A method and system for generating photovoltaic power output extreme scenarios based on physically guided diffusion

Abstract

The application discloses a photovoltaic output extreme scene generation method and system based on physical guidance diffusion, and belongs to the technical field of photovoltaic power generation of electric power systems. The method comprises the following steps: acquiring time series of photovoltaic theoretical maximum power and photovoltaic actual power respectively, and constructing a dimensionless index sequence based on the ratio of the two; constructing a change rate curve of the dimensionless index sequence, screening an extreme fluctuation segment, and generating an extreme seed database; according to the photovoltaic actual power time sequence and the obtained maximum ramp rate of an inverter, a physically feasible region is demarcated; the physically feasible region and the seed data are taken as double conditions, and are input into a conditional latent diffusion model with a cross attention mechanism to generate a plurality of extreme scenes; and the plurality of generated extreme scenes are progressively clustered by using a dynamic confidence threshold to obtain a plurality of sets of extreme scenes. The application effectively solves the problems of extreme sample scarcity and physical distortion of scene generation by a pure data model in the prior art.

Description

Technical Field

[0001] This invention belongs to the field of photovoltaic power generation technology in power systems, specifically relating to a method and system for generating extreme photovoltaic output scenarios based on physical guided diffusion. Background Technology

[0002] With the construction of a new power system based on new energy sources, the penetration rate of photovoltaic power generation is increasing. However, photovoltaic power output is highly random and volatile, especially extreme power output scenarios caused by extreme weather (such as rapid cloud cover, sandstorms, etc.), which pose a huge challenge to the safe and stable operation of the power grid.

[0003] Currently, a large amount of extreme scenario data is needed to assess the ultimate carrying capacity of power systems. However, existing photovoltaic data generation methods have the following drawbacks: First, extreme samples in measured data exhibit a long-tail distribution, resulting in scarce samples, especially for newly built power plants, making it difficult for purely data-driven methods to capture their distribution patterns. Second, data generated by existing generative adversarial networks or standard diffusion models often only possess statistical similarity, ignoring the physical inertia limitations of photovoltaic inverters, and easily generating distorted data that violates physical laws, such as "zero output one second, full output the next." Third, the generated scenarios lack effective classification and screening mechanisms, making them difficult to directly apply to engineering scenarios. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and system for generating extreme photovoltaic power output scenarios based on physical guided diffusion. It utilizes inverter ramp-up constraints to construct dynamic physical boundaries and injects these boundaries as hard constraints into the potential diffusion model, forcing the model to generate extreme scenarios that conform to physical operating conditions. Finally, through a progressive clustering strategy, it achieves accurate clustering of typical extreme operating conditions.

[0005] This invention provides the following technical solution:

[0006] Firstly, a method for generating extreme photovoltaic power output scenarios based on physically guided diffusion is provided, including the following steps: The time series of the theoretical maximum power and the actual power of photovoltaics were obtained respectively, and a dimensionless exponential sequence was constructed based on the ratio of the two. Construct the rate of change curve of the dimensionless exponential sequence, screen extreme fluctuation segments, and generate an extreme seed database; Based on the actual photovoltaic power time series and the obtained maximum ramp rate of the inverter, the dynamic physical allowable boundary sliding with time is calculated, and the physical feasible region is delineated. The seed data from the physical feasible region and the extreme seed database are used as dual conditions and input into the conditional latent diffusion model that introduces a cross-attention mechanism to generate several extreme scenarios. By using dynamic confidence thresholds, several extreme scenarios are progressively clustered to obtain several sets of extreme scenarios.

[0007] Optionally, the formula for obtaining the theoretical maximum power of photovoltaic power is: ; in, for At any given moment, the theoretical maximum power of photovoltaics This refers to the rated installed capacity of the photovoltaic power station. and They are respectively Direct and diffuse irradiance at all times The angle of incidence of sunlight. The irradiance reference value under standard test conditions. This is the atmospheric quality correction function.

[0008] Optionally, the step of obtaining the time series of the theoretical maximum power and the actual power of photovoltaic power respectively, and constructing a dimensionless exponential sequence based on the ratio of the two, specifically involves: Preprocessing is performed on the time series of theoretical maximum power and actual power of photovoltaics. Set the effective start-up power threshold for photovoltaic power generation , , This is an empirical coefficient. This refers to the rated installed capacity of the photovoltaic power station. The maximum exponent at all times is calculated using the following formula. and the calculated maximum exponent Arranged in chronological order, a dimensionless exponent sequence is obtained; ; in, for At any given moment, the theoretical maximum power of photovoltaics for Actual photovoltaic power at any given time.

[0009] Optionally, the step of constructing the rate of change curve of the dimensionless exponential sequence, screening extreme fluctuation segments, and generating an extreme seed database specifically involves: The first difference of the dimensionless exponential sequence is calculated and the absolute value is taken to obtain the rate of change sequence between adjacent time points. After arranging the rate of change series in descending order of numerical value, construct the rate of change curve; The inflection point detection algorithm is used to identify the position with the maximum curvature in the rate of change curve as the extreme judgment threshold. Determine the rate of change at each time point If the extreme threshold is not exceeded, the current moment is marked as a normal fluctuation moment and is not entered into the extreme seed database. If the threshold is exceeded, the actual photovoltaic power sequence with a preset window length before and after the current moment is used as a seed and stored in the extreme seed database.

[0010] Optionally, the step of calculating the dynamic physical allowable boundary that slides over time based on the actual photovoltaic power time series and the obtained maximum ramp rate of the inverter, and delineating the physical feasible region, specifically involves: Obtain the actual photovoltaic power at the previous moment and the inverter's maximum ramp rate ; Based on the physical inertia of the equipment, the physical allowable upper and lower power limits at the current moment are calculated according to the following formula; ; ; in, and Each is the current The physical upper and lower limits of power allowed at any given time. This represents the time difference between the current moment and the previous moment. By combining the physically permissible upper and lower power limits at each moment with the theoretical maximum photovoltaic power, the physically feasible region at each moment is generated. , for The theoretical maximum power of photovoltaic power at any given time.

[0011] Optionally, the step of using seed data from the physically feasible region and the extreme seed database as dual conditions, inputting them into a conditional latent diffusion model that incorporates a cross-attention mechanism, to generate several extreme scenarios, specifically: A seed is selected from the extreme seed database, and its fluctuation features are extracted by the VAE encoder and then mapped to a low-dimensional latent variable. Divide the current physical feasible region by the current theoretical maximum photovoltaic power to transform the physical feasible region into a relative maximum exponential constraint. After verifying and correcting the relative maximum exponential constraint, map it into a physical state vector. A cross-attention mechanism is introduced on the bottleneck layer of the U-net denoising network for the conditional latent diffusion model, using low-dimensional latent variables as query vectors and physical state vectors as key and value vectors. Several extreme scenarios are generated using a conditional latent diffusion model that incorporates a cross-attention mechanism.

[0012] Optionally, the conditional latent diffusion model introduces a physical out-of-bounds penalty into the overall objective function during training; The overall objective function The formula is: ; ; in, The loss is used to reconstruct the distribution of all training samples, with a value representing the mean squared error. To impose penalties for exceeding the limits, As a physical boundary violation penalty, The number of training samples, For the current extreme scenario The actual power of photovoltaics, and Each is the current The physical upper and lower limits of power allowed at any given time. for The theoretical maximum power of photovoltaic power at any given time.

[0013] Optionally, the step of progressively clustering the generated extreme scenarios using a dynamic confidence threshold to obtain several sets of extreme scenarios is as follows: Construct a temporal convolutional network consisting of several residual blocks, each residual block comprising a causal dilated convolutional layer with an exponentially increasing dilation coefficient. The generated extreme scenarios are input into the temporal convolutional network to extract features for each extreme scenario. Initialize cluster centers and confidence thresholds; The cluster centers are updated through several iterations, and the cluster centers of the last iteration and the extreme scenarios for each cluster are output. In each iteration, the features of each extreme scenario are calculated and compared with the current cluster centers. The comprehensive distance is calculated, and based on the comprehensive distance, the extreme scenario features belonging to the cluster centers are calculated using the Student t-distribution kernel function. The algorithm calculates the soft-segment confidence level; it then determines whether the current soft-segment confidence level is greater than the confidence threshold. If it is, it updates the cluster centers using the current extreme scenario features; otherwise, it temporarily ignores the current extreme scenario features and updates the cluster centers in the current iteration. Finally, it determines whether the current iteration round meets the clustering convergence condition. If it does, it outputs the updated cluster centers and the extreme scenarios for each cluster. If it does not, it determines whether the current iteration round has reached the preset threshold update cycle. If it does, it dynamically reduces the confidence threshold according to the preset decay coefficient; otherwise, it maintains the current confidence level for the next iteration.

[0014] Optionally, the extreme scene features are calculated to belong to the cluster center. The formula for the soft-segment reliability is: ; in, Features of extreme scenarios Belongs to cluster center Soft-score configuration confidence, Features of extreme scenarios With cluster center The combined distance includes Euclidean distance and Pearson distance; Features of extreme scenarios With the Cluster centers The overall distance The weights are the weights corresponding to the overall distance.

[0015] Secondly, a photovoltaic power output extreme scenario generation system based on physical guided diffusion is provided, including: The feature reconstruction module is used to obtain the time series of the theoretical maximum power of photovoltaics and the actual power of photovoltaics, and to construct a dimensionless exponential sequence based on the ratio of the two. The seed library generation module is used to construct the rate of change curve of the dimensionless exponential sequence, screen extreme fluctuation segments, and generate an extreme seed database. The physical feasible region generation module is used to calculate the dynamic physical allowable boundary that slides over time based on the actual photovoltaic power time series and the obtained maximum ramp rate of the inverter, and to delineate the physical feasible region. The scene generation module is used to take the seed data from the physical feasible domain and the extreme seed database as dual conditions and input them into the conditional latent diffusion model that introduces a cross-attention mechanism to generate several extreme scenes. The scene identification module is used to progressively cluster several generated extreme scenes using a dynamic confidence threshold to obtain several sets of extreme scenes.

[0016] Compared with the prior art, the beneficial effects of the present invention are: This invention decouples actual photovoltaic power data from theoretical maximum power, eliminating the interference of deterministic solar radiation trends on random fluctuation characteristics. It solves the problems of traditional methods where fluctuation characteristics are masked by solar radiation periodicity and inconsistent sample distribution across seasons. This allows the conditional latent diffusion model to focus on learning high-frequency random fluctuation textures, significantly improving the fidelity of extreme scenario generation. Furthermore, this invention utilizes inverter ramp-up constraints to construct dynamic physical boundaries and injects these boundaries as hard constraints into the latent diffusion model, forcing it to generate extreme scenarios that conform to physical operating conditions, ensuring that the generated extreme scenarios strictly follow inverter ramp-up inertia. Finally, this invention employs a progressive clustering strategy to achieve accurate reproduction and summarization of high-risk extreme operating conditions (such as V-shaped drops and sawtooth oscillations) under small sample conditions. Attached Figure Description

[0017] Figure 1This is a flowchart of the photovoltaic power output extreme scenario generation method based on physical guided diffusion of the present invention; Figure 2 This is a flowchart of the construction of the dimensionless exponential sequence of the present invention; Figure 3 This is a flowchart of the extreme seed database generation process of the present invention; Figure 4 This is a flowchart of the physical feasible domain generation process of the present invention; Figure 5 This is a flowchart of the present invention that uses a conditional latent diffusion model with an introduced cross-attention mechanism to generate several extreme scenarios; Figure 6 This is a flowchart of the progressive clustering method of the present invention; Figure 7 This is a performance index comparison chart comparing the generation method of this invention with existing technologies; Figure 8 This is a schematic diagram comparing the fluctuations of the dimensionless exponential sequence of the present invention with the original actual power. Figure 9 This is a schematic diagram of the extreme determination threshold of the present invention; Figure 10 This is a schematic diagram illustrating the comparison and verification of the physical compliance of the generated scene in this invention; Figure 11 This is a computational example diagram of extreme scenarios generated using the method of this invention; Figure 12 This is a schematic diagram of the clustering results in a typical extreme scenario of the present invention. Detailed Implementation

[0018] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be used to limit the scope of protection of the present invention. It should be noted that the term "comprising" and any variations thereof in the specification, claims and the above-mentioned drawings of the present invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products or devices.

[0019] Example 1: like Figure 1 As shown, a method for generating extreme photovoltaic power output scenarios based on physical guided diffusion is provided, including the following steps: S1: Obtain the time series of the theoretical maximum power and the actual power of photovoltaics respectively, and construct a dimensionless exponential series based on the ratio of the two. S2: Construct the rate of change curve of the dimensionless exponential sequence, screen extreme fluctuation segments, and generate an extreme seed database; S3: Based on the actual photovoltaic power time series and the obtained maximum ramp rate of the inverter, calculate the dynamic physical allowable boundary that slides with time and delineate the physical feasible region. S4: The seed data from the physical feasible region and the extreme seed database are used as dual conditions and input into the conditional latent diffusion model that introduces a cross-attention mechanism to generate several extreme scenarios. S5: Use dynamic confidence thresholds to progressively cluster several generated extreme scenarios to obtain several sets of extreme scenarios.

[0020] This invention decouples the actual power data of photovoltaics from the theoretical maximum power of photovoltaics, eliminating the interference of deterministic solar radiation trends on random fluctuation characteristics. It solves the problems of the fluctuation characteristics being masked by the periodicity of solar radiation and the inconsistent distribution of samples across seasons in traditional methods. This allows the conditional latent diffusion model to focus on learning high-frequency random fluctuation textures, thereby significantly improving the fidelity of extreme scene generation.

[0021] In this embodiment, as Figure 2 As shown, step S1 specifically includes: S11: Obtain the theoretical maximum power of photovoltaics.

[0022] Step S11 specifically includes: S111: Based on the latitude, longitude, and time information in the geographic location data, calculate the solar declination angle, hour angle, and solar altitude angle, and combine this with the tilt angle of the photovoltaic panel installation to calculate the angle of incidence of sunlight. The calculation method refers to existing technology.

[0023] S112: Calculate the theoretical maximum power of photovoltaic power according to the following formula: ; in, for At any given moment, the theoretical maximum power of photovoltaics This refers to the rated installed capacity of the photovoltaic power station. and They are respectively Direct and diffuse irradiance at all times The angle of incidence of sunlight. This is the baseline irradiance value under standard test conditions (usually 1000 W / m²). This is the atmospheric quality correction function.

[0024] S12: Construct a dimensionless exponential sequence based on the time series of the theoretical maximum power and the actual power of photovoltaics.

[0025] Step S12 specifically includes: S121: Preprocess the photovoltaic theoretical maximum power time series and the photovoltaic actual power time series; including timestamp alignment and removal of outlier data points with mismatched timestamps.

[0026] S122: Set the effective start-up power threshold for photovoltaic power generation , , This is an empirical coefficient, typically ranging from 0.01 to 0.03. This refers to the rated installed capacity of the photovoltaic power station.

[0027] Setting an effective start-up power threshold can be used to distinguish between effective and ineffective operating periods of photovoltaic panels.

[0028] S123: Calculate the maximum exponent at all times according to the following formula. and the calculated maximum exponent Arranged in chronological order, a dimensionless exponent sequence is obtained; ; in, for At any given moment, the theoretical maximum power of photovoltaics for The actual power output of the photovoltaic system at any given time is obtained based on historical data.

[0029] Depend on Figure 8 As can be seen, the present invention decouples the actual power data of photovoltaics from the theoretical maximum power of photovoltaics, thereby eliminating the interference of deterministic solar radiation trends on random fluctuation characteristics.

[0030] In this embodiment, as Figure 3 As shown, step S2 specifically includes: S21: Perform first-order difference calculation on the dimensionless exponential sequence and take the absolute value to obtain the rate of change sequence of adjacent time points.

[0031] S22: After arranging the rate of change series in descending order of numerical value, construct the rate of change curve.

[0032] S23: Use the inflection point detection algorithm to identify the position with the largest curvature in the rate of change curve as the extreme judgment threshold.

[0033] The point of maximum curvature in the rate of change curve is the physical boundary of the long-tailed distribution, and this point is used as the extreme threshold. A schematic diagram illustrating the determination of the extreme threshold is shown below. Figure 9 As shown.

[0034] S24: Determine the rate of change at each time point Does it exceed the extreme threshold? If the time limit is not exceeded, the current time will be marked as a normal fluctuation time and will not be included in the extreme seed database; If the limit is exceeded, the actual photovoltaic power sequence with a preset window length before and after the current moment will be used as a seed and stored in the extreme seed database.

[0035] In this embodiment, as Figure 4 As shown, step S3 specifically includes: S31: Obtain the actual photovoltaic power at the previous moment. and the inverter's maximum ramp rate .

[0036] S32: Based on the physical inertia of the equipment, calculate the physical allowable upper and lower power limits at the current moment according to the following formula.

[0037] ; ; in, and Each is the current The physical upper and lower limits of power allowed at any given time. This represents the time difference between the current moment and the previous moment.

[0038] S33: Combine the physically permissible upper and lower power limits at each moment with the theoretical maximum photovoltaic power to generate the physically feasible region at each moment. , for The theoretical maximum power of photovoltaic power at any given time.

[0039] In this embodiment, as Figure 5 As shown, step S4 specifically includes the following steps: S41: Select a seed from the extreme seed database, extract the fluctuation features through the VAE encoder, and then map them into a low-dimensional latent variable.

[0040] The extracted fluctuation features are mapped to low-dimensional latent variables using reparameterized sampling.

[0041] S42: Divide the current physical feasible region by the current theoretical maximum photovoltaic power to transform the physical feasible region into a relative maximum exponential constraint. After verifying and correcting the relative maximum exponential constraint, map it into a physical state vector.

[0042] The maximum exponential constraint ensures that physical constraints can effectively apply to the dimensionless conditional latent diffusion model. The physical feasible region at the current time step. Dividing by the current moment's theoretical maximum photovoltaic power is equivalent to spatial alignment projection.

[0043] The relative maximum exponential constraint is verified and corrected, specifically by verifying the physical feasible region. Does it exist as a cause? Minimal values ​​may cause numerical overflow or interval inversion; if such conditions exist, the preset physical extreme values ​​(0 and 1) are used for clamping correction to ensure the mathematical convergence of the constraints. If no such conditions exist, the relative maximum exponential constraint is directly mapped to the physical state vector.

[0044] S43: A cross-attention mechanism is introduced on the bottleneck layer of the U-net denoising network for the conditional latent diffusion model, using low-dimensional latent variables as query vectors and physical state vectors as key and value vectors.

[0045] The denoising network U-net is structured in accordance with existing techniques for predicting and removing Gaussian noise from latent variables.

[0046] S44: Generate several extreme scenarios using a conditional latent diffusion model that incorporates a cross-attention mechanism.

[0047] In this embodiment, the conditional latent diffusion model introduces a physical out-of-bounds penalty into the overall objective function during training.

[0048] Overall objective function The formula is: ; ; in, The loss is used to reconstruct the distribution of all training samples, with a value representing the mean squared error. To impose penalties for exceeding the limits, As a physical boundary violation penalty, The number of training samples, For the current extreme scenario The actual power of photovoltaics, and Each is the current The physical upper and lower limits of power allowed at any given time. for The theoretical maximum power of photovoltaic power at any given time.

[0049] The gradient is calculated using the comprehensive objective function, and the parameters of the denoising network are updated in reverse to force the model to learn the physical boundary. This process is repeated iteratively until convergence, where the convergence condition is that any of the following formulas are satisfied: ; ; ; in, To generate the number of physically feasible regions violated in several extreme scenarios using a conditional latent diffusion model, To generate the number of extreme scenarios using a conditional latent diffusion model, This represents the maximum number of iterations for the conditional latent diffusion model. For a moment The comprehensive objective function, and These are the two threshold values ​​that are set.

[0050] Depend on Figure 10 It can be seen that using the physical guidance diffusion generation in step S4 to generate extreme scenarios can avoid scenarios that violate the laws of physics, such as "zero output one second and full output the next second".

[0051] In this embodiment, as Figure 6 As shown, step S5 specifically includes: S51: Construct a temporal convolutional network consisting of several residual blocks, each residual block containing a causal dilated convolutional layer with an exponentially increasing dilation coefficient.

[0052] The structures of residual blocks and temporal convolutional networks can refer to existing techniques.

[0053] S52: Input the generated extreme scenarios into the temporal convolutional network and extract the features of each extreme scenario.

[0054] S53: Initialize cluster centers and confidence thresholds.

[0055] S54: Update the cluster centers through several iterations and output the cluster centers of the last iteration as well as the extreme scenarios for each cluster.

[0056] Each iteration specifically includes the following steps: E1: Calculate the features of each extreme scenario and the current cluster center. The overall distance.

[0057] ; in, Features of extreme scenarios With cluster center The combined distance includes Euclidean distance and Pearson distance; The weights are based on the overall distance. For the Pearson correlation coefficient, considering the difference in numerical dimensions between the Euclidean distance term and the correlation coefficient term, and to avoid the large Euclidean distance masking the waveform correlation characteristics, this invention prioritizes features from extreme scenarios. Standardization is performed. Based on this, and considering that the identification of extreme photovoltaic scenarios focuses more on waveform trends (such as V-shaped drops and oscillations), the weighting coefficients are adjusted. Set it to a smaller value (e.g., 0.3) to give the correlation metric a higher weight, thereby enhancing the model's ability to distinguish between different failure modes.

[0058] E2: Based on the comprehensive distance, the extreme scene features are calculated using the Student t-distribution kernel function to belong to the cluster centers. The soft-score configuration reliability.

[0059] The specific formula is as follows: ; in, Features of extreme scenarios Belongs to cluster center Soft-score configuration confidence, Features of extreme scenarios With the Cluster centers The overall distance The weights are the weights corresponding to the overall distance.

[0060] E3: Determine whether the current soft segmentation configuration confidence is greater than the confidence threshold. If it is, update the cluster center using the current extreme scenario features. Otherwise, temporarily ignore the current extreme scenario features in the current iteration and update the cluster center.

[0061] E4: Determine whether the current iteration round meets the clustering convergence condition. If it does, output the updated cluster centers and the extreme scenarios for each cluster. If it does not meet the condition, determine whether the current iteration round has reached the preset threshold update cycle. If it does, dynamically reduce the confidence threshold according to the preset decay coefficient. If not, maintain the current confidence level for the next iteration round.

[0062] This application Figure 11 In the text, (a) represents the original extreme sample. Figure 11 (b) in the figure represents the generation result of the photovoltaic power output extreme scenario generation method using the physical guided diffusion of the present invention, and the dashed line represents the typical extreme scenario of each cluster. The comparison shows that the typical extreme scenarios generated by the present invention effectively simulate the fluctuation characteristics of the original extreme samples. Figure 12 It can be seen that clustering using the method of step S5 of the present invention can achieve accurate reproduction and summarization of high-risk extreme working conditions (such as V-shaped drop and sawtooth oscillation) under small sample conditions.

[0063] Depend on Figure 7 It can be seen that, when using the method of this invention, existing method 1 (WGAN-GP method) and existing method 2 (Standard-LDM method) to generate extreme scenarios, the present invention is significantly superior to existing methods 1 and 2 in terms of generation quality, physical compliance and extreme coverage.

[0064] Example 2: A photovoltaic power output extreme scenario generation system based on physical guided diffusion includes: The feature reconstruction module is used to obtain the time series of the theoretical maximum power of photovoltaics and the actual power of photovoltaics, and to construct a dimensionless exponential sequence based on the ratio of the two. The seed library generation module is used to construct the rate of change curve of the dimensionless exponential sequence, screen extreme fluctuation segments, and generate an extreme seed database. The physical feasible region generation module is used to calculate the dynamic physical allowable boundary that slides over time based on the actual photovoltaic power time series and the obtained maximum ramp rate of the inverter, and to delineate the physical feasible region. The scene generation module is used to take the seed data from the physical feasible domain and the extreme seed database as dual conditions and input them into the conditional latent diffusion model that introduces a cross-attention mechanism to generate several extreme scenes. The scene identification module is used to progressively cluster several generated extreme scenes using a dynamic confidence threshold to obtain several sets of extreme scenes.

[0065] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0066] Those skilled in the art will clearly understand that the techniques in the embodiments of the present invention can be implemented using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or certain parts of the embodiments of the present invention.

[0067] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A method for generating extreme photovoltaic power output scenarios based on physical guided diffusion, characterized in that, Includes the following steps: The time series of the theoretical maximum power and the actual power of photovoltaics were obtained respectively, and a dimensionless exponential sequence was constructed based on the ratio of the two. Construct the rate of change curve of the dimensionless exponential sequence, screen extreme fluctuation segments, and generate an extreme seed database; Based on the actual photovoltaic power time series and the obtained maximum ramp rate of the inverter, the dynamic physical allowable boundary sliding with time is calculated, and the physical feasible region is delineated. The seed data from the physical feasible region and the extreme seed database are used as dual conditions and input into the conditional latent diffusion model that introduces a cross-attention mechanism to generate several extreme scenarios. By using dynamic confidence thresholds, several extreme scenarios are progressively clustered to obtain several sets of extreme scenarios.

2. The method for generating extreme photovoltaic power output scenarios based on physical guided diffusion according to claim 1, characterized in that, The formula for obtaining the theoretical maximum power of photovoltaic power is: ； in, for At any given moment, the theoretical maximum power of photovoltaics This refers to the rated installed capacity of the photovoltaic power station. and They are respectively Direct and diffuse irradiance at all times The angle of incidence of sunlight. The irradiance reference value under standard test conditions. This is the atmospheric quality correction function.

3. The method for generating extreme photovoltaic power output scenarios based on physical guided diffusion according to claim 1, characterized in that, The process involves obtaining the time series of the theoretical maximum power and the actual power of photovoltaic power, respectively, and constructing a dimensionless exponential sequence based on the ratio of the two, specifically as follows: Preprocessing is performed on the time series of theoretical maximum power and actual power of photovoltaics. Set the effective start-up power threshold for photovoltaic power generation , , This is an empirical coefficient. This refers to the rated installed capacity of the photovoltaic power station. The maximum exponent at all times is calculated using the following formula. and the calculated maximum exponent Arranged in chronological order, a dimensionless exponent sequence is obtained; ； in, for At any given moment, the theoretical maximum power of photovoltaics for Actual photovoltaic power at any given time.

4. The method for generating extreme photovoltaic power output scenarios based on physical guided diffusion according to claim 1, characterized in that, The process of constructing the rate of change curve of the dimensionless exponential sequence, screening extreme fluctuation segments, and generating an extreme seed database specifically involves: The first difference of the dimensionless exponential sequence is calculated and the absolute value is taken to obtain the rate of change sequence between adjacent time points. After arranging the rate of change series in descending order of numerical value, construct the rate of change curve; The inflection point detection algorithm is used to identify the position with the maximum curvature in the rate of change curve as the extreme judgment threshold. Determine the rate of change at each time point If the extreme threshold is not exceeded, the current moment is marked as a normal fluctuation moment and is not entered into the extreme seed database. If the threshold is exceeded, the actual photovoltaic power sequence with a preset window length before and after the current moment is used as a seed and stored in the extreme seed database.

5. The method for generating extreme photovoltaic power output scenarios based on physical guided diffusion according to claim 1, characterized in that, The process involves calculating the dynamic physical allowable boundary that slides over time based on the actual photovoltaic power time series and the obtained maximum ramp rate of the inverter, and delineating the physical feasible region. Specifically: Obtain the actual photovoltaic power at the previous moment and the inverter's maximum ramp rate ; Based on the physical inertia of the equipment, the physical allowable upper and lower power limits at the current moment are calculated according to the following formula; ； ； in, and Each is the current The physical upper and lower limits of power allowed at any given time. This represents the time difference between the current moment and the previous moment. By combining the physically permissible upper and lower power limits at each moment with the theoretical maximum photovoltaic power, the physically feasible region at each moment is generated. , for The theoretical maximum power of photovoltaic power at any given time.

6. The method for generating extreme photovoltaic power output scenarios based on physical guided diffusion according to claim 1, characterized in that, The process involves using seed data from both the physically feasible region and the extreme seed database as dual conditions, inputting them into a conditional latent diffusion model that incorporates a cross-attention mechanism to generate several extreme scenarios, specifically: A seed is selected from the extreme seed database, and its fluctuation features are extracted by the VAE encoder and then mapped to a low-dimensional latent variable. Divide the current physical feasible region by the current theoretical maximum photovoltaic power to transform the physical feasible region into a relative maximum exponential constraint. After verifying and correcting the relative maximum exponential constraint, map it into a physical state vector. A cross-attention mechanism is introduced on the bottleneck layer of the U-net denoising network for the conditional latent diffusion model, using low-dimensional latent variables as query vectors and physical state vectors as key and value vectors. Several extreme scenarios are generated using a conditional latent diffusion model that incorporates a cross-attention mechanism.

7. The method for generating extreme photovoltaic power output scenarios based on physical guided diffusion according to claim 1, characterized in that, The conditional latent diffusion model introduces a physical out-of-bounds penalty into the overall objective function during training; The overall objective function The formula is: ； ； in, The loss is used to reconstruct the distribution of all training samples, with a value representing the mean squared error. To impose penalties for exceeding the limits, As a physical boundary violation penalty, The number of training samples, For the current extreme scenario The actual power of photovoltaics, and Each is the current The physical upper and lower limits of power allowed at any given time. for The theoretical maximum power of photovoltaic power at any given time.

8. The method for generating extreme photovoltaic power output scenarios based on physical guided diffusion according to claim 1, characterized in that, The process involves progressively clustering the generated extreme scenarios using a dynamic confidence threshold to obtain several sets of extreme scenarios, specifically: Construct a temporal convolutional network consisting of several residual blocks, each residual block comprising a causal dilated convolutional layer with an exponentially increasing dilation coefficient. The generated extreme scenarios are input into the temporal convolutional network to extract features for each extreme scenario. Initialize cluster centers and confidence thresholds; The cluster centers are updated through several iterations, and the cluster centers of the last iteration and the extreme scenarios for each cluster are output. In each iteration, the features of each extreme scenario are calculated and compared with the current cluster centers. The comprehensive distance is calculated, and based on the comprehensive distance, the extreme scene features belonging to the cluster centers are calculated using the Student t-distribution kernel function. The algorithm calculates the soft-segment confidence level; it then determines whether the current soft-segment confidence level is greater than the confidence threshold. If it is, it updates the cluster centers using the current extreme scenario features; otherwise, it temporarily ignores the current extreme scenario features and updates the cluster centers in the current iteration. Finally, it determines whether the current iteration round meets the clustering convergence condition. If it does, it outputs the updated cluster centers and the extreme scenarios for each cluster. If it does not, it determines whether the current iteration round has reached the preset threshold update cycle. If it does, it dynamically reduces the confidence threshold according to the preset decay coefficient; otherwise, it maintains the current confidence level for the next iteration.

9. The method for generating extreme photovoltaic power output scenarios based on physical guided diffusion according to claim 8, characterized in that, Calculating extreme scene features belongs to cluster centers The formula for the soft-segment reliability is: ； in, Features of extreme scenarios Belongs to cluster center Soft-score configuration confidence, Features of extreme scenarios With cluster center The combined distance includes Euclidean distance and Pearson distance; Features of extreme scenarios With the Cluster centers The overall distance The weights are the weights corresponding to the overall distance.

10. A photovoltaic power output extreme scenario generation system based on physical guided diffusion, characterized in that, include: The feature reconstruction module is used to obtain the time series of the theoretical maximum power of photovoltaics and the actual power of photovoltaics, and to construct a dimensionless exponential sequence based on the ratio of the two. The seed library generation module is used to construct the rate of change curve of the dimensionless exponential sequence, screen extreme fluctuation segments, and generate an extreme seed database. The physical feasible region generation module is used to calculate the dynamic physical allowable boundary that slides over time based on the actual photovoltaic power time series and the obtained maximum ramp rate of the inverter, and to delineate the physical feasible region. The scene generation module is used to take the seed data from the physical feasible domain and the extreme seed database as dual conditions and input them into the conditional latent diffusion model that introduces a cross-attention mechanism to generate several extreme scenes. The scene identification module is used to progressively cluster several generated extreme scenes using a dynamic confidence threshold to obtain several sets of extreme scene classes.

Images