A method and system for generating ultra-short-term output scenarios for new renewable energy stations

Abstract

The present invention discloses a method and system for generating ultra-short-term output scenarios for newly built new energy stations, the method comprising: selecting a target station and a source station according to the geographical location of the target station, and further constructing a target domain data set and a source domain data set; training an improved denoising diffusion probability model on the source domain data set to obtain a source domain model; comprehensively selecting the source domain optimal model from the source domain model according to the model evaluation index; fine-tuning the model parameters on the source domain optimal model in combination with the model migration mechanism and the target domain data set to obtain multiple groups of target domain models; comprehensively selecting the target domain optimal model from the obtained multiple groups of target domain models in combination with the model evaluation index, thereby completing the generation of ultra-short-term output scenarios for newly built new energy stations. The present invention can complete the high-quality generation of output scenarios for newly built new energy stations under small sample conditions, and can effectively support the planning and scheduling decisions of subsequent new power systems, thereby improving the system's new energy absorption capacity.

Description

Technical Field

[0001] The present invention relates to a method and system for generating ultra-short-term output scenarios of new energy stations based on an improved denoising diffusion probability model and model migration, and belongs to the field of power grid operation analysis and dispatching. Background Art

[0002] With the advancement of the construction of new power systems, the penetration rate of new energy represented by photovoltaic and wind power continues to increase, and the uncertainty of their output has brought huge challenges to the safe and stable operation of the system. Therefore, accurate modeling of the randomness and uncertainty of new energy output can effectively support the planning and scheduling decisions of subsequent new power systems, thereby improving the system's new energy absorption capacity.

[0003] At present, the modeling of renewable energy output scenarios is mainly divided into two categories: model-driven probabilistic modeling methods and data-driven probabilistic modeling methods. The model-driven probabilistic modeling method assumes the uncertainty of renewable energy output as a known probability distribution, and uses historical scenario data to fit the probability model parameters, thereby generating output scenarios based on the probability distribution in combination with the sampling method. However, due to the dynamic time-varying weather and the complex spatiotemporal coupling relationship between multiple stations, the model-driven probabilistic modeling method is difficult to apply directly and has poor generalization. In addition, this type of method assumes the data uncertainty distribution as a known probability distribution, and the accuracy of the model depends on prior experience, which also limits the diversity of generated scenarios. In recent years, with the advancement of artificial intelligence algorithms, data-driven probabilistic modeling methods have received extensive attention and research in the field of power system sample generation, such as system typical operation mode generation, temporary stability sample generation, etc. Generative models based on deep neural networks have also been widely used in the field of renewable energy scenario generation.

[0004] However, due to the limitations of its own model, the variational autoencoder cannot accurately infer the log-likelihood of actual samples, resulting in low quality of generated scenes. The convergence and stability of the generative adversarial network during training are difficult to guarantee, and the generative adversarial network has mode collapse phenomenon. The diversity of generated samples is insufficient and it is difficult to cover all distributions of real data. At the same time, data-driven models based on deep neural networks require a large amount of historical data as support during the training process, which is difficult to apply to the training of neural network models in the case of insufficient data. Summary of the invention

[0005] Based on the above analysis, the present invention provides a method for generating ultra-short-term output scenarios of new energy stations based on an improved denoising diffusion probability model and model migration, which selects the target station and the source station according to the geographical location of the target station, and further constructs a target domain data set and a source domain data set; trains the improved denoising diffusion probability model on the source domain data set to obtain a source domain model; comprehensively selects the source domain optimal model from the source domain model according to the model evaluation index; fine-tunes the model parameters on the source domain optimal model in combination with the model migration mechanism and the target domain data set to obtain multiple groups of target domain models; comprehensively selects the target domain optimal model from the obtained multiple groups of target domain models in combination with the model evaluation index, thereby completing the generation of ultra-short-term output scenarios of new energy stations. The present invention can ultimately complete the construction of a scenario generation model for new energy stations under small sample conditions.

[0006] The object of the present invention is to provide a method for generating ultra-short-term output scenarios of new energy stations based on an improved denoising diffusion probability model and model migration.

[0007] The purpose of the present invention can be achieved by adopting the following technical solution: A method for generating ultra-short-term output scenarios of new energy stations based on an improved denoising diffusion probability model and model migration, comprising the following steps:

[0008] S1, select the target station and the source station according to the geographical location of the target station, and further construct the target domain dataset and the source domain dataset;

[0009] S2, train the improved denoising diffusion probability model on the source domain dataset to obtain the source domain model;

[0010] The improved denoising diffusion probability model includes an embedding module, an attention encoding module and a projection module, and its processing process is as follows: the noisy new energy output scene in the input source domain data set is first positionally encoded in the embedding module, the original input data and the position-encoded data are spliced ​​and input into the attention encoding module, and the attention feature is obtained through processing of several attention mechanisms and one-dimensional convolutional neural network layers, and finally the dimension of the attention feature is adjusted to the size of the original input through the fully connected neural network in the projection module; the diffusion step number is set , repeatedly execute the embedding module, attention encoding module and projection module Finally, a new renewable energy output scenario is generated from the renewable energy output scenario containing noise;

[0011] S3, comprehensively selecting the optimal source domain model from the source domain models based on the model evaluation indicators;

[0012] S4, combining the model transfer mechanism with the target domain dataset, fine-tuning the model parameters on the optimal model of the source domain to obtain multiple sets of target domain models;

[0013] S5, combining the model evaluation index to comprehensively select the optimal model of the target domain from the multiple groups of target domain models obtained in step S4, thereby completing the generation of the ultra-short-term output scenario of the newly built new energy station.

[0014] Furthermore, the specific implementation of the embedded module is as follows;

[0015] First, positionally encode the original input data to input the timestamp information of the data into the model:

[0016] (1)

[0017] (2)

[0018] In the formula, It is a new energy ultra-short-term output scenario with noise. is the position embedding and After splicing, is the position embedding, Is the data in the output scene The absolute position in is the data vector dimension, Represents the first Dimension.

[0019] Furthermore, the processing of the attention mechanism is as follows;

[0020] First, through three weight matrices and Multiply them together to get the query, key, and value vector needed to calculate the attention value, that is, ,in represents the output of the embedding module;

[0021] Second, calculate the attention value:

[0022] (3)

[0023] In the formula, is the normalized exponential function, represents a sparse query, Take the first u valuable value, The value of the value is passed Calculate, where , is the set constant coefficient, is the vector dimension of the input data, The calculation formula is:

[0024] (4)

[0025] In the formula, yes Middle A value, yes Middle A value, is the dimension of the feature vector after embedding.

[0026] Furthermore, the model evaluation indicators include probability density function, cumulative distribution function, time correlation and accuracy, and the final evaluation indicators are defined as:

[0027] (10)

[0028] In the formula, and The power distribution range of the real new energy output scenario set and the new energy output scenario set generated by the source domain model The indicator probability density function under and The power distribution range of the real new energy output scenario set and the new energy output scenario set generated by the source domain model The cumulative distribution function under ; and The real new energy output scenario set and the new energy output scenario set generated by the source domain model are respectively j Time correlation under j Depends on actual needs; is the coverage rate of the renewable energy output scenario set generated by the source domain model; , Generate the average width of the power interval of the new energy output scenario set for the real new energy output scenario set and the source domain model respectively; , , , , They are the weight coefficients of each indicator, which are determined according to actual needs;

[0029] The source domain model corresponding to the minimum value is the optimal model of the source domain.

[0030] Furthermore, the calculation process of the probability density function is as follows:

[0031] First, the power distribution is divided into intervals of equal width, each interval has a width of , calculate the frequency of data points in each interval, denoted as ,in , convert the frequency into probability density, that is:

[0032] (5)

[0033] In the formula, is the center point of the interval, is the number of samples in the renewable energy output scenario set generated by the source domain model, Indicates that the new energy output scenario set generated by the source domain model is The probability density value at ;

[0034] The cumulative distribution function is calculated as follows:

[0035] (6)

[0036] In the formula, Indicates that the new energy output scenario set generated by the source domain model is The cumulative distribution value.

[0037] Furthermore, the calculation formula of time correlation is as follows:

[0038] (7)

[0039] In the formula, New energy output scenarios generated for source domain models exist The output value at the moment, is the time interval, for The mean of for The variance of .

[0040] Furthermore, accuracy includes coverage and the average width of the power interval ,in The calculation formula is:

[0041] (8)

[0042] Where: Indicates the number of moments when the actual output falls into the set of renewable energy output scenarios generated by the source domain model; Indicates the total number of moments of concentrated output scenarios of new energy output scenarios generated by the source domain model;

[0043] The calculation formula is:

[0044] (9)

[0045] In the formula, It represents the average width of the power interval of the concentrated output scenario of the new energy output scenario generated by the source domain model; Indicates The upper bound of the concentrated output scenario of the new energy output scenario generated by the source domain model at each moment; Indicates The lower bound of the concentrated output scenario of the new energy output scenario generated by the source domain model at each moment.

[0046] Furthermore, let the source domain dataset and the target domain dataset be , ,but:

[0047] (11)

[0048] In the formula, , for , Samples in

[0049] Suppose the tasks of generating new energy output scenarios under the source domain dataset and the target domain dataset are:

[0050] (12)

[0051] In the formula, To generate new energy output scenarios from the source domain dataset, the model input To target output The mapping is realized by the denoising diffusion probability model. To generate new energy output scenarios in the target domain dataset from the model input To target output The mapping is realized by the denoising diffusion probability model. , are the model parameters respectively;

[0052] Model migration is to Some parameters in Migrate to the target domain, and then Learn the best , the specific formula is:

[0053] (13)

[0054] In the formula, As the loss function, Huber loss function is used. , is the optimal parameter obtained after training, Indicates that the initial value of the optimal model parameter in the source domain is the parameter obtained by random initialization , Indicates that the initial value of the target domain model parameter is the parameter migrated from the optimal model in the source domain .

[0055] Furthermore, during the parameter fine-tuning process, the parameters of the attention encoding module in the optimal model of the source domain are frozen, and only the parameters of the embedding module and the projection module are fine-tuned on the optimal model of the source domain.

[0056] The present invention also provides a system for generating ultra-short-term output scenarios for a newly built new energy station, comprising a processor and a memory, the memory being used to store program instructions, and the processor being used to call the stored instructions in the memory to execute a method for generating ultra-short-term output scenarios for a newly built new energy station as described in the above technical solution.

[0057] Compared with the prior art, the present invention has the following beneficial effects:

[0058] The present invention can complete the high-quality generation of new energy station output scenarios under small sample conditions, which can effectively support the planning and scheduling decisions of subsequent new power systems, thereby improving the system's new energy absorption capacity. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying creative work.

[0060] Figure 1 It is a schematic diagram of the process of the present invention.

[0061] Figure 2 Schematic diagram of the denoising diffusion probability model designed for the present invention.

[0062] Figure 3 This is a schematic diagram of the model migration method proposed in the present invention. DETAILED DESCRIPTION

[0063] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0064] like Figure 1The method flow diagram of the method of the present invention is shown as follows: The present invention discloses a method for generating ultra-short-term output scenarios of new energy stations based on an improved denoising diffusion probability model and model migration, comprising the following steps:

[0065] S1. Select the target station and the source station according to the geographical location of the target station, and further construct the target domain dataset and the source domain dataset;

[0066] S2. Train the improved denoising diffusion probability model on the source domain dataset;

[0067] S3. Comprehensively select the optimal model in the source domain based on model evaluation indicators;

[0068] S4. Combine the model transfer mechanism with the target domain dataset and fine-tune the model parameters based on the optimal model in the source domain;

[0069] S5. Combine the model evaluation indicators to comprehensively select the optimal model for the target domain, so as to complete the generation of ultra-short-term output scenarios for new renewable energy stations.

[0070] Step S1 specifically includes the following steps:

[0071] Determine the geographical location of the target station, the longitude range is , the latitude range is , a total of 111 stations;

[0072] According to the geographical location of the target station, conventional new energy stations with rich historical output scene data are selected as source stations, with a longitude range of , the latitude range is , a total of 99 stations.

[0073] The target station and source station data are normalized to construct the target domain dataset and the source domain dataset.

[0074] In step S2, the improved denoising diffusion probability model is trained on the source domain dataset, such as Figure 2 As shown, the present invention designs a neural network architecture for scene generation to replace the U-Net network of the original denoising probability diffusion model, which specifically includes the following steps:

[0075] Step S21: The input of the model is the noisy new energy output scene in the source domain data set, and the feature dimension is consistent with the length of the new energy output scene sequence to be generated, that is, ; After entering the model, the position is first embedded through the Embedding module, and then concatenated with the original input as the input of the Encoder module;

[0076] First, positionally encode the original input data to input the timestamp information of the data into the model:

[0077] (1)

[0078] (2)

[0079] In the formula, It is a new energy ultra-short-term output scenario with noise. After position embedding, The spliced ​​data. is the position embedding, Is the data in the output scene The absolute position in is the data vector dimension, Represents the first Dimension.

[0080] Step S22, in the Encoder module (attention encoding module), firstly, the improved self-attention mechanism is calculated in combination with the input data, and a one-dimensional convolutional neural network layer is superimposed. The calculation process is repeated multiple times in series, 5 times in this embodiment, and the final output result is used as the input of the Projection module;

[0081] Through three weight matrices and Multiply them together to get the Query, Keys, and Values ​​vectors needed to calculate the attention value, that is, .

[0082] Second, calculate the attention value:

[0083] (3)

[0084] In the formula, is the normalized exponential function, represents a sparse query. Considering that in the implementation process, the attention values ​​of the self-attention mechanism are usually sparse, not all values ​​are worth paying attention to. Therefore, The value contains only the first u valuable value to reduce the computational complexity of the model. The value of the value is passed calculate, , The calculation formula is:

[0085] (4)

[0086] In the formula, is the set constant coefficient, which is 5 in this embodiment. yes Middle A value, yes Middle A value, is the dimension of the feature vector after embedding, is the vector dimension of the input data.

[0087] Step S23, in the Projection module, the data dimension is adjusted to , to complete the noise estimation of this step; in the set diffusion step number After arriving, finally from the noise Generate new energy output scenarios .

[0088] Step S3 comprehensively selects the optimal model of the source domain based on the model evaluation index, which specifically includes the following steps:

[0089] 10,000 output scenarios were generated based on the trained model to evaluate the model performance. The indicator calculation results are as follows:

[0090] (1) Probability density function

[0091] The probability density function (PDF) is a function that describes the probability of ultra-short-term output values ​​of renewable energy near a certain value point. By calculating and drawing the PDF curve, the distribution differences of different scenario sets at different output values ​​can be intuitively compared.

[0092] First, the power distribution interval is divided into intervals of equal width, each interval has a width of . Calculate the frequency of data points in each interval, denoted as ,in . Convert the frequency to probability density, that is:

[0093] (5)

[0094] In the formula, is the center point of the interval, The number of samples in the generated new energy output scenario set, Indicates that the scene set is The probability density value at .

[0095] Through the above steps, the probability density value of the center point of each interval can be obtained, and then a probability density function curve can be drawn to compare the distribution differences between the real scene and the generated scene at different output values.

[0096] (2) Cumulative distribution function

[0097] The Cumulative Distribution Function (CDF) is the integral of the probability density function, which can fully describe the probability distribution of a set of output scenarios, thereby comparing the distribution differences between scenario sets. Its calculation formula is as follows:

[0098] (6)

[0099] In the formula, Indicates that the scene set is The cumulative distribution value.

[0100] (3) Time correlation

[0101] The output of renewable energy is affected by wind speed and sunlight, and the output at each moment has a certain correlation. The scene generated by the model needs to fit the time correlation of the real output of renewable energy. The autocorrelation coefficient is selected to evaluate the quality of the generated scene. The calculation formula is as follows:

[0102] (7)

[0103] In the formula, E represents expectation, For this scene exist The output value at the moment, is the time interval, for The mean of for The variance of .

[0104] (4) Accuracy

[0105] The scenario set generated by the model needs to be as comprehensive as possible and fit the distribution of the actual output scenario set in order to accurately model its output uncertainty and select the coverage rate. Average width of power interval Two metrics evaluate the accuracy of the model.

[0106] The calculation formula is:

[0107] (8)

[0108] Where: Indicates the number of moments when the actual output falls within the generated scenario set; Represents the total number of moments in the output sequence.

[0109] The calculation formula is:

[0110] (9)

[0111] In the formula, Indicates the average width of the power interval of the output scenario; Indicates The upper bound of the set of output scenarios at each moment; Indicates The lower bound of the set of scenarios that can be output at a given moment.

[0112] The two indicators of coverage rate and average width of power interval restrict each other. Only when the average width of power interval is as close as possible to the average width of power interval in the real scene, the high coverage rate can reflect the accuracy of the model. When the average width of power interval is much larger than the average width of power interval in the real scene, even if the coverage rate is 100%, it cannot indicate that the model accurately models the probability distribution of the real scene.

[0113] The final indicator is defined as:

[0114] (10)

[0115] In the formula, and The power distribution range of the real new energy output scenario set and the new energy output scenario set generated by the source domain model The smaller the difference between the indicator probability density function and the better the model performance is, and The power distribution range of the real new energy output scenario set and the new energy output scenario set generated by the source domain model The smaller the difference between the two, the better the model performance. and The real scene set and the model generated scene set are respectively j The smaller the difference between the two, the better the model performance. j According to actual needs, the value of j in this embodiment is 1 to 5; It is the coverage of the scene set generated by the model. The larger the coverage, the better the model. , are the average widths of the power intervals for the real scene set and the scene set generated by the model, respectively. The smaller the difference between the two, the better the model performance; , , , , They are the weight coefficients of each indicator, which are determined according to actual needs. As the final indicator, The smaller the value, the better the model performance.

[0116] In step S4, the model migration mechanism is combined with the target station data set to fine-tune the model parameters on the optimal model in the source domain, such as Figure 3 As shown, the specific steps include:

[0117] The generative model trained on ultra-short-term output scenarios of conventional renewable energy stations is used as the source domain model, and the structural parameters of the source domain model are selectively transferred to the ultra-short-term scenario generation task of newly built renewable energy stations, thereby providing common knowledge for the construction of the target domain model.

[0118] Let the ultra-short-term output scenario set of conventional new energy stations and the ultra-short-term output scenario set of newly built new energy stations be , ,but:

[0119] (11)

[0120] In the formula, , for , The samples in .

[0121] The scenario generation tasks under the ultra-short-term output of conventional new energy stations and the ultra-short-term output of newly built new energy stations are as follows:

[0122] (12)

[0123] In the formula, To generate tasks for ultra-short-term output scenarios of conventional renewable energy stations from model input To target output The mapping of To generate tasks for ultra-short-term output scenarios of new energy stations from model input To target output The mapping of , are their model parameters respectively.

[0124] Model migration is to Some parameters in Migrate to the target domain, and then Learn the best , the specific formula is:

[0125] (13)

[0126] In the formula, is the loss function, and the present invention adopts the Huber loss function, , is the optimal parameter obtained after model training, Indicates that the initial value of the optimal model parameter in the source domain is the parameter obtained by random initialization , Indicates that the initial value of the target domain model parameter is the parameter migrated from the optimal model in the source domain In order to reduce the cost of model retraining and improve model training efficiency, this paper freezes the model parameters of the Encoder module of the optimal model in the source domain, and only fine-tunes the model parameters of the Embedding module and the Projection module on the optimal model in the source domain. By adjusting the model hyperparameters, multiple groups of models are obtained to optimize the model performance.

[0127] Step S5 combines the model evaluation index to comprehensively select the optimal model of the target domain, thereby completing the generation of the ultra-short-term output scenario of the newly built new energy station, which specifically includes the following steps:

[0128] The model evaluation index refers to step S3. After selecting the optimal model in the target domain, the output scenario data can be generated.

[0129] The experimental results of the present invention on the open source wind and solar ultra-short-term output scenario dataset are as follows:

[0130] The source domain model obtained by training in the above steps generates 10,000 scenarios, which are combined with the real scene set to calculate the indicators. In order to reflect the advantages of the model, the generative adversarial network WGAN-GP and the variational autoencoder model VAE are selected for comparison. The accuracy indicators of each model are shown in Table 1. As can be seen from Table 1, the proposed model achieves a higher coverage rate while being closer to the average width of the real power interval. In the photovoltaic scenario, the average power interval width of the improved DDPM (denoising diffusion probability model) is 32.38MW, which is closer to the average power interval width of the real sample set of 36.68MW, and the coverage rate is 99.15%, which is much higher than WGAN-GP's 76.5% and VAE's 66.05%. In the wind power scenario, the improved DDPM better models the characteristic of the maximum output of 16MW in the real scene set. The coverage rate of the proposed model is 100%, which is better than WGAN-GP and VAE.

[0131] Table 1 Accuracy indicators of different models

[0132]

[0133] Based on the source domain model, the parameters are fine-tuned in combination with the target domain dataset. The optimal model in the target domain is obtained through model evaluation, and 4000 new scenarios are generated. Compared with the model without migration, the performance of each model in terms of accuracy is shown in Table 2. As shown in Table 2, in terms of accuracy indicators, in both photovoltaic and wind power scenarios, the proposed model migration method is significantly better than the comparison model. The coverage rate of the proposed method reaches nearly 100% when it is closer to the average width of the actual power range.

[0134] Table 2 Model accuracy indicators

[0135]

[0136] The above is only a preferred embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical solution and inventive concept of the present invention within the scope disclosed by the present invention, which shall fall within the protection scope of the present invention.

Claims

1. A method for generating ultra-short-term output scenarios for new energy stations, characterized in that: The steps include: S1, select the target station and the source station according to the geographical location of the target station, and further construct the target domain dataset and the source domain dataset; S2, train the improved denoising diffusion probability model on the source domain dataset to obtain the source domain model; The improved denoising diffusion probability model includes an embedding module, an attention encoding module and a projection module, and its processing process is as follows: the noisy new energy output scene in the input source domain data set is first positionally encoded in the embedding module, the original input data and the position-encoded data are spliced ​​and input into the attention encoding module, and the attention feature is obtained through processing of several attention mechanisms and one-dimensional convolutional neural network layers, and finally the dimension of the attention feature is adjusted to the size of the original input through the fully connected neural network in the projection module; the diffusion step number is set , repeatedly execute the embedding module, attention encoding module and projection module Finally, a new renewable energy output scenario is generated from the renewable energy output scenario containing noise; S3, comprehensively selecting the optimal source domain model from the source domain models based on the model evaluation indicators; The model evaluation indicators include probability density function, cumulative distribution function, time correlation and accuracy. The final evaluation indicators are defined as: (10) In the formula, and The power distribution range of the real new energy output scenario set and the new energy output scenario set generated by the source domain model The indicator probability density function under and The power distribution range of the real new energy output scenario set and the new energy output scenario set generated by the source domain model The cumulative distribution function under ; and The real new energy output scenario set and the new energy output scenario set generated by the source domain model are respectively j Time correlation under j Depends on actual needs; is the coverage rate of the renewable energy output scenario set generated by the source domain model; , Generate the average width of the power interval of the new energy output scenario set for the real new energy output scenario set and the source domain model respectively; , , , , They are the weight coefficients of each indicator, which are determined according to actual needs; The source domain model corresponding to the minimum value is the optimal model of the source domain; S4, combining the model transfer mechanism with the target domain dataset, fine-tuning the model parameters on the optimal model of the source domain to obtain multiple sets of target domain models; S5, combining the model evaluation index to comprehensively select the optimal model of the target domain from the multiple groups of target domain models obtained in step S4, thereby completing the generation of the ultra-short-term output scenario of the newly built new energy station.

2. A method for generating ultra-short-term output scenarios for a newly built new energy station as claimed in claim 1, characterized in that: The specific implementation of the embedded module is as follows; First, positionally encode the original input data to input the timestamp information of the data into the model: (1) (2) In the formula, It is a new energy ultra-short-term output scenario with noise. is the position embedding and After splicing, is the position embedding, Is the data in the output scene The absolute position in is the data vector dimension, Represents the first Dimension.

3. A method for generating ultra-short-term output scenarios for a newly built new energy station as claimed in claim 1, characterized in that: The processing of the attention mechanism is as follows; First, through three weight matrices and Multiply them together to get the query, key, and value vector needed to calculate the attention value, that is, ,in represents the output of the embedding module; Second, calculate the attention value: (3) In the formula, is the normalized exponential function, represents a sparse query, Take the first u valuable value, The value of the value is passed Calculate, where , is the set constant coefficient, is the vector dimension of the input data, The calculation formula is: (4) In the formula, yes Middle A value, yes Middle A value, is the dimension of the feature vector after embedding.

4. A method for generating ultra-short-term output scenarios for a newly built new energy station as claimed in claim 1, characterized in that: The probability density function is calculated as follows: First, the power distribution is divided into intervals of equal width, each interval has a width of , calculate the frequency of data points in each interval, denoted as ,in , convert the frequency into probability density, that is: (5) In the formula, is the center point of the interval, is the number of samples in the renewable energy output scenario set generated by the source domain model, Indicates that the new energy output scenario set generated by the source domain model is The probability density value at ; The cumulative distribution function is calculated as follows: (6) In the formula, Indicates that the new energy output scenario set generated by the source domain model is The cumulative distribution value.

5. A method for generating ultra-short-term output scenarios for a newly built new energy station as claimed in claim 1, characterized in that: The time correlation is calculated as follows: (7) In the formula, New energy output scenarios generated for source domain models exist The output value at the moment, is the time interval, for The mean of for The variance of .

6. A method for generating ultra-short-term output scenarios for a newly built new energy station as claimed in claim 1, characterized in that: Accuracy includes coverage and the average width of the power interval ,in The calculation formula is: (8) Where: Indicates the number of moments when the actual output falls into the set of renewable energy output scenarios generated by the source domain model; Indicates the total number of moments of concentrated output scenarios of new energy output scenarios generated by the source domain model; The calculation formula is: (9) In the formula, It represents the average width of the power interval of the concentrated output scenario of the new energy output scenario generated by the source domain model; Indicates The upper bound of the concentrated output scenario of the new energy output scenario generated by the source domain model at each moment; Indicates The lower bound of the concentrated output scenario of the new energy output scenario generated by the source domain model at each moment.

7. A method for generating ultra-short-term output scenarios for a newly built new energy station as claimed in claim 1, characterized in that: Let the source domain dataset and the target domain dataset be , ,but: (11) In the formula, , for , Samples in Suppose the tasks of generating new energy output scenarios under the source domain dataset and the target domain dataset are: (12) In the formula, To generate new energy output scenarios from the source domain dataset, the model input To target output The mapping is realized by the denoising diffusion probability model. To generate new energy output scenarios in the target domain dataset from the model input To target output The mapping is realized by the denoising diffusion probability model. , are the model parameters respectively; Model migration is to Some parameters in Migrate to the target domain, and then Learn the best , the specific formula is: (13) In the formula, As the loss function, Huber loss function is used. , is the optimal parameter obtained after training, Indicates that the initial value of the optimal model parameter in the source domain is the parameter obtained by random initialization , Indicates that the initial value of the target domain model parameter is the parameter migrated from the optimal model in the source domain .

8. A method for generating ultra-short-term output scenarios for a newly built new energy station as claimed in claim 1, characterized in that: During parameter fine-tuning, the parameters of the attention encoding module in the optimal model of the source domain are frozen, and only the parameters of the embedding module and the projection module are fine-tuned on the optimal model of the source domain.

9. A new ultra-short-term output scenario generation system for a new energy station, characterized by: It includes a processor and a memory, the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute a method for generating ultra-short-term output scenarios for new energy stations as described in any one of claims 1 to 8.

Images