Method, device, system and medium for generating long-term scene in wind and light multi-scene station

Abstract

The application discloses a kind of wind light multi-station medium and long-term scene generation method, device, system and medium, belong to electrical engineering field, method includes: extracting state component and fluctuation component in wind light multi-station historical scene and fitting fluctuation component;State component is segmented after normalization and input conditional generation confrontation network;For each section input, its mean and variance are used as condition item input network, output multi-station generation state component, the fitted fluctuation component is superimposed to multi-station generation state component, obtains generation day scene;History day scene set and generation day scene set are clustered to be converted into discrete typical day state, according to typical day state, determine cumulative state transition matrix and generate corresponding day state transition sequence, select day scene in day state transition sequence and carry out splicing to meet preset fluctuation component criterion, obtain wind light multi-station medium and long-term scene.For power system dispatching provides stable boundary.

Description

Technical Field

[0001] This invention belongs to the field of electrical engineering, and more specifically, relates to a method, apparatus, system and medium for generating long-term scenarios at multiple wind and solar power stations. Background Technology

[0002] Since the beginning of the 21st century, global fossil fuel power generation capacity has been declining, leading to a new production and consumption pattern with new energy sources as a significant component. With the increasing proportion of new energy sources in the power system and the corresponding new energy power generation equipment, the problem of wind and solar power curtailment is becoming increasingly apparent. Furthermore, with a large amount of wind and solar power being integrated into the grid, accurately describing the uncertainty of new energy output has become a crucial prerequisite for ensuring the stability of the power system and a vital guarantee for its operation, planning, and dispatch.

[0003] Existing technologies primarily employ statistical methods and deep learning to study probabilistic modeling techniques for generating new energy power output scenarios. Deep learning relies on neural networks, trained solely on the input dataset. During training, no likelihood function is involved; the generated data automatically learns the underlying physical characteristics of historical data to accurately generate time series. However, existing technologies for generating long-term scenarios from multiple wind and solar power plants have shortcomings: for geographically proximate wind and solar power plants, the fitting of power output fluctuations does not consider the correlation between plants; and the fitting accuracy for the statistical characteristics and spatial correlations of historical scenarios from multiple wind and solar power plants is relatively low. Summary of the Invention

[0004] To address the shortcomings and improvement needs of existing technologies, this invention provides a method, apparatus, system, and medium for generating long-term scenes of multiple wind and solar power stations. The purpose is to solve the problem that existing technologies do not consider the spatial correlation between multiple stations at the same location, resulting in low accuracy of the generated long-term scenes of multiple wind and solar power stations.

[0005] To achieve the above objectives, according to one aspect of the present invention, a method for generating medium- and long-term scenes from multiple wind and solar power stations is provided, characterized by comprising: extracting state components and fluctuation components from historical scenes of multiple wind and solar power stations; fitting the fluctuation components; normalizing the state components and then inputting them into a conditional generative adversarial network (GAN); for each normalized state component, inputting its mean and variance as conditional terms into the GAN, outputting multi-station generated state components; superimposing the fitted fluctuation components onto the multi-station generated state components to obtain generated daily scenes; each generated daily scene forming a generated daily scene set; clustering the historical daily scene set and the generated daily scene set of the multiple wind and solar power stations to transform them into discrete typical daily states; determining a cumulative state transition matrix based on the typical daily states; generating a daily state transition sequence corresponding to the cumulative state transition matrix using the MCMC method; selecting daily scenes that satisfy a preset fluctuation component criterion from the daily state transition sequence and splicing them together to obtain medium- and long-term scenes from multiple wind and solar power stations.

[0006] Furthermore, the fitting of the fluctuation component specifically includes: selecting the optimal R-vine Copula first-layer tree structure, using the nonparametric kernel density estimation method to estimate the binary Copula function of the fluctuation component in the optimal R-vine Copula first-layer tree structure; and combining the probability distribution function and marginal distribution of each station in the wind and solar multi-station system, using the binary Copula function to establish the joint probability density function of the fluctuation component.

[0007] Furthermore, the step of superimposing the fitted fluctuation component onto the multi-site generated state component specifically includes: combining the joint probability density function of the fluctuation component, randomly selecting the fluctuation component, and superimposing the selected fluctuation component onto the multi-site generated state component.

[0008] Furthermore, before using the conditional generative adversarial network to output multi-site generated state components, the method further includes: training the conditional generative adversarial network using a generator loss function and a discriminator loss function, wherein the generator loss function and the discriminator loss function are respectively:

[0009] Loss D =-E X～p(x) [log(D(X|c))]-E X～p(x') [log(1-D(X|c))]

[0010] Loss G =-E X'～p(x') [log(D(X'|c)))]-λ e L e -λ var L var

[0011]

[0012]

[0013] Among them, Loss D For the discriminator loss function, Loss G Let X be the generator loss function, X be the generator-generated sample, X' be the historical sample of the scenic spot, x be the value of the random variable of the historical scenic spot, x' be the value of the random variable of the generated scenic spot, c be the conditional input, p() be the probability distribution of the random variable, D(X|c) be the output value of X input to the discriminator under condition c, D(X'|c) be the output value of X' input to the discriminator under condition c, G(z,c) be the generator-generated sample under condition c, and E[] be the expected value of the discriminator output. x [] represents the conditional term metric for generating samples, λ e λ is the coefficient of the mean-conditional measure term. var L is the coefficient of the variance conditional measure. e For the loss of the mean-conditional term, L var For the variance conditional measure of loss, c N,e Generate the mean of the samples for the Nth station, c N,v Let z be the variance of the samples generated for the Nth station, and z be Gaussian noise.

[0014] Furthermore, determining the cumulative state transition matrix based on the typical daily state specifically includes: for any typical daily state, based on the typical daily state and the typical daily state of the next day, adjusting matrix S... N×N element S in mn The process involves incrementing by 1. The typical daily state is m, the typical daily state for the next day is n, N is the number of states for the typical daily state, and S... N×N Initially, a matrix of all zeros is used; traverse each typical daily state and apply S... N×N Increment the corresponding elements by 1 to obtain the final matrix S′. N×N ; will S′ N×N Each element is divided by the sum of the elements in its row to construct a state transition matrix; a cumulative state transition matrix is ​​constructed, wherein the value in the i-th row and j-th column of the cumulative state transition matrix is ​​equal to the sum of the elements in the i-th row and the first j-th column of the state transition matrix.

[0015] Furthermore, the step of generating the daily state transition sequence corresponding to the cumulative state transition matrix using the MCMC method specifically includes: comparing the random number a with each element in the h-th row of the cumulative state transition matrix until the random number a is between two adjacent elements in the h-th row, generating the typical daily state for the next day and assigning it a value k; where h is the value of the typical daily state for the current day, k is the column number of the smaller column element among the two adjacent elements, and the typical daily state for the first day is randomly generated; repeating the above operation until the typical daily state for the H-th day is generated, where H is a set value.

[0016] Furthermore, the preset fluctuation component criterion is:

[0017] ΔP data，min ≤ΔP new ≤ΔP data，max

[0018] Wherein, ΔP new ΔP represents the fluctuation component of the period connecting the daily scene of the next typical day and the daily scene of the current typical day. data，min ΔP data，max These are the minimum and maximum values ​​of the historical data fluctuation components, respectively.

[0019] According to another aspect of the present invention, a device for generating medium- and long-term scenes of multiple wind and solar power stations is provided, used to execute the above-described method for generating medium- and long-term scenes of multiple wind and solar power stations, comprising: an extraction and fitting module, used to extract state components and fluctuation components from historical scenes of multiple wind and solar power stations, and to fit the fluctuation components; a first generation module, used to normalize the state components and input them into a conditional generative adversarial network in segments; for each normalized state component, its mean and variance are used as condition terms to input the conditional generative adversarial network, outputting multi-station generated state components, and superimposing the fitted fluctuation components onto the multi-station generated state components to obtain generated daily scenes; each generated daily scene constitutes a set of generated daily scenes; a clustering module, used to cluster the historical daily scene set and the generated daily scene set of multiple wind and solar power stations to transform them into discrete typical daily states, and to determine a cumulative state transition matrix based on the typical daily states; a second generation module, used to generate a daily state transition sequence corresponding to the cumulative state transition matrix using the MCMC method, and to select daily scenes that satisfy a preset fluctuation component criterion from the daily state transition sequence and splice them together to obtain medium- and long-term scenes of multiple wind and solar power stations.

[0020] According to another aspect of the present invention, a long-term scene generation system for multiple wind and solar power sites is provided, comprising: a processor; and a memory storing a computer-executable program, wherein when the program is executed by the processor, the processor performs the long-term scene generation method for multiple wind and solar power sites as described above.

[0021] According to another aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the method for generating long-term scenes of multiple wind and solar power stations as described above.

[0022] In summary, the above-described technical solutions conceived in this invention can achieve the following beneficial effects:

[0023] (1) A method for generating medium- and long-term scenarios of multiple wind and solar power stations is provided. When fitting the power output fluctuation of wind and solar power stations with similar geographical locations, the correlation between the stations is considered. On the one hand, the spatial correlation between multiple stations in the same location is satisfied, and on the other hand, the effectiveness of the generated scenarios is fully guaranteed. A scenario set that conforms to the spatiotemporal characteristics and statistical characteristics is obtained, which accurately simulates the power output characteristics and spatial correlation of wind and solar power, and provides data support for power system scheduling and operation.

[0024] (2) Using the MCMC method to generate discrete typical daily transition states can better express the randomness and time correlation of new energy sequences such as wind power and photovoltaics. The output at the next moment is determined only by the output value at the previous moment, which is consistent with the vertical correlation of the output of new energy scenarios and further improves the authenticity of generating long-term scenarios of multiple wind and solar power stations. Attached Figure Description

[0025] Figure 1 A flowchart of a method for generating medium- and long-term scenes across multiple wind and solar power stations provided in an embodiment of the present invention;

[0026] Figure 2 An explanatory diagram illustrating the generation of daily scene sets based on the C-WGAN-multi model provided in an embodiment of the present invention;

[0027] Figure 3 The spatial correlation matrix heatmap of historical wind power scenarios in different seasons provided in this embodiment of the invention;

[0028] Figure 4 This invention provides a spatial correlation matrix heatmap of historical photovoltaic scenes in different seasons, as provided in an embodiment of the invention.

[0029] Figure 5 This is a heatmap of the spatial correlation matrix of wind power generation scenarios in different seasons provided in an embodiment of the present invention.

[0030] Figure 6 This is a heatmap of the spatial correlation matrix of photovoltaic generation scenarios in different seasons provided in an embodiment of the present invention;

[0031] Figure 7 This is a block diagram of a long-term scene generation device for multiple wind and solar power stations provided in an embodiment of the present invention. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0033] In this invention, the terms "first," "second," etc. (if present) in the invention and the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0034] Figure 1 This is a flowchart illustrating a method for generating medium- to long-term scenes across multiple wind and solar power stations, as provided in an embodiment of the present invention. (See also...) Figure 1 , combined Figures 2-6 The method for generating long-term scenes of multiple wind and solar power stations in this embodiment is described in detail. The method includes operations S1-S4.

[0035] Operation S1 extracts the state components and fluctuation components from the historical scenes of multiple wind and solar power stations, and fits the fluctuation components.

[0036] The hourly power data from historical annual scenarios at multiple wind and solar power plants are divided into state components and fluctuation components. The possible power values ​​at each moment are discretized, with each power interval representing a state component, ranging from 0 to the maximum output. For example, the maximum output of a wind farm is P. max If N states are defined, then the power range P represented by each state component is... s =P max / N. The fluctuation component is defined as the change in output ΔP = P between two consecutive time points. n -P n-1 n = 2, 3, 4…N, where P n Let be the wind and solar power at the nth time point.

[0037] State components for each moment in the historical annual scenes of the multi-site wind and solar power system are extracted, as well as fluctuation components for the historical annual scenes of the multi-site wind and solar power system. The fluctuation components are then fitted using the R-Teng Copula model.

[0038] Copula functions model the marginal distributions and correlations of random variables separately, thereby constructing the joint probability distribution of multivariate random variables. The R-Teng Copula model transforms the correlation problem of multidimensional variables into the correlation problem of multiple two-dimensional variables, based on the combination of joint probability distribution functions and Copula theory. The R-Teng Copula function characterizes the correlation of wind and solar field fluctuations at the same geographical location by establishing a joint distribution model of fluctuations at multiple fields, obtaining the joint probability density function of the fluctuation components at multiple fields.

[0039] According to an embodiment of the present invention, fitting the fluctuation component specifically includes the following sub-operations S11-S12.

[0040] In suboperation S11, the optimal R-vine Copula first-level tree structure is selected, and the bivariate Copula function of the fluctuation component in the optimal R-vine Copula first-level tree structure is estimated using the nonparametric kernel density estimation method.

[0041] In this embodiment, the maximum spanning tree method is used to determine the optimal vine structure. The maximum spanning tree method requires that the sum of the weights of the edges connecting each node in the structure is maximized. The Pearson rank correlation coefficient is selected to characterize spatial correlation, and using the Pearson rank correlation coefficient as the weight, a vine structure is searched that maximizes the sum of the weights of all edges in the first-level tree. The expression for the Pearson rank correlation coefficient is as follows:

[0042]

[0043] Where, ρ X,Y Let be the Pearson rank correlation coefficient between X and Y, where X and Y are historical annual scene sequences from two different stations, cov(X,Y) is the covariance of X and Y, and σ is the Pearson rank correlation coefficient between X and Y. X σ Y Let μ be the standard deviation of X and Y, respectively. X μ Y Let X and Y be the means of X and Y respectively, and E[] be the expected value function.

[0044] Furthermore, using the nonparametric kernel density estimation method, the bivariate Copula functions of the fluctuation components in the first-level tree structure of the optimal R-vine Copula are estimated, for example, C(x1,x2), C(x2,x3), C(x3,x4), and C(x3,x5).

[0045] In suboperation S12, the joint probability density function of the wave components is established by combining the probability distribution function and marginal distribution of each station in the multi-station wind and solar power system with the binary Copula function.

[0046] The sampling points of the conditional distributions C(x1,x3|x2), C(x2,x4|x3), and C(x4,x5|x3) can be obtained according to the following formula. Taking C(x1,x3|x2) as an example, the conditional probability distribution C(x1,x3|x2) is established as follows:

[0047]

[0048]

[0049] Wherein, P(x i ), F(xi ) are random variables x i The probability distribution function and marginal distribution.

[0050] Taking C(x1,x4|x2,x3) as an example, establish the conditional probability distribution C(x1,x4|x2,x3):

[0051]

[0052]

[0053] Taking C(x1,x5|x2,x3,x4) as an example, we establish the conditional probability distribution C(x1,x5|x2,x3,x4):

[0054]

[0055]

[0056] Taking a five-dimensional random variable as an example, its joint probability density function f(X) is shown in the following equation, which is the product of the probability density functions of all random variables and all edges in the vine structure:

[0057]

[0058] Where, f(x) i ) is F(x i The derivative of ), i,j|D(e); i,j|D(e) represents the edge of the vine structure; i and j correspond to the variable x of the conditional probability distribution to be determined. i and x j D(e) is the common conditional value for these two variables. C(x) i ,x j |D(e)) is x i and x j The conditional probability distribution.

[0059] After fitting, the information of the R-copula function (i.e., the joint probability density function of each fluctuation component) is saved for use in operating the random sampling fluctuation components in S2.

[0060] Operation S2 normalizes the state components and then inputs them into a segmented conditional generative adversarial network. For each normalized state component, its mean and variance are used as conditional terms to input the conditional generative adversarial network, which outputs multi-site generated state components. The fitted fluctuation components are then superimposed on the multi-site generated state components to obtain the generated daily scene. All generated daily scenes constitute the generated daily scene set.

[0061] In this embodiment, a multi-input Wasserstein distance-conditional generative adversarial network (C-WGAN-multi model) is constructed to fit the state components. Before using the conditional generative adversarial network to output the multi-site generated state components, the method further includes training the conditional generative adversarial network using the generator loss function and the discriminator loss function.

[0062] The loss functions for the synthesizer and the discriminator are as follows:

[0063] Loss D =-E X～p(x) [log(D(X|c))]-E X～p(x') [log(1-D(X|c))]

[0064] Loss G =-E X'～p(x') [log(D(X'|c)))]-λ e L e -λ var L var

[0065]

[0066]

[0067] Among them, Loss D For the discriminator loss function, Loss G Let X be the generator loss function, X be the generator-generated sample, X' be the historical sample of the scenic spot, x be the value of the random variable of the historical scenic spot, x' be the value of the random variable of the generated scenic spot, c be the conditional input, p() be the probability distribution of the random variable, D(X|c) be the output value of X input to the discriminator under condition c, D(X'|c) be the output value of X' input to the discriminator under condition c, G(z,c) be the generator-generated sample under condition c, and E[] be the expected value of the discriminator output. x [] represents the conditional term metric for generating samples, λ e λ is the coefficient of the mean-conditional measure term. var L is the coefficient of the variance conditional measure. e For the loss of the mean-conditional term, L var For the variance conditional measure of loss, c N,e Generate the mean of the samples for the Nth station, c N,v Let λ be the variance of the samples generated for the Nth station, and z be Gaussian noise. Preferably, λ e and λ var Both are 0.5.

[0068] See Figure 2Conditional Generative Adversarial Networks (GANs) consist of a discriminator and a generator. In the early stages of training, there is no overlap between real and generated samples, causing the discriminator's loss function gradient to vanish. This prevents the parameters of both the discriminator and generator from being updated, thus hindering the generation of samples that accurately fit the real data. To accurately describe the similarity between samples, Wasserstein distance is introduced instead of JS divergence.

[0069] In practical calculations, Wasserstein distance is often difficult to compute. The dual form of the Wasserstein distance is frequently used to describe the similarity between real and generated samples. The dual form is defined as the lower bound of the integral of the distance measure over two probability distributions, and can be expressed as:

[0070]

[0071] Where d(X,Y) is the distance measure between the two probability distributions.

[0072] When using duality for computation, the discriminator function needs to satisfy 1-Lipschitz continuity, a sufficient condition being that the gradient of the discriminator function satisfies 1-Lipschitz continuity. Therefore, the gradient of the discriminator function D(x) is normalized, and the normalized discriminator function value... The gradient normalization is used in the objective function calculation to ensure that the discriminator function absolutely satisfies 1-Lipschitz continuity. The gradient normalization is defined as follows:

[0073]

[0074] The training process for conditional generative adversarial networks is as follows:

[0075] 1) Normalize the state components of the annual historical scenes of multiple stations within one year. The normalized state components are as follows:

[0076]

[0077] Where, x N,t X N,t These are the state components of station N at time t before and after normalization, respectively, X. N,min X N,max These represent the minimum and maximum values ​​of the state components in the annual historical scenarios of station N.

[0078] 2) Input the normalized state components into the conditions in batches to generate the adversarial network. The input matrix has a dimension of N*T, where N is the number of sites; taking a resolution of 1 hour and T as 8760 hours as an example; divide the state components into historical state component scene sets by day, with a dimension of N*365*24, and input them into the conditions in batches to generate the adversarial network. The input matrix is ​​the historical state component daily scene, with a dimension of N*24. The input matrix format is as follows:

[0079]

[0080] 3) Extract the mean and variance of each day from the historical state component scene set, and use them as input conditions. The mean and variance are as follows:

[0081]

[0082]

[0083] Among them, c N,e c N,v Hereinafter, denoted as mean and variance, respectively, for station N.

[0084] 4) Select Gaussian noise that conforms to a standard normal distribution, and randomly sample condition terms, then concatenate them horizontally as input to the generator. The generator outputs the generated scene. The matrix form of the input to the generator is as follows:

[0085]

[0086] 5) Input the historical state component scene with conditional terms and the generated scene into the discriminator. The discriminator outputs the realism evaluation value of the generated scene, which is between 0 and 1.

[0087] 6) To adapt the generator to the requirements of the condition terms, reduce the deviation between the generated scene and the condition terms, and generate a generated scene that meets the requirements of the condition terms, a mean conditional metric loss L is introduced into the discriminator loss function. e And variance conditional measure loss L var .

[0088] 7) Calculate the discriminator loss function and the generator loss function, and update the weights of the discriminator and generator using the RMSprop optimization algorithm. The learning rate is, for example, 0.0001.

[0089] 8) Training ends when the sum of squared residuals of the mean, variance, and spatial correlation matrix error between the generated scene set and the historical state component scene set is less than 0.05. If the sum of squared residuals of the mean, variance, and spatial correlation matrix error is greater than 0.05, the error is backpropagated through the RMSprop optimizer to update the network parameters, and training continues. After training, the generated scenes are extracted, and after inverse normalization, the daily scene set of the multi-site generated state components is obtained.

[0090] According to an embodiment of the present invention, superimposing the fitted fluctuation component onto the multi-site generated state component specifically includes: combining the joint probability density function of the fluctuation component, randomly selecting the fluctuation component, and superimposing the selected fluctuation component onto the multi-site generated state component.

[0091] In this embodiment, considering that the photovoltaic power station output is extremely low during periods of low irradiance (such as night and early morning), the fluctuation component is only superimposed when the output of the photovoltaic power station is not zero, while the wind power station superimposes the fluctuation component throughout the entire time period.

[0092] Operation S3 clusters the historical daily scene set and generated daily scene set of the wind and solar multi-site to transform them into discrete typical daily states, and determines the cumulative state transition matrix based on the typical daily states.

[0093] In this embodiment, the nearest neighbor propagation clustering algorithm is used to take all samples in the initial day scene as initial sample centers. Attraction and membership functions are defined respectively, and the initial sample centers continuously compete to form the final cluster centers. The nearest neighbor propagation algorithm does not require specifying the number of clusters, and each cluster center is taken from the sample value, resulting in stable clustering results.

[0094] According to an embodiment of the present invention, determining the cumulative state transition matrix based on the typical daily state specifically includes the following sub-operations S31-S34.

[0095] In suboperation S31, for any typical day state, based on the typical day state and the typical day state of the next day, the matrix S is modified. N×N element S in mn Increment by 1. The typical daily state is m, the typical daily state of the next day is n, N is the number of typical daily states, and S N×N It is initially a matrix of all zeros.

[0096] In sub-operation S32, sub-operation S31 is executed multiple times to traverse each typical day state and process S. N×N Increment the corresponding elements by 1 to obtain the final matrix S′. N×N .

[0097] In suboperation S33, S′ N×N Divide each element in the matrix by the sum of the elements in its row to construct the state transition matrix P. r P r The element P in the i-th row and j-th column r,ij It can be represented as:

[0098]

[0099] Among them, S′ ij S′ N×N The element in the i-th row and j-th column.

[0100] In suboperation S34, the cumulative state transition matrix P is constructed. cum Wherein, the value P in the i-th row and j-th column of the cumulative state transition matrix cum,ij Equal to the state transition matrix P rThe sum of the elements in the first j columns of the i-th row. P cum,ij It can be represented as:

[0101]

[0102] Operation S4 uses the MCMC method to generate a daily state transition sequence corresponding to the cumulative state transition matrix. Daily scenes that meet the preset fluctuation component criteria are selected from the daily state transition sequence and spliced ​​together to obtain the medium- and long-term scenes of multiple wind and solar power stations.

[0103] According to an embodiment of the present invention, the daily state transition sequence corresponding to the cumulative state transition matrix is ​​generated using the MCMC method, specifically including the following sub-operations S41-S43.

[0104] In sub-operation S41, the typical daily state of the first day is randomly generated.

[0105] In suboperation S42, the random number a is compared one by one with the elements in the h-th row of the cumulative state transition matrix until the random number a lies between two adjacent elements in the h-th row (P). cum，hk <a<P cum，h(k+1) Generate the typical daily state for the next day and assign it the value k; where h is the value of the typical daily state for the current day, k is the column number of the smaller column element among two adjacent elements, and the initial value of h is 1, let h = h + 1.

[0106] In sub-operation S43, sub-operation S42 is executed repeatedly until the typical daily state of day H is generated, where H is a set value.

[0107] This embodiment generates typical daily states according to the seasons, defining spring as February to April, summer as May to July, autumn as August to October, and winter as November to December and January of the following year. That is, H=89 for spring, and H=92 for summer, autumn, and winter.

[0108] According to an embodiment of the present invention, the preset fluctuation component criterion is:

[0109] ΔP data，min ≤ΔP new ≤ΔP data，max

[0110] Wherein, ΔP new The fluctuation component of the period connecting the daily scene of the next typical day and the daily scene of the current typical day is equal to the difference between the initial value of the next day (i.e., 1:00 AM) and the final value of the current day (i.e., 12:00 AM); ΔP data，max ΔP data，max These represent the minimum and maximum values ​​of the historical data fluctuation components, respectively. Let ΔP... new Between ΔP data，min and ΔP data，maxBy stitching together the daytime scenes between them, we can obtain long-term scenes of multiple scenic stations.

[0111] More preferably, daily scenes that simultaneously meet the preset fluctuation component criterion, mean and variance criterion, and spatial correlation criterion can be stitched together to obtain a medium-to-long-term scene for multiple wind and solar power stations. The mean and variance criterion means that the mean and variance errors of the selected daily scene and the corresponding typical historical scene must both be less than a certain value (e.g., 0.1). The spatial correlation criterion means that the average absolute percentage error between the spatial correlation matrix of the selected daily scene and the historical scene is less than a certain value (e.g., 0.05).

[0112] The following uses wind and solar power output data from a location in Jiangsu Province in 2020 and January 2021 to verify the effectiveness of the method in simulating spatial correlation in this invention. The historical scenario set includes output data from 5 wind farms and 5 solar power plants. Spatial correlation is characterized by the Pearson coefficient; a higher Pearson coefficient indicates a higher spatial correlation between power plants. A positive Pearson coefficient indicates a positive correlation between power plant outputs; a negative Pearson coefficient indicates a negative correlation. The spatial correlation within wind farms and solar power plants is high, while the spatial correlation between wind farms and solar power plants is low.

[0113] In this example, the historical annual scene set has a dimension of T*N. Here, T represents 8760 hours, the resolution is 1 hour, and after normalization, the historical daily scene set is obtained and used as the model input. To better extract the spatiotemporal correlations between multiple sites, the generator and discriminator network model parameters of the C-WGAN-multi model are set as shown in Tables 1 and 2, where Table 1 shows the generator model parameters and Table 2 shows the discriminator model parameters.

[0114] Table 1

[0115]

[0116]

[0117] Table 2

[0118]

[0119] Figure 3 , Figure 4 These are spatial correlation matrices of historical scenarios for wind power and solar power in different seasons, respectively. Figure 3 and Figure 4 It can be seen that the spatial correlation between power stations varies in different seasons. The spatial correlation between wind farms and photovoltaic farms is relatively weak. Figure 5 and Figure 6The spatial correlation of the generated scene is shown in this method. The low spatial correlation fitting error indicates that the generated sample can fit the spatial correlation of the real sample well.

[0120] Figure 7 This is a block diagram of a long-term scene generation device for multiple wind and solar power stations provided in an embodiment of the present invention. (See also...) Figure 7 The long-term scene generation device 700 for multiple wind and solar power stations includes an extraction and fitting module 710, a first generation module 720, a clustering module 730, and a second generation module 740.

[0121] The extraction and fitting module 710, for example, performs operation S1 to extract the state components and fluctuation components in the historical scene of the wind and solar multi-site, and fits the fluctuation components.

[0122] The first generation module 720, for example, performs operation S2 to normalize the state components and then input them into the conditional generative adversarial network. For each normalized state component, its mean and variance are used as conditional terms to input the conditional generative adversarial network, outputting multi-site generated state components. The fitted fluctuation components are superimposed on the multi-site generated state components to obtain the generated daily scene. The generated daily scenes together form the generated daily scene set.

[0123] Clustering module 730, for example, performs operation S3 to cluster the historical daily scene set and generated daily scene set of wind and solar multi-sites to transform them into discrete typical daily states, and determines the cumulative state transition matrix based on the typical daily states.

[0124] The second generation module 740, for example, performs operation S4, which uses the MCMC method to generate a daily state transition sequence corresponding to the cumulative state transition matrix, selects daily scenes that meet the preset fluctuation component criteria in the daily state transition sequence and splices them together to obtain the long-term scenes of the wind and solar multi-site.

[0125] The long-term scene generation device 700 for multiple wind and solar power stations is used to execute the above-mentioned long-term scene generation method for multiple wind and solar power stations, which will not be described in detail here.

[0126] This invention also provides a long-term scene generation system for multiple wind and solar power stations, comprising: a processor; and a memory storing a computer-executable program, wherein when the program is executed by the processor, the processor performs the aforementioned long-term scene generation method for multiple wind and solar power stations.

[0127] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for generating long-term scenes at multiple wind and solar power stations.

[0128] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for generating long-term scenes from multiple wind and solar power stations, characterized in that, include: Extract state components and fluctuation components from historical scenes of multiple wind and solar power stations, and fit the fluctuation components; After normalizing the state components, segmented input conditions are used to generate an adversarial network. For each normalized state component, its mean and variance are used as conditional terms to input the conditional generative adversarial network, and the multi-site generated state component is output. The fitted fluctuation component is superimposed on the multi-site generated state component to obtain the generated daily scene. The scenes generated on each day constitute a set of scenes generated on each day; The historical daily scene set and generated daily scene set of the wind and solar multi-site are clustered to transform them into discrete typical daily states, and the cumulative state transition matrix is ​​determined based on the typical daily states; The daily state transition sequence corresponding to the cumulative state transition matrix is ​​generated using the MCMC method. Daily scenes that meet the preset fluctuation component criteria are selected from the daily state transition sequence and spliced ​​together to obtain the medium- and long-term scenes of multiple wind and solar power stations. The fitting of the fluctuation component specifically includes: Select the optimal R-vine Copula first-layer tree structure, and use the nonparametric kernel density estimation method to estimate the bivariate Copula function of the fluctuation component in the optimal R-vine Copula first-layer tree structure; By combining the probability distribution function and marginal distribution of each station in the multi-station wind and solar power system, the joint probability density function of the fluctuation components is established using the binary Copula function. The step of superimposing the fitted fluctuation component onto the multi-station generated state component specifically includes: By combining the joint probability density function of the fluctuation components, the fluctuation components are randomly selected and superimposed onto the multi-site generated state components.

2. The method for generating long-term scenes from multiple wind and solar power stations as described in claim 1, characterized in that, The step of determining the cumulative state transition matrix based on the typical daily state specifically includes: For any typical day state, based on the typical day state and the typical day state of the next day, the matrix S is... N×N element S in mn The process involves incrementing by 1. The typical daily state is m, the typical daily state for the next day is n, N is the number of states for the typical daily state, and S... N×N Initially, it is an all-zero matrix; Iterate through each typical daily state and perform S N×N Increment the corresponding elements by 1 to obtain the final matrix. ; Will Divide each element in the matrix by the sum of the elements in its row to construct the state transition matrix; Construct a cumulative state transition matrix, wherein the value in the i-th row and j-th column of the cumulative state transition matrix is ​​equal to the sum of the elements in the i-th row and the first j-th column of the state transition matrix.

3. The method for generating long-term scenes from multiple wind and solar power stations as described in claim 1, characterized in that, The step of generating the daily state transition sequence corresponding to the cumulative state transition matrix using the MCMC method specifically includes: The random number a is compared with each element in the h-th row of the cumulative state transition matrix until the random number a is between two adjacent elements in the h-th row. The typical daily state of the next day is generated and assigned the value k. Here, h is the value of the typical daily state of the current day, and k is the column number of the smaller column element among the two adjacent elements. The typical daily state of the first day is generated randomly. Repeat the above operations until the typical daily status of day H is generated, where H is a set value.

4. The method for generating medium- and long-term scenes using multiple wind and solar power stations as described in any one of claims 1-3, characterized in that, The preset fluctuation component criterion is: in, This refers to the fluctuation component during the period connecting the daily scene of the next typical day and the daily scene of the current typical day. , These are the minimum and maximum values ​​of the historical data fluctuation components, respectively.

5. A device for generating long-term scenes from multiple wind and solar power stations, characterized in that, The method for generating medium- and long-term scenes using multiple wind and solar power stations as described in any one of claims 1-4 includes: The extraction and fitting module is used to extract the state components and fluctuation components in the historical scenes of multiple wind and solar power stations, and to fit the fluctuation components. The first generation module is used to normalize the state components and then input them into the conditional generative adversarial network in segments; for each normalized state component, its mean and variance are used as condition terms to input the conditional generative adversarial network, outputting multi-site generated state components, and superimposing the fitted fluctuation components onto the multi-site generated state components to obtain the generated day scene; the generated day scenes constitute the generated day scene set. The clustering module is used to cluster the historical daily scene set and the generated daily scene set of multiple wind and solar power stations to transform them into discrete typical daily states, and to determine the cumulative state transition matrix based on the typical daily states. The second generation module is used to generate a daily state transition sequence corresponding to the cumulative state transition matrix using the MCMC method, select daily scenes that meet the preset fluctuation component criteria from the daily state transition sequence and splice them together to obtain the medium and long-term scenes of the wind and solar multi-site.

6. A long-term scene generation system for multiple wind and solar power sites, comprising: processor; A memory storing a computer-executable program, which, when executed by the processor, causes the processor to perform the long-term scene generation method for multiple wind and solar power sites as described in any one of claims 1-4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method for generating long-term scenes of multiple wind and solar power stations as described in any one of claims 1-4.

Images