CN-121997699-A - Controllable interpretable new energy station operation scene generation method based on improvement cWGAN-GP

CN121997699ACN 121997699 ACN121997699 ACN 121997699ACN-121997699-A

Abstract

The invention relates to the technical field of power systems, and discloses a controllable interpretable new energy station operation scene generation method based on improvement cWGAN-GP, which comprises the steps of acquiring historical operation data and corresponding meteorological data of a new energy station, and preprocessing the data; the method comprises the steps of constructing an external control vector and an internal control vector based on preprocessed data, encoding and fusing the external control vector and the internal control vector through an auxiliary network, constructing a wind-light scene generating model of cWGAN-GP, performing multi-stage training on the wind-light scene generating model by utilizing historical operation data and corresponding meteorological data, inputting noise and condition vectors based on the wind-light scene generating model after training, and generating a new energy output scene conforming to the specified characteristics of the condition vectors. The method has the advantages that the physical rationality and engineering practicability of the generated scene are obviously improved, and the defects of the existing method in the aspects of controllability, interpretability, space-time fidelity and training stability are integrally overcome.

Inventors

Zhang zhongdan
CAO ZHE
HE YALAN
WANG XIAOWEN
WANG TAO
XU MIN
WANG XIANG
NI LEI
ZHANG JUN
DING KUN
LI HAIBO
JIANG JINGRUI

Assignees

国网甘肃省电力公司经济技术研究院

Dates

Publication Date: 20260508
Application Date: 20251205

Claims (9)

1. The controllable interpretable new energy station operation scene generation method based on the improvement cWGAN-GP is characterized by comprising the following steps of, Acquiring historical operation data and corresponding meteorological data of a new energy station, and preprocessing the data; Constructing an external control vector and an internal control vector based on the preprocessed data, wherein the external control vector is a key meteorological feature screened according to the correlation coefficient of the meteorological data and the new energy output data, and the internal control vector is a statistical feature of the new energy output data; constructing a wind-light scene generation model based on an improved cWGAN-GP, wherein the wind-light scene generation model comprises a generator and a discriminator; performing multi-stage training on the wind-solar scene generation model by utilizing the historical operation data and the corresponding meteorological data, wherein the multi-stage training at least comprises a pre-training stage and a combined countermeasure training stage, and introducing a gradient punishment item in the combined countermeasure training stage; And (3) based on the wind-light scene generation model after training, inputting noise and a condition vector, and generating a new energy output scene conforming to the specified characteristics of the condition vector.
2. The method for generating the controllable and interpretable new energy station operation scene based on the improvement cWGAN-GP of claim 1, wherein the external control vector calculates the correlation between each meteorological feature and new energy output data by adopting a Szelman correlation coefficient, and selects the meteorological features with the key meteorological factors larger than a preset threshold as the external control vector, wherein the calculation formula of the key meteorological factors is as follows: ; in the formula, The i weather characteristic value; is the average value of the weather characteristic values; The value is the ith wind and light output value; is the average value of wind and light output; Total number of data points; The internal control vector includes at least one of a maximum power P max , a minimum power P min , and a maximum fluctuating power P fluct .
3. The method for generating the controllable and interpretable new energy station operation scene based on the improvement cWGAN-GP of claim 1, wherein 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, Activating functions for the convolution layer and the ReLU; is a full connection layer; Is a high-dimensional condition vector; Is a conditional vector.
4. The method for generating the controllable and interpretable new energy station operation scene based on the improvement cWGAN-GP according to claim 3, wherein the objective function of the scene generation model is: ; in the formula, Cross entropy functions for gaming between the arbiter D and the generator G; based on control vector for a discriminator A sample output to the real data x; to distribute real data Sample x above is expected, S is the noise distribution Obtaining expectations by the samples; the generator generates a vector according to the input noise z and the condition A generated sample; For a random interpolation point between the real sample and the generated sample, , Is a random number; penalty coefficients for gradients; for the purpose of expecting the distribution of the interpolated samples x; for determining output relative to input Is a gradient of (a).
5. The method for generating the controllable and interpretable new energy station operation scene based on the improvement cWGAN-GP according to claim 1 is characterized in that the loss function of the wind-solar scene generation model is as follows: ; ; in the formula, Loss for the arbiter; Generator loss; Is a gradient penalty coefficient; a gradient penalty term; is random noise; Is a condition vector; a generator-generated sample; the probability of judging the authenticity of the generated sample and the condition matching degree is determined for the discriminator.
6. The method for generating the controllable and interpretable new energy station operating scene based on the improvement cWGAN-GP as claimed in claim 1, wherein the input of the generator is noise and condition vectors, and the condition vectors are injected into each feature layer of the generator through a condition batch normalization method: ; in the formula, And The parameter is obtained by the condition vector c through network learning, x is a sample; And The mean and variance of the current batch of samples, respectively.
7. The method for generating the controllable and interpretable new energy station operating scene based on the improvement cWGAN-GP of claim 1, further comprising the steps of constructing multi-zone wind-solar power data into a zone-time matrix: ; wherein N is the number of regions, T is the time sequence length; A force output value of the ith area at the jth time point; Extracting space-time characteristics by adopting a two-dimensional convolutional neural network, uniformly sliding a convolutional kernel in time and space dimensions, and capturing a local space-time mode: ; Introducing a space-time consistency loss: ; in the formula, Is a loss of time correlation; is a spatial correlation loss; And Respectively the time and space correlation coefficients; The cross-channel feature fusion mechanism integrates time, space and meteorological condition features.
8. The method for generating the controllable and interpretable new energy station operation scene based on the improvement cWGAN-GP of claim 1 is characterized in that the pre-training stage is used for independently training a generator to ensure that the generator can generate a stable convergence result; The combined training stage is used for carrying out continuous deterioration countermeasure training on the generator and the discriminator, introducing gradient penalty to ensure gradient stability of the generator and the discriminator, and simultaneously encoding and fusing an external control vector and an internal control vector through an auxiliary network; And the fine tuning stage performs controllable optimization generation of the scene through the control vector.
9. The method for generating the controllable and interpretable new energy station operation scene based on the improvement cWGAN-GP of claim 7, further comprising constructing a comprehensive evaluation index system to evaluate training effects, wherein the comprehensive evaluation index system comprises a deterministic evaluation index and a probabilistic evaluation index; the deterministic evaluation index includes: friedman distance: ; average absolute error: ; Root mean square error: ; in the formula, The distribution of the real data in a certain characteristic space is realized; to generate a distribution of data under the same feature space; And Respectively mean and covariance of the real samples; And Respectively generating a mean value and a covariance of the sample; performing trace operation for the matrix; the number of training samples; Is an actual value; To generate a value; The probabilistic evaluation index includes: Quantile loss: ; ; Continuous hierarchical probability scoring: ; in the formula, For the number of test samples; The index loss of the ith sample under the index point u is calculated, wherein u is an accumulated distribution function value; is the predicted value of sample i at quantile u.

Description

Controllable interpretable new energy station operation scene generation method based on improvement cWGAN-GP Technical Field The invention relates to the technical field of power systems, in particular to a controllable interpretable new energy station operation scene generation method based on improvement cWGAN-GP. Background With the large-scale access of renewable energy sources such as wind energy, solar energy and the like to a power grid, the inherent intermittence and uncertainty of the renewable energy sources bring serious challenges to the planning, scheduling and operation of a power system. To address these challenges, scenario generation techniques are widely used to simulate renewable energy output, support grid risk assessment, optimize scheduling and operational decisions. The existing scene generation method mainly comprises a traditional probability model method, such as a Monte Carlo method, gaussian distribution, latin hypercube sampling and the like, and is difficult to accurately describe high-dimensional nonlinear characteristics of wind and light output probability distribution depending on the assumption of the wind and light output probability distribution, a Copula function method, such as a generation countermeasure network (GAN), a condition generation countermeasure network (CGAN), a WASSERSTEIN GAN (WGAN) and the like, wherein the Copula function method can consider space-time correlation, but has strong dependence on edge distribution and structural functions and is difficult to adapt to complex meteorological conditions, and the depth generation model method is free of dependence on priori distribution to a certain extent, so that the generation quality is improved. The method has the following main problems that a generation scene is difficult to adjust according to specific meteorological conditions or operation requirements, customized requirements of power grid dispatching cannot be met, interpretability is poor, internal characteristics and physical meanings of a model are disjointed, causes of a generation result are difficult to interpret, application of the method in a power system is limited, a characteristic screening mechanism is imperfect, the relation between high-dimensional meteorological characteristics and output is not effectively mined, adaptability of the generation scene to the meteorological conditions is insufficient, space-time correlation modeling is insufficient, a plurality of methods do not fully consider cooperative change rules of wind-solar output in time and space dimensions, deviation exists between the generation scene and actual operation, training stability is poor, the problems of mode collapse, non-convergence of training and the like exist in a traditional GAN, and generation quality and reliability are affected. Disclosure of Invention The invention aims to overcome the defects of the prior art and provides a controllable interpretable new energy station operation scene generation method based on improvement cWGAN-GP. The invention aims at realizing the technical scheme that the method for generating the controllable interpretable new energy station operation scene based on the improvement cWGAN-GP comprises the following steps of, Acquiring historical operation data and corresponding meteorological data of a new energy station, and preprocessing the data; Constructing an external control vector and an internal control vector based on the preprocessed data, wherein the external control vector is a key meteorological feature screened according to the correlation coefficient of the meteorological data and the new energy output data, and the internal control vector is a statistical feature of the new energy output data; constructing a wind-light scene generation model based on an improved cWGAN-GP, wherein the wind-light scene generation model comprises a generator and a discriminator; performing multi-stage training on the wind-solar scene generation model by utilizing the historical operation data and the corresponding meteorological data, wherein the multi-stage training at least comprises a pre-training stage and a combined countermeasure training stage, and introducing a gradient punishment item in the combined countermeasure training stage; based on the wind-light scene generation model after training, a noise vector and a target control vector are input, and a new energy output scene conforming to the specified characteristics of the target control vector is generated. Specifically, the external control vector calculates the correlation between each meteorological feature and new energy output data by adopting a spearman correlation coefficient, and selects the meteorological features with key meteorological factors larger than a preset threshold as the external control vector, wherein the calculation formula of the key meteorological factors is as follows: ; in the formula, The i weather characteristic value; is the average value