A Day-ahead Spot Clearing Price-Load Data Repair Method Based on Generative Adversarial Networks

Abstract

This invention proposes a day-ahead spot clearing price-load data repair method based on generative adversarial networks. A training dataset is established; K-means clustering is used to cluster the user electricity load dataset, classifying it into typical user groups; a deep convolutional generative adversarial network (DCGAN) model is established; the DCGAN model is trained using historical day-ahead spot market clearing price data and historical user electricity load data after classifying typical user groups, retaining the generator structure for each typical user group; a binary mask matrix is ​​introduced to represent the missing locations of newly collected user electricity load data; then, the newly collected day-ahead spot market clearing price dataset, typical user electricity load dataset, and target user electricity load data with missing data are input into the generator corresponding to the typical user group, outputting new generated data; the missing parts of the target user electricity load data are filled in using the new generated data samples.

Description

Technical Field

[0001] This invention belongs to the field of power system analysis technology, specifically relating to a method for repairing day-ahead spot clearing price-load data based on generative adversarial networks. Background Technology

[0002] In modern power systems, load data from electricity users plays a crucial role. Complete and accurate historical load data is the foundation for various data mining techniques to conduct load forecasting and analyze user electricity consumption behavior. Effective load forecasting and user electricity consumption behavior analysis can improve the management efficiency of the power system, ensure the formulation of reasonable production plans, avoid resource waste, guarantee the safe and reliable operation of the power grid, and provide economic benefits. However, the collection, transmission, and processing of user load data can be affected by factors such as malfunctions in user metering devices, communication interruptions, or signal interference, leading to problems such as missing data and degraded data quality. Incomplete load data cannot accurately represent user electricity consumption behavior, resulting in decreased accuracy in load forecasting and electricity user behavior analysis. Therefore, research on the repair of missing user load data is crucial for the management efficiency and safe and stable operation of modern power systems.

[0003] Current research largely relies on the temporal characteristics of load data collection, the correlation between load data of different types of users, and the patterns of load changes as the basis for missing data repair. It then uses data modeling and big data analytics to uncover the inherent patterns in user load data and its connections with other relevant factors to reconstruct the missing data. However, existing solutions require large amounts of data and diverse data types, lack applicability to multiple scenarios, and struggle to quantify the correlation between the model and the scenario. Consequently, they often fail to achieve satisfactory and stable results in practical applications. Summary of the Invention

[0004] To address the shortcomings and deficiencies of existing technologies, the present invention aims to provide a day-ahead spot clearing price-load data repair scheme based on generative adversarial networks. Its core component is the method for constructing a data repair model. Through a series of model construction and training processes, the model is ultimately adapted to be executed on a computing device in the form of computer software, thereby solving the problem of day-ahead spot clearing price-load data repair.

[0005] The basic construction process of this scheme includes: collecting historical cleared electricity price data and historical user electricity load data from the day-ahead spot market of the power system, retaining the data without missing parts, and establishing a training dataset; using the K-means clustering algorithm to perform cluster analysis on the user electricity load dataset and classify typical user groups; establishing a deep convolutional generative adversarial network (DCGAN) model, introducing Wasserstein distance and gradient penalty functions to improve the performance of the DCGAN model; training the DCGAN model with historical cleared electricity price data from the day-ahead spot market and historical user electricity load data after classifying typical user groups, retaining the generator structure for each typical user group; introducing a binary mask matrix to represent the missing locations of newly collected user electricity load data, and then inputting the newly collected day-ahead spot market cleared electricity price dataset, typical user electricity load dataset, and target user electricity load data with missing parts into the generator corresponding to the typical user group, outputting new generated data; and filling the missing parts of the target user electricity load data with the new generated data samples.

[0006] The present invention specifically adopts the following technical solution:

[0007] A method for repairing day-ahead spot clearing price-load data based on generative adversarial networks (GANs) is proposed. This method involves collecting historical day-ahead spot market clearing price data and historical user load data from the power system, retaining the data without missing parts, and establishing a training dataset. K-means clustering is used to cluster the user load dataset, classifying it into typical user groups. A deep convolutional generative adversarial network (DCGAN) model is established, incorporating Wasserstein distance and gradient penalty functions to improve its performance. The DCGAN model is trained using historical day-ahead spot market clearing price data and historical user load data after classifying typical user groups, retaining the generator structure for each typical user group. A binary mask matrix is ​​introduced to represent the missing locations in newly collected user load data. The newly collected day-ahead spot market clearing price dataset, typical user load dataset, and target user load data with missing parts are then input into the generator corresponding to the typical user group, outputting new generated data. Finally, the missing parts of the target user load data are filled in using the new generated data samples.

[0008] Furthermore, the specific steps include:

[0009] Step S1: Collect historical clearing electricity price data and historical user electricity load data of the day-ahead spot market of the power system, retain the data without missing parts, and establish a training dataset;

[0010] Step S2: Use the K-means clustering algorithm to perform cluster analysis on the user electricity load dataset L and divide it into typical user groups;

[0011] Step S3: Normalize the sample data of typical user groups and the current spot market clearing electricity price;

[0012] Step S4: Establish a deep convolutional generative adversarial network (DCGAN) model;

[0013] Step S5: Introduce Wasserstein distance and gradient penalty function to improve the performance of DCGAN model;

[0014] Step S6: Generate the day-ahead spot market clearing electricity price dataset P and L k Except for L k0 Other load datasets L k-0 As a condition, it is concatenated with random noise data z and then input into the generator, and the generator outputs generated data G. k (z);

[0015] Step S7: Generate the day-ahead spot market clearing electricity price dataset P and L k Except for L k0 Other load datasets L k-0 As a condition, sample G is generated with user load data. k (z) Input the concatenated data into the discriminator; and combine the current day-to-date spot market clearing electricity price datasets P and L. k Except for L k0 Other load datasets L k-0 As a condition, with the selected real load dataset L k0 The concatenated data is also input into the discriminator, causing the discriminator to output a sample G generated from the user load data. k (z) and the selected real load dataset L k0 The judgment situation;

[0016] Step S8: Calculate the loss functions of the generator and discriminator, and use the RMSprop optimization algorithm to optimize and update the network weight parameters of the generator and discriminator; when one round of training is completed, return to step S6 to start the next round of training.

[0017] Step S9: After training is complete, retain the generator model G in DCGAN. k The network structure and parameters;

[0018] Step S10: When newly collected user electricity load data is missing, a binary mask matrix is ​​introduced to represent the missing location of the user electricity load data;

[0019] Step S11: Generate the day-ahead spot market clearing electricity price dataset P and Lk-new Except for L k-new0 Other load datasets L k-knew0 As a condition, it is concatenated with random noise data z and input into the generator G. k Output generated data G k (z);

[0020] Step S12: Generate data sample G k (z) Fill in the target user's electricity load data The missing part.

[0021] Further, in step S1:

[0022] Let P be the complete dataset of historical cleared electricity prices in the day-ahead spot market, and L be the complete dataset of historical user electricity load. The expression for P is as follows:

[0023]

[0024] In the formula, p d,t Let m represent the historical day-ahead spot market clearing electricity price at time t on day d, where D represents the maximum number of days in the dataset, and m represents the number of day-ahead spot market clearing electricity prices per day.

[0025] L consists of N L's n The dataset consists of N users; L n The expression is as follows:

[0026]

[0027] In the formula, l n,d,t This represents the electricity load data of the nth user at time t on day d, where D represents the maximum number of days in the dataset, and m represents the number of electricity load data points collected by the user each day. The time of electricity load collection by the user corresponds one-to-one with the time of clearing in the spot market the day before.

[0028] In step S2:

[0029] Calculate the average load data L for each user n Average load The formula is as follows:

[0030]

[0031] Input average load data The set of:

[0032] Then, the "elbow method" is used to determine the number of clusters K in the cluster analysis;

[0033] Initialize cluster centers, i.e., from the average load data set. K data points are randomly selected as cluster centers. Calculate the average load data for each user Euclidean distance d to each cluster center nk The calculation formula is as follows:

[0034]

[0035] In the formula, n = 1, 2, ..., N; k = 1, 2, ..., K;

[0036] Setting data The clustering label is l n and will The clusters that are closest to the Euclidean distance are calculated using the following formula:

[0037] l n =arg min d nk (5)

[0038] Update the cluster centers, and make l n =k, k = 1, 2, ..., K;

[0039] Set a clustering analysis threshold ε, and determine whether the update amount of the objective function is less than the threshold ε. The formula for calculating the objective function E is as follows:

[0040]

[0041] If the update of the objective function E is not less than the threshold ε, then continue the clustering iteration t = t + 1; when the update of the objective function E is less than the threshold ε, the clustering iteration process ends.

[0042] Ultimately, based on clustering, K typical user groups are formed, and each typical user group is represented by L. k Let k = 1, 2, ..., K; each L k Includes several users;

[0043] In step S3:

[0044] The normalization formula is as follows:

[0045]

[0046]

[0047] The normalized data forms the day-ahead spot market clearing price dataset P and the N user load datasets L. n Based on cluster analysis, the user dataset can be further divided into K typical user datasets L.k .

[0048] Further, in step S4:

[0049] For a complete, unmissing typical user load dataset L k Suppose there are I users in total, and one user L is randomly selected. k0 Define it as real data; define a set of random noise data z as the input of the generator, with p z (z) represents the probability distribution of z, where p data (L k0 ) represents the actual data L k0 The probability distribution; targeting a typical user group L k The characteristics define the data samples G generated by the generator. k (z) represents the generated data, with a probability distribution p. Gk (z); Define the input of the discriminator network as real data L k0 and generated data G k (z), the output is a scalar D(G) k (z) indicates that the input noise z follows the true data distribution p. data (L k0 The probability of ).

[0050] Based on the training objectives of the generator and discriminator, the loss functions L for the generator and discriminator are constructed respectively. G and L D And determine the objective function during GAN training:

[0051]

[0052]

[0053]

[0054] Further, in step S5:

[0055] The Wasserstein distance is defined as follows:

[0056]

[0057] In the formula, Ω(p) data ,p Gk ) is based on p data and p Gk W(p) is the set of joint probability distributions γ of the marginal distributions; data ,p Gk ) is the infimum of the expectation of (u,v)~γ, indicating that the generated data distribution p GkFit to the true data distribution p data We need to move u by a distance v, where u and v represent real data samples L randomly sampled from the joint distribution γ. k0 With the generated data sample G k (z); The distance between generated samples and real samples is described using the Kantorovich-Rubinstein dual form:

[0058]

[0059] In the formula ||f D ||L≤K indicates that the discriminator function D(L) k0 The function must satisfy K-Lipschitz continuity, meaning the upper limit of the absolute value of the gradient is K. To ensure that the gradient does not exceed the limit K, a discriminator function D(L) is introduced into equation (13). k0 The gradient penalty function within the domain makes the discriminator function D(L) k0 The GAN approximates K-Lipschitz continuity to accurately describe the Wasserstein distance. In this case, the GAN's objective function transforms into:

[0060]

[0061] Further, in step S6:

[0062] The concatenated samples input to the generator are in the form of a three-dimensional matrix, consisting of dataset P, random noise data z, and L. k-0 The data is concatenated using the third-dimensional channel, resulting in a sample of size m×D×I; the concatenation order of the third-dimensional matrix is ​​[P, z, L]. k-0 ], L k-0 ={L k1 ,L k2 ,···,L k(I-1)};

[0063] The generator structure of the DCGAN model is shown below:

[0064] Convolutional Layer 1: 32 convolutional kernels, each with a size of 3; stride of 1; padding of 1; normalization function of BatchNorm2D; activation function of LeakyReLU.

[0065] Convolutional layer 2: 64 convolutional kernels, each with a size of 3; stride of 1; padding of 1; normalization function of BatchNorm2D; activation function of LeakyReLU.

[0066] Convolutional layer 3: 128 convolutional kernels, each with a size of 3; stride of 1; padding of 1; normalization function of BatchNorm2D; activation function of LeakyReLU.

[0067] Convolutional layer 4: 256 convolutional kernels, each with a size of 3; stride of 1; padding of 1; normalization function of BatchNorm2D; activation function of LeakyReLU.

[0068] Convolutional layer 5: 16 convolutional kernels, each with a size of 3; stride of 1; padding of 1; normalization function of BatchNorm2D; activation function of ReLU.

[0069] Convolutional layer 6: 4 convolutional kernels, each with a size of 3; stride of 1; padding of 1; normalization function of BatchNorm2D; activation function of ReLU.

[0070] Further, in step S7:

[0071] The recent spot market clearing electricity price dataset P and L k Except for L k0 Other load datasets L k-0 Sample G is generated from user load data. k (z) The samples are stitched together through the third-dimensional channels to form a three-dimensional matrix. The final sample size is m×D×I; the stitching order of the third-dimensional matrix is ​​[P,G]. k (z),L k-0 ], L k-0 ={L k1 ,L k2 ,···,L k(I-1)};

[0072] The recent spot market clearing electricity price dataset P and L k Except for L k0 Other load datasets L k-0 , and the selected real load dataset L k0 The data is stitched together using the third-dimensional channel to form a three-dimensional matrix. The final sample size is m×D×I; the stitching order of the third-dimensional matrix is ​​[P,L]. k0 ,L k-0 ], L k-0 ={L k1 ,L k2 ,···,L k(I-1)};

[0073] The discriminator structure of the DCGAN model is shown below:

[0074] Input layer: 64 convolutional kernels, each with a size of 3; stride of 1.

[0075] Convolutional Layer 1: 128 convolutional kernels, each kernel size is 4; stride is 2; normalization function is BatchNorm2D; activation function is LeakyReLU.

[0076] Convolutional layer 2: 256 convolutional kernels, each with a size of 4; stride of 2; normalization function of BatchNorm2D; activation function of LeakyReLU.

[0077] Convolutional layer 3: 512 convolutional kernels, each kernel size is 4; stride is 2; normalization function is BatchNorm2D; activation function is LeakyReLU.

[0078] Convolutional layer 4: 1024 convolutional kernels, each kernel size is 4; stride is 2; normalization function is BatchNorm2D; activation function is LeakyReLU;

[0079] Convolutional layer 5: 2048 convolutional kernels, each with a size of 4; stride of 2; normalization function of BatchNorm2D; activation function of LeakyReLU;

[0080] Convolutional layer 6: 512 convolutional kernels, each kernel size 3; stride 1; normalization function BatchNorm2D; activation function LeakyReLU;

[0081] Convolutional layer 7: 128 convolutional kernels, each kernel size is 3; stride is 1; normalization function is BatchNorm2D; activation function is LeakyReLU;

[0082] Output layer: Dense layer structure containing 1024 neural network units, with 1 sample as the output.

[0083] Further, in step S8: the learning rate of the RMSprop optimization algorithm is set to 2 × 10⁻⁶. -4 , parameter = 0.9;

[0084] The formula for the RMSprop optimization algorithm is as follows:

[0085] s dw =βs dw +(1-β)dW 2

[0086] s db=βs db +(1-β)db 2

[0087]

[0088]

[0089] In the formula s dw and s db These are the gradient momentum accumulated by the loss function in previous iterations, and β is a parameter representing gradient accumulation. The RMSprop algorithm calculates the differential squared weighted average of the gradient. When a value in dW or db exceeds the preset range, this change is divided by the gradient momentum accumulated in previous iterations to meet the requirements for gradient oscillation amplitude. ε is an auxiliary parameter to prevent singularities caused by a zero denominator.

[0090] In step S10:

[0091] Define the newly collected typical user load dataset of the kth class as L k-new Let L be the target user's electricity load data that is missing in this load dataset. k-new0 A binary mask matrix M of dimension m×D is introduced to mask the target user's electricity load data L. k-new0 The missing values ​​are characterized, with the element corresponding to the missing position having a value of 0, otherwise 1; the target user's electricity load data after missing value characterization is defined as follows: The formula is as follows:

[0092]

[0093] In the formula, ⊙ represents the Hadamard product operation of the matrix.

[0094] Further, in step S11:

[0095] The recent spot market clearing electricity price dataset P and L k-new Except for L k-new0 Other load datasets L k-knew0 The random noise data z is concatenated with the random noise data z through the third dimension channel to form a three-dimensional matrix. The final sample size is m×D×I; the concatenation order of the third dimension matrix is ​​[P,z,L]. k-knew0 ], L k-knew0 ={L k-knew1 ,L k-knew2 ,···,L k-knew(I-1)};

[0096] In step S12:

[0097] definition After the missing data was filled in, the final output user electricity load data was: The original data retains the parts that are not missing, and the missing parts are generated using data G. k (z) is used for filling, and the calculation formula is shown below:

[0098]

[0099] Furthermore, a day-ahead spot clearing price-load data repair device based on generative adversarial networks, which is based on a computer system and uses the day-ahead spot clearing price-load data repair method based on generative adversarial networks as described above during operation.

[0100] A non-transitory computer-readable storage medium storing a computer program, characterized in that: during runtime, it uses the day-ahead spot clearing price-load data repair method based on generative adversarial networks as described above.

[0101] Compared to related technologies, this invention and its preferred embodiment are based on generative adversarial network algorithms and utilize big data modeling and analysis techniques to repair missing load data of electricity users. Compared to existing methods, this method is primarily applicable to repairing missing load data in the electricity spot market. It is simple, efficient, and requires fewer data types. It is highly versatile, suitable for various electricity spot markets with different spatiotemporal conditions and for various electricity users with different electricity consumption behaviors. This method fully considers the electricity supply and demand situation in the electricity spot market, using deep learning modeling to simulate the relationship between market supply and demand and electricity prices. It then constructs the correlation between the day-ahead clearing price and user load data, and finally repairs missing user load data by analyzing the inherent variation patterns of the day-ahead clearing price. Attached Figure Description

[0102] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0103] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0104] Figure 1 This is a flowchart of the overall scheme of an embodiment of the present invention.

[0105] The following detailed description, in conjunction with the accompanying drawings, will further illustrate the present invention. Detailed Implementation

[0106] To make the objectives, technical solutions, and advantages of this application clearer, the application is described and illustrated 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 scope of this application. All other embodiments obtained by those skilled in the art based on the embodiments provided in this application without inventive effort are within the scope of protection of this application.

[0107] Obviously, the accompanying drawings described below are merely some examples or embodiments of this application. Those skilled in the art can apply this application to other similar scenarios based on these drawings without any inventive effort. Furthermore, it is understood that although the efforts made in this development process may be complex and lengthy, for those skilled in the art related to the content disclosed in this application, any changes to design, manufacturing, or production based on the technical content disclosed in this application are merely conventional technical means and should not be construed as insufficient disclosure of the content of this application.

[0108] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment that is mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application may be combined with other embodiments without conflict.

[0109] Unless otherwise defined, the technical or scientific terms used in this application shall have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms “a,” “an,” “an,” “the,” and similar words used in this application do not indicate quantity limitation and may indicate singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or modules (units) is not limited to the listed steps or units, but may also include steps or units not listed, or may include other steps or units inherent to these processes, methods, products, or devices. The terms “connected,” “linked,” “coupled,” and similar words used in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. “Multiple” used in this application refers to two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. The character " / " generally indicates that the preceding and following objects are in an "or" relationship. The terms "first," "second," and "third" used in this application are merely to distinguish similar objects and do not represent a specific ordering of the objects.

[0110] like Figure 1 As shown, the solution process provided in this embodiment of the invention specifically includes the following steps:

[0111] Step S1: Collect historical clearing electricity price data and historical user electricity load data of the day-ahead spot market of the power system, retain the data without missing parts, and establish a training dataset;

[0112] Let P be the complete dataset of historical cleared electricity prices in the day-ahead spot market, and L be the complete dataset of historical user electricity load. The expression for P is as follows:

[0113]

[0114] In the formula, p d,t Let L represent the historical day-ahead spot market cleared electricity price at time t on day d, where D represents the maximum number of days in the dataset, and m represents the number of day-ahead spot market cleared prices each day (i.e., day-ahead spot market cleared m times per day). L consists of N L... n The dataset consists of N users. L n The expression is as follows:

[0115]

[0116] In the formula, l n,d,tThis represents the electricity load data of the nth user at time t on day d, where D represents the maximum number of days in the dataset, and m represents the number of electricity load data points collected by the user each day (consistent with the daily clearing count in the spot market). The time of electricity load data collection by the user corresponds one-to-one with the clearing time in the spot market.

[0117] Step S2: Use the K-means clustering algorithm to perform cluster analysis on the user electricity load dataset L and divide it into typical user groups;

[0118] Calculate the average load data L for each user n Average load The formula is as follows:

[0119]

[0120] Input average load data The set of:

[0121] The "elbow method" is used to determine the number of clusters K in cluster analysis.

[0122] Initialize cluster centers, i.e., from the average load data set K data points are randomly selected as cluster centers. Calculate the average load data for each user Euclidean distance d to each cluster center nk The calculation formula is as follows:

[0123]

[0124] In the formula, n = 1, 2, ..., N; k = 1, 2, ..., K.

[0125] Setting data The clustering label is l n and will The clusters that are closest to the Euclidean distance are calculated using the following formula:

[0126] l n =arg min d nk (5)

[0127] Update the cluster centers, and make l n =k, k = 1, 2, ..., K.

[0128] Set a clustering analysis threshold ε, and determine whether the update amount of the objective function is less than the threshold ε. The formula for calculating the objective function E is as follows:

[0129]

[0130] If the update of the objective function E is not less than the threshold ε, then continue the clustering iteration t = t + 1; when the update of the objective function E is less than the threshold ε, the clustering iteration process ends.

[0131] Ultimately, based on clustering, K typical user groups are formed, and each typical user group is represented by L. k Let k = 1, 2, ..., K. Each L... k It contains several users.

[0132] Step S3: Normalize the sample data of typical user groups and the current spot market clearing electricity price;

[0133] The normalization formula is as follows:

[0134]

[0135]

[0136] The normalized data forms the day-ahead spot market clearing price dataset P and the N user load datasets L. n Based on cluster analysis, the user dataset can be further divided into K typical user datasets L. k .

[0137] Step S4: Establish a deep convolutional generative adversarial network (DCGAN) model;

[0138] For a complete, unmissing dataset L of a typical user load class k Define the dataset L k There are I users in total, and one user L is randomly selected from them. k0 Define it as real data; define a set of random noise data z as the input of the generator, with p z (z) represents the probability distribution of z, where p data (L k0 ) represents the actual data L k0 The probability distribution for this typical user group L; k The characteristics define the data samples G generated by the generator. k (z) represents the generated data, with a probability distribution p. Gk (z); Define the input of the discriminator network as real data L k0 and generated data G k (z), the output is a scalar D(G) k (z) indicates that the input noise z follows the true data distribution p. data (L k0 The probability of ).

[0139] Based on the training objectives of the generator and discriminator, the loss functions L for the generator and discriminator are constructed respectively. G and L D And determine the objective function during GAN training:

[0140]

[0141]

[0142]

[0143] Step S5: Introduce Wasserstein distance and gradient penalty function to improve the performance of DCGAN model;

[0144] The Wasserstein distance is defined as follows:

[0145]

[0146] In the formula, Ω(p) data ,p Gk ) is based on p data and p Gk W(p) is the set of joint probability distributions γ of the marginal distributions; data ,p Gk ) is the infimum of the expectation of (u,v)~γ, indicating that the generated data distribution p Gk Fit to the true data distribution p data We need to move u by a distance v, where u and v represent real data samples L randomly sampled from the joint distribution γ. k0 With the generated data sample G k (z). The distance between generated samples and real samples is described using its Kantorovich-Rubinstein dual form:

[0147]

[0148] In the formula ||f D ||L≤K indicates that the discriminator function D(L) k0 The function must satisfy K-Lipschitz continuity, meaning the upper limit of the absolute value of the gradient is K. To ensure that the gradient does not exceed the limit K, a discriminator function D(L) is introduced into equation (13). k0 The gradient penalty function within the domain makes the discriminator function D(L) k0 The GAN approximates K-Lipschitz continuity to accurately describe the Wasserstein distance. In this case, the GAN's objective function transforms into:

[0149]

[0150] Step S6: Generate the day-ahead spot market clearing electricity price dataset P and L k Except for L k0 Other load datasets L k-0 As a condition, it is concatenated with random noise data z and then input into the generator, and the generator outputs generated data G. k (z);

[0151] The concatenated samples input to the generator are in the form of a three-dimensional matrix, consisting of dataset P, random noise data z, and L. k-0 The data is concatenated using the third-dimensional channel, resulting in a sample of size m×D×I. The concatenation order of the third-dimensional matrix is ​​[P, z, L]. k-0 ], L k-0 ={L k1 ,L k2 ,···,L k(I-1)}

[0152] The generator structure of the DCGAN model is shown below:

[0153] Convolutional Layer 1: 32 convolutional kernels, each with a size of 3; stride of 1; padding of 1; normalization function of BatchNorm2D; activation function of LeakyReLU.

[0154] Convolutional layer 2: 64 convolutional kernels, each with a size of 3; stride of 1; padding of 1; normalization function of BatchNorm2D; activation function of LeakyReLU.

[0155] Convolutional layer 3: 128 convolutional kernels, each with a size of 3; stride of 1; padding of 1; normalization function of BatchNorm2D; activation function of LeakyReLU.

[0156] Convolutional layer 4: 256 convolutional kernels, each with a size of 3; stride of 1; padding of 1; normalization function of BatchNorm2D; activation function of LeakyReLU.

[0157] Convolutional layer 5: 16 convolutional kernels, each with a size of 3; stride of 1; padding of 1; normalization function of BatchNorm2D; activation function of ReLU.

[0158] Convolutional layer 6: 4 convolutional kernels, each with a size of 3; stride of 1; padding of 1; normalization function of BatchNorm2D; activation function of ReLU.

[0159] Step S7: Generate the day-ahead spot market clearing electricity price dataset P and L k Except for L k0 Other load datasets L k-0 As a condition, sample G is generated with user load data. k (z) The concatenated data is input into the discriminator; simultaneously, the day-ahead spot market clearing electricity price datasets P and L are input into the discriminator. k Except for L k0 Other load datasets L k-0 As a condition, with the selected real load dataset L k0 The concatenated data is also input into the discriminator, causing the discriminator to output a sample G generated from the user load data. k (z) and the selected real load dataset L k0 The judgment situation;

[0160] The recent spot market clearing electricity price dataset P and L k Except for L k0 Other load datasets L k-0 Sample G is generated from user load data. k (z) The data is stitched together using the third-dimensional channels to form a three-dimensional matrix. The final sample size is m×D×I. The stitching order of the third-dimensional matrix is ​​[P,G]. k (z),L k-0 ], L k-0 ={L k1 ,L k2 ,···,L k(I-1)}

[0161] The recent spot market clearing electricity price dataset P and L k Except for L k0 Other load datasets L k-0 , and the selected real load dataset L k0 The data is stitched together using the third-dimensional channel to form a three-dimensional matrix. The final sample size is m×D×I. The stitching order of the third-dimensional matrix is ​​[P,L]. k0 ,L k-0 ], L k-0 ={L k1 ,L k2 ,···,L k(I-1)}

[0162] The discriminator structure of the DCGAN model is shown below:

[0163] Input layer: 64 convolutional kernels, each with a size of 3; stride of 1.

[0164] Convolutional Layer 1: 128 convolutional kernels, each kernel size is 4; stride is 2; normalization function is BatchNorm2D; activation function is LeakyReLU.

[0165] Convolutional layer 2: 256 convolutional kernels, each with a size of 4; stride of 2; normalization function of BatchNorm2D; activation function of LeakyReLU.

[0166] Convolutional layer 3: 512 convolutional kernels, each kernel size is 4; stride is 2; normalization function is BatchNorm2D; activation function is LeakyReLU.

[0167] Convolutional layer 4: 1024 convolutional kernels, each kernel size is 4; stride is 2; normalization function is BatchNorm2D; activation function is LeakyReLU;

[0168] Convolutional layer 5: 2048 convolutional kernels, each with a size of 4; stride of 2; normalization function of BatchNorm2D; activation function of LeakyReLU;

[0169] Convolutional layer 6: 512 convolutional kernels, each kernel size 3; stride 1; normalization function BatchNorm2D; activation function LeakyReLU;

[0170] Convolutional layer 7: 128 convolutional kernels, each kernel size is 3; stride is 1; normalization function is BatchNorm2D; activation function is LeakyReLU;

[0171] Output layer: Dense layer structure, containing 1024 neural network units, with 1 sample as output;

[0172] Step S8: Calculate the loss functions of the generator and discriminator, and use the RMSprop optimization algorithm to optimize and update the network weight parameters of the generator and discriminator. Set the learning rate of the RMSprop optimization algorithm to 2 × 10⁻⁶. -4 The parameter is set to 0.9. At this point, one round of training ends, and we return to step S6 to begin the next round of training.

[0173] The formula for the RMSprop optimization algorithm is as follows:

[0174] s dw =βs dw +(1-β)dW 2

[0175] s db =βsdb +(1-β)db 2

[0176]

[0177]

[0178] In the formula s dw and s db These are the gradient momentum accumulated by the loss function in the previous iterations, and β is a parameter representing the gradient accumulation. The RMSprop algorithm calculates a differential squared weighted average of the gradients. When dW or db contains a value exceeding a preset range, this change is divided by the gradient momentum accumulated in previous iterations to meet the requirements for gradient oscillation magnitude; ε is an auxiliary parameter to prevent singularities caused by a zero denominator.

[0179] Step S9: After training is complete, retain the generator model G in DCGAN. k The network structure and parameters.

[0180] Step S10: When newly collected electricity load data for a certain user is missing, a binary mask matrix is ​​introduced to represent the missing location of the user's electricity load data;

[0181] Define the newly collected typical user load dataset of the kth class as L k-new Let L be the target user's electricity load data that is missing in this load dataset. k-new0 A binary mask matrix M of dimension m×D is introduced to mask the target user's electricity load data L. k-new0 The missing values ​​are represented by a value of 0 for the missing element and 1 for the missing element. The target user's electricity load data after missing value representation is defined as follows: The formula is as follows:

[0182]

[0183] In the formula, ⊙ represents the Hadamard product operation of the matrix.

[0184] Step S11: Generate the day-ahead spot market clearing electricity price dataset P and L k-new Except for L k-new0 Other load datasets L k-knew0 As a condition, it is concatenated with random noise data z and input into the generator G. k Output generated data G k (z);

[0185] The recent spot market clearing electricity price dataset P and L k-new Except for L k-new0 Other load datasets Lk-knew0 The random noise data z is concatenated with the random noise data z through the third dimension channel to form a three-dimensional matrix. The final sample size is m×D×I. The concatenation order of the third dimension matrix is ​​[P, z, L]. k-knew0 ], L k-knew0 ={L k-knew1 ,L k-knew2 ,···,L k-knew(I-1)}

[0186] Step S12: Generate data sample G k (z) Fill in the target user's electricity load data The missing part;

[0187] definition After the missing data was filled in, the final output user electricity load data was: The original data retains the parts that are not missing, and the missing parts are generated using data G. k (z) is used for filling, and the calculation formula is shown below:

[0188]

[0189] 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.

[0190] 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.

[0191] 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.

[0192] 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.

[0193] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0194] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

[0195] This patent is not limited to the above-described preferred embodiment. Anyone can derive other forms of day-ahead spot clearing price-load data repair method based on generative adversarial networks under the guidance of this patent. All equivalent changes and modifications made within the scope of this patent application shall fall within the scope of this patent.

Claims

1. A day-ahead spot clearing price-load data restoration method based on a generative adversarial network, characterized in that, The day-ahead spot market historical clearing price data and historical user power load data of a power system are collected, the data parts without missing are retained, and a training data set is established; a K-means clustering algorithm is used to perform clustering analysis on the user power load data set, and typical user groups are divided; The average load of the load data of each user is calculated; Then, the elbow method is used to determine the number of clusters in the clustering analysis; A clustering analysis threshold ε is set, and it is judged whether the update amount of the objective function is less than the threshold ε, so as to determine whether to end the clustering iteration process; A deep convolutional generative adversarial network (DCGAN) model is established, the Wasserstein distance and the gradient penalty function are introduced, and the performance of the DCGAN model is improved; the day-ahead spot market historical clearing price data and the historical user power load data after the typical user groups are divided are used to train the DCGAN model, and the generator structure for each typical user group is retained; Specifically, this involves compiling a complete dataset of historical cleared electricity prices in the day-ahead spot market (P) and a complete dataset of typical user loads (L). k In addition to the selected real load dataset L k0 Other load datasets L k-0 As a condition, it is concatenated with random noise data z and then input into the generator, and the generator outputs generated data G. k (z); The concatenated samples input to the generator are in the form of a three-dimensional matrix, consisting of dataset P, random noise data z, and L. k-0 The data is concatenated using the third dimension channel, resulting in a sample size of m×D×I; where D represents the maximum number of days in the dataset, m represents the number of daily cleared electricity prices in the day-to-day spot market, and I represents the number of users. The concatenation order of the third dimension matrix is ​​[P, z, L]. k-0 ], L k-0 ={L k1 ,L k2 ,···,L k(I-1) }; A binary mask matrix is introduced to represent the missing positions of the newly collected user power load data; Let Lkdenote the newly collected typical user load data set of the k-th type k-new Let Lkdenote the newly collected typical user load data set of the k-th type k-new0 Let M denote the m x D binary mask matrix for characterizing the missing status of the target user load data L k-new0 The element value of the missing position is 0, otherwise it is 1. Then, the newly collected day-ahead spot market historical clearing price data set without missing, the typical user power load data set, and the target user power load data with missing are input into the generator corresponding to the typical user group, and new generated data are output; Finally, the missing parts of the target user power load data are filled by using the new generated data samples. 2.The day-ahead spot-out clearing locational marginal price-load data restoration method based on a generative adversarial network according to claim 1, wherein, Specifically, the method comprises the following steps: Step S1: Collecting the day-ahead spot market historical clearing price data and historical user power load data of a power system, retaining the data parts without missing, and establishing a training data set; Step S2: Using a K-means clustering algorithm to perform clustering analysis on the user power load data set L, and dividing typical user groups; Step S3: Normalizing the sample data of the typical user groups and the day-ahead spot market clearing price; Step S4: Establishing a deep convolutional generative adversarial network (DCGAN) model; Step S5: Introducing the Wasserstein distance and the gradient penalty function to improve the performance of the DCGAN model; Step S6: Concatenate the complete day-ahead spot market historical clearing price dataset P, L k with the load dataset L k0 except L k-0 as a condition, and input the concatenated data into the generator after splicing with random noise data z, and the generator outputs generated data G k (z); Step S7: concatenate the complete day-ahead spot market historical clearing price dataset P, L k except L k0 to generate a load dataset L k-0 as a condition, with the user load data generation sample G k (z) to the discriminator; and concatenate the complete day-ahead spot market historical clearing price dataset P, L k except L k0 to generate a load dataset L k-0 as a condition, with the selected real load dataset L k0 to the discriminator after concatenation, so that the discriminator outputs a discrimination result of the user load data generation sample G k (z) and the selected real load dataset L k0 ; Step S8: Calculating the loss function of the generator and the discriminator, and using the RMSprop optimization algorithm to optimize and update the network weight parameters of the generator and the discriminator; when a round of training is completed, returning to step S6 for the next round of training; Step S9: When the training is finished, the network structure and parameters of the generator model G in the DCGAN are reserved. k ​ Step S10: When the power load data of a certain user is newly collected with missing, a binary mask matrix is introduced to represent the missing positions of the user power load data; Step S11: inputting a complete day-ahead spot market historical clearing price data set P, L k-new except for L k-new0 a load data set L k-knew0 as a condition, splicing with random noise data z and inputting a generator, generator G k output generation data G k (z); Step S12: generating data samples G k (z) filling in missing parts of the target user's electricity load data .

3. The day-ahead spot market clearing price-load data repair method based on a generative adversarial network according to claim 2, characterized in that: In step S1: The day-ahead spot market historical clearing price data set without missing is defined as P, and the historical user power load data set without missing is defined as L, and the expression of P is as follows: In the formula, p d,t represents the historical clearing price of the day-ahead spot market on the dth day at the tth time L is composed of N L n The data set consists of N users in total in the data set; L n The expression is as follows: In the formula, l n,d,t represents the power consumption load data of the nth user at the dth day and the tth time, and the power consumption load collection time corresponds to the day-ahead spot market clearing time. In step S2: calculating an average load of the load data of each user The formula is as follows: Input average load data set : Then, the elbow method is used to determine the number of clusters in the clustering analysis; Initialization of cluster centers, i.e. randomly select K data from the average load data set as cluster centers Initialization of cluster centers, i.e. randomly select K data from the average load data set as cluster centers Calculation of the average load of each user Calculation of the average load of each user nk The calculation formula is as follows: In the formula, n=1, 2, …, N; k=1, 2, …, K; Set average load The cluster label is l n And divide To the cluster with the shortest Euclidean distance, the calculation formula is as follows: l n = arg min d nk (5) Update the cluster centers and make sure that n = k, k = 1, 2, ···, K; A clustering analysis threshold ε is set, and it is judged whether the update amount of the objective function is less than the threshold ε, so as to determine whether to end the clustering iteration process; The calculation formula of the objective function E is as follows: If the target function E update amount is not less than the threshold value ε, continue clustering iteration t=t+1; when the target function E update amount is less than the threshold value ε, end the clustering iteration process; Finally, according to the clustering division, K typical user groups are formed, each of which is represented by L k representing, k = 1, 2,..., K; each L k contains a plurality of users; In step S3: The normalization formula is as follows: The normalized data forms a complete day-ahead spot market historical clearing price dataset P and N user load datasets L n ; according to the cluster analysis, the user dataset is divided into K complete typical user load datasets L k .

4. The day-ahead spot clearing price-load data restoration method based on a generative adversarial network according to claim 2, characterized in that: In step S4: For a typical user load dataset L without missing values k , let I users be contained, randomly select one of them L k0 , define it as real data; define a set of random noise data z as the input of the generator, p z (z) represents the probability distribution of z, p data (L k0 ) represents the probability distribution of real data L k0 ; for the characteristics of the typical user group L k , define the data sample G k (z) generated by the generator as generated data, and the probability distribution is p Gk (z); define the input of the discriminator network as real data L k0 and generated data G k (z), and the output is a scalar D(G k (z)), which represents the probability that the input noise z obeys the real data distribution p data (L k0 ). According to the training target of the generator and the discriminator, loss functions L of the generator and the discriminator are respectively constructed G and L D , and a target function in the GAN training process is determined:

5. The day-ahead spot clearing price-load data restoration method based on a generative adversarial network according to claim 4, characterized in that: In step S5: The definition of the Wasserstein distance is as follows: where Ω(p data ,p Gk ) is a set of joint probability distributions γ with marginal distributions p data and p Gk ; W(p data ,p Gk ) is the lower bound of the expectation of (u, v) ~ γ, representing the distance that u needs to be moved to v in order to fit the generated data distribution p Gk to the real data distribution p data , where u and v represent a real data sample L k0 and a generated data sample G k (z) randomly sampled from the joint distribution γ, respectively; the Kantorovich-Rubinstein dual form is adopted to describe the distance between the generated sample and the real sample: where ||f D ||L≤K indicates that the discriminator function D(L k0 ) satisfies K-Lipschitz continuity, that is, the upper limit of the absolute value of the function gradient is K; in order to ensure that the gradient does not exceed the limit value K, the gradient penalty function of the discriminator function D(L k0 ) in the definition domain is introduced to the formula (13), so that the discriminator function D(L k0 ) approximately satisfies K-Lipschitz continuity, so as to accurately describe the Wasserstein distance, at this time, the objective function of the GAN is transformed into:

6. The day-ahead spot clearing price-load data restoration method based on a generative adversarial network according to claim 5, characterized in that: The generator structure of the DCGAN model is as follows: Convolutional layer 1: 32 convolutional kernels, each with a size of 3; a stride of 1; an edge padding number of 1; a normalization function of BatchNorm2D; and an activation function of LeakyReLU; Convolutional layer 2: 64 convolutional kernels, each with a size of 3; a stride of 1; an edge padding number of 1; a normalization function of BatchNorm2D; and an activation function of LeakyReLU; Convolutional layer 3: 128 convolutional kernels, each with a size of 3; a stride of 1; an edge padding number of 1; a normalization function of BatchNorm2D; and an activation function of LeakyReLU; Convolutional layer 4: 256 convolutional kernels, each with a size of 3; a stride of 1; an edge padding number of 1; a normalization function of BatchNorm2D; and an activation function of LeakyReLU; Convolutional layer 5: 16 convolutional kernels, each with a size of 3; a stride of 1; an edge padding number of 1; a normalization function of BatchNorm2D; and an activation function of ReLU; Convolutional layer 6: 4 convolutional kernels, each with a size of 3; a stride of 1; an edge padding number of 1; a normalization function of BatchNorm2D; and an activation function of ReLU.

7. The day-ahead spot clearing price-load data restoration method based on a generative adversarial network according to claim 6, characterized in that: In step S7: The complete day-ahead spot market historical clearing price dataset P, L k The load dataset L k0 except L k-0 The user load data generating sample G k (z) is spliced through the third dimension channel to form a three-dimensional matrix, and finally spliced to form a sample with a size of m x D x I; wherein the splicing order of the third dimension matrix is [P, G k (z), L k-0 ], L k-0 ={L k1 , L k2 ,..., L k(I-1)}; a complete day-ahead spot market historical clearing price dataset P, L k a load dataset L other than L k0 k-0 a selected real load dataset L k0 splicing through the third dimension channel to form a three-dimensional matrix, and finally splicing to form a sample with a size of m x D x I;​ wherein the concatenation order of the 3rd dimension matrix is [P, L k0 , k-0 ], L k-0 = {L k1 , L k2 , ···, L k(I-1)} ; The discriminator structure of the DCGAN model is as follows: Input layer: 64 convolutional kernels, each with a size of 3; a stride of 1; Convolutional layer 1: 128 convolutional kernels, each with a size of 4; a stride of 2; Normalization function: BatchNorm2D; and activation function: LeakyReLU; Convolutional layer 2: 256 convolutional kernels, each with a size of 4; a stride of 2; Normalization function: BatchNorm2D; and activation function: LeakyReLU; Convolutional layer 3: 512 convolutional kernels, each with a size of 4; a stride of 2; Normalization function: BatchNorm2D; and activation function: LeakyReLU; Convolutional layer 4: 1024 convolution kernels, each with a size of 4; stride of 2; The normalization function is BatchNorm2D; the activation function is LeakyReLU; Convolutional layer 5: 2048 convolution kernels, each with a size of 4; stride of 2; The normalization function is BatchNorm2D; the activation function is LeakyReLU; Convolutional layer 6: 512 convolution kernels, each with a size of 3; stride of 1; The normalization function is BatchNorm2D; the activation function is LeakyReLU; Convolutional layer 7: 128 convolution kernels, each with a size of 3; stride of 1; The normalization function is BatchNorm2D; the activation function is LeakyReLU; Output layer: dense layer structure, containing 1024 neural network units, 1 sample as output.

8. The day-ahead spot market clearing price-load data restoration method based on a generative adversarial network according to claim 7, characterized in that: In step S8: Set the learning rate of the RMSprop optimization algorithm = 2 x 10 -4 , parameter = 0.9; The formula of the RMSprop optimization algorithm is as follows: In the formula s dw and s db are the momentum of the gradient accumulated in the previous iteration process, and β is a parameter representing the gradient accumulation; the RMSprop algorithm calculates the differential square weighted average of the gradient; when the value in dW or db exceeds the preset range, the change is divided by the momentum of the gradient accumulated in the previous iteration process to meet the requirement of the gradient swing amplitude value. Epsilon is an auxiliary parameter to prevent singularity caused by a zero denominator; In step S10: Definition of the newly collected typical user load data set of the kth type as L k-new Definition of the target user power load data with missing phenomenon in the load data set as L k-new0 Introducing a binary mask matrix M with dimension m x D to represent the missing situation of the target user power load data L k-new0 The element value corresponding to the missing position is 0, otherwise it is 1; definition of the target user power load data after the missing value is represented as The formula is as follows: In the formula, is the Hadamard product operation of the matrix; In step S11: The complete day-ahead spot market historical clearing price data set P, L k-new The load data set L k-new0 except L k-knew0 , and random noise data z through the third dimension channel, forming a three-dimensional matrix, and finally splicing the sample size m x D x I; wherein the splicing order of the third dimension matrix is [P, z, L k-knew0 ] k-knew0 = {L k-knew1 , L k-knew2 ,..., L k-knew(I-1)} In step S12: Definition The final output of the user electricity load data is The original data is retained without missing, and the missing part is filled by generating data G k (z) is filled, and the calculation formula is as follows: 9.A day-ahead spot clearing price-load data restoration device based on a generative adversarial network, characterized by: In runtime, use the day-ahead spot market clearing price-load data restoration method based on a generative adversarial network according to any one of claims 1-8. 10.A non-transitory computer-readable storage medium having stored thereon a computer program. In runtime, use the day-ahead spot market clearing price-load data restoration method based on a generative adversarial network according to any one of claims 1-8.

Images