CN-122020149-A - Method, system and equipment for enhancing and predicting power consumption small sample of extreme scene based on PC-GAN

CN122020149ACN 122020149 ACN122020149 ACN 122020149ACN-122020149-A

Abstract

A method for predicting the electric load of an extreme scene based on PC-GAN includes such steps as obtaining the electric load-meteorological data set of target region, identifying the extreme scene sample, building multi-dimensional feature vector, building physical constraint to generate an antagonistic network structure model, using multi-dimensional feature vector as input to implement three-stage constraint strategy to obtain high-quality synthesized sample, merging the extreme scene sample with high-quality synthesized sample to obtain an enhanced training data set, and implementing strategy and physical and service dual constraint mechanism by three-stage constraint and matching with scene classification auxiliary task.

Inventors

BIE FANGMEI
XU SHUANG
DING HANG
WAN JING
CHEN XI

Assignees

国网湖北省电力有限公司经济技术研究院

Dates

Publication Date: 20260512
Application Date: 20251210

Claims (10)

1. The method for enhancing and predicting the power consumption of the extreme scene by using the small sample based on the PC-GAN is characterized by comprising the following steps of: Acquiring an electricity load-meteorological data set of a target area, identifying an extreme scene sample in the electricity load-meteorological data set, and constructing a multidimensional feature vector based on the electricity load-meteorological data set; Constructing physical constraints comprising a generator, a discriminator and a scene classifier to generate a PC-GAN model of an countermeasure network architecture; Generating an antagonism network architecture PC-GAN model based on physical constraint, and implementing a three-stage constraint strategy by taking a multidimensional feature vector as input to obtain a high-quality synthetic sample; And combining the extreme scene sample and the high-quality synthetic sample into an enhanced training data set, and training a load prediction model based on the enhanced training data set to realize the electricity load prediction of the target area.
2. The PC-GAN based extreme scene electricity small sample enhancement prediction method of claim 1, characterized by: the implementation of the three-stage constraint strategy specifically comprises the following steps: Setting general physical loss, specific scene loss, characteristic distribution matching loss, time sequence consistency loss, generator total loss and discriminator total loss as constraints in a training stage, and alternately training and guiding a countering network architecture model to perform sample learning to obtain initial candidate samples; Aiming at the initial candidate sample in the generation stage, carrying out step-by-step forced light post-processing correction and dispersion processing on general physical loss, electric quantity consistency constraint and specific scene loss to obtain a post-processing candidate sample; and in the verification stage, aiming at the post-processing candidate samples, calculating the passing rate of each constraint and the comprehensive quality score to perform layered quality inspection, and obtaining a high-quality synthesized sample.
3. The PC-GAN based extreme scene electricity small sample enhancement prediction method of claim 1, characterized by: the generic physical loss includes: Temperature relationship constraints The loss function is as follows: ; Wherein: Is the lowest temperature; is the average temperature; is the highest temperature; rainfall relationship constraint The loss function is as follows: ; Wherein: Is the maximum rainfall; To accumulate rainfall; Wind speed relationship constraints The loss function is as follows: ; Wherein: Is the average wind speed; Is the maximum wind speed; humidity relationship constraints The loss function is as follows: ; Wherein: is in a humidity range; electric quantity consistency constraint The expression is as follows: ; Wherein: Is the sum of the time-sharing electric quantity; Is the daily electricity quantity; Is a time segment sequence number; The specific scene loss includes: extreme high temperature scene business constraints The loss function is as follows: ; Wherein: is the highest temperature; temperature-load related constraints The loss function is as follows: ; ; Wherein: Is the sum of time-sharing electric quantity; is a temperature sensitivity coefficient; is the humidity sensitivity coefficient; Is a temperature response function; is the reference load; Is the actual temperature; Is the reference temperature; extreme storm scene traffic constraints The loss function is as follows: ; Wherein: is daily rainfall; Rainfall-load association constraints The loss function is as follows: ; Wherein: Is a rainfall load suppression coefficient; Is that Rainfall at the moment; Is an extra electric load in a heavy rain scene; Extreme cold tide scene business constraint The loss function is as follows: ; Wherein: The temperature is reduced for 24 hours; Cold tide-load association constraint The loss function is as follows: ; Wherein: is the cold tide load coefficient; Is the reference temperature; is the influence coefficient of the temperature reduction amplitude on the load.
4. The PC-GAN based extreme scene electricity small sample enhancement prediction method of claim 3, wherein: The characteristic distribution matching loss The loss function is as follows: ; Wherein: 、 respectively the first The mean value of the real sample and the synthesized sample of the dimensional characteristics; 、 respectively the first Standard deviation of real sample and synthesized sample of dimensional characteristics; Is a feature dimension; sequence number of feature dimension; the loss of timing consistency The loss function is as follows: ; Wherein: An average load change pattern for a real sample; total loss of the generator The loss function is as follows: ; ; ; Wherein: an counterloss function for generator loss; classifying a loss function for a scene; The scene category number; Is a real scene label; a scene category sequence number; Is a mathematical expectation; Is a discriminator For false samples And conditions An output of (2); Total loss of said discriminant The loss function is as follows: ; ; Wherein: an counterloss function for the discriminator loss; Is a discriminator For real samples And conditions An output of (2); Is a mathematical expectation.
5. The PC-GAN based extreme scene electricity small sample enhancement prediction method of claim 4, wherein: aiming at an initial candidate sample in a generation stage, carrying out step-by-step forced light post-processing correction on general physical loss, electric quantity consistency constraint and specific scene loss, wherein the method specifically comprises the following steps of: the first level of correction is a forced correction for a general physical relationship, comprising: the samples violating the temperature magnitude relation are forcedly corrected, and the expression is as follows: ; ; ; Wherein: 、、 correction values of the minimum temperature, the maximum temperature and the average temperature respectively; The samples violating the wind speed relation are forcedly corrected, and the expression is as follows: ; Wherein: a correction value for the average wind speed; And (3) forcedly correcting the samples violating the rainfall relation, wherein the expression is as follows: ; Wherein: A correction value for the maximum rainfall; The samples violating the humidity range are subjected to forced correction, and the expression is as follows: ; Wherein: a correction value for the humidity range; the second-stage correction is carried out for the consistency of the electric quantity, the time-sharing electric quantity sum and the deviation between the time-sharing electric quantity sum and the daily electric quantity of each sample are calculated to carry out forced correction, and the expression is as follows: ; Wherein: Correction value for time-sharing electric quantity sum; the third stage of correction is forced correction for specific scenes, and specifically comprises the following steps: For extremely high temperature scenes, if the highest temperature is not in the range of 35-40 ℃, the forced correction is carried out, and the expression is as follows: ; Wherein: correction values for extreme high temperatures; For samples with temperature-load correlation anomalies, theoretical load values are calculated, expressed as follows: ; Wherein: Theoretical load values for temperature-load related anomalies; If the actual load deviates from the theoretical load by more than 20%, the forced correction is performed by adopting weighted average, and the expression is as follows: ; Wherein: Correction values for theoretical load values for temperature-load related anomalies; For a storm scene, if the daily cumulative rainfall is less than 50 mm, the forced correction is carried out, and the expression is as follows ; Wherein: a correction value for daily accumulated rainfall; for samples of rainfall-load correlation anomalies, theoretical load values are calculated, expressed as follows: ; Wherein: theoretical load values for rainfall-load correlation anomalies; is a rainfall saturation threshold; if the actual load deviates from the theoretical load by more than 25%, the forced correction is performed by using a weighted average, and the expression is as follows: ; Wherein: Correction values for theoretical load values for rainfall-load related anomalies; For a chill scene, ensuring that the minimum temperature is not higher than 4 ℃, the expression is as follows: ; Wherein: A correction value for the lowest temperature; For samples with cold-load-related anomalies, a theoretical load value is calculated, expressed as follows: ; Wherein: theoretical load values for cold-load related anomalies; Is the sensitivity coefficient of the cooling rate; If the actual load deviates from the theoretical load by more than 20%, the forced correction is performed by adopting weighted average, and the expression is as follows: ; Wherein: correction values for theoretical load values for cold-load related anomalies.
6. The PC-GAN based extreme scene electricity small sample enhancement prediction method of claim 5, wherein: The step of calculating the passing rate of each constraint for the post-processing candidate sample in the verification stage specifically comprises the following steps: Calculating the passing rate of the temperature relation The expression is as follows: ; Wherein: is the total number of temperature samples; calculating the rainfall relation passing rate The expression is as follows: ; Wherein: the total number of the rainfall samples; Calculating the passing rate of wind speed relation The expression is as follows: ; Wherein: The total number of wind speed samples; Calculating the passing rate of electric quantity consistency The expression is as follows: ; Wherein: the total number of the electric quantity samples; calculating the passing rate of the humidity range The expression is as follows: ; Wherein: The total number of humidity samples; Based on the temperature relationship, rainfall relationship, wind speed relationship, electric quantity consistency and humidity range passing rate, calculating the general physical relationship total passing rate The expression is as follows: ; Based on the determined extreme high temperature, heavy rain and cold tide scene service constraint, the corresponding passing rate is calculated, and the method specifically comprises the following steps: calculating service constraint passing rate of extremely high temperature scene The expression is as follows: ; Wherein: Is the total number of extremely high temperature samples; Calculating temperature-load associated pass rate The expression is as follows: ; Wherein: total number of temperature-load samples; calculating the total pass rate of service constraint of extremely high temperature scene The expression is as follows: ; calculating service constraint passing rate of extreme storm scene The expression is as follows: ; Wherein: The total number of the storm scene samples is; Calculating rainfall-load correlation passing rate The expression is as follows: ; Wherein: Total number of rainfall-load samples; calculating total pass rate of service constraint of extreme storm scene The expression is as follows: ; Calculating extreme cold tide scene business constraints The expression is as follows: ; Wherein: the total number of the cold tide scene samples is; Calculating the cold tide-load association passing rate The expression is as follows: ; Wherein: total number of cold-load samples; calculating total pass rate of extreme cold tide scene business constraint The expression is as follows: 。
7. the PC-GAN based extreme scene electricity small sample enhancement prediction method of claim 6, wherein: The calculation mode of the comprehensive quality score specifically comprises the following steps: Calculating the distribution coverage rate of generated samples and real samples in the post-processing candidate samples, dividing a feature space into grids, and counting grid cells covered by the real samples and grid cells covered by the generated samples, wherein the expression is as follows: ; Wherein: Is the distribution coverage rate; generating a set of sample covered grid cells; A set of grid cells covered for a real sample; Calculating Wasserstein distance measurement distribution similarity The expression is as follows: ; Wherein: is the distribution of the real samples; generating a distribution of samples; Is distributed in a joint way; Is that Lower real sample And generating a sample Is a distance expectation of (2); to take the minimum expected distance under all possible joint distributions; calculating matching degree of real sample and generated sample The expression is as follows: ; Wherein: the first to be a true sample A moment of order; To generate the sample of A moment of order; calculating the overall pass rate The expression is as follows: ; Wherein: is the total number of samples; Is the total number of constraints; is the first The number of times the individual samples violate the constraint; calculating a weighted comprehensive quality score based on the general physical relationship total pass rate, the specific scene constraint total pass rate, the distribution coverage rate and the matching degree The expression is as follows: ; Wherein: The total pass rate is the general physical relationship; constraining the total pass rate for a particular scene; Is a normalized index of Wasserstein distance, Distance for real samples/generated sample distribution; is the upper distance limit.
8. A PC-GAN based extreme scene electricity small sample enhancement prediction system, characterized in that it is applied to the method of any of claims 1-7, said system comprising: the multi-dimensional feature vector construction module (1) is used for acquiring an electric load-meteorological data set of a target area, identifying an extreme scene sample in the electric load-meteorological data set and constructing a multi-dimensional feature vector based on the electric load-meteorological data set; The network architecture model building module (2) is used for building physical constraints comprising a generator, a discriminator and a scene classifier to generate an antagonistic network architecture PC-GAN model; The synthetic sample acquisition module (3) is used for generating an antagonistic network architecture PC-GAN model based on physical constraint, and implementing a three-stage constraint strategy by taking the multidimensional feature vector as input to obtain a high-quality synthetic sample; And the data enhancement prediction module (4) is used for combining the extreme scene samples and the high-quality synthetic samples into an enhancement training data set and training a load prediction model based on the enhancement training data set so as to realize the power utilization load prediction of the target area.
9. An extreme scene electricity consumption small sample enhancement prediction device based on PC-GAN is characterized in that: The device comprises a processor (5) and a memory (6); -said memory (6) is adapted to store computer program code (61) and to transmit said computer program code (61) to said processor (5); The processor (5) is configured to execute the PC-GAN based extreme scene small sample enhancement prediction method of any of claims 1-7 according to instructions in the computer program code (61).
10. A computer-readable storage medium, characterized by: The computer-readable storage medium having stored therein computer-executable instructions that, when executed on a computer, implement the PC-GAN-based small sample enhancement prediction method for extreme scenes of any of claims 1-7.

Description

Method, system and equipment for enhancing and predicting power consumption small sample of extreme scene based on PC-GAN Technical Field The invention relates to a data processing means, belongs to the technical field of data processing, and particularly relates to a method, a system and equipment for enhancing and predicting an electric small sample of an extreme scene based on PC-GAN. Background In the running process of the power system, although the occurrence probability of an extreme scene is low, the influence on the safe and stable running and the power supply reliability of the power grid is extremely remarkable, the low-probability high-influence event covers various types such as extreme weather, serious emergencies, power grid faults and the like, the core characteristics are shown as data scarcity, and the training requirement of a deep learning model is difficult to meet due to the extremely small historical data sample size of the extreme scene, so that the data distribution of a normal scene and the extreme scene is in a serious unbalanced state. The traditional statistical method and the machine learning model have the problem of insufficient generalization capability under the condition of small samples, the phenomenon of over fitting is easy to occur, and the lack of the training samples in extreme scenes further leads to weak robustness of the risk prediction model, and the support for decision making is difficult to effectively provide. The simple sampling method in the prior art only carries out linear interpolation operation in a feature space, complex nonlinear features contained in extreme scenes cannot be accurately captured, the traditional generation of samples generated by an countermeasure network lacks physical interpretability and possibly contradicts with actual business logic and operation rules of an electric power system, the numerical simulation method relies on accurate physical model construction, so that the calculation cost is high, and various extreme cases are difficult to fully cover. Therefore, a means for generating high quality, physical and business constraint compliant electrosynthetic samples for extreme scenarios is needed to address the above-mentioned shortcomings of the prior art. Disclosure of Invention The invention aims to overcome the defects and problems in the prior art and provide a method, a system and equipment for enhancing and predicting small electricity consumption samples of extreme scenes based on PC-GAN so as to generate high-quality samples and meet the prediction requirements. In order to achieve the purpose, the technical scheme of the invention is that the method for enhancing and predicting the power consumption of the small sample in the extreme scene based on PC-GAN comprises the following steps: Acquiring an electricity load-meteorological data set of a target area, identifying an extreme scene sample in the electricity load-meteorological data set, and constructing a multidimensional feature vector based on the electricity load-meteorological data set; Constructing physical constraints comprising a generator, a discriminator and a scene classifier to generate a PC-GAN model of an countermeasure network architecture; Generating an antagonism network architecture PC-GAN model based on physical constraint, and implementing a three-stage constraint strategy by taking a multidimensional feature vector as input to obtain a high-quality synthetic sample; And combining the extreme scene sample and the high-quality synthetic sample into an enhanced training data set, and training a load prediction model based on the enhanced training data set to realize the electricity load prediction of the target area. Optionally, the implementation of the three-stage constraint strategy specifically includes: Setting general physical loss, specific scene loss, characteristic distribution matching loss, time sequence consistency loss, generator total loss and discriminator total loss as constraints in a training stage, and alternately training and guiding a countering network architecture model to perform sample learning to obtain initial candidate samples; Aiming at the initial candidate sample in the generation stage, carrying out step-by-step forced light post-processing correction and dispersion processing on general physical loss, electric quantity consistency constraint and specific scene loss to obtain a post-processing candidate sample; and in the verification stage, aiming at the post-processing candidate samples, calculating the passing rate of each constraint and the comprehensive quality score to perform layered quality inspection, and obtaining a high-quality synthesized sample. Optionally, the general physical loss includes: Temperature relationship constraints The loss function is as follows: ; Wherein: Is the lowest temperature; is the average temperature; is the highest temperature; rainfall relationship constraint The loss function