Controllable scene generation method based on adversarial network
By introducing interpretable control vector C and random noise vector Z in the GAN model, the potential space input in the structure of the constructor solves the problems that the scene generation method in the existing technology cannot meet the unpredictable needs of multiple needs and potential space. The adaptive learning of network models and the generation of new scene models reduces the operating cost of the power system and improves stability.
Patent Information
- Application Number
- CN202411239408.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-05
- Publication Date
- 2025-01-28
AI Technical Summary
The existing GAN-based renewable energy scenario generation method cannot meet the multi-demand scenario generation task, and the potential space of the adversarial network is not interpretable, and it is impossible to generate non-existent new scenario modes based on historical data, resulting in stable operation of the power system. Reduced performance and increased operating costs.
By constructing a potential space input in the form of streaming, including control vector C with characteristic indicators with physical significance, and random noise vector Z, iterative training generates a confrontation network, adjust the control vector C along the explanatory features, to achieve controllable scenarios generating the generating scene generation Essence
Realizing data -driven adaptive learning against network models can generate new scenario models that meet the statistical characteristics, reduce the operating cost of the power system and improve stability.
Smart Images
Figure CN119378354A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of integrated energy systems, and in particular to a controllable scenario generation method based on an adversarial network. Background Art
[0002] The volatility and intermittency of renewable energy power generation have a significant impact on the stable operation and economic management of the power system. Therefore, it is necessary to study effective renewable energy scenario generation methods. Through the scenario generation method, the power system can identify potential supply and demand imbalance risks in advance and take corresponding scheduling measures to ensure the stability and reliability of power supply. At the same time, the scenario generation method helps the power system to more accurately predict the power generation output of renewable energy, thereby optimizing resource allocation, reducing unnecessary reserve capacity investment, and reducing operating costs.
[0003] Generative Adversarial Network (GAN) and its variants have gradually attracted great attention for renewable energy scenario generation. This deep generative model can directly capture the uncertainty and spatiotemporal correlation in historical data through semi-supervised learning without tedious manual annotation. Studies have shown that the generation method based on multi-antigen generative network GAN can reflect the dynamic characteristics of renewable energy output and the rich and diverse patterns it contains, thereby providing more accurate scenario generation results.
[0004] Although the GAN-based scenario generation method can be easily applied to renewable energy output data sets under different resource endowments, how to better utilize deep networks to achieve scenario generation tasks that meet more needs remains a challenging problem. The current GAN-based generation model uses random noise as input, so the pattern of the generated scenario cannot be controlled. The relationship between the latent layer and the output result is unexplainable and weakly correlated. In practice, complete and sufficient historical power data of renewable energy is usually difficult to obtain. Seasonal differences result in some historical data failing to cover uncaptured scenario patterns. However, information such as geography, climate, and local experience can describe some key characteristics of renewable energy resource endowments. Taking the continental monsoon climate in northern China in spring as an example, wind power generation has the characteristics of large average output, obvious peak-to-valley difference, and strong volatility. These interpretable characteristics can be further used as expert prior knowledge that can be used when generating unknown pattern scenarios.
[0005] Conditional GAN (cGAN) can overcome the above problems to a certain extent. However, the conditional features of interest cannot be adjusted continuously and can only be set to relatively rough categories. Due to the limitation of this discrete classification, the model cannot set too many conditions and the pattern preferences available for selection are very limited. If new adjustable features are added during the generation process, the entire GAN model must be retrained. The latent space of existing generative adversarial networks is not interpretable, so the input features of the model cannot directly guide the generation of patterns. Summary of the invention
[0006] The purpose of the present invention is to overcome the shortcomings that the scenario generation method in the prior art cannot meet the scenario generation tasks with multiple requirements, and the potential space of the adversarial network is not interpretable. Therefore, the input features of the model cannot directly guide the generation of the pattern, and it is impossible to generate new scenario patterns that do not exist based on historical data, resulting in reduced operating stability and increased operating costs of the power system. A controllable scenario generation method based on an adversarial network is provided.
[0007] The purpose of the present invention is achieved through the following technical solutions: A controllable scene generation method based on an adversarial network comprises the following steps: Step 1: Collect the historical time series data of renewable energy power generation as a training set, construct a latent space input on the construction manifold to optimize the historical data, and the latent space input includes a control vector c and a random noise vector z with characteristic indicators that can be interpreted physically; Step 2: Generate an adversarial network through iterative training. In each iteration, the generator G, the discriminator D, and the auxiliary network Q are trained in sequence through the back propagation optimization algorithm until the convergence criterion is reached. Step 3: In the process of generating adversarial networks, the control vector c is adjusted along one or more interpretable features to achieve controllable scene generation.
[0008] Preferably, the adversarial network is an explainable adversarial network based on the minimax game theory, specifically: Among them, V represents the loss function, which is used to measure the performance of the generator G and the discriminator D in adversarial training. It represents the ability of the discriminator D to judge the real data x. It represents the ability of the discriminator D to judge the fake data G(z) generated by the generator G. It represents the minimum matching result of the Euclidean distance between the corresponding characteristic index F and the control vector c in the historical real sample, I(c, G(z, c) is the mutual information, and λ is the hyperparameter.
[0009] Preferably, by estimating the posterior distribution Solve the maximization of mutual information I(c, G(z, c); define an auxiliary distribution Approximate the posterior distribution The infimum of is as follows: By KL divergence The lower bound of mutual information can be obtained, and the variational information maximization is shown as follows: According to the corresponding axioms, we can get: From this axiom, we can define the variational infimum L of the mutual information I(c; G(z,c)) I (G)Q), as shown below: Where L I (G)Q) is approximated by a Monte Carlo method and maximized by reparameterizing Q and G.
[0010] Preferably, the auxiliary distribution Q is parameterized as a neural network, so the explainable adversarial network is updated as follows:
[0011] Preferably, the interpretable features include average power, peak power, valley power, average fluctuation power and maximum climbing power.
[0012] Preferably, the renewable energy power generation historical time series data includes wind power or photovoltaic power generation historical time series data.
[0013] The beneficial effects of the present invention are: the adversarial network model of the present invention is data-driven and can adaptively learn the random distribution and implicit dynamic associations inherent in the historical data of renewable energy output. Therefore, the adversarial network model of the present invention can be easily extended to renewable energy output data sets under various resource endowments; the construction of the model network input space contains interpretable features as control vectors on the manifold. These adjustable features have practical physical meanings and can directly reflect the statistical characteristics of the scene pattern; for the controllable generation function of the pattern, the proposed model utilizes mutual information maximization and matching sampling techniques. The generated scene can be intentionally modified and adjusted according to the statistical characteristic preferences, and new scene patterns that do not exist in the historical data can be realized. BRIEF DESCRIPTION OF THE DRAWINGS
[0014] Figure 1 is a flow chart of the present invention; Figure 2 It is a two-dimensional scatter plot composed of each feature of the real sample of the present invention under the relationship between the two; Figure 3 It is a controllable scene sample pattern distribution generated by the k-fold cross-validation method of the present invention. DETAILED DESCRIPTION
[0015] The present invention is further described below in conjunction with the accompanying drawings and embodiments.
[0016] Example: A controllable scene generation method based on adversarial networks, such as Figure 1 As shown, the following steps are included: Step 1: collect historical time series data of wind power or photovoltaic power generation as a training set, construct a latent space input on the construction manifold to optimize the historical data, and the latent space input includes a control vector c and a random noise vector z with characteristic indicators that can be interpreted physically; Step 2: Generate an adversarial network through iterative training. In each iteration, the generator G, the discriminator D, and the auxiliary network Q are trained in sequence through the back propagation optimization algorithm until the convergence criterion is reached. Step 3: In the process of generating adversarial networks, the control vector c is adjusted along one or more interpretable features to achieve controllable scene generation.
[0017] Most typical generative adversarial networks use unsupervised learning methods, which generate corresponding high-dimensional data by inputting potential layer vectors (. The advantage of unsupervised learning is that it can classify disordered data at the machine level, which is very valuable for data analysis. A famous example of an unsupervised learning generative model is the autoencoder, which encodes high-dimensional data by training a neural network, thereby extracting real and effective feature information from the original data in the low-dimensional encoded latent space. Related experiments show that there are usually two good properties in the latent space: (1) most points in the manifold can generate reasonable data; (2) due to the continuity of the manifold, the interpolation between two points on the manifold can still make the output data have a smooth transition. Unsupervised learning methods may be useful for certain specific tasks. Invalid. In the scenario generation process based on unsupervised learning, the selection of scenario mode preferences cannot be achieved due to the lack of interpretability of data representation in the latent space. The characteristics of the generated scenarios cannot be adjusted additionally because the scenario modes only depend on the distribution of existing historical data. However, a wide range of operating scenarios (such as some extreme scenarios) need to be considered in the research of power system planning and operation. Therefore, it is necessary to coordinate expert knowledge to generate scenarios that tend to have certain preferences or generate completely new modes, which requires an interpretable and controllable scenario generation method. The method proposed in this chapter uses time series interpretable characteristic indicators to predescribe the characteristics of the expected generated scenarios, thereby controlling the preferences of the scenario generation mode, and then using domain expert knowledge to generate unknown distributions.
[0018] In order to make the input features interpretable and make full use of the controllability on the manifold in the latent space, the traditional GAN random input vector is divided into two parts: the first part is z, which represents the random noise vector; the second part is c, which represents the control vector composed of a set of interpretable features. These features are selected as the characteristic indicators of the time series scene introduced in Section 2.1. In order to ensure that GAN does not ignore the potential controllable encoding information c on the manifold, referring to existing work, the concept of mutual information (MI) is introduced to evaluate the relationship between the latent vector c and the generated distribution G(z)c). Maximizing the mutual information between the two can reduce the information loss of the control vector c during the generation process. At the same time, the discriminator D needs to distinguish between true and false to ensure that the generated data can be close enough to the real data.
[0019] The adversarial network is an explainable adversarial network based on the minimax game theory, specifically: Among them, V represents the loss function, which is used to measure the performance of the generator G and the discriminator D in adversarial training. It represents the ability of the discriminator D to judge the real data x. It represents the ability of the discriminator D to judge the fake data G(z) generated by the generator G. It represents the minimum matching result of the Euclidean distance between the corresponding characteristic index F and the control vector c in the historical real sample, I(c, G(z, c) is the mutual information, and λ is the hyperparameter.
[0020] Generative adversarial network is a well-known deep generative network training framework based on minimax game theory. Its purpose is to learn the distribution of P with real data. data (x) matches the generator distribution P G (x). Instead of trying to explicitly specify the probability of each random variable x in the data distribution, GANs transform the random noise variables z~P noise (z) is converted into a generated sample G(z) to train the generator network G. The samples generated by the generator G obey the generator distribution P G This generator is trained by competing with its adversarial discriminator network D, which attempts to distinguish the true data distribution P data Samples and generator distribution P G Therefore, for a given generator, the optimal discriminator output satisfies: D(x) = P data (x) / (P data (x)+P G(x)). A well-trained generator network G can fool the discriminator into treating the generated data as real data. The game process is to iteratively train the high-dimensional parameters of the two deep learning networks. Between the generator network G and the discriminator network D, the game eventually reaches a Nash equilibrium.
[0021] By estimating the posterior distribution Solve the maximization of mutual information I(c, G(z, c); define an auxiliary distribution Approximate the posterior distribution The infimum of is as follows: By KL divergence The lower bound of mutual information can be obtained, and the variational information maximization is shown as follows: According to the corresponding axioms, we can get: From this axiom, we can define the variational infimum L of the mutual information I(c; G(z,c)) I (G)Q), as shown below: Where L I (G)Q) is approximated by a Monte Carlo method and maximized by reparameterizing Q and G.
[0022] The auxiliary distribution Q is parameterized as a neural network, so the explainable adversarial network is updated as:
[0023] The control vector c is composed of selected interpretable features. During the training process, since a large amount of input data is required, the noise variables can be obtained in batches from the Gaussian distribution. The value of the control vector can also be assigned by calculating the characteristic index from the historical samples, without manually setting the control vector corresponding to each training sample. After the model training is completed, the generated scene mode can be consciously modified by adjusting the control vector c of the input space of the generator G. The specific training process algorithm of the model is given in the following table: Table 1: Controllable and interpretable adversarial generation network algorithm process The interpretable features include average power, peak power, valley power, average fluctuation power and maximum climbing power.
[0024] The following takes the wind power generation scenario as an example to verify the effectiveness of the controllable generation process of this embodiment.
[0025] During the generation process, each interpretable feature of each control vector, namely the average power, peak power, valley power, average fluctuation power and maximum climbing power, is adjusted in turn, and then the value of each controllable feature is changed from small to large to generate several corresponding scenario curves.
[0026] The comparison between the preset expected values of each characteristic index in the control vector and the actual statistics of the generated samples is shown in the following table. The numerical results show that under the condition of reasonable correlation between the statistical characteristics, the proposed controllable generation model can independently and accurately control the quantitative size of specific statistical characteristics, thereby achieving interpretable generation of specified preference patterns.
[0027] Table 2 Comparison between preset control values of each interpretable feature and actual statistical indicators In addition, in order to verify the control generation ability of the proposed model for unknown scene patterns, the k-fold cross validation method is used to train the model and generate results. The interpretable features can represent the pattern distribution of the reproducible scene. Therefore, the training samples can be based on the historical time series samples {x1,x2,...,x N}, The characteristic index of F1 ,F2,...,FN}, In this embodiment, the k-means clustering algorithm is used to divide the data set into subsets with similar patterns through an unsupervised learning process. Then, the corresponding sequence samples {x1,x2,...,x N} is divided into subsets of nearest neighbor clusters. The sequence scenes in each subset have similar patterns. Then each pattern is selected as the validation set in turn, and the other 4 patterns are used as training sets to train the proposed model. In this way, the simulation of the scene generation task under the new mode can be achieved.
[0028] This embodiment uses k-fold cross validation for training and generation to verify the proposed model's ability to control and generate unknown scene patterns. The two-dimensional scatter plot of each feature of the real sample under the pairwise relationship is as follows: Figure 2 The trained model generates new patterns by presetting the control vector of the target feature. The number of generated scenes is consistent with the number of scene curves included in the validation set. The generation results of the model under 5-fold cross validation are shown in Figure 3 shown.
[0029] By comparison Figure 2 and Figure 3The results show that the pattern morphology of the generated scene is well matched with the pattern morphology of the target scene, but the characteristic distribution cannot be completely consistent. When the cross-validated training data set does not contain the relevant features of the scene to be generated, the controllable scene generation model proposed in this chapter can still generate the scene pattern distribution under the expected preference. The interpretable features of the new target pattern are embedded as control vectors in the controllable model generation process, which can be used to use expert knowledge to generate new pattern scenes when data is insufficient. Since the latent layer on the manifold space can be used to pre-evaluate the statistical characteristics of the generated scene, the model can complete generation tasks that traditional GAN networks cannot do.
[0030] The above-described embodiment is only a preferred solution of the present invention and does not limit the present invention in any form. There are other variations and modifications without exceeding the technical solution described in the claims.
Claims
1. A controllable scene generation method based on adversarial network, characterized in that: The following steps are involved: Step 1: Collect the historical time series data of renewable energy power generation as a training set, construct a latent space input on the construction manifold to optimize the historical data, and the latent space input includes a control vector c and a random noise vector z with characteristic indicators that can be interpreted physically; Step 2: Generate an adversarial network through iterative training. In each iteration, the generator G, the discriminator D, and the auxiliary network Q are trained in sequence through the back propagation optimization algorithm until the convergence criterion is reached. Step 3: In the process of generating adversarial networks, the control vector c is adjusted along one or more interpretable features to achieve controllable scene generation.
2. The controllable scene generation method based on adversarial network according to claim 1 is characterized in that: The adversarial network is an explainable adversarial network based on the minimax game theory, specifically: Among them, V represents the loss function, which is used to measure the performance of the generator G and the discriminator D in adversarial training. It represents the ability of the discriminator D to judge the real data x. It represents the ability of the discriminator D to judge the fake data G(z) generated by the generator G. It represents the minimum matching result of the Euclidean distance between the corresponding characteristic index F and the control vector c in the historical real sample, I(c, G(z, c) is the mutual information, and λ is the hyperparameter.
3. The controllable scene generation method based on adversarial network according to claim 2 is characterized in that: By estimating the posterior distribution Solve the maximization of mutual information I(c, G(z, c); define an auxiliary distribution Approximate the posterior distribution The infimum of is as follows: By KL divergence The lower bound of mutual information can be obtained, and the variational information maximization is shown as follows: According to the corresponding axioms, we can get: From this axiom, we can define the variational infimum L of the mutual information I(c; G(z,c)) I (G,Q), as shown below: Where L I (G,Q) is approximated by Monte Carlo methods and maximized by reparameterizing Q and G.
4. The controllable scene generation method based on adversarial network according to claim 3 is characterized in that: The auxiliary distribution Q is parameterized as a neural network, so the explainable adversarial network is updated as:
5. The controllable scene generation method based on adversarial network according to claim 1 is characterized in that: The interpretable features include average power, peak power, valley power, average fluctuation power and maximum climbing power.
6. The controllable scene generation method based on adversarial network according to claim 1 is characterized in that: The renewable energy power generation historical time series data includes wind power or photovoltaic power generation historical time series data.