Controllable explainable new energy station operation scene generation method based on improved cWGAN-GP

Abstract

The invention relates to the technical field of power systems, and discloses a controllable explainable new energy station operation scene generation method based on improved cWGAN-GP, and the method comprises the steps: obtaining historical operation data and corresponding meteorological data of a new energy station, and carrying out the preprocessing of the data; constructing an external control vector and an internal control vector based on the preprocessed data; the external control vector and the internal control vector are coded and fused through an auxiliary network; constructing a wind and light scene generation model of the improved cWGAN-GP; performing multi-stage training on a wind and light scene generation model by using the historical operation data and the corresponding meteorological data; and on the basis of the trained wind and light scene generation model, inputting noise and a condition vector, and generating a new energy output scene according with specified features of the condition vector. The method has the advantages that the physical rationality and engineering practicability of a generated scene are remarkably improved, and the defects of an existing method in the aspects of controllability, interpretability, space-time fidelity and training stability are overcome on the whole.

Description

Technical Field

[0001] This invention relates to the field of power system technology, and in particular to a method for generating controllable and interpretable new energy power plant operation scenarios based on an improved cWGAN-GP. Background Technology

[0002] With the large-scale integration of renewable energy sources such as wind and solar power into the power grid, their inherent intermittency and uncertainty pose severe challenges to the planning, scheduling, and operation of the power system. To address these challenges, scenario generation technology is widely used to simulate renewable energy output, supporting grid risk assessment, optimized scheduling, and operational decision-making.

[0003] Existing scene generation methods mainly include: traditional probabilistic model methods, such as Monte Carlo methods, Gaussian distribution, and Latin hypercube sampling, which rely on assumptions about the probability distribution of wind and solar power output and are difficult to accurately characterize their high-dimensional nonlinear features; Copula function methods, which can consider spatiotemporal correlations, but are highly dependent on edge distributions and structure functions and are difficult to adapt to complex meteorological conditions; and deep generative model methods, such as Generative Adversarial Networks (GANs), Conditional Generative Adversarial Networks (CGANs), and Wasserstein GANs (WGANs), which to some extent get rid of the dependence on prior distributions and improve the generation quality.

[0004] Existing methods suffer from the following main problems: Lack of controllability: The generated scenarios are difficult to adjust according to specific meteorological conditions or operational needs, failing to meet the customized requirements of power grid dispatching; Poor interpretability: The internal features of the model are disconnected from their physical meaning, making it difficult to explain the causes of the generated results and limiting their application in power systems; Imperfect feature selection mechanism: The relationship between high-dimensional meteorological features and power output has not been effectively explored, resulting in insufficient adaptability of the generated scenarios to meteorological conditions; Insufficient spatiotemporal correlation modeling: Most methods do not fully consider the coordinated changes in wind and solar power output in time and space dimensions, leading to deviations between the generated scenarios and actual operation; Poor training stability: Traditional GANs suffer from problems such as pattern collapse and training non-convergence, affecting the quality and reliability of generation. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for generating controllable and interpretable new energy power station operation scenarios based on improved cWGAN-GP.

[0006] The objective of this invention is achieved through the following technical solution: a method for generating controllable and interpretable new energy power plant operation scenarios based on an improved cWGAN-GP, the method comprising,

[0007] Acquire historical operational data and corresponding meteorological data of new energy power stations, and preprocess the data;

[0008] Based on the preprocessed data, an external control vector and an internal control vector are constructed. The external control vector consists of key meteorological features selected based on the correlation coefficient between meteorological data and renewable energy output data, while the internal control vector consists of statistical features of renewable energy output data. The external and internal control vectors are then encoded and fused using an auxiliary network.

[0009] A landscape scene generation model based on an improved cWGAN-GP is constructed, which includes a generator and a discriminator.

[0010] Using the historical operational data and corresponding meteorological data, the landscape scene generation model is trained in multiple stages. The multi-stage training includes at least a pre-training stage and a joint adversarial training stage, in which a gradient penalty term is introduced.

[0011] Based on the trained wind and solar scene generation model, input noise vector and target control vector to generate new energy power output scene that conforms to the characteristics specified by the target control vector.

[0012] Specifically, the external control vector uses the Spearman correlation coefficient to calculate the correlation between various meteorological characteristics and new energy output data, and selects meteorological characteristics with key meteorological factors greater than a preset threshold as the external control vector; the calculation formula for the key meteorological factors is as follows:

[0013] ;

[0014] In the formula, Let i be the i-th meteorological feature value; The mean of the meteorological characteristic values; This represents the i-th wind and solar power output value; The average output of wind and solar power; This represents the total number of data points.

[0015] The internal control vector includes the maximum power P. max Minimum power P min and maximum fluctuation power P fluct At least one of them.

[0016] Specifically, the auxiliary network fuses the external control vector and the internal control vector into a high-dimensional condition vector. The formula is:

[0017] ;

[0018] In the formula, For convolutional layers and ReLU activation functions; It is a fully connected layer; It is a high-dimensional conditional vector; This is the condition vector.

[0019] Specifically, the objective function of the landscape scene generation model is:

[0020] ;

[0021] In the formula, Let be the cross-entropy function of the game between the discriminator D and the generator G; For the discriminator based on the control vector The output sample for the real data x; To determine the distribution of real data Calculate the expectation of a sample x; S is the sum of the values ​​of the noise distribution. Calculate the expectation of the samples above; The generator is based on the input noise z and the condition vector The generated samples; These are random interpolation points between real and generated samples. , It is a random number; This is the gradient penalty coefficient; To calculate the expectation of the distribution of the interpolated sample x; The discriminator output is relative to the input. The gradient.

[0022] Specifically, the loss function of the landscape scene generation model is:

[0023] ;

[0024] ;

[0025] In the formula, For discriminator loss; For generator loss; It is the gradient penalty coefficient; This is a gradient penalty term; It is random noise; It is a conditional vector; Samples generated by the generator; This represents the probability that the discriminator judges the authenticity and condition matching degree of the generated sample.

[0026] Specifically, the generator's input consists of noise and a conditional vector, which is injected into each feature layer of the generator using a conditional batch normalization method.

[0027] ;

[0028] In the formula, and The parameters are obtained by learning the conditional vector c through the network; x is the sample. and These are the mean and variance of the current batch of samples, respectively.

[0029] Specifically, this also includes constructing a regional-time matrix from multi-regional wind and solar power output data:

[0030] ;

[0031] In the formula, N is the number of regions; T is the length of the time series; The output value of the i-th region at the j-th time point;

[0032] A two-dimensional convolutional neural network is used to extract spatiotemporal features. The convolutional kernel slides uniformly in both time and space dimensions to capture local spatiotemporal patterns.

[0033] ;

[0034] Introducing spatiotemporal consistency loss:

[0035] ;

[0036] In the formula, This is the time-related loss; This represents spatial correlation loss; and These are the temporal and spatial correlation coefficients, respectively.

[0037] The cross-channel feature fusion mechanism integrates temporal, spatial, and meteorological features.

[0038] Specifically, the pre-training stage trains the generator separately to ensure that the generator can generate stable and convergent results; and at the same time trains the feature extraction to extract key meteorological features and spatiotemporal correlation information.

[0039] The joint training phase involves adversarial training between the generator and the discriminator, introducing gradient penalties to ensure the gradient stability of the generator and the discriminator; at the same time, an auxiliary network is used to encode and fuse the external control vector and the internal control vector.

[0040] The fine-tuning stage generates a controllable optimization of the scene using control vectors.

[0041] Specifically, this also includes constructing a comprehensive evaluation index system to assess training effectiveness, including deterministic evaluation indicators and probabilistic evaluation indicators;

[0042] The deterministic evaluation indicators include:

[0043] Friedman distance:

[0044] ;

[0045] Mean absolute error:

[0046] ;

[0047] Root mean square error:

[0048] ;

[0049] In the formula, The distribution of real data in a certain feature space; To generate the distribution of data in the same feature space; and These are the mean and covariance of the real sample, respectively; and These are the mean and covariance of the generated samples, respectively. This is the trace operation of a matrix; This represents the number of training samples; This is the actual value; To generate a value;

[0050] Probabilistic evaluation indicators include:

[0051] Quantile loss:

[0052] ;

[0053] ;

[0054] Continuous graded probability scoring:

[0055] ;

[0056] In the formula, This represents the number of test samples; Let u be the quantile loss of the i-th sample at quantile u; u is the cumulative distribution function value. Let be the predicted value of sample i at quantile u.

[0057] The present invention has the following advantages:

[0058] This invention utilizes an improved cWGAN-GP framework, combined with a conditional vector fused from meteorological features and power output statistics, and innovatively employs an auxiliary network for high-dimensional encoding and conditional batch normalization techniques. This enables the high-quality, fine-grained, and controllable generation of new energy power output scenarios. Simultaneously, through region-time matrix modeling and spatiotemporal consistency constraints, the invention effectively maintains the real correlation of power output scenarios in the temporal and spatial dimensions, significantly improving the physical rationality and engineering practicality of the generated scenarios. Overall, it addresses the shortcomings of existing methods in terms of controllability, interpretability, spatiotemporal fidelity, and training stability. Attached Figure Description

[0059] Figure 1 This is a schematic diagram of the cWGAN-GP model structure of the present invention;

[0060] Figure 2 This is the auxiliary network fusion vector indicator diagram of the present invention;

[0061] Figure 3 This is a schematic diagram of spatiotemporal feature fusion according to the present invention;

[0062] Figure 4 This is a schematic diagram of the model training process of the present invention;

[0063] Figure 5 This is a schematic diagram illustrating the generation of a wind and solar power output scenario according to the present invention;

[0064] Figure 6 This is a schematic diagram of a typical daily wind power output curve;

[0065] Figure 7 A schematic diagram of a typical daily photovoltaic power output curve;

[0066] Figure 8 This is a schematic diagram of the generator loss variation curve;

[0067] Figure 9 This is a schematic diagram of the discriminator loss variation curve;

[0068] Figure 10 A schematic diagram illustrating the power output of the generated landscape scene;

[0069] Figure 11 This is a schematic diagram of the spatial correlation coefficients of GANs;

[0070] Figure 12 This is a schematic diagram of the spatial correlation coefficient of CGAN;

[0071] Figure 13 This is a schematic diagram of the spatial correlation coefficient of WGAN;

[0072] Figure 14 A schematic diagram of the spatial correlation coefficients of DCGAN;

[0073] Figure 15 This is a schematic diagram of the spatial correlation coefficient of cWGAN-GP. Detailed Implementation

[0074] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention; that is, the described embodiments are merely some embodiments of the invention, and not all embodiments. The components of the embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0075] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0076] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0077] The present invention will be further described below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description.

[0078] like Figures 1 to 15 As shown, a method for generating controllable and interpretable new energy power plant operation scenarios based on improved cWGAN-GP is presented. This method includes:

[0079] Acquire historical operational data and corresponding meteorological data of new energy power stations, and preprocess the data;

[0080] Based on the preprocessed data, an external control vector and an internal control vector are constructed. The external control vector consists of key meteorological features selected based on the correlation coefficient between meteorological data and renewable energy output data, while the internal control vector consists of statistical features of renewable energy output data. The external and internal control vectors are then encoded and fused using an auxiliary network.

[0081] The external control vector uses the Spearman correlation coefficient to calculate the correlation between various meteorological features and new energy output data, and selects meteorological features with key meteorological factors greater than a preset threshold as the external control vector; the calculation formula for the key meteorological factors is as follows:

[0082] ;

[0083] In the formula, Let be the i-th meteorological characteristic value, such as wind speed, light intensity, temperature, etc. The mean of the meteorological characteristic values; This represents the i-th wind and solar power output value; The average output of wind and solar power; The total number of data points represents the number of samples used in the calculation;

[0084] The internal control vector includes the maximum power P. max Minimum power P min and maximum fluctuation power P fluct At least one of them.

[0085] ;

[0086] To enhance the model's understanding and response to internal and external control vectors, this invention designs an auxiliary network as a high-level encoding and fusion module for internal and external control vectors. Its core function is:

[0087] Dynamic feature mapping: Mapping internal and external control vectors to a high-dimensional semantic space to enhance their expressive power.

[0088] Nonlinear relationship modeling: learning the complex nonlinear relationship between internal and external control vectors and target output.

[0089] Weight adaptation: dynamically adjusts the weights of different features in the generation process through a feedback mechanism;

[0090] The auxiliary network fuses the external control vector and the internal control vector into a high-dimensional condition vector. The formula is:

[0091] ;

[0092] In the formula, For convolutional layers and ReLU activation functions; It is a fully connected layer; It is a high-dimensional conditional vector; The condition vector is composed of the concatenation of the internal control vector and the external control vector.

[0093] A landscape scene generation model based on an improved cWGAN-GP is constructed, which includes a generator and a discriminator.

[0094] The objective function of the landscape scene generation model is:

[0095] ;

[0096] In the formula, Let be the cross-entropy function of the game between the discriminator D and the generator G; For the discriminator based on the control vector The output sample for the real data x; To determine the distribution of real data The expected value of sample x is calculated to reflect the discriminator's output on the real samples; S is the expression for the noise distribution. The expected value of the generated samples reflects the discriminator's ability to make judgments on the generated samples; The generator is based on the input noise z and the condition vector The generated samples; These are random interpolation points between real and generated samples. , It is a random number; This is the gradient penalty coefficient; To calculate the expectation of the distribution of the interpolated sample x, which is used to measure the deviation of the discriminator gradient; The discriminator output is relative to the input. The gradient.

[0097] A gradient penalty mechanism is employed to improve training stability. The generator G takes random noise z and a conditional vector c as input. By fusing information from both, it generates new energy output data consistent with the conditional vector c. The discriminator takes a sample x (real data or data generated by generator G) and the conditional vector c as input. It needs to determine whether sample x is data that fits the current load conditions based on both the authenticity of the sample and its matching degree with the conditional vector c. The model's loss function is designed with two parts: the generator loss and the discriminator loss. and discriminator loss The generator loss guides the generator G to generate data that the discriminator D identifies as a real sample conforming to the condition vector c. The discriminator loss guides the discriminator D to accurately distinguish between real and generated samples and ensures that the judgment is strongly correlated with the condition vector c. Additionally, a gradient penalty term is introduced to ensure the gradient stability of the discriminator D, avoid gradient anomalies (such as gradient explosion and vanishing) during training, and improve the model's training stability. The loss function of the landscape scene generation model is:

[0098] ;

[0099] ;

[0100] In the formula, For discriminator loss; For generator loss; It is the gradient penalty coefficient; This is a gradient penalty term; It is random noise; It is a conditional vector; Samples generated by the generator; This represents the probability that the discriminator judges the authenticity and condition matching degree of the generated sample.

[0101] The generator takes noise and a conditional vector as input. Conditional batch normalization (CBN) injects the conditional vector *c* into each feature layer of the generator. Traditional GANs typically only concatenate the conditional vector *c* and noise *z* at the input layer when generating conditional samples. This coarse-grained fusion method results in the influence of conditional information (the information contained in the high-dimensional conditional vector) gradually diminishing and weakening with increasing network depth during the forward propagation of the generator network, leading to insufficient control over the generation process. CBN technology aims to inject conditional information into each feature layer of the generator, thereby achieving comprehensive and deep-level conditional control from coarse to fine granular.

[0102] ;

[0103] In the formula, and The parameters are obtained by learning the conditional vector c through the network; x is the sample. and These are the mean and variance of the current batch of samples, respectively.

[0104] CBN introduces the condition vector c into each batch normalization (BatchNorm) layer of the generator. It uses a small neural network (usually a fully connected layer) to dynamically generate the normalization parameters—scaling factors—for the current layer from the condition vector c. and offset factor This means that for different conditional vectors c, the feature distribution of each layer of the generator will be scaled and shifted differently. The conditional information c directly determines the scale and location of the activation of neurons in each layer, thus continuously and powerfully regulating the feature generation process throughout the forward propagation of the entire network. This achieves fine-grained control from global statistical features to local details.

[0105] The generator receives not only noise z, but also a condition vector. The batch-normalized condition vector is:

[0106]

[0107] In the formula, h represents the output feature of the convolutional layer; , For the reason Scaling and bias parameters generated by a linear layer; and These represent the mean and variance of the current batch.

[0108] The auxiliary network W shares convolutional layer parameters with the discriminator D, but its final layer is a separate fully connected layer used to output a high-dimensional conditional vector. Its training objective is to minimize the following auxiliary loss:

[0109]

[0110] This loss ensures that the statistical characteristics of the generated samples (such as mean and variance) approximate the target value specified by the control vector, thereby achieving accurate and controllable generation.

[0111] This also includes constructing a region-time matrix from multi-regional wind and solar power output data:

[0112] ;

[0113] In the formula, N is the number of regions; T is the length of the time series; The output value of the i-th region at the j-th time point;

[0114] A two-dimensional convolutional neural network is used to extract spatiotemporal features. The convolutional kernel slides uniformly in both time and space dimensions to capture local spatiotemporal patterns.

[0115] ;

[0116] Introducing spatiotemporal consistency loss:

[0117] ;

[0118] In the formula, This is the time-related loss; This represents spatial correlation loss; and These are the temporal and spatial correlation coefficients, respectively.

[0119] Then, a cross-channel feature fusion mechanism is adopted to integrate multi-regional information and enhance the spatial collaborative representation capability. First, temporal and spatial features are concatenated through two sampling channels. Then, the key meteorological condition features extracted from the text are convolved with the spatiotemporally concatenated features through a third sampling channel using a 1×1 convolution operation to achieve the fusion of spatiotemporal and weather features, and finally output a complete feature sequence. By constructing a regional-time matrix from the wind and solar power output sequence, spatiotemporal features are extracted through convolution and deconvolution operations. Finally, a cross-channel feature fusion mechanism is adopted to integrate multi-regional information, thereby enhancing the spatial collaborative representation capability of the feature sequence.

[0120] Using the historical operational data and corresponding meteorological data, the landscape scene generation model is trained in multiple stages. The multi-stage training includes at least a pre-training stage and a joint adversarial training stage, in which a gradient penalty term is introduced.

[0121] The pre-training phase trains the generator separately to ensure that the generator can generate stable and convergent results; and at the same time trains the feature extraction to extract key meteorological features and spatiotemporal correlation information.

[0122] The joint training phase involves adversarial training between the generator and the discriminator, introducing gradient penalties to ensure the gradient stability of the generator and the discriminator; at the same time, an auxiliary network is used to encode and fuse the external control vector and the internal control vector.

[0123] The fine-tuning stage generates a controllable optimization of the scene using control vectors;

[0124] The RMSprop optimizer is used to assist training and improve training efficiency. The optimizer learning rate is 0.0001, the gradient penalty coefficient is 10, the batch size is 32, and the number of training epochs is 3000.

[0125] To evaluate the training effect, a comprehensive evaluation index system is constructed, including deterministic evaluation indexes and probabilistic evaluation indexes. In order to eliminate the randomness of the results, this invention trains ten times and calculates the average value of all training results as the final evaluation index.

[0126] The deterministic evaluation indicators include:

[0127] Friedman distance, a metric used to measure the similarity between generated data and real data:

[0128] ;

[0129] Mean absolute error, a metric used to evaluate the error between generated data and training data:

[0130] ;

[0131] Root mean square error (RMSE) is a metric that represents the sample standard deviation of the difference between the generated data and the training data.

[0132] ;

[0133] In the formula, The distribution of real data in a certain feature space; To generate the distribution of data in the same feature space; and These are the mean and covariance of the real sample, respectively; and These are the mean and covariance of the generated samples, respectively. This is the trace operation of a matrix, which is the sum of the diagonal elements; This represents the number of training samples; This is the actual value; The generated value is FID; when FID approaches 0, it indicates that the generated sample is similar to the distribution of the real sample, so the smaller the FID, the better.

[0134] Probabilistic evaluation indicators include:

[0135] Quantile loss:

[0136] ;

[0137] ;

[0138] Continuous graded probability scoring:

[0139] ;

[0140] In the formula, This represents the number of test samples; Let u be the quantile loss of the i-th sample at quantile u; u is the cumulative distribution function value. Let be the predicted value of sample i at quantile u.

[0141] FID and PS are dimensionless metrics that measure the similarity of generated scenes in feature distribution and the degree of pattern matching, respectively. MAE, RMSE, and CRPS are all calculated based on normalized data, and the results are per-unit values, which can effectively reflect the similarity, error degree, and generation accuracy between the generated scene and the real data.

[0142] Based on the trained wind and solar scene generation model, input noise vector and target control vector to generate new energy power output scene that conforms to the characteristics specified by the target control vector.

[0143] To verify the effectiveness of the method of the present invention, actual wind and solar power output data and corresponding meteorological data for a total of three years from January 1, 2023 to December 31, 2023 in a certain province in the northwest of China were selected for experimental verification. The dataset covers 6 wind and solar power generation stations, including 3 wind farms and 3 photovoltaic power stations, with a geographical distribution radius of about 50 kilometers, which can effectively reflect the spatio-temporal distribution characteristics of regional wind and solar resources.

[0144] In the time dimension, the data acquisition time resolution is 1h, with 24 time points per day, fully covering the cycle of day and night alternation and weather changes. A total of about 17,520 effective data records were obtained, which were divided into a training set (0.8 years) and a test set (0.2 years) in chronological order to ensure the seasonal coverage integrity of model training and testing.

[0145] In terms of meteorological data, 9 key meteorological elements such as temperature, humidity, wind speed, wind direction, solar radiation intensity, and precipitation were collected, all from the automatic meteorological stations supporting each station. The data quality control adopts a three-level verification mechanism: first, outlier detection is carried out, and obvious error data are removed based on the 3σ principle; second, the spatio-temporal KNN interpolation method is used to fill in missing values; finally, data standardization processing is carried out to eliminate the influence of dimension.

[0146] As Figure 6 and Figure 7 , it can be seen from the typical daily output curves that the wind power output does not have obvious distribution characteristics, while the photovoltaic power output is strictly restricted by the sunshine time and reaches the peak at noon. This characteristic provides an optimization space for power grid dispatching, but also poses higher requirements for the accuracy of scenario generation.

[0147] Based on the improved cWGAN-GP framework proposed in the present invention, a controllable and interpretable wind and solar scenario generation model was constructed, and the deep fusion of multi-level conditional information of meteorological characteristics and output characteristics was achieved through conditional batch normalization technology.

[0148] The generator network adopts a deep convolutional architecture. The input layer receives a 100-dimensional Gaussian noise and an 8-dimensional control vector (including 5 external meteorological characteristics and 3 internal power characteristics). The network contains 5 deconvolution layers, each followed by conditional batch normalization and ReLU activation function. The final output layer uses the Tanh activation function to constrain the data range between [-1,1]. The number of generator parameters reaches 4.5M, with strong feature expression ability.

[0149] The discriminator network adopts a symmetric convolutional architecture, including 4 convolutional layers and 2 fully connected layers. All convolutional layers use the LeakyReLU activation function (negative slope is 0.2) to avoid the problem of gradient disappearance. The discriminator not only judges the authenticity of the input data, but also shares the first three convolutional layer parameters with the auxiliary network to achieve collaborative optimization of feature extraction.

[0150] The auxiliary network, the core innovation of this invention, employs a unique parameter-sharing design. The first three convolutional layers completely share parameters with the discriminator, reducing model complexity; followed by two independent fully connected layers, specifically designed to learn the high-dimensional representation of the control vector. This design ensures both consistency in feature extraction and specialization in conditional coding.

[0151] The training process employs a multi-stage optimization strategy, involving a total of 1000 training epochs with a batch size of 32. The optimizer is RMSprop, with learning rates set to 0.0001 and 0.0004 for the generator and discriminator, respectively, to ensure training stability. A gradient penalty coefficient λ=10 effectively constrains the Lipschitz continuity of the discriminator.

[0152] like Figure 8 and Figure 9 As can be seen from the training curves, the model converges rapidly in the first 500 epochs, with the discriminator loss and generator loss reaching a dynamic balance. No mode collapse occurred during the entire training process, validating the effectiveness of the gradient penalty mechanism.

[0153] In the scene generation stage, four other GAN-based generative models are considered: GAN, CGAN, WGAN, and DCGAN. By comparing them with the proposed improved cWGAN-GP method, a basis is provided for selecting the optimal generative model.

[0154] During the scene generation process, each feature of the control vector (meteorological features, power output features) is adjusted sequentially, and then 100 sets of data are generated for each set of modulated control vectors for subsequent statistical evaluation.

[0155] The evaluation system employs multi-dimensional quantitative indicators, including distribution similarity indicators (FID, KS test), point prediction accuracy indicators (MAE, RMSE), probability prediction indicators (PS, CRPS), and spatiotemporal correlation indicators (SCC, ACF). All indicators are averaged through 10 repeated experiments to ensure statistical significance.

[0156] Through a comprehensive analysis of the generated scenarios and a comparison with six other GAN-based generation models—GAN, CGAN, WGAN, and DCGAN—the proposed improved cWGAN-GP method fully validates the technical advantages of this invention. In terms of generation quality, the proposed method achieves an FID score of 15.3, significantly outperforming traditional GAN ​​(35.0), CGAN (28.5), WGAN (26.2), and DCGAN (23.6), indicating that the generated data distribution most closely approximates the real data distribution.

[0157] In terms of point prediction accuracy, the MAE and RMSE indices reached 1.42 and 2.65, respectively, representing improvements of 17.4% and 11.9% compared to DCGAN. This improvement is mainly attributed to the efficient encoding of the control vector by the auxiliary network and the fine adjustment of the conditional batch normalization, which enabled the generated scene to better capture the subtle changes in actual output while maintaining statistical characteristics.

[0158] Table 1: Comparison of Overall Performance of Different Methods

[0159]

[0160] In terms of probabilistic prediction performance, the PS and CRPS indices reached 0.48 and 0.50 respectively, representing improvements of 12.7% and 15.2% compared to DCGAN. This indicates that the generated scenario not only has high point prediction accuracy but also performs well in probability distribution fitting, providing a more reliable basis for power system risk assessment.

[0161] like Figures 11-15 The ability to maintain spatiotemporal correlation is a significant advantage of this invention. The temporal correlation coefficient of the generated scene reaches 0.91, and the spatial correlation coefficient reaches 0.88, both significantly higher than the comparison methods. This indicates that the model has successfully learned the spatiotemporal variation patterns of wind and solar power output, and the generated scene maintains a high degree of consistency in both time and space dimensions.

[0162] The comparison between the average power, maximum power, minimum power, average fluctuating power and actual indicators of wind power shows that the average power control error is 3.7%, the maximum power control error is 4.2%, the minimum power control error is 5.1%, and the fluctuating power control error is 5.3%. The overall error is small, which reflects the accuracy and stability of the present invention in terms of characteristic controllability.

[0163] Table 2 Comparison of Explainable Eigenvectors and Actual Indicators for Wind Power

[0164]

[0165] In summary, this invention, through an improved cWGAN-GP framework and innovative auxiliary network design, achieves significant improvements in all aspects of wind and solar scenario generation, providing a reliable technical means for the optimized operation of power systems with high proportions of renewable energy integration. This method not only generates high-quality data with strong controllability but also possesses good interpretability and practicality, showing broad prospects for widespread application.

[0166] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the present invention. Any person skilled in the art can make many possible variations and modifications to the technical solution of the present invention, or modify it into equivalent embodiments, without departing from the scope of the present invention. Therefore, any modifications, equivalent changes, and alterations made to the above embodiments based on the technology of the present invention without departing from the scope of the present invention are within the protection scope of the present invention.

Claims

1. A method for generating controllable and interpretable new energy power plant operation scenarios based on improved cWGAN-GP, characterized by: The method includes, Acquire historical operational data and corresponding meteorological data of new energy power stations, and preprocess the data; Based on the preprocessed data, an external control vector and an internal control vector are constructed. The external control vector consists of key meteorological features selected based on the correlation coefficient between meteorological data and renewable energy output data, while the internal control vector consists of statistical features of renewable energy output data. The external and internal control vectors are then encoded and fused using an auxiliary network. A landscape scene generation model based on an improved cWGAN-GP is constructed, which includes a generator and a discriminator. Using the historical operational data and corresponding meteorological data, the landscape scene generation model is trained in multiple stages. The multi-stage training includes at least a pre-training stage and a joint adversarial training stage, in which a gradient penalty term is introduced. Based on the trained wind and solar scene generation model, input noise and condition vectors to generate new energy power output scenes that meet the characteristics specified by the condition vectors.

2. The method for generating controllable and interpretable new energy power station operation scenarios based on improved cWGAN-GP according to claim 1, characterized in that: The external control vector uses the Spearman correlation coefficient to calculate the correlation between various meteorological features and new energy output data, and selects meteorological features with key meteorological factors greater than a preset threshold as the external control vector; the calculation formula for the key meteorological factors is as follows: ； In the formula, Let i be the i-th meteorological feature value; The mean of the meteorological characteristic values; This represents the i-th wind and solar power output value; The average output of wind and solar power; This represents the total number of data points. The internal control vector includes the maximum power P. max Minimum power P min and maximum fluctuation power P fluct At least one of them.

3. The method for generating controllable and interpretable new energy power station operation scenarios based on improved cWGAN-GP according to claim 1, characterized in that: The formula for encoding and fusing the external control vector and the internal control vector through the auxiliary network is as follows: ； In the formula, For convolutional layers and ReLU activation functions; It is a fully connected layer; It is a high-dimensional conditional vector; This is the condition vector.

4. The method for generating controllable and interpretable new energy power station operation scenarios based on improved cWGAN-GP according to claim 3, characterized in that: The objective function of the landscape scene generation model is: ； In the formula, Let be the cross-entropy function of the game between the discriminator D and the generator G; For the discriminator based on the control vector The output sample for the real data x; To determine the distribution of real data Calculate the expectation of a sample x; S is the sum of the values ​​of the noise distribution. Calculate the expectation of the samples above; The generator is based on the input noise z and the condition vector The generated samples; These are random interpolation points between real and generated samples. , It is a random number; This is the gradient penalty coefficient; To calculate the expectation of the distribution of the interpolated sample x; The discriminator output is relative to the input. The gradient.

5. The method for generating controllable and interpretable new energy power station operation scenarios based on improved cWGAN-GP according to claim 1, characterized in that: The loss function of the landscape scene generation model is: ； ； In the formula, For discriminator loss; For generator loss; It is the gradient penalty coefficient; This is a gradient penalty term; It is random noise; It is a conditional vector; Samples generated by the generator; This represents the probability that the discriminator judges the authenticity and condition matching degree of the generated sample.

6. The method for generating controllable and interpretable new energy power station operation scenarios based on improved cWGAN-GP according to claim 1, characterized in that: The generator takes noise and a conditional vector as input, and the conditional vector is injected into each feature layer of the generator using a conditional batch normalization method. ； In the formula, and The parameters are obtained by learning the conditional vector c through the network; x is the sample. and These are the mean and variance of the current batch of samples, respectively.

7. The method for generating controllable and interpretable new energy power station operation scenarios based on improved cWGAN-GP according to claim 1, characterized in that: This also includes constructing a region-time matrix from multi-regional wind and solar power output data: ； In the formula, N is the number of regions; T is the length of the time series; The output value of the i-th region at the j-th time point; A two-dimensional convolutional neural network is used to extract spatiotemporal features. The convolutional kernel slides uniformly in both time and space dimensions to capture local spatiotemporal patterns. ； Introducing spatiotemporal consistency loss: ； In the formula, This is the time-related loss; This represents spatial correlation loss. and These are the temporal and spatial correlation coefficients, respectively. The cross-channel feature fusion mechanism integrates temporal, spatial, and meteorological features.

8. The method for generating controllable and interpretable new energy power station operation scenarios based on improved cWGAN-GP according to claim 1, characterized in that: The generator is trained separately during the pre-training phase to ensure that the generator can produce stable and convergent results. Simultaneously, feature extraction is trained to extract key meteorological features and spatiotemporal correlation information; The joint training phase involves adversarial training between the generator and the discriminator, introducing gradient penalties to ensure the gradient stability of the generator and the discriminator; at the same time, an auxiliary network is used to encode and fuse the external control vector and the internal control vector. The fine-tuning stage generates a controllable optimization of the scene using control vectors.

9. The method for generating controllable and interpretable new energy power station operation scenarios based on improved cWGAN-GP according to claim 7, characterized in that: It also includes constructing a comprehensive evaluation index system to evaluate the training effect, including deterministic evaluation indicators and probabilistic evaluation indicators; The deterministic evaluation indicators include: Friedman distance: ； Mean absolute error: ； Root mean square error: ； In the formula, The distribution of real data in a certain feature space; To generate the distribution of data in the same feature space; and These are the mean and covariance of the real sample, respectively; and These are the mean and covariance of the generated samples, respectively. This is the trace operation of a matrix; This represents the number of training samples; This is the actual value; To generate a value; Probabilistic evaluation indicators include: Quantile loss: ； ； Continuous graded probability scoring: ； In the formula, This represents the number of test samples; Let u be the quantile loss of the i-th sample at quantile u; u is the cumulative distribution function value. Let be the predicted value of sample i at quantile u.

Images