent layers in the dilated convolutional layer group, adding the input features of the previous layer to the output features of the current layer according to their corresponding channels to form inter-layer fusion features. This ensures that the initial perturbation patterns of the rate of frequency change and the system frequency remain traceable in the deep convolutional features. The output of the last layer of the dilated convolutional layer group is compressed into a time-dimensional convergence layer to obtain a temporal feature vector of length 64. This temporal feature vector is then input into a fully connected mapping layer to obtain a perturbation representation vector of dimension 32, which is used as the sole input to the joint distribution head of the conditional Copula for the perturbation latent variables.

[0044] The specific operation of step 3) is as follows: Let the perturbation input sequence be denoted as . From continuous The real-time wind power, system frequency, and rate of change of frequency at each sampling time are obtained in chronological order and organized into a three-channel time-series signal matrix within the model, constructed using the following formula:

[0045] in, The input sequence is a perturbation sequence; The preset sequence length; The sampling sequence number; This represents the real-time wind power output at the corresponding sampling number; This represents the system frequency at the corresponding sampling number; This represents the frequency change rate at the corresponding sampling number.

[0046] Will When inputting the channel-separated convolutional layer of Causal-TCN, the matrix column direction is kept consistent with the sampling time sequence, and the time length is maintained by padding only on the left side of the time axis during convolution calculation. Make the time index as The convolution output consists only of time indices not greater than 1. The input columns are used in the calculation; among them, This is a time index, with values ​​ranging from one to... Integers.

[0047] Channel-separating convolutional layers are applied to the real-time wind power channel, system frequency channel, and frequency change rate channel. The kernel length of the channel-separating convolutional layer is denoted as [missing information]. ,in, The number of temporal sampling points covered by the convolutional kernel is a positive integer; each channel-separated convolutional layer contains sixteen one-dimensional convolutional kernels and corresponding biases, outputting sixteen convolutional channels. For any channel-separated convolutional layer, the time indices are from one to... Perform causal convolution calculations sequentially: for time indices... The output only applies to time indices not greater than [value]. The input sampling points are weighted and summed using convolution kernels, and biases are added to obtain the convolution output of that channel at that time index, thus forming three sets of channel-separated convolution output matrices. , , ,in, This is the channel-separated convolution output matrix for real-time wind power channels. This is the channel-separated convolution output matrix for the system's frequency channels. The channel-separated convolution output matrix for the frequency change rate channel; , , All are sixteen lines The columns of the matrix represent the convolution channel indices, and the columns represent the time indices.

[0048] Will , , The channel splicing feature matrix is ​​formed by concatenating the data along the channel dimension. ,in The number of channels is forty-eight and the time length is .Will The input channel fusion convolutional layer performs channel fusion convolution calculations, and outputs sixty-four convolutional channels to obtain the fused feature matrix. ,in Sixty-four lines List, Used to express the coupling relationship of three-channel perturbation information within a unified feature space.

[0049] Will The dilated convolutional layers are input sequentially according to the inter-layer causal connection method. The dilated convolutional layer group consists of four dilated convolutional layers in sequence, each with 64 output channels. The four dilation factors are denoted as follows: , , , ,in, , , , All are positive integers and satisfy the following conditions: Less than , Less than , Less than Dilated convolutional layers introduce a time interval determined by the dilation factor between adjacent sampling positions of the convolutional kernel. This allows smaller dilation factors to correspond to response representations with shorter time spans, and larger dilation factors to correspond to response representations with longer time spans. As a result, the same perturbation input sequence can form response representations with different time spans at different dilation factors.

[0050] Bypass addition connections are set between adjacent layers in the dilated convolutional layer group, using Indicates the index of the dilated convolutional layer. The value is an integer from one to four, for the first... The dilated convolutional layer adds the input feature matrix from the previous layer to the output matrix of the current layer element-wise, channel by channel, to obtain the inter-layer fusion feature matrix, and uses the inter-layer fusion feature matrix as the first... The input of each dilated convolutional layer ensures that the rate of frequency change carried by the shallow features and the initial morphology of the system frequency perturbation remain traceable along the network depth. After completing four dilated convolutional layers and connecting them with the bypass, the final layer output feature matrix is ​​obtained. ,in, Sixty-four lines List.

[0051] Will The input temporal convergence layer performs temporal dimension compression, compressing each convolutional channel. The values ​​at each time index are averaged and aggregated to obtain sixty-four scalars, which are then concatenated in the order of the convolution channel indices to form a temporal feature vector. ,in The length is sixty-four, using Indicates the convolution channel index. The value can be an integer from one to sixty-four. The Each component is The The result of compressing each convolutional channel in the time dimension is the temporal feature vector. The input is a fully connected mapping layer, which is configured with thirty-two neurons, each neuron having a pair of... Perform a weighted summation and bias concatenation to output a scalar. The thirty-two scalars are concatenated in neuron index order to form a perturbation representation vector. ,in, The dimension is thirty-two, and the perturbation representation vector is... As the sole input to the conditional Copula joint distribution head of the perturbation latent variable, it ensures that the mapping link from the perturbation input sequence to the perturbation representation vector maintains a single connection with the generation link of the support quantity-duration joint distribution.

[0052] 4) Input the perturbation representation vector into the perturbation latent variable conditional Copula joint distribution head. The perturbation latent variable conditional Copula joint distribution head sequentially generates the perturbation mechanism latent variable, the virtual inertia support quantity probability parameter, the support duration probability parameter, and the Copula correlation parameter. Based on the Copula correlation parameter, the virtual inertia support quantity probability parameter and the support duration probability parameter are combined to form the support quantity-duration joint distribution. During training, the perturbation latent variable conditional Copula joint distribution head updates the Causal-TCN network parameters based on the matching results of the support quantity-duration joint distribution on the historical virtual inertia support quantity and historical support duration. It should be noted that the reference Figure 5The perturbation representation vector is input into the joint distribution head of the conditional Copula for perturbation latent variables. This joint distribution head consists of a latent variable generation layer, a marginal parameter generation layer, and a correlation parameter generation layer connected sequentially. The latent variable generation layer performs a fully connected mapping on the perturbation representation vector, with sixteen neurons, outputting a perturbation underlying latent variable of eight dimensions. The perturbation underlying latent variable is then input into the first fully connected network of the marginal parameter generation layer to generate a virtual inertia support probability parameter of two dimensions. The perturbation underlying latent variable is further input into the second fully connected network of the marginal parameter generation layer to generate a support duration probability parameter of two dimensions. Finally, the perturbation underlying latent variable is input into the fully connected network of the correlation parameter generation layer to generate a Copula. a. Correlation parameters, wherein the Copula correlation parameter has a dimension of one; the marginal probability expression of virtual inertia support quantity is determined based on the virtual inertia support quantity probability parameter, the marginal probability expression of support duration is determined based on the support duration probability parameter, and the marginal probability expression of virtual inertia support quantity and the marginal probability expression of support duration are correlated and combined according to the Copula correlation parameter to obtain the support quantity-duration joint distribution; during the joint discrimination model training, the support demand pair composed of historical virtual inertia support quantity and historical support duration is jointly distributed and matched based on the support quantity-duration joint distribution, the network parameters of the perturbation latent variable conditional Copula joint distribution head are updated according to the matching result, and the matching result is backpropagated through the perturbation latent variable conditional Copula joint distribution head to update the network parameters of Causal-TCN.

[0053] The specific operation of step 4) is as follows: The perturbation representation vector output by Causal-TCN is denoted as... Perturbation representation vector A 32-dimensional vector, organized in component index order as follows: ,in For the first The value of each component. Take values ​​from one to thirty-two. Transform the perturbation representation vector... As the sole input to the joint distribution head of the perturbation latent variable conditional Copula, the joint distribution head is composed of a latent variable generation layer, a marginal parameter generation layer, and a correlation parameter generation layer connected sequentially, making the perturbation representation vector... In the same forward link, the generation of disturbance mechanism hidden variables, virtual inertia support probability parameters, support duration probability parameters, and Copula correlation parameters are driven sequentially.

[0054] The latent variable generation layer uses a fully connected mapping with sixteen neurons to represent the perturbation representation vector. By performing mapping, we obtain the intermediate vector of latent variables. ,in, It is a 16-dimensional vector organized according to the neuron index order. , For the first The output value of each neuron Take values ​​from one to sixteen. The latent variable generation layer will generate the intermediate vector of latent variables. Further mapping yields the hidden variables of the perturbation mechanism. Hidden variables of disturbance mechanism An eight-dimensional vector organized as ,in For the first The values ​​of each latent variable component Take numbers one through eight.

[0055] Edge parameter generation layer receives hidden variables of disturbance mechanism As input, the virtual inertia support probability parameters are output through two sets of independent fully connected networks. With support duration probability parameter Virtual inertia support probability parameters It is a two-dimensional vector, organized as ,in The mean parameter value is the marginal probability expression of the virtual inertia support quantity. The scale parameter value is the marginal probability expression of the virtual inertia support quantity; the support duration probability parameter is... It is a two-dimensional vector, organized as ,in To support the mean parameter values ​​for the duration marginal probability expression, The scale parameter values ​​are used to support the representation of the marginal probability of duration. To ensure... and To generate positive-scaled values ​​that can be directly used for numerical calculations, the network output for the scale parameter is generated using an exponential activation method. and ,make and It is a positive number.

[0056] The correlation parameter generation layer receives the hidden variables of the disturbance mechanism. As input, and output Copula correlation parameters through a set of fully connected networks. ,in It is a one-dimensional scalar value. To ensure... To obtain correlation coefficient values ​​that can be directly used in Gaussian Copula correlation structures, the network output of the correlation parameters is generated using a hyperbolic tangent activation method. ,make lie in .

[0057] Based on the probability parameter of virtual inertia support quantity Determine the marginal probability representation of the virtual inertia support quantity based on the support duration probability parameter. The marginal probability expression for determining the support duration is established. The marginal probability expression for the virtual inertia support quantity adopts a normal distribution form: for any virtual inertia support quantity value... First calculate the standardized values. Then, based on standardized numerical values, the normal distribution probability density value is calculated as... And based on standardized numerical calculations, the cumulative distribution values ​​of the normal distribution are used as... The marginal probability expression for support duration adopts a normal distribution form: for any support duration value First calculate the standardized values. Then calculate the normal distribution probability density value as And calculate the cumulative distribution value of the normal distribution as The cumulative distribution values ​​of the normal distribution are obtained through numerical integration or table lookup interpolation, making... and These are unit interval values ​​that can be directly used in Copula mappings.

[0058] Marginal probability expressions based on virtual inertia support quantity and support duration, and based on Copula correlation parameters. By performing correlation and combination, a joint distribution of support quantity and duration is constructed. Support quantity-duration joint distribution The calculation is performed using the combinatorial method of Copula decomposition, and the result is obtained by the following formula:

[0059] in, The support quantity and duration are jointly distributed across the virtual inertia support quantity values. With support duration value The joint probability density value at; The numerical value of the virtual inertia support quantity; The numerical value used to support the duration; Probabilistic parameters supported by virtual inertia The marginal probability density of the determined virtual inertia support quantity is in The value at that location; For the probability parameter of the duration of support The marginal probability density of the determined support duration is in The value at that location; To and Consistent virtual inertia support quantity edge cumulative distribution in The value at that location; To and Consistent support duration edge cumulative distribution The value at that location; Values ​​for virtual inertia support The cumulative probability variable value corresponding to the cumulative distribution at the edge of the virtual inertia support quantity; To support the value of duration The cumulative probability variable value corresponding to the cumulative distribution at the edge of the support duration; For parameters associated with Copula The determined Copula density is in The value at that location; For Copula-related parameters; For virtual inertia support probability parameters; To support the duration probability parameter.

[0060] Copula density The numerical procedure using Gaussian Copula density is employed: the cumulative probability variable is calculated. and By inputting the inverse cumulative distribution function of the standard normal distribution, two standard normal quantiles are obtained; using the correlation coefficient as... Calculate the density values ​​of the two-dimensional standard normal distribution at the two quantiles; calculate the density values ​​of the two one-dimensional standard normal distributions at the corresponding quantiles; divide the two-dimensional density value by the product of the two one-dimensional density values ​​to obtain the result. The value.

[0061] During the training of the joint discriminant model, the perturbation representation vector obtained by passing the historical perturbation input sequence through Causal-TCN is used. Inputting the latent variable conditional Copula joint distribution head yields the support-duration joint distribution. The historical virtual inertia support value is denoted as... The duration of historical support is recorded as and will This serves as a process for matching the input joint distribution to support requirements. The joint distribution matching process involves the support quantity-duration joint distribution. The following computational support requirements The corresponding joint log-likelihood value is used, and the inverse of the joint log-likelihood is used as the training loss value. Backpropagation is performed on the training loss to update the network parameters of the perturbation latent variable conditional Copula joint distribution head, making the virtual inertia support probability parameter... Support duration probability parameter Copula related parameters Synchronous updates are performed under the same training objective; the gradient of the training loss is backpropagated to the perturbation representation vector through the joint distribution head of the perturbation latent variable conditional Copula. And it is backpropagated to the network parameters of Causal-TCN, so that the joint distribution matching result of the supporting demand pairs constrains the perturbation representation vector. The formation process and output information are shown in Table 2: Table 2

[0062] 5) Determine the trigger state based on the frequency change rate and the preset frequency change rate threshold. In the trigger state, obtain the initial virtual inertia support amount and initial support duration based on the joint distribution of support amount and duration. In the non-trigger state, obtain zero support amount and zero duration. It should be noted that the reference Figure 6 The frequency change rate of the last sampling point in the perturbation input sequence is selected as the determination frequency change rate. The amplitude of the determination frequency change rate is taken and compared with a preset frequency change rate threshold. If the amplitude is not less than the preset frequency change rate threshold, it is determined as a triggered state; if the amplitude is less than the preset frequency change rate threshold, it is determined as a non-triggered state. In the triggered state, the joint distribution of support quantity and duration is used as the discrimination criterion for support demand pairs. The support demand pair with the highest joint probability is determined, and the support quantity in the support demand pair is determined as the initial virtual inertia support quantity, and the duration in the support demand pair is determined as the initial support duration. In the non-triggered state, the initial virtual inertia support quantity is determined as zero support quantity, and the initial support duration is determined as zero duration.

[0063] The specific operation of step 5) is as follows: Let the perturbation input sequence be denoted as , It is a three-channel timing signal matrix with a time dimension length of . The three channels, in order, represent real-time wind power, system frequency, and frequency change rate. The frequency change rate channel is indexed by time... The value of the end sampling point is selected as the determination frequency change rate. Determine the rate of change of frequency It is a single scalar value.

[0064] Determine the rate of change of frequency The amplitude value is obtained by determining the amplitude. Determine the amplitude It is a non-negative scalar value. The preset frequency change rate threshold is denoted as... Preset frequency change rate threshold The value is a pre-defined non-negative scalar. The amplitude will be determined. Compared with the preset frequency change rate threshold Comparison yields trigger flags Triggering flag A binary label with a value of or ;when Not less than season And it is determined to be in a triggered state, when Less than season And it was determined to be in a non-triggered state.

[0065] In the trigger flag When the support quantity-duration joint distribution is denoted as Support quantity-duration joint distribution This refers to the joint probability density calculation object output by the Copula joint distribution head under perturbation latent variable conditions. The preset lower bound for the virtual inertia support quantity search is denoted as... The upper bound of the virtual inertia support quantity search is denoted as... The preset virtual inertia support search step size is denoted as ;in , , All are directly readable scalar values ​​and satisfy the following conditions: Not less than , It is a positive number. The preset support duration search lower bound is denoted as . The upper bound of the preset support duration search is denoted as The preset support duration search step size is denoted as ;in, , , All are directly readable scalar values ​​and satisfy the following conditions: Not less than , It is a positive number. According to... , , Candidate support vectors are generated using an equally spaced, incremental method. ,in For the first Candidate values ​​for virtual inertia support quantity Candidate index and Take one to , The number of candidate support quantities is a positive integer; according to , , Generate candidate duration vectors in an equally spaced incrementing manner. ,in, For the first One support duration candidate value Candidate index and Take one to , The number of candidate durations is a positive integer.

[0066] Candidate support vector With candidate duration vector Perform grid combination to form a candidate support requirement pair matrix. Candidate support requirements for the matrix Include Two-dimensional data sets For each group Perform joint probability density evaluation: and As the input support quantity-duration joint distribution at the value point The evaluation process yields the joint probability density value at the given point, and all evaluation results are organized into a joint probability matrix according to the candidate indices. Joint probability matrix for OK Column matrix; joint probability matrix elements For the joint distribution of support quantity and duration exist The joint probability density values ​​at [location]. A traversal comparison method is used in the joint probability matrix. Search for the maximum element in the middle to obtain the candidate index corresponding to the maximum element. ,in This is the optimal candidate index for virtual inertia support. To support the optimal candidate index for duration, the optimal support requirement pair is represented as follows: and will Support volume Determined as the initial virtual inertia support quantity ,Will Duration in Determined as the initial support duration The initial virtual inertia support amount Duration of initial support All of these are scalar values ​​that can be directly used to control execution.

[0067] In the trigger flag At that time, the initial virtual inertia support amount Determine the support quantity to be zero and set the value to a numerical value. The initial support duration The duration is set to zero and the value is a numerical value. The virtual inertia support process is shown in Table 3: Table 3

[0068] 6) Under the constraints of the maximum exchange power of the energy storage system and the available energy of the energy storage system, the initial virtual inertia support quantity and the initial support duration are converted into virtual inertia support commands, and the virtual inertia support commands are sent to the energy storage control execution terminal.

[0069] It should be noted that the reference Figure 7 The system acquires the maximum exchange power constraint and available energy constraint of the energy storage system, and uses them as constraint inputs for the virtual inertia support command. It determines the effective duration of the virtual inertia support command based on the initial support duration and the start time of the virtual inertia support command. Within the effective duration of the virtual inertia support command, it generates an energy storage active power setpoint based on the initial virtual inertia support amount and the real-time frequency change rate, and limits the energy storage active power setpoint to the power range allowed by the maximum exchange power constraint of the energy storage system. Based on the limited energy storage active power setpoint and the effective duration of the virtual inertia support command, it determines the required energy demand. If the energy demand exceeds the available energy constraint of the energy storage system, it adjusts the energy storage active power setpoint or the effective duration to obtain an executable virtual inertia support command, and sends the virtual inertia support command to the energy storage control execution terminal.

[0070] The specific operation of step 6) is as follows: The maximum exchange power constraint of the energy storage system is denoted as: , The scalar value sent by the energy storage control actuator, in units of or This is used to limit the absolute upper limit of the given active power of the energy storage system. , The scalar value sent by the energy storage control actuator, in units of or This is used to limit the total energy consumed by the virtual inertia support command within its effective duration. The initial virtual inertia support quantity is denoted as... The initial support duration is denoted as , and All of these are scalar values ​​generated in the previous discrimination step.

[0071] The effective duration of the virtual inertia support command is denoted as... The duration-based clipping method is used. get Read the preset minimum duration With preset maximum duration , and All are scalar values ​​and satisfy Not less than ;when Less than season ,when Greater than season ,when Located in the interval season The start time of the virtual inertia support command is recorded as... , Retrieve the current control cycle timestamp , This is a timestamp scalar value that is consistent between the energy storage control execution end and the controller where the joint discrimination model is located.

[0072] During the effective duration of the virtual inertia support command Within, based on the sampling period Discrete generation of control sequences, This is a scalar value. Calculate the number of control points. , For the reason and A definite positive integer. Organize the real-time frequency change rate sampling sequence over the effective duration into a vector. ,in For the first The real-time frequency change rate value at each sampling time. Take one to Read the preset power conversion factor. , These are scalar values, with units consistent with the power per unit or actual power scale used by the system. For each sampling index... Perform point-by-point calculation: first calculate intermediate quantities. and The product value is then used to calculate the given active power component of the unconfined energy storage. for and The product of these components is inverted to ensure that the active power given by the energy storage system is in the opposite direction to the real-time rate of change of frequency. All components of the unconstrained active power given by the energy storage system are then organized by time index into an unconstrained active power given vector. .

[0073] Given a vector of active power for unconfined energy storage By applying the maximum exchange power constraint limit, the constrained active power give vector of the energy storage is obtained. The cropping rule uses point-by-point cropping: when Greater than season ,when Less than season In other cases This ensures that the active power of the energy storage system at any sampling time meets the maximum exchange power constraint of the energy storage system.

[0074] Based on the given vector of the constrained energy storage active power Effective duration of virtual inertia support commands Determine the required energy demand, denoted as . , The value is a scalar. It is calculated using a discrete accumulation method. For each sampling index Calculate energy fragments Energy fragments Pick and The product of these factors, summed over all energy segments in sampling index order, yields the result. Organize all energy fragments into energy fragment vectors. And organize the cumulative energy into a cumulative energy vector. ,in for In index one to The cumulative value within the range.

[0075] When the required energy demand Not exceeding the available energy constraint of the energy storage system At that time, the virtual inertia support instructions are organized into instruction data groups. When the required energy demand Exceeding the available energy constraints of energy storage systems At that time, energy constraint satisfaction processing is performed, and a deterministic choice is made between adjusting the given active power of the energy storage and the effective duration of the adjustment: the preset adjustment mode flag is read. , For the value to be or A binary scalar; when Adjust the active power setpoint of energy storage in real time and calculate the scaling factor. for and The ratio, and Each component multiplied by Obtain the adjusted power command vector At the same time, maintain Unchanged; when The effective duration is adjusted in time, using the cumulative energy vector. Search minimum index make No more than and Exceed Update the effective duration to and update the power give vector to Organize the executable virtual inertia support instructions into instruction data sets. and the instruction data group Send to the energy storage control execution terminal, which then executes the command according to the instructions. Start and Internal point-to-point tracking .

[0076] This invention has the following characteristics: This invention replaces the traditional two independent regressions or fixed mappings for virtual inertia support quantity and support duration with a unified modeling of support demand pairs. A Copula joint distribution head for perturbation latent variables is introduced at the model output: first, a perturbation mechanism latent variable is generated from the perturbation representation vector to accommodate mechanistic differences in the initial observation signal that are difficult to distinguish with a single definite value; then, marginal probability parameters for support quantity and duration are generated separately, and Copula correlation parameters are generated to construct their joint distribution. During the training phase, the joint distribution matching results (joint likelihood) of historical support demand pairs are used to simultaneously constrain the marginal and correlation parameters, enabling the model to express the sources of uncertainty and their interrelationships when observations are similar but support demands differ. This avoids inconsistencies in the combination of support quantity and duration caused by independent outputs, and from the perspective of discrimination results, it is more in line with the actual use of paired support demands in frequency modulation scenarios.

[0077] This invention proposes a novel end-to-end learning link for disturbance characterization extraction and joint distribution-driven learning. It achieves this by using only three channels of time-series data: real-time wind power, system frequency, and frequency change rate, through Causal... TCN enables the causalization and multi-timescale representation of initial disturbance information. This Causal TCN employs channel-separated causal convolution to process three types of physical quantities separately, reducing the impact of shallow aliasing of signals with different dimensions on feature formation. Subsequently, channel-fusion convolution forms coupled features, and by expanding convolutional layers, it simultaneously captures short-term abrupt changes in the rate of frequency change and long-term continuous shifts in the system frequency within the same time window. Bypass-addition connections ensure the traceability of shallow perturbation patterns in deep features. More importantly, the joint distribution matching loss uses the Copula distribution head to inversely constrain Causal... The TCN parameters ensure that the learning objective of the perturbation representation vector directly targets the support quantity. The requirement for pairwise discrimination based on duration, rather than relying on general statistical fitting of the original frequency or power signals, forms a feature extraction path oriented towards business discrimination.

[0078] This invention integrates the joint judgment results with the on-site triggering logic and energy storage execution constraints. During operation, the triggering state is determined based on the frequency change rate threshold, and triggering is based on the support quantity. The duration joint distribution selects the support demand pair with the highest joint probability. When not triggered, it outputs zero support and zero duration, ensuring consistency between the discrimination and startup logic. Subsequently, under the constraints of maximum energy storage exchange power and available energy, the initial support quantity and duration are converted into a power command sequence and subjected to limiting and energy consistency adjustments, allowing the output command to be directly executed within the power boundary and energy budget. Compared to methods that only provide a single support quantity or empirical duration, this closed-loop link unifies the discrimination, triggering, and constraint-based command formation of paired support demands within the same technical framework, enabling the virtual inertia support demand in the initial stage of disturbance to be implemented at the control execution end in a form that meets on-site constraints.

[0079] Example 2 The frequency regulation virtual inertia discrimination system for thermal power plants described in this invention includes: The acquisition module is used to acquire real-time wind power, system frequency and frequency change rate during grid-connected frequency regulation operation of thermal power plants, and to use these to perturb the input sequence. The judgment module is used to train a joint discrimination model using historical perturbation input sequences and corresponding historical virtual inertia support quantities and historical support durations. The joint discrimination model includes the Causal-TCN and the perturbation latent variable conditional Copula joint distribution head. The perturbation module is used to input the perturbation input sequence into the Causal-TCN of the joint discriminant model to obtain the perturbation representation vector. The generation module is used to input the disturbance characterization vector into the disturbance latent variable conditional Copula joint distribution head. The disturbance latent variable conditional Copula joint distribution head generates the disturbance mechanism latent variable, the virtual inertia support quantity probability parameter, the support duration probability parameter, and the Copula correlation parameter in sequence. Based on the Copula correlation parameter, the virtual inertia support quantity probability parameter and the support duration probability parameter are combined to form the support quantity-duration joint distribution. The triggering module is used to determine the triggering state based on the frequency change rate and the preset frequency change rate threshold. In the triggering state, the initial virtual inertia support amount and the initial support duration are obtained based on the joint distribution of support amount and duration. The conversion module is used to convert the initial virtual inertia support quantity and initial support duration into virtual inertia support commands under the constraints of the maximum exchange power of the energy storage system and the available energy of the energy storage system.

[0080] The module division in this embodiment is illustrative and represents only one logical functional division. In actual implementation, other division methods may be used. Furthermore, the functional modules in each embodiment of this application can be integrated into a single processor, exist as separate physical entities, or be integrated into a single module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0081] Example 3 A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the virtual inertia discrimination method for frequency regulation in thermal power plants. For example, the method includes: acquiring real-time wind power, system frequency, and frequency change rate during grid-connected frequency regulation operation of a thermal power plant, and using these as a disturbance input sequence; training a joint discrimination model using historical disturbance input sequences and corresponding historical virtual inertia support quantities and historical support durations, the joint discrimination model including a Causal-TCN and a joint distribution head of disturbance latent variable conditional Copula; inputting the disturbance input sequence into the Causal-TCN of the joint discrimination model to obtain a disturbance representation vector; and inputting the disturbance representation vector into the disturbance... In the Copula joint distribution head for the dynamic latent variable condition, the Copula joint distribution head for the disturbance latent variable condition sequentially generates the disturbance mechanism latent variable, the probability parameter of the virtual inertia support quantity, the probability parameter of the support duration, and the Copula correlation parameter. Based on the Copula correlation parameter, the probability parameter of the virtual inertia support quantity and the probability parameter of the support duration are combined to form a support quantity-duration joint distribution. The trigger state is determined based on the frequency change rate and a preset frequency change rate threshold. In the trigger state, the initial virtual inertia support quantity and the initial support duration are obtained based on the support quantity-duration joint distribution. Under the constraints of the maximum exchange power of the energy storage system and the available energy of the energy storage system, the initial virtual inertia support quantity and the initial support duration are converted into virtual inertia support commands. The memory may include main memory, such as high-speed random access memory, or it may also include non-volatile memory, such as at least one disk storage device. The processor, network interface, and memory are interconnected via an internal bus, which can be an industry standard architecture bus, a peripheral component interconnection standard bus, an extended industry standard architecture bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. The memory is used to store programs; specifically, the program may include program code, which includes computer operation instructions. The memory may include main memory and non-volatile memory, and provides instructions and data to the processor.

[0082] Example 4 A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the virtual inertia discrimination method for frequency regulation in thermal power plants. For example, the method includes: acquiring real-time wind power, system frequency, and frequency change rate during grid-connected frequency regulation operation of a thermal power plant, and using these as a disturbance input sequence; training a joint discrimination model using historical disturbance input sequences and corresponding historical virtual inertia support quantities and historical support durations, the joint discrimination model including a Causal-TCN and a disturbance latent variable conditional Copula joint distribution head; inputting the disturbance input sequence into the Causal-TCN of the joint discrimination model to obtain a disturbance representation vector; and inputting the disturbance representation vector into the disturbance latent variable conditional Copula. In the Copula joint distribution head, the perturbation latent variable conditional Copula joint distribution head sequentially generates the perturbation mechanism latent variable, the virtual inertia support probability parameter, the support duration probability parameter, and the Copula correlation parameter. Based on the Copula correlation parameter, the virtual inertia support probability parameter and the support duration probability parameter are combined to form a support quantity-duration joint distribution. A trigger state is determined based on the frequency change rate and a preset frequency change rate threshold. In the trigger state, the initial virtual inertia support quantity and initial support duration are obtained based on the support quantity-duration joint distribution. Under the constraints of the maximum exchange power of the energy storage system and the available energy of the energy storage system, the initial virtual inertia support quantity and initial support duration are converted into virtual inertia support commands. Specifically, the computer-readable storage medium includes, but is not limited to, volatile memory and / or non-volatile memory. The volatile memory may include random access memory (RAM) and / or cache memory, etc. The non-volatile memory may include read-only memory (ROM), hard disk, flash memory, optical disk, magnetic disk, etc.

[0083] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0084] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0085] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0086] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0087] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and disclosure of the invention. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the following claims.

[0088] It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

[0089] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the present invention. Any simple modifications, alterations, or equivalent structural changes made to the above embodiments based on the technical essence of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for discriminating virtual inertia of frequency modulation of a thermal power plant, characterized in that, include: 1) During the grid-connected frequency regulation operation of thermal power plants, real-time wind power, system frequency, and frequency change rate are obtained, and the input sequence is perturbed accordingly; 2) A joint discrimination model is trained using the historical perturbation input sequence and the corresponding historical virtual inertia support quantity and historical support duration. The joint discrimination model includes the Causal-TCN and the perturbation latent variable conditional Copula joint distribution head. 3) Input the perturbation input sequence into the Causal-TCN of the joint discriminant model to obtain the perturbation representation vector; 4) Input the disturbance characterization vector into the disturbance latent variable conditional Copula joint distribution head. The disturbance latent variable conditional Copula joint distribution head generates the disturbance mechanism latent variable, the virtual inertia support quantity probability parameter, the support duration probability parameter, and the Copula correlation parameter in sequence. Based on the Copula correlation parameter, the virtual inertia support quantity probability parameter and the support duration probability parameter are combined to form the support quantity-duration joint distribution. 5) Determine the trigger state based on the frequency change rate and the preset frequency change rate threshold. In the trigger state, obtain the initial virtual inertia support amount and the initial support duration based on the joint distribution of support amount and duration. 6) Under the constraints of the maximum exchange power of the energy storage system and the available energy of the energy storage system, the initial virtual inertia support quantity and the initial support duration are converted into virtual inertia support commands.

2. The method of claim 1, wherein the method further comprises: The specific operation of step 1) is as follows: Set the sampling period and preset sequence length of the perturbation input sequence; During grid-connected frequency regulation operation of thermal power plants, real-time wind power and system frequency are obtained at the same moment according to the sampling period. The frequency change rate is calculated based on the system frequency of adjacent sampling periods, and the frequency change rate is associated with the system frequency samples involved in the calculation at the same time. The real-time wind power, system frequency, and frequency change rate at the same moment are used as sampling points. A series of sampling points of a preset length are arranged in chronological order to form a disturbance input sequence. The disturbance input sequence is then organized into a three-channel time-series signal within the model. The three channels are, in order, real-time wind power, system frequency, and frequency change rate.

3. The method of claim 1, wherein the method further comprises: Using the historical perturbation input sequence as input, forward computation is performed to output the joint distribution of support quantity and duration. The historical virtual inertia support quantity and historical support duration are used as support demand pairs for joint distribution matching, and the network parameters of the joint discriminant model are updated in reverse accordingly. The matching result synchronously constrains the probability parameters of virtual inertia support quantity, support duration, and Copula correlation parameters. The network parameters of Causal-TCN are updated in reverse through the perturbation latent variable conditional Copula joint distribution head, resulting in the trained joint discriminant model.

4. The method of claim 1, wherein the method further comprises: The loss function of the joint discriminant model during training is: in, This is due to batch training loss; This refers to the batch sample size. The sample number; This is the historical virtual inertia support quantity; The duration of historical support; For virtual inertia support probability parameters; To support the duration probability parameter; For Copula-related parameters; For the reason The normal distribution probability density corresponding to the marginal probability expression of the determined virtual inertia support quantity; For the reason The marginal probability expression of the determined support duration corresponds to the normal distribution probability density; To and A consistent normal cumulative distribution function; To and A consistent normal cumulative distribution function; This represents the cumulative probability variable of historical virtual inertia support quantity under the marginal probability expression of virtual inertia support quantity; The cumulative probability variable representing the historical support duration under the marginal probability expression of the support duration; For the reason Determined Gaussian Copula density.

5. The method of claim 1, wherein the method further comprises: The perturbation input sequence is: wherein, is a preset sequence length; is a sampling sequence number; is a real-time wind power value at the corresponding sampling sequence number; is a system frequency value at the corresponding sampling sequence number; is a frequency change rate value at the corresponding sampling sequence number.

6. The method of claim 1, wherein the method further comprises: The specific operation of step 4) is as follows: The perturbation representation vector is input into the perturbation latent variable conditional Copula joint distribution head, which is composed of a latent variable generation layer, a marginal parameter generation layer and a correlation parameter generation layer connected in sequence. A latent variable generation layer is used to perform a fully connected mapping on the perturbation representation vector. The number of neurons in the fully connected mapping is sixteen, and the output perturbation mechanism latent variable has a dimension of eight. The hidden variables of the disturbance mechanism are input into the first set of fully connected networks of the edge parameter generation layer to generate virtual inertia support probability parameters, wherein the virtual inertia support probability parameters have a dimension of two. The hidden variables of the disturbance mechanism are input into the second set of fully connected networks of the edge parameter generation layer to generate the support duration probability parameter, which has a two-dimensional dimension. The hidden variables of the disturbance mechanism are input into the fully connected network of the association parameter generation layer to generate Copula association parameters, wherein the Copula association parameters have a dimension of one. The marginal probability expression of virtual inertia support quantity is determined based on the probability parameter of virtual inertia support quantity, and the marginal probability expression of support duration is determined based on the probability parameter of support duration. The marginal probability expression of virtual inertia support quantity and the marginal probability expression of support duration are associated and combined according to the Copula correlation parameter to obtain the joint distribution of support quantity-duration.

7. The method of claim 1, wherein the method further comprises: The specific operation of step 5) is as follows: The frequency change rate of the last sampling point in the perturbation input sequence is selected as the determination frequency change rate. The amplitude of the determined frequency change rate is taken and compared with a preset frequency change rate threshold. If the amplitude is not less than the preset frequency change rate threshold, it is determined to be a triggered state; if the amplitude is less than the preset frequency change rate threshold, it is determined to be a non-triggered state. In the triggered state, the joint distribution of support quantity and duration is used as the criterion for determining support demand pairs. The support demand pair with the highest joint probability is determined, and the support quantity in the support demand pair is determined as the initial virtual inertia support quantity, and the duration in the support demand pair is determined as the initial support duration.

8. A system for discriminating virtual inertia of frequency modulation of a thermal power plant, characterized in that, include: The acquisition module is used to acquire real-time wind power, system frequency, and frequency change rate during grid-connected frequency regulation operation of thermal power plants, and to use these to perturb the input sequence. The judgment module is used to train a joint discrimination model using historical perturbation input sequences and corresponding historical virtual inertia support quantities and historical support durations. The joint discrimination model includes the Causal-TCN and the perturbation latent variable conditional Copula joint distribution head. The perturbation module is used to input the perturbation input sequence into the Causal-TCN of the joint discriminant model to obtain the perturbation representation vector. The generation module is used to input the disturbance characterization vector into the disturbance latent variable conditional Copula joint distribution head. The disturbance latent variable conditional Copula joint distribution head generates the disturbance mechanism latent variable, the virtual inertia support quantity probability parameter, the support duration probability parameter, and the Copula correlation parameter in sequence. Based on the Copula correlation parameter, the virtual inertia support quantity probability parameter and the support duration probability parameter are combined to form the support quantity-duration joint distribution. The triggering module is used to determine the triggering state based on the frequency change rate and the preset frequency change rate threshold. In the triggering state, the initial virtual inertia support amount and the initial support duration are obtained based on the joint distribution of support amount and duration. A conversion module is configured to convert the initial virtual inertia support amount and the initial support duration into a virtual inertia support instruction under a maximum exchange power constraint of the energy storage system and an available energy constraint of the energy storage system.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor implements the steps of the method for generating and sending the trusted state alarm information according to any one of claims 1-7 when executing the computer program.

10. A computer-readable storage medium storing a computer program, the computer-readable storage medium comprising: The computer program is executed by the processor to implement the steps of the method for generating and sending the trusted state alarm information according to any one of claims 1-7.

Images