Power load extreme scenario generation method based on self-representation constrained diffusion model

Abstract

The present application relates to the technical field of power system scenario generation, and discloses a power load extreme scenario generation method based on a self-representation constraint diffusion model, comprising: obtaining historical power load data, and preprocessing the historical power load data to obtain load samples; performing typical mode clustering on the load samples to obtain a plurality of typical mode clusters, calculating an extreme load intensity index, and screening out extreme load samples according to the extreme load intensity index; constructing a self-representation constraint diffusion model; extracting a self-representation constraint condition vector for each load sample; jointly training the self-representation constraint diffusion model by using the load samples and the corresponding self-representation constraint condition vectors; inputting a current historical load condition sequence into the trained self-representation constraint diffusion model, and screening out a power load extreme scenario. The present application has the advantage of realizing credible generation of a power load extreme scenario under the condition that historical extreme load samples are insufficient, and providing effective support for power system interval prediction, risk analysis and dispatching decision-making.

Description

Technical Field

[0001] This invention relates to the field of power system scenario generation technology, and in particular to a method for generating extreme power load scenarios based on a self-representation constrained diffusion model. Background Technology

[0002] With the continuous advancement of new power system construction, power load is increasingly influenced by multiple factors, including meteorological conditions, user behavior, holiday effects, industrial production arrangements, and unforeseen events, leading to greater randomness, volatility, and uncertainty. Particularly under special scenarios such as high temperatures, cold waves, extreme weather, or major social events, power load is prone to extreme characteristics such as peak surges, short-term spikes, rapid ramp-ups, and abnormal fluctuations. Although these extreme load scenarios occur infrequently in historical data, they have a significant impact on day-ahead dispatching, reserve capacity allocation, demand response arrangements, and operational risk early warning for the power system. Therefore, constructing realistic and diverse extreme power load scenarios is a key issue in the safe and stable operation and risk analysis of the power system.

[0003] Existing methods for power load forecasting or scenario generation mainly include statistical modeling methods, perturbation generation methods, bootstrap resampling methods, and deep learning-based generation methods. Traditional statistical modeling methods typically rely on Gaussian distributions, linear regression, autoregressive models, or probability distribution fitting to describe load uncertainty. While suitable for typical load fluctuation scenarios, they struggle to accurately characterize the non-Gaussian, peak-and-tailed, and locally abrupt characteristics of extreme loads. Perturbation generation methods usually superimpose random noise or proportional perturbations onto historical load curves. Although this expands the sample size, the generated results often lack realistic temporal evolution patterns, easily leading to distorted peak-and-valley structures, discontinuous local fluctuations, and unreasonable sudden increases and decreases. Bootstrap resampling methods construct scenarios by repeatedly sampling historical samples, but their essence remains a combination and reuse of existing samples, making it difficult to generate extreme load patterns that may occur in the future but were not fully reflected in historical data.

[0004] In recent years, deep generative models such as generative adversarial networks, variational autoencoders, and diffusion models have been gradually applied to the generation of power load scenarios. These methods can learn load variation patterns from historical data and generate load curves with a certain degree of diversity. However, ordinary generative models still have shortcomings in generating extreme load scenarios: on the one hand, extreme load samples account for a relatively small proportion of historical data, and the model training process is prone to bias towards conventional load patterns, resulting in insufficient representation of extreme features such as peak rise, rapid ramp-up, and abnormal fluctuations in the generated results; on the other hand, ordinary generative models usually lack explicit constraints on the temporal structure, local morphology, peak and valley positions, and typical patterns of the load sequence itself, which may lead to problems such as weak temporal correlation, local morphology deviating from the real samples, and insufficient retention of peak and valley features in the generated curves. Therefore, there is an urgent need for a method for generating extreme power load scenarios that can fully maintain the original load temporal structure and typical peak and valley morphology while enhancing the realism of extreme pattern generation, even under conditions of insufficient historical extreme load samples. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for generating extreme scenarios of power load based on a self-characteristic constrained diffusion model.

[0006] The objective of this invention is achieved through the following technical solution: a method for generating extreme scenarios of power load based on a self-representation constrained diffusion model, the method comprising the following steps:

[0007] S1. Obtain historical power load data, perform preprocessing, and construct a load sample consisting of historical load condition sequences and target load scenario sequences;

[0008] S2. Perform typical pattern clustering on the load samples based on dynamic time warping distance to obtain several typical pattern clusters, and calculate the extreme load intensity index of each load sample. Based on the extreme load intensity index, select extreme load samples from the load samples.

[0009] S3. Construct a self-representational constraint diffusion model, which includes a forward diffusion process and a conditional backward denoising process.

[0010] S4. For each load sample, extract its self-representation constraint vector, which includes at least the sample's time series representation, statistical characteristics, representation of its typical pattern cluster, and extreme load intensity index.

[0011] S5. The self-representation constraint vector is used as guiding information and introduced into the conditional reverse denoising process so that the diffusion model is constrained by the structural characteristics of the original load when generating the load scenario.

[0012] S6. The self-representation constraint diffusion model is jointly trained using the load sample and its corresponding self-representation constraint vector to learn the mapping relationship between historical load conditions and target load scenarios.

[0013] S7. Input the current historical load condition sequence into the trained self-representation constrained diffusion model to generate multiple candidate load scenarios, and select the extreme power load scenarios that meet the extreme intensity requirements.

[0014] Specifically, the typical pattern clustering of the load samples based on the dynamic time warp distance in S2 includes:

[0015] The similarity between any two load samples is measured by the dynamic time warping distance, and a load sample distance matrix is ​​constructed.

[0016] Based on the load sample distance matrix, combined with K-means or K-medoids clustering algorithms, the load samples are divided into several typical pattern clusters, which include at least regular load, peak load, sudden increase load, and abnormal fluctuation load.

[0017] Specifically, the calculation of the extreme load intensity index for each load sample includes:

[0018] Peak load intensity, ramp intensity, fluctuation intensity, and peak-to-valley difference characteristics are extracted from the load curve, and an extreme load intensity index is constructed using a weighted fusion method as follows:

[0019] ;

[0020] ;

[0021] ;

[0022] ;

[0023] ;

[0024] In the formula, For the i-th load sample or target load scenario sequence; For the sample The load value at time step t; For the sample Maximum load value; For the sample The minimum load value; T is the time length of the target load scenario sequence or the number of prediction time steps; For the sample Average load value; For the sample Maximum climbing intensity; For the sample The standard deviation of load fluctuation; For the sample Extreme load intensity index; , , , These are the weighting coefficients; This is the extreme load screening threshold; This is the set of extreme load samples obtained through screening;

[0025] When the extreme load intensity index is greater than the preset threshold, the load sample is identified and screened as an extreme load sample.

[0026] Specifically, the self-representation constraint vector is constructed as follows:

[0027] Extracting the temporal representation and statistical feature vector from the load sample:

[0028] ;

[0029] ;

[0030] ;

[0031] In the formula, For the temporal representation of the sample itself; For the i-th load sample or target load scenario sequence; For the temporal representation of the sample itself, It is a temporal feature encoder; This is the statistical feature vector of the sample; This is a self-representational constraint vector; This is the representation vector of the typical pattern cluster to which the sample belongs; This indicates a feature splicing operation.

[0032] Specifically, S5 further includes introducing a self-representation alignment constraint, specifically:

[0033] In diffusion denoising networks, the latent representations of early network layers are aligned with the stable target representations of later network layers, achieved through the following alignment loss function:

[0034] ;

[0035] ;

[0036] ;

[0037] In the formula, For the latent representation of the l-th layer of the diffusion denoising network; Let L be the feature extraction mapping of the l-th layer; L is the number of layers in the denoising network. A stable objective representation obtained from a moving exponential average or a teacher network; Indicates an exponential moving average operation; Represents a lightweight nonlinear mapping function; This is the self-representation alignment loss; This represents the difference between the features mapped at the l-th layer and the stable target features; To constrain all intermediate layers from layer 1 to layer L-1;

[0038] Specifically, the joint training in S6 includes constructing a shape-preserving loss function and an extreme feature enhancement loss function; the total loss function of the joint training is obtained by weighted summation of the diffusion model's basic noise prediction loss function, the shape-preserving loss function, and the extreme feature enhancement loss function, as specifically expressed as:

[0039] ;

[0040] ;

[0041] ;

[0042] ;

[0043] In the formula, This is the basic noise prediction loss for the diffusion model; To simulate real load scenarios diffusion step t and Gaussian noise Seeking expectations; The noise in the model prediction; For peak feature extraction function; This is the valley value feature extraction function; This refers to the load change rate or ramp-up characteristics; Loss due to shape preservation; Enhance loss for extreme features; This is the total training loss function for the model; , , These are the weighting coefficients; This is a sequence of real load scenarios; For the generated load scenario; This is an index of extreme load intensity for generating samples; The extreme intensity index is the result of screening the actual extreme load sample set.

[0044] Specifically, S7 includes:

[0045] Input the current historical load condition sequence, determine its typical mode category, and construct a self-representation constraint vector; starting from Gaussian noise, perform a conditional inverse denoising process using a trained diffusion model to generate S candidate load scenarios; calculate the extreme load intensity index of each candidate load scenario, and select scenarios that exceed a preset threshold to form the final set of extreme power load scenarios, represented as:

[0046] ;

[0047] ;

[0048] ;

[0049] ;

[0050] In the formula, The current historical load condition sequence is entered online; The typical pattern category to which the current load sequence belongs; This is the self-representational constraint vector corresponding to the current sample; A set of candidate load scenarios generated for the diffusion model; This is the s-th generated candidate load sequence; It is a conditional diffusion generative model (with parameters) Input noise and condition vector Output S candidate load sequences. S represents the number of candidate load scenarios; This is the set of extreme load scenarios obtained through screening; It is Gaussian noise; This is the extreme load screening threshold.

[0051] The present invention has the following advantages:

[0052] 1. This invention introduces a self-representation constraint mechanism, which uses the temporal morphology, local peak-valley structure, ramp-up change characteristics, and typical load patterns of the load sequence itself as constraint information in the reverse denoising process. This enables the diffusion model to not only rely on random noise sampling when generating extreme load scenarios, but also to be guided by the structural characteristics of the original load itself, thereby enhancing the model's ability to express extreme peaks, sudden increases and decreases, and abnormal fluctuation patterns.

[0053] 2. This invention transforms the traditional method of generating extreme power load scenarios, which relies on Gaussian assumptions, simple perturbations, or historical sample resampling, into a method driven by DTW typical pattern clustering, extreme intensity screening, and a self-representational constrained diffusion model. This method, while preserving the temporal structure, peak-valley shape, and extreme fluctuation characteristics of the load curve, can generate extreme load scenarios with realism, diversity, and risk analysis value. It is suitable for day-ahead dispatching, reserve capacity allocation, extreme load risk early warning, and operational safety analysis of power systems. Attached Figure Description

[0054] Figure 1 This is a schematic diagram of the scene generation method of the present invention;

[0055] Figure 2 This is a schematic diagram of the algorithm flow for the scene generation method of the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention; that is, the described embodiments are merely some embodiments of the invention, and not all embodiments. The components of the embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0057] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0058] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0059] The present invention will be further described below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description.

[0060] like Figures 1-2 As shown, a method for generating extreme scenarios of power load based on a self-representation constrained diffusion model is presented. This method includes the following steps:

[0061] S1. Obtain historical power load data, perform preprocessing, and construct a load sample consisting of historical load condition sequences and target load scenario sequences;

[0062] To eliminate dimensional differences and improve model training stability, the original load sequences are processed by outlier removal, missing value imputation, time alignment, and normalization to obtain standardized load sequences. Based on this, a sliding window approach is used to construct load samples, using historical load segments as input conditions and load curves within corresponding future time windows as target scenarios to be generated or learned. The above load sequence preprocessing and sample construction process can be uniformly represented by the following grouped formulas:

[0063] ;

[0064] ;

[0065] ;

[0066] ;

[0067] ;

[0068] In the formula, This is a historical power load time series; For the first The load value at each sampling time; The length of the historical load sequence; This is the t-th load value after normalization; and These represent the maximum and minimum load values ​​in the historical load sample, respectively. Let be the i-th historical load condition sequence, used to represent the historical load segment of the model input; Let L be the sequence of the i-th target load scenario; L be the length of the historical input window; T be the length of the target scenario window; and M be the number of load training samples constructed from the sliding window. It is a training sample set consisting of historical load condition sequences and target load scenario sequences.

[0069] S2. Perform typical pattern clustering on the load samples based on dynamic time warping distance to obtain several typical pattern clusters, and calculate the extreme load intensity index of each load sample. Based on the extreme load intensity index, select extreme load samples from the load samples.

[0070] Because power load curves may exhibit peak time shifts, local time scaling, and load pattern misalignments under different dates or operating conditions, directly using Euclidean distance for similarity measurement can easily overlook the consistency of the overall load curve shape. Therefore, this embodiment introduces Dynamic Time Warped Distance (DTW) to characterize the morphological similarity between different load scenario sequences, and constructs a load sample distance matrix based on this. Subsequently, combining K-means, K-medoids, or improved clustering algorithms, historical load samples are divided into several typical pattern clusters to obtain typical patterns such as regular load, peak load, sudden load surge, and abnormal fluctuation load. The Dynamic Time Warped Distance (DTW) similarity measurement and typical pattern clustering process can be represented by the following grouped formulas:

[0071] ;

[0072] ;

[0073] ;

[0074] ;

[0075] ;

[0076] In the formula, For load sequence In the Load values ​​at each time index; For load sequence In the Load values ​​at each time index; A set of matching points in the optimal alignment path. represents the optimal time alignment path between two load sequences; a and b are the time indices in the two sequences, respectively; D is the load sample distance matrix. For the i-th load sample With the j-th load sample The morphological distance between them; It is a collection of typical load pattern clusters; This represents the Kth typical load pattern cluster; K is the number of typical load pattern clusters. This is the center curve or representative curve of the kth typical pattern cluster; For the sample The typical pattern category to which it belongs; The representation vector representing the typical pattern cluster to which the sample belongs; For the sample The center curve of the model cluster; It is a temporal feature encoder;

[0077] After completing the clustering of typical patterns, to further identify extreme load samples with risk analysis value, this embodiment extracts features such as peak intensity, ramp intensity, fluctuation intensity, and peak-to-valley difference from the load curve to construct an extreme load intensity index. This index measures the degree of deviation of the load curve from the normal operating state and can comprehensively reflect extreme characteristics such as high peak values, rapid ramps, abnormal fluctuations, and widening peak-to-valley differences. When the extreme intensity index of a load sample exceeds a set threshold, it is determined to be an extreme load sample. The above extreme load intensity evaluation and screening process can be represented by the following grouped formula:

[0078] ;

[0079] ;

[0080] ;

[0081] ;

[0082] ;

[0083] In the formula, For the i-th load sample or target load scenario sequence; For the sample The load value at time step t; For the sample Maximum load value; For the sample The minimum load value; T is the time length of the target load scenario sequence or the number of prediction time steps; For the sample Average load value; For the sample Maximum climbing intensity; For the sample The standard deviation of load fluctuation; For the sample Extreme load intensity index; , , , These are the weighting coefficients; This is the extreme load screening threshold; This is the set of extreme load samples obtained through screening;

[0084] When the extreme load intensity index is greater than the preset threshold, the load sample is determined and screened as an extreme load sample.

[0085] S3. Construct a self-representational constraint diffusion model, which includes a forward diffusion process and a conditional backward denoising process.

[0086] In this embodiment, DTW typical pattern clustering and extreme intensity indicators can identify typical patterns and extreme samples from historical load data. However, since extreme load scenarios occur with low frequency in historical data, relying solely on historical sample screening is insufficient to meet the requirements for sample quantity and diversity in scenario generation, interval prediction, and risk analysis. Therefore, this embodiment introduces a diffusion generation model and incorporates self-representation constraints during the reverse denoising process. The model uses the temporal representation of the load sequence itself, the local peak-valley structure, the typical pattern cluster representation, and extreme intensity features as generation guidance information, thereby achieving reliable generation of extreme power load scenarios.

[0087] Let the sequence of real load scenarios be... The diffusion model first gradually moves towards the point through a forward diffusion process. Gaussian noise is added to degenerate it into intermediate states under different noise levels. Subsequently, the reverse denoising process starts from the high-noise state and gradually restores the load scenario with the true load time sequence structure. The forward denoising process of the above diffusion model can be represented by the following set of formulas:

[0088] ;

[0089] ;

[0090] ;

[0091] ;

[0092] ;

[0093] In the formula, This refers to the forward noise addition process in the diffusion model. This represents the noise load state at the t-th diffusion step. It is the load state of the previous diffusion step; It follows a Gaussian distribution. This represents the decayed portion of the original state. Indicates noise scheduling parameters; It is the noise attenuation coefficient for the current step; It is the cumulative decay coefficient from the initial step to the current step, used to directly calculate Y at any step from the initial state Y0. t ; It is Gaussian noise; It is the identity matrix;

[0094] S4. For each load sample, extract its self-representation constraint vector, which includes at least the sample's time series representation, statistical characteristics, representation of its typical pattern cluster, and extreme load intensity index.

[0095] In the reverse denoising stage, traditional diffusion models typically utilize neural networks to approximate the reverse transition distribution, gradually recovering the target sample from a high-noise state. However, ordinary diffusion models mainly rely on random noise sampling, and the generation process lacks explicit constraints on the temporal morphology and extreme characteristics of the load curve itself, easily leading to problems such as peak-valley structure shifts, local ramp distortion, and insufficient representation of abnormal fluctuations in the generated load scenario. Therefore, this embodiment constructs a self-representational constraint vector and introduces it into the reverse denoising process. The above-mentioned construction of the self-representational constraint vector and the conditional reverse denoising process can be represented by the following set of formulas:

[0096] ;

[0097] ;

[0098] ;

[0099] ;

[0100] ;

[0101] In the formula, For the temporal representation of the sample itself; This is the statistical feature vector of the sample; This is a self-representational constraint vector; This is the representation vector of the typical pattern cluster to which the sample belongs; For feature splicing operations; This is a conditional inverse denoising distribution with self-representation constraints introduced; The current noise load state is represented by t, which is the diffusion step; the mean is the mean value. Decision in the current state Predict the expected value of the previous time step; covariance Indicates the level of uncertainty or noise in the prediction; For parameters The diffusion generation model; For the generated load scenario;

[0102] S5. The self-representation constraint vector is used as guiding information and introduced into the conditional reverse denoising process so that the diffusion model is constrained by the structural characteristics of the original load when generating the load scenario.

[0103] In this embodiment, to ensure the consistency of the load temporal structure under different noise levels, a self-representation alignment constraint is further introduced. Specifically, the early layers of the diffusion denoising network tend to learn the global trend of the load curve under high noise conditions, while the later layers tend to learn local peaks and valleys and detailed fluctuations under low noise conditions. By aligning the latent representations of the early layers with the stable representations of the later layers, the model's ability to generate fine-grained morphologies from coarse-grained trends can be enhanced.

[0104] In diffusion denoising networks, the latent representations of early network layers are aligned with the stable target representations of later network layers, achieved through the following alignment loss function:

[0105] ;

[0106] ;

[0107] ;

[0108] In the formula, For the latent representation of the l-th layer of the diffusion denoising network; Let L be the feature extraction mapping of the l-th layer; L is the number of layers in the denoising network. A stable objective representation obtained from a moving exponential average or a teacher network; Indicates an exponential moving average operation; Represents a lightweight nonlinear mapping function; This is the self-representation alignment loss; This represents the difference between the features mapped at the l-th layer and the stable target features; To constrain all intermediate layers from layer 1 to layer L-1.

[0109] S6. The self-representation constraint diffusion model is jointly trained using the load sample and its corresponding self-representation constraint vector to learn the mapping relationship between historical load conditions and target load scenarios.

[0110] The joint training in S6 includes constructing a shape-preserving loss function and an extreme feature enhancement loss function; the total loss function of the joint training is obtained by weighted summation of the diffusion model's basic noise prediction loss function, the shape-preserving loss function, and the extreme feature enhancement loss function.

[0111] To further ensure the rationality of the generated load scenarios in terms of peak, trough, and ramp patterns, this embodiment constructs a pattern preservation loss and an extreme feature enhancement loss. The pattern preservation loss is used to constrain the generated curve to maintain consistency with the real curve in terms of peak and trough positions, peak and trough amplitudes, and load change rates; the extreme feature enhancement loss is used to constrain the generated samples to have extreme intensities that match extreme load samples. The above total loss function can be expressed by the following set of formulas:

[0112] ;

[0113] ;

[0114] ;

[0115] ;

[0116] In the formula, This is the basic noise prediction loss for the diffusion model; To simulate real load scenarios diffusion step t and Gaussian noise Seeking expectations; The noise in the model prediction; For peak feature extraction function; This is the valley value feature extraction function; This refers to the load change rate or ramp-up characteristics; Loss due to shape preservation; Enhance loss for extreme features; This is the total training loss function for the model; , , These are the weighting coefficients; This is a sequence of real load scenarios; For the generated load scenario; This is an index of extreme load intensity for generating samples; The extreme intensity index is the result of screening the actual extreme load sample set.

[0117] S7. Input the current historical load condition sequence into the trained self-representation constrained diffusion model to generate multiple candidate load scenarios, and select the extreme power load scenarios that meet the extreme intensity requirements.

[0118] During the offline training phase, historical load samples, typical pattern labels obtained from DTW clustering, and extreme intensity indices are used as training information to jointly train the self-representational constraint diffusion model, enabling it to learn the mapping relationship between "historical load conditions—typical pattern representation—extreme features—target load scenario". After training, the current historical load condition sequence is input, its typical pattern category is determined, and a self-representational constraint vector is constructed. Starting with Gaussian noise, a conditional inverse denoising process is performed using the trained diffusion model to generate S candidate load scenarios. The extreme load intensity index of each candidate load scenario is calculated, and scenarios exceeding a preset threshold are selected to form the final set of extreme power load scenarios, represented as:

[0119] ;

[0120] ;

[0121] ;

[0122] ;

[0123] In the formula, The current historical load condition sequence is entered online; The typical pattern category to which the current load sequence belongs; This is the self-representational constraint vector corresponding to the current sample; A set of candidate load scenarios generated for the diffusion model; Let s be the s-th generated candidate load sequence. Generate a conditional diffusion model (with parameters) Input noise and condition vector Output S candidate load sequences. S represents the number of candidate load scenarios; This is the set of extreme load scenarios obtained through screening; It is Gaussian noise; This is the extreme load screening threshold.

[0124] This invention transforms the traditional method of generating extreme power load scenarios, which relies on Gaussian assumptions, simple perturbations, or historical sample resampling, into a generation method driven by DTW typical pattern clustering, extreme intensity screening, and a self-representational constrained diffusion model. While preserving the temporal structure, peak-valley morphology, and extreme fluctuation characteristics of the load curve, this method can generate extreme load scenarios with realism, diversity, and risk analysis value, applicable to day-ahead dispatching, reserve capacity allocation, extreme load risk early warning, and operational safety analysis of power systems.

[0125] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the present invention. Any person skilled in the art can make many possible variations and modifications to the technical solution of the present invention, or modify it into equivalent embodiments, without departing from the scope of the present invention. Therefore, any modifications, equivalent changes, and alterations made to the above embodiments based on the technology of the present invention without departing from the scope of the present invention are within the protection scope of the present invention.

Claims

1. A method for generating extreme scenarios of power load based on a self-representation constrained diffusion model, characterized in that: The method includes the following steps: S1. Obtain historical power load data, perform preprocessing, and construct a load sample consisting of historical load condition sequences and target load scenario sequences; S2. Perform typical pattern clustering on the load samples based on dynamic time warping distance to obtain several typical pattern clusters, and calculate the extreme load intensity index of each load sample. Based on the extreme load intensity index, select extreme load samples from the load samples. S3. Construct a self-representational constraint diffusion model, which includes a forward diffusion process and a conditional backward denoising process. S4. For each load sample, extract its self-representation constraint vector, which includes at least the sample's time series representation, statistical characteristics, representation of its typical pattern cluster, and extreme load intensity index. S5. The self-representation constraint vector is used as guiding information and introduced into the conditional reverse denoising process so that the diffusion model is constrained by the structural characteristics of the original load when generating the load scenario. S6. The self-representation constraint diffusion model is jointly trained using the load sample and its corresponding self-representation constraint vector to learn the mapping relationship between historical load conditions and target load scenarios. S7. Input the current historical load condition sequence into the trained self-representation constrained diffusion model to generate multiple candidate load scenarios, and select the extreme power load scenarios that meet the extreme intensity requirements.

2. The method for generating extreme scenarios of power load based on a self-representational constrained diffusion model according to claim 1, characterized in that: The typical pattern clustering of the load samples based on dynamic time warp distance in S2 specifically includes: The similarity between any two load samples is measured by the dynamic time warping distance, and a load sample distance matrix is ​​constructed. Based on the load sample distance matrix, combined with K-means or K-medoids clustering algorithms, the load samples are divided into several typical pattern clusters, which include at least regular load, peak load, sudden increase load, and abnormal fluctuation load.

3. The method for generating extreme scenarios of power load based on a self-representational constrained diffusion model according to claim 1, characterized in that: The calculation of the extreme load intensity index for each load sample specifically includes: Peak load intensity, ramp intensity, fluctuation intensity, and peak-to-valley difference characteristics are extracted from the load curve, and an extreme load intensity index is constructed using a weighted fusion method as follows: ； ； ； ； ； In the formula, For the i-th load sample or target load scenario sequence; For the sample The load value at time step t; For the sample Maximum load value; For the sample The minimum load value; T is the time length of the target load scenario sequence or the number of prediction time steps; For the sample Average load value; For the sample Maximum climbing intensity; For the sample The standard deviation of load fluctuation; For the sample Extreme load intensity index; , , , These are the weighting coefficients; This is the extreme load screening threshold; This is the set of extreme load samples obtained through screening; When the extreme load intensity index is greater than the preset threshold, the load sample is identified and screened as an extreme load sample.

4. The method for generating extreme power load scenarios based on a self-representational constraint diffusion model according to claim 3, characterized in that: The self-representation constraint vector is constructed as follows: Extracting the temporal representation and statistical feature vector from the load sample: ； ； ； In the formula, For the i-th load sample or target load scenario sequence; For the temporal representation of the sample itself; It is a temporal feature encoder; This is the statistical feature vector of the sample; This is a self-representational constraint vector; This is the representation vector of the typical pattern cluster to which the sample belongs; This is a feature splicing operation.

5. The method for generating extreme power load scenarios based on a self-representational constrained diffusion model according to claim 1, characterized in that: The S5 also includes introducing a self-representation alignment constraint, specifically: In diffusion denoising networks, the latent representations of early network layers are aligned with the stable target representations of later network layers, achieved through the following alignment loss function: ； ； ； In the formula, For the latent representation of the l-th layer of the diffusion denoising network; Let L be the feature extraction mapping of the l-th layer; L is the number of layers in the denoising network. A stable objective representation obtained from a moving exponential average or a teacher network; Indicates an exponential moving average operation; Represents a lightweight nonlinear mapping function; This is the self-representation alignment loss; This represents the difference between the features mapped at the l-th layer and the stable target features; To constrain all intermediate layers from layer 1 to layer L-1.

6. The method for generating extreme power load scenarios based on a self-representational constrained diffusion model according to claim 1, characterized in that: The joint training in S6 includes constructing a shape-preserving loss function and an extreme feature enhancement loss function; the total loss function of the joint training is obtained by weighted summation of the diffusion model's basic noise prediction loss function, the shape-preserving loss function, and the extreme feature enhancement loss function, specifically expressed as follows: ； ； ； ； In the formula, This is the basic noise prediction loss for the diffusion model; To simulate real load scenarios diffusion step t and Gaussian noise Seeking expectations; The noise in the model prediction; For peak feature extraction function; This is the valley value feature extraction function; This refers to the load change rate or ramp-up characteristics; Loss due to shape preservation; Enhance loss for extreme features; This is the total training loss function for the model; , , These are the weighting coefficients; This is a sequence of real load scenarios; For the generated load scenario; This is an index for the extreme load intensity of the generated samples; The extreme intensity index is the result of screening the actual extreme load sample set.

7. The method for generating extreme scenarios of power load based on a self-representational constrained diffusion model according to claim 1, characterized in that: Specifically, S7 includes: Input the current historical load condition sequence, determine its typical mode category, and construct a self-representation constraint vector; starting from Gaussian noise, perform a conditional inverse denoising process using a trained diffusion model to generate S candidate load scenarios; calculate the extreme load intensity index of each candidate load scenario, and select scenarios that exceed a preset threshold to form the final set of extreme power load scenarios, represented as: ； ； ； ； In the formula, The current historical load condition sequence is entered online; The typical pattern category to which the current load sequence belongs; The vector of self-representational constraints corresponding to the current sample; S represents the number of candidate load scenarios; A set of candidate load scenarios generated for the diffusion model; This is the s-th generated candidate load sequence; It is a conditional diffusion generation model, with input noise. and condition vector Output S candidate load sequences; This is the set of extreme load scenarios obtained through screening; This is the extreme load screening threshold.

Images