Extreme scene generation method, system and device and storage medium

Abstract

The invention discloses an extreme scene generation method, system and device and a storage medium. The method comprises the steps of obtaining historical operation data of a power system; inputting the historical operation data into a preset scene generation model to obtain a first scene sample of the power system; normal distribution of the first scene sample in a hidden variable space of the scene generation model is determined, the normal distribution is iteratively optimized through a KL divergence and importance sampling method to obtain first importance distribution of the extreme scene, and hidden variables are potential feature representation of the extreme scene; and performing iterative optimization on the first importance distribution until an iteration termination condition is met, determining second importance distribution, and determining an extreme scene sample corresponding to the power system based on the second importance distribution. According to the invention, the accuracy and efficiency of power system extreme scene generation can be improved.

Description

Technical Field

[0001] This application relates to the field of power system operation, and particularly to a method, system, device and storage medium for generating extreme scenarios. Background Art

[0002] Extreme scenarios cover various situations that may cause significant impacts on power systems, such as natural disasters (including typhoons, earthquakes, etc.), extreme weather (such as cold snaps, heatwaves, etc.), and large-scale equipment failures. Once these events occur, they will pose serious challenges to the stable operation of power systems and may even trigger large-scale power outages. Therefore, accurately simulating these extreme scenarios is of crucial significance for evaluating the reliability and security of power systems. It can not only help anticipate potential risks in advance but also provide a scientific basis for formulating countermeasures.

[0003] Traditional scenario generation methods mainly rely on historical data and statistical analysis. By statistically modeling historical events, they generate extreme scenarios that may occur in the future. However, when faced with complex and diverse extreme events, traditional methods often prove inadequate. Especially in cases where the data exhibits non-linear characteristics and is in a high-dimensional environment, the limitations of traditional methods are more obvious. In addition, the scenarios generated by traditional methods also have deficiencies in terms of quantity and quality, making it difficult to comprehensively cover all possible extreme situations, which may affect the accurate evaluation of the reliability and security of power systems.

[0004] Application Content

[0005] This application provides a method, system, device and storage medium for generating extreme scenarios, which improves the accuracy and efficiency of generating extreme scenarios for power systems.

[0006] In a first aspect, this application provides a method for generating extreme scenarios, including:

[0007] Obtain the historical operation data of the power system;

[0008] Input the historical operation data into a preset scenario generation model to obtain a first scenario sample of the power system;

[0009] Determine the normal distribution of the first scenario sample in the latent variable space of the scenario generation model, and iteratively optimize the normal distribution through the KL divergence and importance sampling method to obtain the first importance distribution of the extreme scenario, where the latent variable is the potential feature representation of the extreme scenario;

[0010] Iteratively optimize the first importance distribution until the iterative termination condition is met, determine the second importance distribution, and based on the second importance distribution, determine the extreme scenario samples corresponding to the power system, where the second importance distribution includes the mean and variance.

[0011] In the embodiments of the present application, by obtaining the historical operation data of the power system, it is convenient to input it into the subsequent scenario generation model, so that the scenario generation model can learn the operation characteristics and rules of the power system, thereby generating scenario samples that conform to the actual situation and providing a reliable basis for the subsequent steps; by iteratively optimizing the normal distribution through the KL divergence and importance sampling methods, the true distribution approaching the extreme scenario can be gradually obtained, ensuring that the generated extreme scenario is more accurate; by iteratively optimizing the first importance distribution, the characteristics of the extreme scenario can be further refined, ensuring that the simulated extreme scenario is closer to the actual situation. At the same time, by selecting appropriate iterative termination conditions and optimization algorithms, while ensuring accuracy, the consumption of computing resources can be minimized as much as possible, improving the overall efficiency. Compared with the prior art, the present application can improve the accuracy and efficiency of power system extreme scenario generation.

[0012] Further, the inputting the historical operation data into a preset scenario generation model to obtain the first scenario samples of the power system is specifically:

[0013] Input the historical operation data into the causal convolution layer of the scenario generation model to obtain the first time series features;

[0014] Input the first time series features into the dilated convolution unit to obtain the second time series features, and summarize the first time series features and the second time series features through skip connections to obtain the third time series features;

[0015] The third time series features pass through the activation function to predict the noise at the current time, and based on the noise, determine the first scenario samples of the power system.

[0016] In this way, by inputting historical data into the trained model, the first scenario samples that conform to the actual situation can be generated, providing a reliable basis for the subsequent steps.

[0017] Further, the iterative optimization of the normal distribution through the KL divergence and importance sampling methods is specifically:

[0018] Determine the corresponding density function based on the first scenario samples, and determine the likelihood ratio function based on the density function and the normal distribution;

[0019] Estimate the probability expectation value of the occurrence of the extreme scenario based on the likelihood ratio function, and solve the probability expectation value with the goal of minimizing the variance of the probability expectation value to obtain the third importance distribution;

[0020] Iteratively optimize the third importance distribution through the KL divergence to obtain the first importance distribution.

[0021] In this way, through the iterative optimization method of the KL divergence and the importance sampling method, the true distribution of the extreme scenario can be gradually approximated while maximizing the computational efficiency.

[0022] Furthermore, the calculation formula for obtaining the first importance distribution is specifically:

[0023] min h D(g * , h) = min v (∫g(x)ln g(x)μ(dx) - ∫g(x)ln h(x)μ(dx));

[0024]

[0025] In the formula, g * is the third importance distribution, h is the normal distribution, h and f are functions of the same family, that is, multivariate independent Gaussian distributions, denoted as The parameter θ = (μ, σ) is the mean and standard deviation parameters of the normal distribution, and are the mean and variance parameter vectors of the normal distribution h, g(x) is the importance distribution function of the xth power system scenario; u is the measure function of the occurrence probability of the scenario random variable x; H(x) is the discriminant function for whether the xth power system scenario is a set extreme scenario; v k and v k-1 are the first importance distributions of the kth and (k - 1)th iterations respectively; L(X i ) is the likelihood ratio function, X i is the ith scenario sample sampled from the power system scenario distribution ; N is the total number of samples.

[0026] Furthermore, the first importance distribution is a mixture Gaussian model, and the iterative formula of the KL divergence is specifically:

[0027]

[0028] In the formula, is the random variable that the ith sample comes from the mth component of the mixture Gaussian model, y j is the ith sample; μ m is the mean parameter of the mth component of the mixture Gaussian model; is the label of the Gaussian function; Xi is the scenario sample obtained by sampling from the distribution of the i-th power system scenario in; α m is the weight of the m-th Gaussian distribution in the Gaussian mixture model; θ m is the parameter of the m-th component of the Gaussian mixture model; M is the number of components of the Gaussian mixture model; N is the total number of samples; N m is a temporary variable; L(X i ) is the likelihood ratio function; H(X i ) is the discriminant function; and are the parameters of the updated Gaussian mixture model.

[0029] Further, the iterative optimization of the first importance distribution is specifically as follows:

[0030] Obtain the initial mean and initial variance of the first importance distribution, and use maximum likelihood estimation to iteratively update the initial mean;

[0031] Calculate the Euclidean distance between the means of two iterations, determine the inflation coefficient according to the Euclidean distance and the preset distribution dimension, and update the initial variance based on the inflation coefficient.

[0032] In this way, by iteratively optimizing the first importance distribution, the characteristics of extreme scenarios can be further refined to ensure that the simulated extreme scenarios are closer to the actual situation.

[0033] Further, the calculation formula for the iterative optimization of the first importance distribution is specifically as follows:

[0034] d = |μ (k+1) - μ (k) |;

[0035]

[0036] S(k + 1) = δS(k);

[0037] In the formula, d is the Euclidean distance; μ (k+1) and μ (k) are the means of the first importance distribution in the (k + 1)-th iteration and the k-th iteration respectively; δ is the inflation coefficient; τ is the speed-up factor; J is the Gaussian distribution dimension, is a constant introduced to avoid the curse of dimensionality; S(k + 1) and S(k) are the variances in the (k + 1)-th iteration and the k-th iteration respectively.

[0038] In a second aspect, the present application provides an extreme scenario generation system, including: an acquisition module, a generation module, a first optimization module, and a second optimization module;

[0039] The obtaining module is configured to obtain historical operation data of the power system;

[0040] The generating module is configured to input the historical operation data into a preset scenario generation model to obtain a first scenario sample of the power system;

[0041] The first optimization module is configured to determine the normal distribution of the first scenario sample in the latent variable space of the scenario generation model, and iteratively optimize the normal distribution by using the KL divergence and importance sampling method to obtain a first importance distribution of the extreme scenario, where the latent variable is a potential feature representation of the extreme scenario;

[0042] The second optimization module is configured to iteratively optimize the first importance distribution until an iteration termination condition is met, determine a second importance distribution, and determine an extreme scenario sample corresponding to the power system based on the second importance distribution, where the second importance distribution includes a mean and a variance.

[0043] In the embodiment of the present application, by obtaining the historical operation data of the power system, it is convenient to input it into the scenario generation model subsequently, so that the scenario generation model can learn the operation characteristics and rules of the power system, thereby generating scenario samples that conform to the actual situation and providing a reliable basis for subsequent steps; by iteratively optimizing the normal distribution through the KL divergence and importance sampling method, the true distribution approaching the extreme scenario can be gradually obtained, ensuring that the generated extreme scenario is more accurate; by iteratively optimizing the first importance distribution, the characteristics of the extreme scenario can be further refined, ensuring that the simulated extreme scenario is closer to the actual situation. At the same time, by selecting appropriate iteration termination conditions and optimization algorithms, while ensuring accuracy, the consumption of computing resources can be minimized as much as possible, improving the overall efficiency. Compared with the prior art, the present application can improve the accuracy and efficiency of generating extreme scenarios of the power system.

[0044] In a third aspect, the present application further provides a terminal device, which includes: one or more processors; a memory coupled to the processors for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the extreme scenario generation method as described in the present application.

[0045] In a fourth aspect, the present application further provides a computer-readable storage medium, on which a computer program is stored, and the computer program, when executed by a processor, implements the extreme scenario generation method as described in the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a schematic flowchart of an embodiment of the extreme scenario generation method provided by the present application;

[0047] Figure 2 It is a schematic diagram of the network structure of the WaveNet model provided by this application;

[0048] Figure 3 It is a schematic diagram of the training process of the DiffWave model provided by this application;

[0049] Figure 4 It is a schematic diagram of the forward diffusion process of the 96-point daily wind power output curve provided by this application;

[0050] Figure 5 It is a schematic diagram of the historical wind power curve data sample provided by this application;

[0051] Figure 6 It is a schematic diagram of the wind power curve scenario generated from DiffWave provided by this application;

[0052] Figure 7 It is the iterative process of the KL divergence importance sampling method for two extreme scenarios provided by this application;

[0053] Figure 8 It is the generated extreme scenario 1 sample sampled from DiffWave after the sample ratio boosting algorithm provided by this application;

[0054] Figure 9 It is the historical wind power curve data sample of extreme scenario 1 provided by this application;

[0055] Figure 10 It is the generated extreme scenario 2 sample sampled from DiffWave after the sample ratio boosting algorithm provided by this application;

[0056] Figure 11 It is the historical wind power curve data sample of extreme scenario 2 provided by this application;

[0057] Figure 12 It is a schematic diagram of the structure of an embodiment of the extreme scenario generation system provided by this application;

[0058] Figure 13 It is a schematic diagram of the structure of an embodiment of the terminal device provided by this application. Detailed implementation manners

[0059] Next, the technical solutions in the embodiments of this application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of this application. Obviously, the described embodiments are only a part of the embodiments of this application, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in this application without making creative efforts shall fall within the scope of protection of this application.

[0060] It should be understood that the step numbers used in the text are only for convenience of description and do not limit the order of execution of the steps.

[0061] It should be understood that the terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the specification of the present invention and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include the plural forms.

[0062] The terms "comprising" and "including" indicate the presence of the described features, wholes, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or their combinations.

[0063] The term "and / or" refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0064] Extreme scenarios, including natural disasters (typhoons, earthquakes), extreme weather (cold snaps, heatwaves), and large-scale equipment failures, pose serious challenges to the stable operation of power systems. Therefore, accurately simulating these scenarios is crucial for evaluating the reliability and security of power systems. Traditional scenario generation methods rely on historical data and statistical analysis to generate possible future extreme scenarios through statistical modeling. However, in the face of complex and diverse extreme events, traditional methods have obvious limitations. In addition, the scenarios generated by traditional methods are insufficient in quantity and quality, making it difficult to comprehensively cover all possible extreme situations and affecting the accurate evaluation of the reliability and security of power systems.

[0065] Next, the nouns involved in this application are analyzed:

[0066] The Kullback-Leibler Divergence is an index for measuring the difference between two probability distributions. By minimizing the KL divergence between the generated samples and the real sample distribution, the generated scenarios can be made closer to the real situation.

[0067] Importance sampling is a method of sampling from a probability distribution. By adjusting the sample generation strategy, the number of samples of rare events can be increased, thereby better simulating extreme scenarios.

[0068] Based on this, the embodiments of this application provide an extreme scenario generation method, system, device and storage medium, which can improve the accuracy and efficiency of generating extreme scenarios for power systems.

[0069] A method, system, device, and storage medium for generating extreme scenarios provided by an embodiment of the present application will be specifically described through the following embodiments. First, a method for generating extreme scenarios in an embodiment of the present application will be described.

[0070] A method for generating extreme scenarios provided by an embodiment of the present application relates to the field of power system operation. The method for generating extreme scenarios provided by an embodiment of the present application can be applied to a terminal, a server, or software running on a terminal or a server. In some embodiments, the terminal can be a smart phone, a tablet computer, a notebook computer, a desktop computer, etc.; the server can be configured as an independent physical server, a server cluster or a distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms; the software can be an application implementing a method for generating extreme scenarios, etc., but is not limited to the above forms.

[0071] The present application can also be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multi-processor systems, microprocessor-based systems, set-top boxes, programmable consumer electronic devices, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, and so on. The present application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. The present application can also be practiced in a distributed computing environment where tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media including storage devices.

[0072] Embodiment 1

[0073] Please refer to Figure 1 , Figure 1 , which is a schematic flowchart of an embodiment of the method for generating extreme scenarios provided by the present application, including steps S101 to S104;

[0074] Step S101, obtain historical operation data of the power system;

[0075] In some embodiments, historical operation data of the power system is obtained from the database of the power company, the smart grid system, or a publicly available power dataset, where the historical operation data includes, but is not limited to, historical weather data, wind power output curves, photovoltaic power output curves, and load curve data.

[0076] In some embodiments, after obtaining the historical operation data, data preprocessing needs to be performed on the historical operation data, where the data preprocessing includes, but is not limited to, data cleaning, outlier removal, missing value filling, data normalization, standardization processing, etc., to ensure the quality and consistency of the input data and facilitate subsequent input into the model for analysis.

[0077] Step S102, input the historical operation data into a preset scenario generation model to obtain a first scenario sample of the power system;

[0078] In some embodiments, after obtaining the preprocessed historical operation data, the historical operation data needs to be input into a preset scenario generation model. The schematic diagram of the network structure of the WaveNet model is as Figure 2 shown. WaveNet introduces a causal convolutional layer and an extended convolutional unit, and the extended convolutional unit is composed of a single-layer extended convolutional layer, a gated activation unit, and a residual connection.

[0079] In some embodiments, the input historical operation data first passes through the causal convolutional layer of the scenario generation model to obtain a first temporal feature; subsequently, the first temporal feature is input into the dilated convolutional unit to obtain a second temporal feature, and the first temporal feature and the second temporal feature are aggregated through a skip connection to obtain a third temporal feature; finally, the third temporal feature passes through an activation function to predict the noise at the current time, and based on the noise, a first scenario sample of the power system is determined.

[0080] In some embodiments, the predicted noise is passed through an inverse diffusion process to generate a first scenario sample of the power system.

[0081] It should be noted that after the first temporal feature passes through the first extended convolutional layer and through the first gated activation unit, it is directly connected to the historical operation data and passed to the next extended convolutional unit. When all the gated activation units corresponding to the extended convolutional layers are completed, the second temporal data is output.

[0082] It should be noted that to determine the first scenario sample of the power system based on the noise, prediction needs to be continuously repeated until t = 0, and finally, the first scenario sample of the power system generated by DiffWave is obtained, where the first scenario sample includes possible states of the power system under different operating conditions, including load, power generation, faults, etc.

[0083] It should be noted that the scenario generation model can be, but is not limited to, the DiffWave model. The DiffWave model is a model pre-trained based on power system scenario samples, and its model parameters are WaveNet neural network parameters, which are obtained through the standard diffusion model training method. Among them, the schematic diagram of the training process of the DiffWave model is as Figure 3 shown. The training uses the Adam optimizer, sets the learning rate to 0.001, trains for 5000 rounds, and the training batch size is 2048. In the figure, both the training and test errors gradually decrease with the number of training rounds, indicating that the model is well-trained and there is no overfitting phenomenon. The specific training process is as follows: First, the diffusion model gradually adds Gaussian noise to the power system scenario sample x0 through the forward diffusion process, mapping the target sample x0 to the standard normal distribution space The formula for the forward diffusion process is: In the formula, x t and x t-1 are the power system scenario samples after t and t - 1 diffusions respectively; α t and β t are both pre-set constants, α t = 1 - β t , β t can take values from 0.0001 to 0.002 as t increases; t and T are both the steps of the forward diffusion; ∈~N(0, I) is the random noise introduced by the forward diffusion process to the sample; z θ (x t , t) is the noise added during the prediction process of the DiffWave model; q is the conditional probability notation; the power system scenario sample x t satisfies In the formula, When t → ∞, x t will be equivalent to the isotropic Gaussian distribution and has nothing to do with the original power system scenario x0, x ∞ ~N(0, I). Subsequently, reverse diffusion is required. The process of recovering the previous-step sample x t from x t-1 is to learn a neural network z θ (x t , t) to predict the noise z t added during the forward process, so as to calculate q(x t-1 |x t ) to approximate the above forward process q(x t-1 |x t ), and recover x t from x t-1 . The formula for the reverse diffusion process is: In the formula, x t and xt-1 are the power system scenario samples after t and t-1 diffusions respectively; z θ (x t ,t) is the noise added in the prediction process of the DiffWave model. The noise z is optimized by the neural network. t The prediction error of ||z t -z θ (x t ,t)||2, the DiffWave model can be trained. Among them, the above is the generation process of the most typical probability denoising diffusion model (Denoising Diffusion Probabilistic Model, DDPM) in the diffusion model. In addition, there are many diffusion model variants such as the implicit denoising diffusion model (Denoising Diffusion Implicit Model, DDIM), which are not limited in this application.

[0084] In this way, by inputting historical data into the trained model, a first scenario sample that matches the actual situation can be generated, providing a reliable foundation for subsequent steps.

[0085] Step S103, determining the normal distribution of the first scene sample in the latent variable space of the scene generation model, and iteratively optimizing the normal distribution by KL divergence and importance sampling method to obtain a first importance distribution of the extreme scene, wherein the latent variable is a potential feature representation of the extreme scene;

[0086] In some embodiments, in the scene generation model, the first scene sample can be mapped to the latent variable space through the encoder of the scene generation model, that is, the features of the first scene sample are represented as points in the latent variable space, and the distribution of these points can be approximated as a normal distribution.

[0087] It should be noted that latent variables refer to hidden variables or potential factors that are not directly observed in the model but affect the observed data. In the scene generation model, latent variables can be regarded as potential feature representations of scene samples.

[0088] In some embodiments, after determining the normal distribution of the first scene sample and the latent variable, in order to obtain the first importance distribution of the extreme scene, the normal distribution of the latent variable space can be iteratively optimized using the KL divergence (Kullback-Leibler divergence) and importance sampling method.

[0089] In some embodiments, the normal distribution is iteratively optimized by the KL divergence and importance sampling method, specifically: determining a corresponding density function based on the first scenario samples, and determining a likelihood ratio function based on the density function and the normal distribution; estimating the expected value of the probability of the occurrence of an extreme scenario based on the likelihood ratio function, and solving the expected value of the probability with the goal of minimizing the variance of the expected value of the probability to obtain a third importance distribution; and iteratively optimizing the third importance distribution by the KL divergence to obtain a first importance distribution.

[0090] In some embodiments, determining a corresponding density function based on the first scenario samples, and determining a likelihood ratio function based on the density function and the normal distribution, specifically: after determining the first scenario samples, the density function f(x) of the random variable x of the power system operation scenario under the measure function μ can be determined, and then, based on the density function f(x) and the normal distribution g(x), the likelihood ratio function L(x) is determined, where the formula of the likelihood ratio function is: In the formula, L(x) is the likelihood ratio function, f(x) is the density function, and g(x) is the normal distribution of the latent variable.

[0091] In some embodiments, estimating the expected value of the probability of the occurrence of an extreme scenario based on the likelihood ratio function, and solving the expected value of the probability with the goal of minimizing the variance of the expected value of the probability to obtain a third importance distribution, specifically: estimating the expected value of the probability of the occurrence of an extreme scenario based on the likelihood ratio function L(x) where the expected value of the probability has the following related formula: In the formula, l is the expected value of the probability, H(x) is the discriminant function for whether the x-th power system scenario is a set extreme scenario, f(x) is the density function, g(x) is the normal distribution of the latent variable, μ is the measure function of the occurrence probability of the scenario random variable x, and L(x) is the likelihood ratio function. After determining the expected value of the probability then, with the goal of minimizing the variance of the expected value of the probability the expected value of the probability is solved, that is In the formula, f(x) is the density function, g(x) is the normal distribution of the latent variable, and H(x) is the discriminant function for whether the x-th power system scenario is a set extreme scenario; this process can minimize the variance of the expected value of the probability by selecting an appropriate g, and thus solving can obtain the third importance distribution g * and its related formula is: In the formula, g * (x) is the importance distribution function, g *is the third importance distribution, i.e., the optimal solution; H(x) is a discriminant function for whether the x-th power system scenario is a set extreme scenario; f(x) is a density function; μ is a measure function of the occurrence probability of the scenario random variable x; l is the probability expectation value; thus, the probability expectation value of the occurrence of extreme scenarios is estimated by sampling from this distribution of the variance Var X～g (l) = 0.

[0092] In some embodiments, the third importance distribution is iteratively optimized through KL divergence to obtain the first importance distribution. Specifically, since the third importance distribution g * cannot be directly solved, an iterative method can only be used to find a new probability distribution h (i.e., a normal distribution) of the power system scenario to approximately replace g * , and the KL divergence between the normal distribution h and g * is minimized. The relevant formula is: In the formula, h is a normal distribution, g * is the third importance distribution, D(g * , h) represents calculating the KL divergence between the probability distributions g * and h. g(x) is the importance distribution function, and h(x) is the probability of the sample x appearing under the normal distribution h; μ is the measure function of the occurrence probability of the scenario random variable x. Since the problem under study is a multivariate independent Gaussian distribution, for simplicity, it is assumed that h and f are functions of the same family, i.e., a multivariate independent Gaussian distribution, denoted as The parameter θ = (μ, σ) is the mean and standard deviation parameters of the normal distribution, and are the mean and variance parameter vectors of the normal distribution h, Therefore, the original problem is transformed into: min h D(g * , h) = min v (∫g(x)ln g(x)μ(dx) - ∫g(x)ln h(x)μ(dx)), where the first term is independent of h, so only the maximum value of the second term (and ignoring the constant term) needs to be found: that is, find From the perspective of iteration, at this time, the previous v k-1 has been obtained. Similarly, using Monte Carlo simulation, sampling and calculating from the power system scenario distribution : Subsequently, by setting the first derivative to 0, the next v k is updated. Since is a multivariate independent Gaussian distribution, the KL divergence and the above optimization process can be solved through very simple explicit expressions: In the formula, X i is the i-th sampling from the power system scenario distribution Scene samples obtained by medium sampling; N is the total number of samplings; and are the mean and variance parameter vectors of the normal distribution h.

[0093] It should be noted that in the initial iteration, it is very likely that, like direct Monte Carlo simulation, no samples falling into the small sample subspace Z sub can be obtained. At this time, the subset simulation method can be used for reference. By constructing a series of nested subsets Z i , i = 1, 2,..., K, where and Z0 = Z sub , Z k = Z. As the number of iterations increases, the subset is gradually shrunk until it converges to the target subset.

[0094] In some embodiments, the calculation formula for obtaining the first importance distribution is specifically:

[0095] min h D(g * , h) = min v (∫g(x)ln g(x)μ(dx) - ∫g(x)ln h(x)μ(dx));

[0096]

[0097] In the formula, g * is the third importance distribution, h is the normal distribution, h and f are functions of the same family, that is, multivariate independent Gaussian distributions, denoted as The parameter θ = (μ, σ) is the mean and standard deviation parameters of the normal distribution, and are the mean and variance parameter vectors of the normal distribution h, g(x) is the importance distribution function of the xth power system scenario; u is the measure function of the occurrence probability of the scenario random variable x; H(x) is the discriminant function for whether the xth power system scenario is a set extreme scenario; v k and v k-1 are the first importance distributions of the kth and (k - 1)th iterations respectively; L(X i ) is the likelihood ratio function, X i is the scenario sample obtained by sampling from the ith power system scenario distribution ; N is the total number of samplings.

[0098] In some embodiments, the above method can be extended to a Gaussian Mixture Model (GMM), that is, when the first importance distribution is a Gaussian mixture model, the first importance distribution is a Gaussian mixture model, and the iterative formula of the KL divergence is specifically:

[0099]

[0100] where is a random variable indicating that the i-th sample comes from the m-th component of the Gaussian mixture model, y j is the i-th sample; μ m is the mean parameter of the m-th component of the Gaussian mixture model; is the label of the Gaussian function; X i is the i-th scenario sample sampled from the power system scenario distribution φ(x; v k-1 ); α m is the weight of the m-th Gaussian distribution in the Gaussian mixture model; θ m is the parameter of the m-th component of the Gaussian mixture model; M is the number of components of the Gaussian mixture model; N is the total number of samplings; N m is a temporary variable; L(X i ) is the likelihood ratio function; H(X i ) is the discriminant function; and are the parameters of the updated Gaussian mixture model.

[0101] It should be noted that the main problem faced by the original importance function KL divergence iterative algorithm is that the calculation of |Σ -1 / 2 | in the Gaussian mixture function involves a product, and in high dimensions (for example, in a 50-step diffusion model generated by a 96-point daily curve, the number of variables can reach 4800), Figure 3 Figure 395 is a schematic diagram of the forward diffusion process of the 96-point daily wind power output curve. This calculation step is very likely to exceed the range of common 64-bit / 32-bit floating-point numbers (10-308 to 10308).

[0102] It should be noted that in common deep learning generative models, the dimensions in the latent variable sampling space are usually independent of each other. Therefore, there are conditions for decoupling and splitting each dimension for iteration. Derive the parameter update expression of the decoupled iterative method for the KL divergence of this importance function: where and are the mean and variance parameter vectors of the normal distribution h respectively, j is the dimension, N is the total number of samplings, L j is the likelihood ratio function considering only the j-th dimension, H(x) is the discriminant function for whether the x-th power system scenario is a set extreme scenario, and X ijis the component of the j-th dimension of the i-th normal distribution. Under the standard normal distribution studied: The above algorithm does not need to calculate the product of consecutive multiplications during iteration, and under the assumption of the prototype of the importance function of the single-component independent multivariate Gaussian distribution, it is theoretically close to the result of the original algorithm.

[0103] In this way, through the KL divergence and the importance sampling method for iterative optimization, the true distribution of the extreme scenario can be gradually approximated while maximizing the computational efficiency.

[0104] Step S104: Iteratively optimize the first importance distribution until the iteration termination condition is met, determine the second importance distribution, and based on the second importance distribution, determine the extreme scenario samples corresponding to the power system, where the second importance distribution includes the mean and variance.

[0105] In some embodiments, the iterative optimization of the first importance distribution is specifically as follows: Obtain the initial mean and initial variance of the first importance distribution, and use the maximum likelihood estimation to iteratively update the initial mean; calculate the Euclidean distance between the means of two consecutive iterations, determine the inflation coefficient according to the Euclidean distance and the preset distribution dimension, and update the initial variance based on the inflation coefficient. Specifically: First, after determining the first importance distribution, it is necessary to first calculate the initial mean μ 0 and the initial variance S 0 . Then, the sample proportion improvement method that estimates by maximizing the likelihood function is used to further improve the parameters of the first importance distribution. During each iteration, continuously calculate and update the mean μ (k+1) and μ (k) corresponding to the current sample set X. Second, after determining the means of two consecutive iterations, calculate the Euclidean distance d between the means of two consecutive iterations. The relevant formula is: d = |μ (k+1) - μ (k) |, where d is the Euclidean distance; μ (k+1) and μ (k) are the means of the (k + 1)-th iteration and the k-th iteration, respectively. When iterating the variance, to avoid the variance parameter S of the first importance distribution collapsing to 0 during the update process, which would cause the loss of diversity in sampling the extreme scenario, a certain proportion of increment δ (inflation coefficient) is given to the variance parameter during iteration. The inflation coefficient is taken as a certain multiple of the Euclidean distance between the means of two consecutive iterations. The relevant calculation formula is: where δ is the inflation coefficient; τ is the speed factor; J is the dimension of the Gaussian distribution, Constants introduced to avoid the curse of dimensionality. Finally, after determining the expansion coefficient δ, the variance S(k + 1) can be iteratively updated. The variance S(k + 1) of the next sampling is the product of the variance S(k) of the previous sample and the expansion coefficient δ. The relevant calculation formula is: S(k + 1) = δS(k), where δ is the expansion coefficient, and S(k + 1) and S(k) are the variances of the (k + 1)-th iteration and the k-th iteration, respectively.

[0106] In some embodiments, the method for estimating the sample proportion improvement by maximizing the likelihood function further improves the parameters of the first importance distribution. Specifically, by continuously generating samples that meet the target constraint conditions from the first importance distribution and using the maximum likelihood estimation method to update the parameters of the distribution h, the relevant calculation formula is: where N is the total number of samplings, H(X i ) is the discriminant function for whether the X i -th power system scenario is a set extreme scenario, and φ(X i ; v) is the expression of the normal distribution h function. By continuously iteratively updating the mean through the above method until the final mean parameter converges or the target sample proportion reaches the set value, the iteration terminates at this time. Since as the iteration progresses, the central mean point of the distribution will gradually stabilize at the arithmetic average center of the sample region, the target convergence condition is that the absolute value of the difference between the means of two adjacent iterations |μ (k+1) - μ (k) | ∞ < ε, where μ (k+1) and μ (k) are the means of the (k + 1)-th iteration and the k-th iteration, respectively, and ε is a preset threshold.

[0107] In some embodiments, the calculation formula for iteratively optimizing the first importance distribution is specifically:

[0108] d = |μ (k+1) - μ (k) |;

[0109]

[0110] S(k + 1) = δS(k);

[0111] where d is the Euclidean distance; μ (k+1) and μ (k) are the means of the first importance distribution of the (k + 1)-th iteration and the k-th iteration, respectively; δ is the expansion coefficient; τ is the speed-up factor; J is the dimension of the Gaussian distribution, is a constant introduced to avoid the curse of dimensionality; S(k + 1) and S(k) are the variances of the (k + 1)-th iteration and the k-th iteration, respectively.

[0112] It should be noted that since the goal of the importance function iteration method is to find the optimal importance function, when the iteration finally terminates, the proportion of target samples in the generated samples may still not be very high. Therefore, it is still necessary to further increase the proportion of generated samples through this step.

[0113] By iteratively optimizing the first importance distribution in this way, the characteristics of extreme scenarios can be further refined to ensure that the simulated extreme scenarios are closer to the actual situation.

[0114] In some embodiments, after the mean and variance are iteratively updated, the second importance distribution is determined. At this time, only sampling from the second importance distribution is needed to obtain extreme scenario samples.

[0115] It should be noted that extreme scenario samples refer to high-risk and severe scenarios in the power system that may cause system instability, power flow over-limit, curtailment of wind and solar power, etc.

[0116] To verify the effectiveness of this method, a case study selects the 15-minute resolution data of a provincial wind power output in 2019 for analysis. However, this method is not limited to generating wind power extreme scenarios, and can also be used to generate photovoltaic, load, hydropower, and combined wind-solar-hydro-load extreme scenarios. The programming language is Python 3.10, the deep learning framework is Pytorch 2.0, and the computer hardware configuration is: NVIDIA Geforce RTX 3070Ti Laptop with a video memory capacity of 8GB. The case study generates a curve with a 1-hour resolution for 7 days a week, with a total length of 168. The WaveNet model sets the number of channels in the residual layer to 32, the number to 10, and the residual loop base number to 10; the Diffusion model uses linear diffusion, with α set from 0.0001 to 0.08 and 50 diffusion steps. In the algorithm, the dimension of the standard Gaussian distribution is 168×50 = 8400, attached Figure 4 Figure 15 shows a schematic diagram of the forward diffusion process of the 96-point daily wind power output curve, sampling and observing once every 10 steps. The last one is a bar chart of the data distribution after the diffusion of 128 samples, and an orange standard normal distribution curve is plotted, indicating that the data approximately follows the standard normal distribution after diffusion.

[0117] Figure 5 Figure 19 is a schematic diagram of the historical wind power curve data samples, Figure 6 Figure 21 is a schematic diagram of the wind power curve scenario generated from DiffWave; from the two pictures, it can be seen that the data samples of the two pictures are relatively close in characteristics, so it is proved that DiffWave can generate realistic scenarios after training.

[0118] Based on the above parameters, the proposed extreme scenario sampling method is tested below. Two extreme scenarios are set in the numerical example: 1) Continuous low output scenario: Denote the scenario where the output is continuously lower than 0.2 p.u. within a week as the continuous low output scenario; 2) High volatility scenario: Denote the scenario where the output fluctuation exceeds 0.7 p.u. within a week as the high volatility scenario.

[0119] It should be noted that the definitions of the above extreme scenarios can be arbitrarily selected theoretically. The proposed algorithm can adapt to most extreme scenario definitions by appropriately modifying H(x), not limited to the above two.

[0120] The probabilities of the above two extreme scenarios occurring in the original data and the samples randomly generated by the trained DiffWave model are measured and shown in Table 1 respectively. Among them, since the selected scenarios are all small probability scenarios, in order to measure the scenario occurrence probability as accurately as possible, sampling is repeated 10 times, and at the same time, the number of samples in a single sampling is controlled to be the same as the number of historical scenarios. As can be seen from the table, whether in the historical data or the generated scenarios, the probability of the selected extreme scenarios does not exceed 5%, which is a small probability event of the data.

[0121] Table 1 Probabilities of two extreme scenarios occurring in historical and generated data

[0122]

[0123]

[0124] The above two scenarios are evaluated and sampled respectively using the aforementioned KL divergence importance sampling method and the sample ratio improvement method. The initial threshold of Scenario 1 is set to 0.5, the termination threshold is 0.2, and it is reduced by 0.05 in each iteration; the initial threshold of Scenario 2 is set to 0.4, the termination threshold is set to 0.7, and it is increased by 0.05 in each iteration. The number of samples for single importance sampling is 2048, and the number of samples for single sample ratio improvement algorithm sampling is 1024. Attached Figure 7 Shows the change of the target value -lg|E x～g H(X)L(X; u, v )lnf(X; v)| during the iterative process of the KL divergence importance sampling method for two extreme scenarios. Under the learning rate setting of 0.001, the algorithm has basically converged after the threshold reaches the target value in the 6th iteration. Although the objective function value still has small fluctuations thereafter, it still basically remains at the optimal value level.

[0125] Figure 8 Is the generated extreme scenario 1 sample obtained by sampling from DiffWave after the sample ratio improvement algorithm provided in this application, Figure 9It is the historical wind power curve data sample of Extreme Scenario 1 provided by this application. Since the extreme scenario samples are few, the diffusion model is prone to show certain underfitting characteristics during learning. Attached Figure 8 Some of the generated scenarios are a bit abrupt.

[0126] Figure 10 It is the generated extreme scenario 2 sample obtained by sampling from DiffWave after the sample ratio enhancement algorithm provided by this application, Figure 11 It is the historical wind power curve data sample of Extreme Scenario 2 provided by this application. Compared with Extreme Scenario 1, the occurrence probability of Extreme Scenario 2 in the historical data is about 5%, which is about 2 times higher. Thus, the generated scenarios obtained by sampling from DiffWave after sample ratio enhancement are significantly more consistent with the historical extreme scenarios, more realistic, and basically do not show the phenomenon of underfitting.

[0127] In the embodiment of this application, by obtaining the historical operation data of the power system, it is convenient to input it into the scenario generation model subsequently, so that the scenario generation model can learn the operation characteristics and rules of the power system, thereby generating scenario samples that conform to the actual situation and providing a reliable basis for subsequent steps; by iteratively optimizing the normal distribution through the KL divergence and importance sampling methods, the true distribution approaching the extreme scenario can be gradually obtained to ensure that the generated extreme scenario is more accurate; by iteratively optimizing the first importance distribution, the characteristics of the extreme scenario can be further refined to ensure that the simulated extreme scenario is closer to the actual situation. At the same time, by selecting appropriate iteration termination conditions and optimization algorithms, while ensuring accuracy, the consumption of computing resources can be minimized and the overall efficiency can be improved. Compared with the prior art, this application can improve the accuracy and efficiency of generating extreme scenarios of the power system.

[0128] Embodiment Two

[0129] Please refer to Figure 12 , Figure 12 It is a schematic structural diagram of an embodiment of the extreme scenario generation system provided by this application, including: an acquisition module 100, a generation module 200, a first optimization module 300, and a second optimization module 400;

[0130] The acquisition module 100 is used to acquire the historical operation data of the power system;

[0131] The generation module 200 is used to input the historical operation data into a preset scenario generation model to obtain the first scenario sample of the power system;

[0132] The first optimization module 300 is configured to determine the normal distribution of the first scenario sample in the latent variable space of the scenario generation model, and iteratively optimize the normal distribution through the KL divergence and importance sampling method to obtain the first importance distribution of the extreme scenario, where the latent variable is the potential feature representation of the extreme scenario;

[0133] The second optimization module 400 is configured to iteratively optimize the first importance distribution until the iteration termination condition is met, determine the second importance distribution, and determine the extreme scenario sample corresponding to the power system based on the second importance distribution, where the second importance distribution includes a mean and a variance.

[0134] The information interaction, execution process, etc. between the modules in the above extreme scenario generation system are based on the same concept as the embodiments of the extreme scenario generation method in the first aspect of the present invention, and the achieved technical effects are basically the same. For specific content, reference can be made to the description in Embodiment 1 of the method of the present invention, which will not be elaborated here.

[0135] The device embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separated, that is, they may be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the method of this embodiment.

[0136] Please refer to Figure 13 , Figure 13 which schematically shows the hardware structure of a terminal device in another embodiment. The terminal device includes:

[0137] A processor 1301, which can be implemented in a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, etc., and is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present application;

[0138] A memory 1302, which can be implemented in the form of a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM), etc. The memory 1302 can store an operating system and other application programs. When implementing the technical solutions provided in the embodiments of this specification through software or firmware, the relevant program codes are stored in the memory 1302 and are called by the processor 1301 to execute the large model-based conversation risk assessment method of the embodiments of the present application;

[0139] An input / output interface 1303 for implementing information input and output;

[0140] A communication interface 1304 for implementing communication interaction between this device and other devices, which can achieve communication through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.);

[0141] A bus 1305 for transmitting information between various components of the device (such as a processor 1301, a memory 1302, an input / output interface 1303, and a communication interface 1304);

[0142] Among them, the processor 1301, the memory 1302, the input / output interface 1303, and the communication interface 1304 achieve communication connections with each other inside the device through the bus 1305.

[0143] The present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements an extreme scenario generation method as described in the first embodiment above.

[0144] Those of ordinary skill in the art can understand that all or part of the processes in implementing the above-described embodiment methods can be completed by instructing relevant hardware through a computer program. The program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the above-described method embodiments. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM), etc.

[0145] The above-described specific embodiments have further elaborated on the purpose, technical solutions, and beneficial effects of the present application. It should be understood that the above is only the specific embodiments of the present application and is not used to limit the protection scope of the present application.

[0146] It is particularly pointed out that for those skilled in the art, any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for generating an extreme scene, characterized in that: include: Obtain historical operation data of the power system; Inputting the historical operation data into a preset scenario generation model to obtain a first scenario sample of the power system; Determine the normal distribution of the first scene sample in the latent variable space of the scene generation model, and iteratively optimize the normal distribution by KL divergence and importance sampling method to obtain a first importance distribution of the extreme scene, wherein the latent variable is a potential feature representation of the extreme scene; The first importance distribution is iteratively optimized until an iteration termination condition is met, a second importance distribution is determined, and based on the second importance distribution, extreme scenario samples corresponding to the power system are determined, wherein the second importance distribution includes a mean and a variance.

2. The extreme scenario generation method according to claim 1, characterized in that: The inputting of the historical operation data into a preset scenario generation model to obtain a first scenario sample of the power system is specifically as follows: Inputting the historical operation data into the causal convolution layer of the scenario generation model to obtain a first time series feature; Inputting the first time series feature into a dilated convolution unit to obtain a second time series feature, and aggregating the first time series feature and the second time series feature through a skip connection to obtain a third time series feature; The third time series feature is activated through a function to predict the noise at the current time, and a first scene sample of the power system is determined based on the noise.

3. The extreme scenario generation method according to claim 1, characterized in that: The normal distribution is iteratively optimized by the KL divergence and importance sampling method, specifically: Determining a corresponding density function based on the first scene sample, and determining a likelihood ratio function based on the density function and the normal distribution; estimating the expected probability value of the occurrence of the extreme scenario based on the likelihood ratio function, and solving the expected probability value with the goal of minimizing the variance of the expected probability value to obtain a third importance distribution; The third importance distribution is iteratively optimized through KL divergence to obtain a first importance distribution.

4. The extreme scenario generation method according to claim 3, characterized in that: The calculation formula for obtaining the first importance distribution is specifically: bad h D(g * ,h)=min v (∫g(x)lng(x)μ(dx)-∫g(x)lnh(x)μ(dx)); In the formula, g * is the third importance distribution, h is the normal distribution, h and f are functions of the same family, namely, multivariate independent Gaussian distribution, denoted as Parameters θ = (μ, σ) are the mean and standard deviation parameters of the normal distribution, and is the mean and variance parameter vector of the normal distribution h, g(x) is the importance distribution function of the x-th power system scenario; u is the measurement function of the probability of occurrence of the scenario random variable x; H(x) is the discriminant function of whether the x-th power system scenario is a set extreme scenario; v k and v k-1 are the first importance distributions of the kth and k-1th iterations respectively; L(X i ) is the likelihood ratio function, X i is the distribution of the i-th slave power system scenario The scene samples are obtained by sampling; N is the total number of samples.

5. The extreme scenario generation method according to claim 3, characterized in that: in, The first importance distribution is a mixed Gaussian model, and the iterative formula of the KL divergence is specifically: In the formula, is the random variable of the mth component of the i-th sample from the mixed Gaussian model, y j is the i-th sample; μ m is the mean parameter of the mth Gaussian mixture model component; is the label of the Gaussian function; X i is the distribution of the i-th slave power system scenario The scene samples obtained by sampling; α m is the weight of the mth Gaussian distribution in the mixed Gaussian model; θ m is the parameter of the mth Gaussian mixture model component; M is the number of components of the Gaussian mixture model; N is the total number of samples; N m is a temporary variable; L(X i ) is the likelihood ratio function; H(X i ) is the discriminant function; and are the parameters of the updated Gaussian mixture model.

6. The extreme scenario generation method according to claim 1, characterized in that: The iterative optimization of the first importance distribution is specifically: Obtaining an initial mean and an initial variance of the first importance distribution, and iteratively updating the initial mean using maximum likelihood estimation; The Euclidean distance between two iterative means is calculated, an expansion coefficient is determined according to the Euclidean distance and a preset distribution dimension, and the initial variance is updated based on the expansion coefficient.

7. The extreme scenario generation method according to claim 6, characterized in that: The calculation formula for iteratively optimizing the first importance distribution is specifically: d=|μ (k+1) -m (k) |; S(k+1)=δS(k); Where d is the Euclidean distance; μ (k+1) and μ (k) are the means of the first importance distributions of the k+1th iteration and the kth iteration respectively; δ is the expansion coefficient; τ is the multiplication factor; J is the Gaussian distribution dimension, A constant introduced to avoid the curse of dimensionality; S(k+1) and S(k) are the variances of the k+1th iteration and the kth iteration, respectively.

8. An extreme scene generation system, characterized in that: include: An acquisition module, a generation module, a first optimization module, and a second optimization module; The acquisition module is used to acquire historical operation data of the power system; The generating module is used to input the historical operation data into a preset scenario generating model to obtain a first scenario sample of the power system; The first optimization module is used to determine the normal distribution of the first scene sample in the latent variable space of the scene generation model, and iteratively optimize the normal distribution by KL divergence and importance sampling method to obtain a first importance distribution of the extreme scene, wherein the latent variable is a potential feature representation of the extreme scene; The second optimization module is used to iteratively optimize the first importance distribution until the iteration termination condition is met, determine the second importance distribution, and determine the extreme scenario samples corresponding to the power system based on the second importance distribution, wherein the second importance distribution includes a mean and a variance.

9. A terminal device, characterized in that: include: one or more processors; A memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the extreme scenario generation method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the extreme scenario generation method according to any one of claims 1 to 7 is implemented.

Images