An anti-learning-based wind power scenario set construction method, system and device for short-circuit capacity evaluation

Abstract

The application discloses a wind power scenario set construction method, system and equipment based on an adversarial learning for short-circuit capacity evaluation, and relates to the field of power systems; training data is acquired; a neural network model is constructed; day-ahead prediction power data and random noise data of a wind power plant in the training data are vertically spliced as input of the neural network model, and day-ahead real power data of the wind power plant is taken as output of the neural network model, and the neural network model is trained by using a novel adversarial learning method; current prediction power data and random noise data of the wind power plant are vertically spliced and input into the trained neural network model, and a wind power scenario set is obtained. The application can improve the construction quality of a wind power output scenario set.

Description

Technical Field

[0001] This invention relates to the field of power systems, and in particular to a method, system, and device for constructing wind power scenario sets based on adversarial learning for short-circuit capacity assessment. Background Technology

[0002] With the increasing depletion of fossil fuels and the intensifying greenhouse effect, the transformation from a traditional carbon-dominated energy structure to a clean and low-carbon energy system has become imperative. Therefore, my country has proposed the strategic goal of "dual carbon," with the energy sector serving as the main battleground for carbon emission reduction, and the power industry as the main force in achieving this goal. Developing clean and low-carbon energy and power systems can effectively reduce carbon dioxide emissions and make a positive contribution to achieving the "dual carbon" objective.

[0003] To achieve the low-carbon transformation of the energy and power system, vigorously developing renewable energy and replacing traditional coal-fired power generation with clean energy generation can not only alleviate the energy crisis but also promote the low-carbon process of the energy system. While wind power and other new energy sources bring zero-carbon electricity to the energy system, the randomness and volatility of their output also pose serious challenges to the safe and stable operation of the power system. Short-circuit capacity, as a key indicator for measuring the grid's voltage support capability and fault current level, is an important basis for system planning, protection settings, and operation mode arrangements. After large-scale wind power grid connection, the output prediction error will cause uncertainty deviations in the system's operating state, which will directly affect the accuracy of short-circuit capacity assessment. This makes it difficult for traditional assessment methods based on deterministic operating conditions to truly reflect the system's short-circuit withstand capability under complex operating conditions. Therefore, how to effectively characterize the uncertainty of wind power output and construct a set of wind power output scenarios that can cover the actual operating boundary has become an urgent need to improve the credibility of short-circuit capacity assessment.

[0004] Currently, there is considerable research on methods for constructing wind power output scenario sets. However, most of these studies are based on statistical methods, assuming that the prediction error of wind power follows a known probability distribution. However, historical wind power output data exhibits high-dimensional nonlinear characteristics, making statistical modeling difficult and hindering the acquisition of a suitable probability distribution for accurate modeling of wind power prediction errors. This results in the generated scenario sets often failing to adequately reflect the actual output fluctuation range, thus affecting the effective coverage of short-circuit capacity assessment under the worst or critical operating conditions. Furthermore, traditional adversarial learning, which uses cross-entropy loss functions, suffers from gradient vanishing during backpropagation, making network training difficult and causing pattern collapse, further limiting the diversity and representativeness of the generated scenarios. Therefore, there is an urgent need to develop a scenario set construction method that can adaptively learn the high-dimensional distribution of wind power output and achieve stable and efficient training, to support accurate scenario generation for short-circuit capacity assessment. Summary of the Invention

[0005] The purpose of this invention is to provide a method, system, and device for constructing wind power scenario sets based on adversarial learning for short-circuit capacity assessment, which can improve the construction quality of wind power output scenario sets.

[0006] To achieve the above objectives, the present invention provides the following solution: A method for constructing a wind power scenario set based on adversarial learning includes: Acquire training data; the training data includes the day-ahead predicted power data and the day-ahead actual power data of the wind farm; Build a neural network model; The predicted day-ahead power data and random noise data of wind farms in the training data are vertically concatenated as the input of the neural network model, and the actual day-ahead power data of wind farms are used as the output of the neural network model. A novel adversarial learning method is used to train the neural network model. The current predicted power data and random noise data of the wind farm are vertically concatenated and input into the trained neural network model to obtain the wind power scenario set.

[0007] Optionally, the current predicted power data and random noise data of the wind farm are concatenated vertically and input into the trained neural network model to obtain a wind power scenario set, which then includes: The prediction error of wind power is characterized based on a set of wind power scenarios.

[0008] Optionally, the day-ahead predicted power data and random noise data of the wind farm in the training data are vertically concatenated as input to the neural network model, and the day-ahead actual power data of the wind farm is used as output to the neural network model. A novel adversarial learning method is used to train the neural network model, specifically including: Constructing generative neural networks and discriminative neural networks; The day-ahead predicted power data and random noise data of wind farms in the training data are vertically concatenated as input to the generative neural network, and the day-ahead real power data of wind farms are used as output to the generative neural network. The generative neural network is trained with the aim of minimizing the first loss function to obtain the trained generative neural network; the first loss function is calculated based on the output of the discriminative neural network. The day-ahead predicted power data and random noise data of wind farms in the training data are vertically concatenated and input into the trained generative neural network to obtain the generated power data. The day-ahead predicted power data of wind farms in the training data is used as the conditional information of the discriminant neural network. The generated power data, the actual power data and the day-ahead predicted power data of wind farms are vertically concatenated as the input of the discriminant neural network. The discriminant neural network is trained with the goal of minimizing the second loss function to obtain the trained discriminant neural network. The second loss function is calculated based on the output of the trained generator neural network.

[0009] Optionally, the first loss function is: ; The second loss function: ; in, Represents the first loss function; This represents the second loss function; Indicates the expected value; Indicates the expected value; This represents the output value of the discriminant neural network; This represents the output value of the discriminant neural network; D This represents a discriminative neural network; G This represents a generative neural network.

[0010] Optionally, the generative neural network includes a first input layer, a first convolutional layer, a first normalization layer, a second convolutional layer, a second normalization layer, a third convolutional layer, a third normalization layer, and a first fully connected layer connected in sequence.

[0011] A wind power scene set construction system based on adversarial learning is applied to the aforementioned wind power scene set construction method based on adversarial learning. The wind power scene set construction system based on adversarial learning includes: The acquisition module is used to acquire training data, which includes the day-ahead predicted power data and the day-ahead actual power data of the wind farm. Modules for building neural network models; The training module is used to vertically concatenate the day-ahead predicted power data and random noise data of wind farms in the training data as the input of the neural network model, and use the day-ahead real power data of wind farms as the output of the neural network model. A novel adversarial learning method is used to train the neural network model. The wind power scene set determination module is used to vertically concatenate the current predicted power data and random noise data of the wind farm and input them into the trained neural network model to obtain the wind power scene set.

[0012] Optionally, the wind power scenario set construction system based on adversarial learning further includes: The characterization module is used to characterize the prediction error of wind power based on the wind power scenario set.

[0013] Optionally, the training module specifically includes: Building units are used to construct generative and discriminative neural networks; A generative neural network training unit is used to vertically concatenate the day-ahead predicted power data and random noise data of the wind farm in the training data as the input of the generative neural network, and use the day-ahead real power data of the wind farm as the output of the generative neural network. The generative neural network is trained with the aim of minimizing the first loss function to obtain the trained generative neural network. The first loss function is calculated based on the output of the discriminative neural network. The power data generation unit is used to vertically concatenate the day-ahead predicted power data and random noise data of the wind farm in the training data and input them into the trained generative neural network to obtain the generated power data. The discriminative neural network training unit is used to take the day-ahead predicted power data of the wind farm in the training data as the conditional information of the discriminative neural network, and to vertically concatenate the generated power data, the actual power data and the day-ahead predicted power data of the wind farm as the input of the discriminative neural network. The discriminative neural network is trained with the aim of minimizing the second loss function to obtain the trained discriminative neural network. The second loss function is calculated based on the output of the trained generative neural network.

[0014] Optionally, the first loss function is: ; The second loss function is: ; in, Represents the first loss function; This represents the second loss function; Indicates the expected value; Indicates the expected value; This represents the output value of the discriminant neural network; This represents the output value of the discriminant neural network; D This represents a discriminative neural network; G This represents a generative neural network.

[0015] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the wind power scenario set construction method based on adversarial learning as described above.

[0016] A computer-readable storage medium having a computer program stored thereon, which is executed by a processor to implement the wind power scenario set construction method based on adversarial learning as described above.

[0017] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: This invention discloses a method, system, and device for constructing wind power scenario sets based on adversarial learning for short-circuit capacity assessment. The method uses a novel adversarial learning method to train the neural network model by vertically concatenating the day-ahead predicted power data and random noise data of the wind farm in the training data, and using the day-ahead actual power data of the wind farm as the output of the neural network model. The current predicted power data and random noise data of the wind farm are then vertically concatenated and input into the trained neural network model to obtain the wind power scenario set. This invention can improve the construction quality of wind power output scenario sets. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of the wind power scenario set construction method based on adversarial learning in an embodiment of the present invention; Figure 2 This is a basic structural diagram of traditional adversarial learning in an embodiment of the present invention; Figure 3 This is a structural diagram of the generative neural network of the novel adversarial learning method in this embodiment of the invention; Figure 4 This is a structural diagram of the discriminative neural network of the novel adversarial learning method in this embodiment of the invention; Figure 5 This is a waveform diagram of the training data in an embodiment of the present invention; Figure 6 This is a schematic diagram of the generative neural network of the novel adversarial learning method in an embodiment of the present invention; Figure 7 This is a schematic diagram of the discriminative neural network of the novel adversarial learning method in an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] The purpose of this invention is to provide a method, system, and device for constructing wind power scenario sets based on adversarial learning for short-circuit capacity assessment, which can improve the construction quality of wind power output scenario sets.

[0022] This invention collects data from renewable energy power plants, including historical and predicted power output data, and performs data cleaning on this data. It uses a novel adversarial learning method instead of traditional statistical and adversarial learning methods, feeding predicted wind power output information as conditional information into a deep neural network for training. This method can uncover the high-dimensional nonlinear characteristics of historical renewable energy power output data, avoiding the difficulties of statistical model building and the training mode collapse problem of traditional adversarial learning. This invention provides a new approach to the problem of constructing wind power output scenario sets and is applicable to various renewable energy power plants in the power system, possessing strong practical significance and serving as an important method for constructing wind power output scenario sets.

[0023] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0024] Example 1

[0025] like Figure 1 As shown, the wind power scenario set construction method based on adversarial learning provided by this invention includes: Step 101: Obtain training data, the training data as follows: Figure 5 As shown. The training data includes the day-ahead predicted power data and the day-ahead actual power data of the wind farm.

[0026] Data collection is conducted for new energy power plants, including day-ahead forecast power data and day-ahead actual power data of wind farms. The above data is then cleaned, including data normalization, outlier filtering, replacement, and blank data filling.

[0027] Based on the processed new energy output data, test sets and training sets will be divided for a novel scenario set construction method based on adversarial learning.

[0028] like Figure 1 As shown, step 102: Construct a neural network model.

[0029] First, the current state of development of adversarial learning methods is collected and summarized to ensure that the method proposed in this invention has a high degree of compatibility and practicality for the future development of power systems. The adversarial learning method consists of two independent deep neural networks: a generative neural network and a discriminative neural network.

[0030] Generative neural networks (GNNs) map random noise into generated data by learning the latent distribution of historical wind power output data, while discriminative neural networks (DNNs) try to determine whether the input data is real or generated. Through the game-like learning between the GNN and the DNN, the realism of the samples generated by the GNN is improved.

[0031] Secondly, based on the problems of adversarial learning methods and the characteristics of wind power output scenario set composition, a novel adversarial learning method is proposed. In both the generative and discriminative neural networks of this novel adversarial learning method, the least squares loss function is used as the first and second loss functions. By penalizing samples far from the decision boundary, more gradients are provided, alleviating the gradient vanishing problem and promoting the generation of higher-quality samples. Then, an evaluation metric is constructed to measure the scenario generation method. The generated scenario set needs to conform to the characteristics of factors affecting wind power output prediction error to the greatest extent possible.

[0032] like Figure 6 As shown, the generative neural network includes a first input layer, a first convolutional layer, a first normalization layer, a second convolutional layer, a second normalization layer, a third convolutional layer, a third normalization layer, and a first fully connected layer connected in sequence.

[0033] like Figure 7 As shown, the discriminative neural network includes a second input layer, a fourth convolutional layer, a fourth normalized layer, a fifth convolutional layer, a fifth normalized layer, a sixth convolutional layer, a sixth normalized layer, and a second fully connected layer connected in sequence.

[0034] Based on a novel adversarial learning approach, the day-ahead predicted power data of wind farms is used as conditional information. This data is then vertically concatenated with random noise data and fed into a generative deep neural network for training. After training, the generative neural network learns the mapping relationship between the noise distribution and the output distribution that satisfy the conditional information. The trained generative neural network can then generate wind farm power data.

[0035] Step 103: Vertically concatenate the day-ahead predicted power data and random noise data of the wind farm in the training data as the input of the neural network model, and use the day-ahead real power data of the wind farm as the output of the neural network model. Use a novel adversarial learning method to train the neural network model.

[0036] Construct generative neural networks and discriminative neural networks.

[0037] The day-ahead predicted power data and random noise data of wind farms in the training data are vertically concatenated as input to the generative neural network, and the day-ahead real power data of wind farms are used as output to the generative neural network. The generative neural network is trained with the aim of minimizing the first loss function to obtain the trained generative neural network. The first loss function is calculated based on the output of the discriminant neural network.

[0038] The day-ahead predicted power data and random noise data of wind farms in the training data are vertically concatenated and input into the trained generative neural network to obtain the generated power data.

[0039] The day-ahead predicted power data of wind farms in the training data is used as the conditional information of the discriminant neural network. The generated power data, the actual power data and the day-ahead predicted power data of wind farms are vertically concatenated as the input of the discriminant neural network. The discriminant neural network is trained with the goal of minimizing the second loss function to obtain the trained discriminant neural network. The second loss function is calculated based on the output of the trained generator neural network.

[0040] like Figure 2 As shown, the traditional adversarial learning method is an unsupervised model proposed by Goodfellow in 2014, consisting of two independent deep neural networks: a generative neural network and a discriminative neural network. The generative neural network maps random noise data to generated data by fitting a nonlinear function to historical wind power output data, while the discriminative neural network tries to determine whether the input data is real or generated. Through the game-like learning between the generative and discriminative neural networks, the authenticity of the samples generated by the generative neural network is improved.

[0041] For ease of description, use Represents a real data sample. Represents the true data distribution. Represents random noise data. This represents the distribution of noisy data. During training, a batch of samples... The samples are fed into a generative neural network to generate samples. It follows a probability distribution The training objective of the generator is to make... as much as possible with same.

[0042] The training objective of the discriminative neural network is to correctly determine whether the input data is a real sample or a generated sample. After the adversarial training described above, the generative and discriminative neural networks reach a Nash equilibrium, meaning the discriminative neural network can no longer determine whether its input sample is a real sample or a generated sample. At this point, it can be considered that the generative neural network has learned the probability distribution of real samples.

[0043] Based on the training objectives of generative and discriminative neural networks, the loss functions of generative and discriminative neural networks are... and The definition is as follows.

[0044] (1) (2) In the formula: This represents the expected value of a batch of samples.

[0045] Based on (1) and (2), the objective function of traditional adversarial learning can be designed as shown in (3).

[0046] (3)

[0047] However, the loss function of discriminative neural networks in traditional adversarial learning is equivalent to the Jensen-Shannon (JS) divergence. When the generated data distribution and the real data distribution do not overlap in all dimensions, the JS divergence is always very close to a constant. At this point, the JS divergence reaches saturation, and the loss of the discriminative neural network becomes zero. Ultimately, due to the use of the cross-entropy loss function, traditional adversarial learning suffers from the gradient vanishing problem during backpropagation, making network training difficult and causing pattern collapse.

[0048] To address the aforementioned issues, an adversarial learning method using ab encoding and a least-squares loss function was proposed. The objective function of this adversarial learning is shown in (4).

[0049] (4)

[0050] In the formula: a and b Labels representing generated data and real data, respectively; c This indicates that the generative neural network aims to determine the trust value of the neural network in the generated data.

[0051] In generative and discriminative neural networks using least-squares adversarial learning methods, the least-squares loss function is used to provide more gradients by penalizing samples that are far from the decision boundary, thus mitigating the vanishing gradient problem. Simultaneously, the loss function is beneficial for generating higher-quality samples because it penalizes spurious samples and "pulls" them towards the decision boundary.

[0052] Furthermore, to generate data specific to a particular event, conditional information is incorporated into the generative and discriminative neural networks of the least-squares adversarial learning method. Therefore, the first loss function of the novel adversarial learning method is: (5) The second loss function is: (6) in, Represents the first loss function; This represents the second loss function; Indicates the expected value; Indicates the expected value; This represents the output value of the discriminant neural network; This represents the output value of the discriminant neural network; D This represents a discriminative neural network; G This represents a generative neural network.

[0053] This invention uses predicted wind power output data as conditional information and feeds both the predicted and actual data into a novel adversarial learning method for training. After training, the generative neural network of this novel adversarial learning method can learn the nonlinear functional relationship between noisy data and output data that satisfies the predicted information. The structure of the conditional least squares generative adversarial network's generative and discriminative neural networks is as follows: Figure 3 and Figure 4 As shown.

[0054] Step 104: Vertically concatenate the current predicted power data and random noise data of the wind farm and input them into the trained neural network model to obtain the wind power scenario set.

[0055] Random noise data is data from different scenarios. The random noise data from different scenarios and the current predicted power data of the wind farm are vertically concatenated and input into the trained neural network model to obtain wind power data in different scenarios. Wind power scenario sets are constructed from the wind power data in different scenarios.

[0056] The prediction error of wind power is characterized based on a set of wind power scenarios. The dispatch center can use a trained generative neural network to characterize the prediction error of wind power output.

[0057] The generated scene set needs to conform to the characteristics of wind power output prediction errors to the greatest extent possible. Therefore, this invention also provides two evaluation metrics to measure the quality of the generated scene set: coverage and envelope area.

[0058] The novel adversarial learning-based method for constructing wind power output scenarios was tested using wind power output test data. The performance of the method on the wind power output scenario construction problem was measured by coverage and envelope area.

[0059] Coverage aims to evaluate whether the generated scene set can cover the actual power generation of wind power. Envelope area aims to evaluate whether the generated scene set can cover the actual power generation data with the smallest possible envelope area. Finally, an example of the novel adversarial learning-based wind power scene set construction method is run to test the feasibility and superiority of the proposed method in the wind power scene set construction problem.

[0060] It is important to note that these two metrics cannot be used alone as evaluation criteria; they must be used together. Generally, the larger the coverage and the smaller the envelope area, the better the method's performance. The specific calculation method is as follows: (7) (8) In the formula: Indicates coverage rate; Indicates the area of ​​the envelope; Indicates the number of time periods; Indicates the output of the generated scene; Indicates actual effort exerted; n Indicates the first n A time period; This indicates whether the actual output is within the output range of the generated scene.

[0061] Example 2

[0062] A wind power scene set construction system based on adversarial learning is applied to the wind power scene set construction method based on adversarial learning described in Example 1. The wind power scene set construction system based on adversarial learning includes: The acquisition module is used to acquire training data, which includes the day-ahead predicted power data and the day-ahead actual power data of the wind farm.

[0063] Modules for building neural network models.

[0064] The training module is used to vertically concatenate the day-ahead predicted power data and random noise data of wind farms in the training data as input to the neural network model, and use the day-ahead real power data of wind farms as output to the neural network model. A novel adversarial learning method is used to train the neural network model.

[0065] Building units are used to construct generative and discriminative neural networks.

[0066] The generative neural network training unit is used to vertically concatenate the day-ahead predicted power data and random noise data of the wind farm in the training data as the input of the generative neural network, and use the day-ahead real power data of the wind farm as the output of the generative neural network. The generative neural network is trained with the aim of minimizing the first loss function to obtain the trained generative neural network. The first loss function is calculated based on the output of the discriminative neural network.

[0067] The power data generation unit is used to vertically concatenate the day-ahead predicted power data and random noise data of the wind farm in the training data and input them into the trained generative neural network to obtain the generated power data.

[0068] The discriminative neural network training unit is used to take the day-ahead predicted power data of the wind farm in the training data as the conditional information of the discriminative neural network, and to vertically concatenate the generated power data, the actual power data and the day-ahead predicted power data of the wind farm as the input of the discriminative neural network. The discriminative neural network is trained with the aim of minimizing the second loss function to obtain the trained discriminative neural network. The second loss function is calculated based on the output of the trained generative neural network.

[0069] The first loss function is: ; The second loss function is: ; in, Represents the first loss function; This represents the second loss function; Indicates the expected value; Indicates the expected value; This represents the output value of the discriminant neural network; This represents the output value of the discriminant neural network; D This represents a discriminative neural network; G This represents a generative neural network.

[0070] The wind power scene set determination module is used to vertically concatenate the current predicted power data and random noise data of the wind farm and input them into the trained neural network model to obtain the wind power scene set.

[0071] As one embodiment, the wind power scenario set construction system based on adversarial learning further includes: The characterization module is used to characterize the prediction error of wind power based on the wind power scenario set.

[0072] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the wind power scenario set construction method based on adversarial learning as described in Embodiment 1.

[0073] A computer-readable storage medium having a computer program stored thereon, which is executed by a processor to implement the wind power scenario set construction method based on adversarial learning as described in Example 1.

[0074] This invention can effectively characterize the prediction error of wind power by generating wind power scenario sets. It avoids the problems of traditional adversarial learning methods, such as difficulty in constructing suitable probability distribution models, model collapse during training, and the inability to fully exploit the high-dimensional nonlinear characteristics of historical wind power output data. It provides new suggestions for solving the problem of wind power output prediction error in power systems and helps improve the security of power system dispatch. Therefore, this invention has significant practical guiding significance and application value.

[0075] The superior effects of the technical solution provided by this invention are: 1. Data collection for new energy scenarios, including historical output data and forecast data, is performed. This data is then cleaned, including data normalization, anomaly filtering, replacement, and blank data filling. This invention meets the actual operational requirements of power systems, ensuring its safety and practicality.

[0076] 2. A novel adversarial learning method is used to generate scenarios that characterize wind power output prediction errors. By fitting the nonlinear function between the noisy data under the prediction information and the generated wind power output data scenario set through a deep neural network, the high-dimensional nonlinear characteristics of historical wind power output data can be extracted, avoiding the difficulties in modeling statistical models. A better characterization of wind power prediction errors can ensure the stability and security of power system operation.

[0077] 3. An evaluation index for the wind power output prediction error scenario set is constructed to measure the superiority of the proposed novel adversarial learning-based method in addressing the problem of constructing wind power output prediction error scenario sets. Coverage aims to assess whether the generated scenario set can cover the actual power generation of the wind power output data. Envelope area aims to assess whether the generated scenario set can cover the actual power generation data of wind power with the smallest possible envelope area.

[0078] 4. Using the least squares loss function as the loss function of the novel adversarial learning method can alleviate the problem of gradient vanishing in traditional adversarial learning methods and generate higher quality samples.

[0079] 5. Design a novel deep neural network structure for adversarial learning. To generate data under specific conditions, conditional information is incorporated into the generative and discriminative neural networks of the novel adversarial learning method. This enables the generative neural network of the novel adversarial learning method to fit the nonlinear functional relationship between noise data and output data that satisfies the wind power output prediction information.

[0080] 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.

[0081] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for constructing a wind power scenario set based on adversarial learning, characterized in that, include: Acquire training data; the training data includes the day-ahead predicted power data and the day-ahead actual power data of the wind farm; Build a neural network model; The predicted day-ahead power data and random noise data of wind farms in the training data are vertically concatenated as the input of the neural network model, and the actual day-ahead power data of wind farms are used as the output of the neural network model. A novel adversarial learning method is used to train the neural network model. The current predicted power data and random noise data of the wind farm are vertically concatenated and input into the trained neural network model to obtain the wind power scenario set.

2. The method for constructing a wind power scenario set based on adversarial learning according to claim 1, characterized in that, The current predicted power data and random noise data of the wind farm are concatenated vertically and input into the trained neural network model to obtain the wind power scenario set, which then includes: The prediction error of wind power is characterized based on a set of wind power scenarios.

3. The method for constructing a wind power scenario set based on adversarial learning according to claim 1, characterized in that, The day-ahead predicted power data and random noise data of wind farms in the training data are vertically concatenated as input to the neural network model, and the day-ahead actual power data of wind farms are used as the output of the neural network model. A novel adversarial learning method is used to train the neural network model, specifically including: Constructing generative neural networks and discriminative neural networks; The day-ahead predicted power data and random noise data of wind farms in the training data are vertically concatenated as input to the generative neural network, and the day-ahead real power data of wind farms are used as output to the generative neural network. The generative neural network is trained with the aim of minimizing the first loss function to obtain the trained generative neural network; the first loss function is calculated based on the output of the discriminative neural network. The day-ahead predicted power data and random noise data of wind farms in the training data are vertically concatenated and input into the trained generative neural network to obtain the generated power data. The day-ahead predicted power data of wind farms in the training data is used as the conditional information of the discriminant neural network. The generated power data, the actual power data and the day-ahead predicted power data of wind farms are vertically concatenated as the input of the discriminant neural network. The discriminant neural network is trained with the goal of minimizing the second loss function to obtain the trained discriminant neural network. The second loss function is calculated based on the output of the trained generator neural network.

4. The method for constructing a wind power scenario set based on adversarial learning according to claim 3, characterized in that, The first loss function is: ； The second loss function is: ； in, Represents the first loss function; This represents the second loss function; Indicates the expected value; Indicates the expected value; This represents the output value of the discriminant neural network; This represents the output value of the discriminant neural network; D This represents a discriminative neural network; G This represents a generative neural network.

5. The method for constructing a wind power scenario set based on adversarial learning according to claim 3, characterized in that, The generative neural network includes a first input layer, a first convolutional layer, a first normalized layer, a second convolutional layer, a second normalized layer, a third convolutional layer, a third normalized layer, and a first fully connected layer connected in sequence.

6. A wind power scenario set construction system based on adversarial learning, characterized in that, The wind power scene set construction system based on adversarial learning, applicable to any one of claims 1-5, comprises: The acquisition module is used to acquire training data, which includes the day-ahead predicted power data and the day-ahead actual power data of the wind farm. Modules are used to build neural network models; The training module is used to vertically concatenate the day-ahead predicted power data and random noise data of wind farms in the training data as the input of the neural network model, and use the day-ahead real power data of wind farms as the output of the neural network model. A novel adversarial learning method is used to train the neural network model. The wind power scene set determination module is used to vertically concatenate the current predicted power data and random noise data of the wind farm and input them into the trained neural network model to obtain the wind power scene set.

7. The wind power scenario set construction system based on adversarial learning according to claim 6, characterized in that, The wind power scenario set construction system based on adversarial learning also includes: The characterization module is used to characterize the prediction error of wind power based on the wind power scenario set.

8. The wind power scenario set construction system based on adversarial learning according to claim 6, characterized in that, The training module specifically includes: Building units are used to construct generative and discriminative neural networks; A generative neural network training unit is used to vertically concatenate the day-ahead predicted power data and random noise data of the wind farm in the training data as the input of the generative neural network, and use the day-ahead real power data of the wind farm as the output of the generative neural network. The generative neural network is trained with the aim of minimizing the first loss function to obtain the trained generative neural network. The first loss function is calculated based on the output of the discriminative neural network. The power data generation unit is used to vertically concatenate the day-ahead predicted power data and random noise data of the wind farm in the training data and input them into the trained generative neural network to obtain the generated power data. The discriminative neural network training unit is used to take the day-ahead predicted power data of the wind farm in the training data as the conditional information of the discriminative neural network, and to vertically concatenate the generated power data, the actual power data and the day-ahead predicted power data of the wind farm as the input of the discriminative neural network. The discriminative neural network is trained with the aim of minimizing the second loss function to obtain the trained discriminative neural network. The second loss function is calculated based on the output of the trained generative neural network.

9. The wind power scenario set construction system based on adversarial learning according to claim 8, characterized in that, The first loss function is: ； The second loss function is: ； in, Represents the first loss function; This represents the second loss function; Indicates the expected value; Indicates the expected value; This represents the output value of the discriminant neural network; This represents the output value of the discriminant neural network; D This represents a discriminative neural network; G This represents a generative neural network.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the wind power scenario set construction method based on adversarial learning as described in any one of claims 1-5.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that, The program is executed by the processor to implement the wind power scenario set construction method based on adversarial learning as described in any one of claims 1-5.

Images