Source load short-time fluctuation scene generation method based on improved generative adversarial network

Abstract

The invention discloses a source load short-time fluctuation scene generation method based on an improved generative adversarial network, and belongs to the technical field of power system data analysis. The method comprises the following steps: firstly, performing sliding window segmentation on a source load power time sequence, extracting climbing features, and identifying a typical fluctuation mode by using fuzzy C-means clustering; then, constructing and training a gradient penalty Wasserstein generative adversarial network model based on mean-standard deviation correction, introducing a gradient penalty term, and performing statistical correction of mean and standard deviation on a generated sample; and finally, generating a high-quality minute-level source load fluctuation scene by using the trained model, and comprehensively evaluating the quality of the generated scene by using multi-dimensional indexes such as principal component analysis, t-SNE, kernel density estimation, Wasserstein distance and the like. According to the method, the high-frequency and non-stationary characteristics of minute-level source load fluctuation can be accurately reconstructed, the generated scene is high in authenticity and good in diversity, and reliable data support is provided for short-time scheduling and safety analysis of a power system under high-proportion new energy access.

Description

Technical Field

[0001] This invention belongs to the field of power system data analysis technology, specifically relating to a method for generating short-term source-load fluctuation scenarios based on an improved generative adversarial network. Background Technology

[0002] With the increasing penetration of renewable energy in new power systems, the power output of renewable energy sources such as wind and solar power is highly volatile due to limitations imposed by natural conditions. Simultaneously, the large-scale integration of distributed solar power and the rapid development of electric vehicles are further amplifying the volatility of power load, making the overall fluctuation characteristics of the power system increasingly complex, exhibiting stronger non-stationarity and rapid changes at minute-level time scales. This places higher demands on system frequency regulation, ramp-up capabilities, reserve allocation, and short-term dispatch strategies. However, most current source-load scenario generation methods still focus on hourly and longer time scales, failing to reflect minute-level fluctuations, thus limiting their dispatch guidance capabilities in high-time-sensitivity scenarios. Therefore, studying source-load fluctuation characteristics and constructing source-load scenarios with minute-level time resolution that can reflect high-frequency fluctuation characteristics is of great significance for short-timescale power grid planning and dispatch operation.

[0003] To address the fluctuation characteristics of source loads, existing literature has studied this issue using multiple indicators such as fluctuation amplitude and tilt rate, and has combined these with scene generation methods to generate a large number of scenes with the same fluctuation characteristics. Currently, common scene generation methods mainly include statistical methods and deep learning methods.

[0004] In terms of statistical methods, most methods use parametric or non-parametric estimation to establish a probability model that conforms to the distribution of the original data, and then generate scenes through sampling and clustering. Existing statistical methods need to consider establishing a suitable probability model to improve the quality of the generated scenes during the scene generation process, and there are problems such as unreasonable scene generation due to differences between the assumed distribution and the actual distribution.

[0005] In practice, deep learning also has certain limitations. For example, the stability and convergence of the traditional GAN ​​training process are difficult to guarantee, resulting in insufficient diversity of generated samples. Furthermore, if the parameter optimization process between the generator and discriminator is not well synchronized, the training results of GAN may be unsatisfactory. To address the problems of gradient vanishing, training instability, and mode collapse in GANs, WGAN uses Wasserstein distance instead of JS divergence as the optimization objective.

[0006] Existing research largely focuses on source-load scenario generation on intraday and longer timescales, with limited research on scenario generation for minute-level short-term source-load fluctuations. Furthermore, minute-level source-load fluctuations often involve critical periods such as sudden and drastic output fluctuations, peak load changes, and renewable energy ramp-up events, directly impacting the system's regulation capacity and safety margin. Compared to hourly data, minute-level source-load fluctuations exhibit greater uncertainty and higher frequency characteristics, placing higher demands on scenario generation methods in terms of distribution reproduction and feature restoration. In addition, traditional generation models driven by standard Gaussian noise often suffer from problems such as sample distribution shifts and scale mismatches when dealing with highly volatile and significantly skewed minute-level data, making it difficult to accurately reconstruct the actual fluctuation characteristics. Summary of the Invention

[0007] This invention takes into account the current status of power system sources and loads and the shortcomings of traditional scenario generation models. By studying scenarios with a high proportion of renewable energy access, this invention constructs high-quality short-term, minute-level source and load fluctuation scenarios and proposes a method for generating minute-level short-term source and load fluctuation scenarios based on a Wasserstein generative adversarial network with gradient penalty and Mean-Standard Deviation Calibration (WGAN-GP-MS).

[0008] To achieve the above method, the present invention adopts the following technical solution:

[0009] A method for generating short-term source load fluctuation scenarios based on an improved generative adversarial network includes the following steps:

[0010] (1) Source load short-term fluctuation segment characteristic analysis: The source load power time series is segmented, the characteristic parameters of each fluctuation segment are extracted, and cluster analysis is performed based on the characteristic parameters to identify typical fluctuation patterns;

[0011] (2) Scene generation model construction and training: Construct a gradient-penalized Wasserstein generative adversarial network model based on mean-standard deviation correction, and train the model using the typical fluctuation pattern data obtained in step (1);

[0012] (3) Scene generation and quality evaluation: The trained model is used to generate short-term fluctuation scenes of source load, and the quality of the generated scenes is comprehensively evaluated using preset evaluation indicators.

[0013] Furthermore, in step (1):

[0014] The segmented processing employs a sliding time window method.

[0015] The characteristic parameters include at least: climbing range, climbing time, climbing rate, and climbing direction;

[0016] The clustering analysis employs the fuzzy c-means clustering algorithm, and the optimal number of clusters is determined by an effectiveness evaluation index function based on intra-cluster compactness and inter-cluster separation.

[0017] Furthermore, the index function is: ,in, The number of clusters; Membership matrix generated for the FCM algorithm; Intra-cluster compactness (intra-cluster variance index) measures the similarity between samples within the same cluster; It is an inter-class separation index, reflecting the separability between different clusters; It is an index of inter-cluster overlap, reflecting the degree of ambiguity of samples among multiple clusters; For adjustment terms (generally taken) To control the weight of indicators.

[0018] Furthermore, the model in step (2) is the WGAN-GP-MS model, which contains a generator and a discriminator;

[0019] The generator is used to receive random noise and generate simulated source load fluctuation segment data;

[0020] The discriminator is used to evaluate real data and generated data and output an authenticity score;

[0021] The training objective of the model is to minimize the Wasserstein distance between the real data distribution and the generated data distribution.

[0022] Furthermore, the objective function of the WGAN-GP-MS model is expressed as:

[0023]

[0024] The first term is the expected score of the real sample; the second term is the expected score of the generated sample; and the third term is the gradient penalty term. This is the penalty coefficient; Interpolated samples between real and generated samples: In the formula, It is a random number that is uniformly distributed between 0 and 1.

[0025] Furthermore, the discriminator loss function of the WGAN-GP-MS model includes a gradient penalty term to force the discriminator to satisfy the Lipschitz continuity constraint.

[0026] Furthermore, the WGAN-GP-MS model performs mean-standard deviation correction on the synthesized samples output by the generator. The specific correction steps are as follows:

[0027] a) The generator generates samples based on Gaussian noise. : , ;

[0028] b) Calculate the mean μ of the real sample across each time dimension. t,real and standard deviation σ t,real ;

[0029] c) Scale the generated samples by their mean and standard deviation to align their statistical properties with those of the real samples: ,in, In time step The first after the upcorrection One generated sample; For real samples, μ t,real σ is the mean of the real sample at time step t; t,real This represents the standard deviation of the real sample at time step t.

[0030] Furthermore, the evaluation indicators in step (3) include:

[0031] Principal component analysis and t-distributed random neighborhood embedding methods are used to perform dimensionality reduction and visualization comparison of high-dimensional fluctuating feature vectors.

[0032] The kernel density function estimation method is used to compare the probability density distributions of real samples and generated samples on key features;

[0033] Wasserstein distance is used to quantify the consistency of key feature distributions between real and generated samples.

[0034] Furthermore, the kernel density function is estimated as follows: ,in, For kernel functions; The bandwidth parameter is x; the feature dimension is x, and the sample set is x. .

[0035] Improvements and effects of the present invention

[0036] (1) This invention combines two dimensionality reduction visualization methods, PCA and t-SNE, as well as statistical similarity evaluation indicators such as kernel density estimation and Wasserstein distance. The WGAN-GP-MS model can effectively reconstruct the distribution and structural characteristics of real samples in various source-load fluctuation scenarios, making the method of this invention have good generalization ability and adaptability, and providing a high-quality data foundation for the analysis and simulation of power system fluctuation characteristics under high proportion of new energy scenarios.

[0037] (2) The improved WGAN-GP-MS model of this invention introduces the statistical information of real samples to post-process the generated samples through the sample standardization correction strategy. While improving the distribution fitting accuracy, it further optimizes the robustness and sample availability of the model. The generated samples are closer to the real samples in terms of numerical indicators, which enhances the practical value of the model in the generation of supplementary scenarios in small sample fluctuation range. Attached Figure Description

[0038] Figure 1 : A flowchart of the scene generation method described in this invention. Detailed Implementation

[0039] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments and specific features in the embodiments are detailed descriptions of the technical solution of the present application, rather than limitations thereof. In the absence of conflict, the embodiments and technical features in the embodiments can be combined with each other.

[0040] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0041] The core challenge in generating high-quality minute-level source-load fluctuation scenarios lies in how to accurately capture and reproduce their high-frequency, non-stationary, and skewed distribution characteristics, while avoiding the distribution shift and scale mismatch problems caused by the mismatch between input noise and the statistical characteristics of real data in traditional generation models.

[0042] To address this challenge, this invention constructs a source-load minute-level scene generation method based on a gradient-penalized Wasserstein generative adversarial network (WGAN-GP-MS) with mean-standard deviation correction.

[0043] like Figure 1 As shown, the specific implementation steps of the method of the present invention are as follows:

[0044] 1. Analysis of the characteristics of short-term fluctuations in source load

[0045] a. Short-term fluctuation segment division and identification

[0046] To characterize the minute-level power fluctuations of the source load, a sliding time window method is used to segment the power time series and extract its key ramp-up features, as shown below to quantify its fluctuation characteristics:

[0047] 1) Gradient The change in power within the fluctuation range.

[0048]

[0049] In the formula, Maximum power value within the fluctuation range; Minimum power value within the fluctuation range.

[0050] 2) Climbing time Δt, the duration of power change within the fluctuation segment.

[0051]

[0052] In the formula, The time when the maximum power value occurs within the fluctuation range; The time when the minimum power value occurs within the fluctuation range.

[0053] 3) Slope rate , the rate of change of power.

[0054]

[0055] 4) Climbing direction: Power increase is uphill climb, recorded as 1; power decrease is downhill climb, recorded as 0.

[0056] b. Cluster analysis of typical fluctuation segments based on fuzzy C-means clustering algorithm

[0057] To more effectively analyze the characteristics of the climbing fluctuation segment of source load power on a minute-scale basis, fuzzy C-means clustering algorithm is used to perform unsupervised clustering analysis on different fluctuation segments based on four types of characteristic parameters of the fluctuation segment in order to identify typical fluctuation patterns.

[0058] In the process of cluster analysis, it is important to reasonably determine the number of clusters. This is crucial for ensuring consistency between clustering results and physical interpretation. An effectiveness evaluation index function based on intra-cluster compactness and inter-cluster separation is introduced. The mathematical expression for determining the optimal number of clusters is as follows:

[0059]

[0060] In the formula, The number of clusters; Membership matrix generated for the FCM algorithm; Intra-cluster compactness (intra-cluster variance index) measures the similarity between samples within the same cluster; It is an inter-class separation index, reflecting the separability between different clusters; It is an index of inter-cluster overlap, reflecting the degree of ambiguity of samples among multiple clusters; For adjustment terms (generally taken) To control the weight of indicators.

[0061] The index function W(c,U) comprehensively considers intra-class consistency and inter-class separability. The smaller the value, the better the clustering effect, that is, the clearer the distribution of fluctuation segments in each category, and the more discriminative the typical fluctuation pattern recognition results. The optimal number of clusters is determined by equation (4). Subsequently, the FCM algorithm was used to perform cluster analysis on the wave segment feature matrix, thereby identifying several representative wave segment types. This provides a foundation for the generation of typical wave pattern scenarios in the future.

[0062] 2. Gradient-penalized generative adversarial network model based on mean-standard deviation correction

[0063] a. Wasserstein Generative Adversarial Network

[0064] Generative Adversarial Networks (GANs) are widely used in tasks such as modeling complex data distributions and synthesizing high-quality samples. Their basic idea is to learn the data distribution through adversarial training between a generator and a discriminator. The generator produces samples by receiving a random noise vector z. (i.e., fake samples), while the discriminator accepts sample x and outputs a probability value. , which represents the probability that the sample is real data.

[0065] During the training process of GAN, the generator is responsible for generating more realistic samples. The discriminator is responsible for distinguishing between real samples (i.e., genuine samples) and other samples. In the alternating training of the generator and discriminator, the samples generated by the generator are... To approximate the distribution of real data as closely as possible, the loss function of a GAN can be expressed in the following form, where the generator loss function is:

[0066]

[0067] In the formula, The input noise distribution for the generator; The generated sample.

[0068] The discriminator loss function is:

[0069]

[0070] In the formula, This represents the true data distribution. For expectations; For the discriminator to judge real samples The score; D(G(z)) is the discriminant's score for the generated samples. The rating.

[0071] However, traditional GANs suffer from problems such as training instability, vanishing gradients, and pattern collapse, especially when modeling high-dimensional, continuous time series. To address the instability and pattern collapse issues of traditional GANs during training, this invention employs a Wasserstein Generative Adversarial Network (WGAN). Unlike the JS divergence used by traditional GANs, WGAN uses Wasserstein distance to measure the difference between the real data distribution and the generated data distribution. Its advantage lies in the fact that even if the two distributions do not overlap, the Wasserstein distance can still reflect their proximity, resulting in smoother gradients and improved training stability.

[0072] Wasserstein distance for:

[0073]

[0074] In the formula, supremum is the supremum for all 1-Lipschitz continuous functions f.

[0075] The generator and discriminator loss functions of WGAN are as follows:

[0076]

[0077]

[0078] In the formula, The output of the discriminator represents the sample. The rating.

[0079] WGAN makes the objective functions of the generator and discriminator smoother by minimizing the Wasserstein distance, thus avoiding the training instability caused by the non-smoothness of the cross-entropy loss function in GAN.

[0080] b. Improvements to the WGAN-GP-MS algorithm

[0081] WGAN still faces some challenges, such as ensuring the discriminator satisfies the K-Lipschitz continuity condition to guarantee efficient computation of the Wasserstein distance. To improve the generation of samples... To improve quality, a gradient penalty term is introduced into the discriminator loss function, which directly constrains the gradient norm of the discriminator output relative to the input, ensuring that it satisfies the K-Lipschitz continuity condition, without the need for weight clipping. This significantly improves training stability and the quality of generated samples.

[0082] Meanwhile, considering that the generator's input is standard Gaussian random noise, which theoretically follows a distribution with zero mean and unit variance, samples generated directly from standard Gaussian noise may deviate from real samples in terms of numerical range, center position, or fluctuation amplitude. This is especially pronounced in the early stages of training or when the sample categories themselves have significant deviation characteristics, which is detrimental to model convergence. To improve the realism and stability of the generated samples in terms of numerical scale and statistical properties, a simple and effective correction strategy is used: during the sample generation output stage, adjustments are made based on the mean and standard deviation of the real samples.

[0083] Specifically, let the synthesized sample output by the generator be... These samples are all derived from Gaussian random noise with a mean of 0 and a variance of 1. During the generation process, they are corrected and adjusted so that their mean and standard deviation are close to the mean of the true data. and standard deviation The specific calibration steps are as follows:

[0084] 1) Generate samples: First, the generator outputs a set of samples. Each sample is generated based on Gaussian noise. Assume these samples have a dimension of 1. ,in For the sample size, The time dimension for each sample.

[0085]

[0086] In the formula, .

[0087] 2) Calculate the mean and standard deviation of the real sample at each time dimension: for each time step (from 1 to ), calculate the mean and standard deviation of all samples:

[0088]

[0089]

[0090] In the formula, For real samples, μ t,real Let σ be the mean of the real sample at time step t. t,real This represents the standard deviation of the real sample at time step t.

[0091] 3) Correcting generated samples: Scaling the generated samples by adjusting their mean and standard deviation to align them with the mean and standard deviation of the real samples.

[0092]

[0093] In the formula, In time step The first after the upcorrection One generated sample.

[0094] The improved WGAN-GP-MS mainly consists of the following two adversarial subnetworks: the generator G uses random noise. As input, it outputs simulated source load fluctuation segment data. Discriminator D compares the real data... With generated data Instead of probability values ​​as in traditional GANs, the evaluation outputs a "realism" score.

[0095] The core idea of ​​WGAN-GP-MS is to transform the game process between the generator G and the discriminator D into minimizing the Wasserstein distance, and its objective function is expressed as:

[0096]

[0097] In the formula, the first term is the expected score of the real sample; the second term is the expected score of the generated sample; and the third term is the gradient penalty term, which is used to force the discriminator to satisfy the Lipschitz constraint. This is the penalty coefficient; Interpolated samples between real and generated samples:

[0098]

[0099] In the formula, It is a random number that is uniformly distributed between 0 and 1.

[0100] 3. Scene quality evaluation indicators

[0101] The authenticity and diversity of the source-load fluctuation scenarios generated by WGAN-GP-MS are comprehensively evaluated. The quality of the generated scenarios is assessed from multiple perspectives, including low-dimensional visualization comparison and consistency of statistical features.

[0102] Firstly, regarding low-dimensional visualization, to observe the similarity between synthetic and real data in their overall distribution, principal component analysis (PCA) and t-distributed random neighborhood embedding (t-SNE) are used to reduce the dimensionality of the high-dimensional fluctuating feature vectors, and the distribution of real and generated samples is visualized in two-dimensional space. PCA extracts principal components through linear mapping, preserving the direction of maximum variance in the data; t-SNE is suitable for visualizing complex nonlinear structures, emphasizing the preservation of local structures, which helps to observe whether the cluster distributions of real and synthetic samples are consistent. If the dimensionality-reduced synthetic samples and real samples have good overlap or similar clustering structures in two-dimensional space, the model can be considered to have good generative capabilities.

[0103] Secondly, to further compare the distribution consistency of the samples across statistical feature dimensions, kernel density estimation is used to estimate and visualize the probability density functions of the real and synthetic sample sets on the key feature ramp-up magnitude. The feature dimension is x, and the sample set is... Then its kernel density function is estimated as follows:

[0104]

[0105] In the formula, For kernel functions; This is the bandwidth parameter.

[0106] In summary, by comparing the kernel density curves of real and synthetic samples, we can observe whether their probability distributions are consistent, thereby evaluating the model's fit to the statistical distributions of each feature.

[0107] Wasserstein distance is used to measure the minimum "transportation cost" between two probability distributions. Under a one-dimensional continuous distribution, the two distributions... and The first-order Wasserstein distance is defined as:

[0108]

[0109] In the formula, For distribution The cumulative distribution function; For distribution The cumulative distribution function of the generated sample. The closer the Wasserstein distance is to 0, the more consistent the two distributions are. The larger the Wasserstein distance is, the more obvious the difference between the distributions, and the less the generated sample resembles the real sample.

[0110] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for generating short-term source load fluctuation scenarios based on an improved generative adversarial network, characterized in that, Includes the following steps: (1) Source load short-term fluctuation segment characteristic analysis: The source load power time series is segmented, the characteristic parameters of each fluctuation segment are extracted, and cluster analysis is performed based on the characteristic parameters to identify typical fluctuation patterns; (2) Scene generation model construction and training: Construct a gradient-penalized Wasserstein generative adversarial network model based on mean-standard deviation correction, and train the model using the typical fluctuation pattern data obtained in step (1); (3) Scene generation and quality evaluation: The trained model is used to generate short-term fluctuation scenes of source load, and the quality of the generated scenes is comprehensively evaluated using preset evaluation indicators.

2. The scene generation method according to claim 1, characterized in that, In step (1): The segmented processing employs a sliding time window method; The characteristic parameters include at least: climbing range, climbing time, climbing rate, and climbing direction; The clustering analysis employs the fuzzy c-means clustering algorithm, and the optimal number of clusters is determined by an effectiveness evaluation index function based on intra-cluster compactness and inter-cluster separation.

3. The scene generation method according to claim 2, characterized in that, The index function is: ,in, The number of clusters; Membership matrix generated for the FCM algorithm; Intra-cluster compactness (intra-cluster variance index) measures the similarity between samples within the same cluster; It is an inter-class separation index, reflecting the separability between different clusters; It is an index of inter-cluster overlap, reflecting the degree of ambiguity of samples among multiple clusters; For adjustment terms (generally taken) To control the weight of indicators.

4. The scene generation method according to claim 1, characterized in that, The model in step (2) is the WGAN-GP-MS model, which contains a generator and a discriminator; The generator is used to receive random noise and generate simulated source load fluctuation segment data; The discriminator is used to evaluate real data and generated data and output an authenticity score; The training objective of the model is to minimize the Wasserstein distance between the real data distribution and the generated data distribution.

5. The scene generation method according to claim 4, characterized in that, The objective function of the WGAN-GP-MS model is expressed as follows: The first term is the expected score of the real sample; the second term is the expected score of the generated sample; and the third term is the gradient penalty term. This is the penalty coefficient; Interpolated samples between real and generated samples: In the formula, It is a random number that is uniformly distributed between 0 and 1.

6. The scene generation method according to claim 4, characterized in that, The discriminator loss function of the WGAN-GP-MS model includes a gradient penalty term, which is used to force the discriminator to satisfy the Lipschitz continuity constraint.

7. The scene generation method according to any one of claims 4 to 6, characterized in that, The WGAN-GP-MS model performs mean-standard deviation correction on the synthetic samples output by the generator. The specific correction steps are as follows: a) The generator generates samples based on Gaussian noise. : , ; b) Calculate the mean μ of the real sample across each time dimension. t,real and standard deviation σ t,real ; c) Scale the generated samples by their mean and standard deviation to align their statistical properties with those of the real samples: ,in, In time step The first after the upcorrection One generated sample; For real samples, μ t,real σ is the mean of the real sample at time step t; t,real This represents the standard deviation of the real sample at time step t.

8. The scene generation method according to claim 1, characterized in that, The evaluation indicators in step (3) include: Principal component analysis and t-distributed random neighborhood embedding methods are used to perform dimensionality reduction and visualization comparison of high-dimensional fluctuating feature vectors. The kernel density function estimation method is used to compare the probability density distributions of real samples and generated samples on key features; Wasserstein distance is used to quantify the consistency of key feature distributions between real and generated samples.

9. The scene generation method according to claim 1, characterized in that, The kernel density function is estimated as follows: ,in, For kernel functions; The bandwidth parameter is x; the feature dimension is x, and the sample set is x. .

Images