WO2019238096A1 - 一种气象敏感负荷功率估算方法及装置 - Google Patents

一种气象敏感负荷功率估算方法及装置 Download PDF

Info

Publication number
WO2019238096A1
WO2019238096A1 PCT/CN2019/091142 CN2019091142W WO2019238096A1 WO 2019238096 A1 WO2019238096 A1 WO 2019238096A1 CN 2019091142 W CN2019091142 W CN 2019091142W WO 2019238096 A1 WO2019238096 A1 WO 2019238096A1
Authority
WO
WIPO (PCT)
Prior art keywords
meteorologically
fully connected
curve
power
connected layer
Prior art date
Application number
PCT/CN2019/091142
Other languages
English (en)
French (fr)
Inventor
尹积军
陈庆
吴争
陆晓
罗建裕
刘林
赵静波
鞠平
陈彦翔
秦川
施佳君
廖诗武
朱鑫要
王大江
Original Assignee
国网江苏省电力有限公司
国家电网有限公司
国网江苏省电力有限公司电力科学研究院
河海大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 国网江苏省电力有限公司, 国家电网有限公司, 国网江苏省电力有限公司电力科学研究院, 河海大学 filed Critical 国网江苏省电力有限公司
Publication of WO2019238096A1 publication Critical patent/WO2019238096A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00Computer-aided design [CAD]
    • G06F30/20Design optimisation, verification or simulation
    • GPHYSICS
    • G01MEASURING; TESTING
    • G01WMETEOROLOGY
    • G01W1/00Meteorology
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/10Interfaces, programming languages or software development kits, e.g. for simulating neural networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06QINFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00Administration; Management
    • G06Q10/04Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06QINFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/06Energy or water supply
    • HELECTRICITY
    • H02GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
    • H02JCIRCUIT ARRANGEMENTS OR SYSTEMS FOR SUPPLYING OR DISTRIBUTING ELECTRIC POWER; SYSTEMS FOR STORING ELECTRIC ENERGY
    • H02J3/00Circuit arrangements for ac mains or ac distribution networks
    • H02J3/003Load forecast, e.g. methods or systems for forecasting future load demand
    • YGENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y04INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
    • Y04SSYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
    • Y04S10/00Systems supporting electrical power generation, transmission or distribution
    • Y04S10/50Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications

Definitions

  • the present application belongs to the field of power system load prediction and load power model, and for example, relates to a method and device for estimating weather-sensitive load power.
  • the patent application number 201810607600.2 proposes a method for estimating meteorologically sensitive load power based on a load-meteor nonlinear correlation model, but the model requires high integrity of the load power and weather sample data.
  • meteorological data especially the meteorological factor change curve at the 10-minute sampling interval, is prone to data missing. If there are many missing meteorological data on that day, the load-meteor nonlinear correlation model cannot be used to estimate the meteorologically sensitive load for that day.
  • this application proposes a method and device for estimating the power of meteorologically sensitive loads, which can directly obtain the power curve of meteorologically sensitive loads from the daily load curve, and is particularly suitable for situations where meteorological data are often missing in practical applications.
  • an embodiment of the present application provides a method for estimating power of a weather-sensitive load, including:
  • the meteorologically sensitive load power is output.
  • obtaining the power estimation model of the weather-sensitive load includes:
  • the weather-sensitive load power estimation model includes a stacked auto-encoder (SAE) model and a fully connected layer;
  • SAE stacked auto-encoder
  • Training to obtain weather-sensitive load power estimation models includes:
  • the output of the meteorologically sensitive load power includes:
  • the fully connected layer outputs the meteorologically sensitive load power based on the dimensionality reduction characteristics of the daily load curve and the meteorologically sensitive load power extracted from the day to be estimated extracted by the SAE model.
  • training the SAE model includes:
  • the training objective function is the average MAPE of the absolute value of the relative percentage error between the output of the SAE model and the daily load curve of the corresponding historical data sample, x i is the actual daily load power, x i ′ is the output of the SAE model, and n is the total number of sampling points.
  • training the first AE of the SAE model satisfies the following formula:
  • h (1) i s f (W 1 x i + b 1 );
  • x i is the input of the first AE of the SAE model,
  • h (1) i is the output of the coding layer of the first AE,
  • W 1 , b 1 is a weight matrix and a bias matrix, and
  • s f is an activation function;
  • W 1 ′ and b 1 ′ are the weight matrix and the offset matrix during reconstruction, and s g is the activation function during reconstruction;
  • the mean square error with x i is the smallest, ⁇ * is the optimal fully connected layer parameter of the coding and decoding layers of the first AE, and N is the number of historical data samples.
  • the fully connected layer is trained to obtain the optimal fully connected layer parameter ⁇ ′ * , Where the dimensionality reduction characteristics of the daily load curve of the historical data sample are the same as the corresponding dates of the power curve of the meteorologically sensitive load;
  • O i is the output of the last fully connected layer of the i-th sample
  • P W i is the meteorologically sensitive load power of the i-th sample
  • N ′ is the number of days of fully connected layer training samples.
  • I and O are the input and output vectors of the fully connected layer
  • W and b are the weight matrix and offset matrix of the fully connected layer, respectively
  • R is the activation function of the fully connected layer.
  • the method further includes:
  • an embodiment of the present application further provides a meteorologically sensitive load power estimation device, including: a stack autoencoder SAE model and a fully connected layer; wherein,
  • the SAE model is set to input the daily load curve of the day to be estimated, extract the dimensionality reduction feature of the daily load curve on the day to be estimated, and input the dimensionality reduction feature of the daily load curve on the day to be estimated into the fully connected layer;
  • the fully connected layer is connected to the output of the SAE model, and is set to output the meteorologically sensitive load power according to the feature of the daily load curve dimension reduction on the day to be estimated, and the mapping between the feature of the daily load curve and the meteorologically sensitive load power.
  • the dimension of the daily load curve on the day to be estimated is the number of sampling points of the daily load curve on the day to be estimated; the dimension of the meteorologically sensitive load power to be estimated is the number of sampling points on the daily load curve on the day to be estimated.
  • the SAE model is formed by stacking multiple autoencoders AE, and each AE includes an encoding layer and a decoding layer;
  • the number of fully connected layers is at least one.
  • an embodiment of the present application provides a method for estimating the power of a meteorologically sensitive load.
  • a multi-layer fully connected layer is added to the output end of the stack autoencoder SAE model to establish a meteorologically sensitive load power estimation model based on the SAE.
  • the estimation model consists of two parts: the first part is a traditional stacked autoencoder SAE, and the second part is multiple fully connected layers superimposed on the SAE output.
  • the input of the estimation model is a daily load curve
  • the input dimension is the number of sampling points of the daily load curve
  • the output is the weather-sensitive load power
  • the output dimension is the number of sampling points of the daily load curve
  • the SAE forward propagation calculation formula is as follows:
  • W 1 and b 1 are weight matrix and bias matrix, respectively, and s f is activation function
  • W 1 ′ and b 1 ′ are respectively a weight matrix and a bias matrix during reconstruction, and s g is an activation function during reconstruction.
  • the SAE unsupervised training method is as follows:
  • the final SAE is formed by stacking multiple AEs.
  • I and O are the input and output vectors of the fully connected layer
  • W and b are the weight matrix and offset matrix of the fully connected layer, respectively
  • R is the activation function of the fully connected layer.
  • the supervised training method of the fully connected layer is as follows:
  • O i is the output of the last fully connected layer of the i-th sample
  • P W i is the weather-sensitive load power of the i-th sample
  • N ′ is the number of days of the weather-sensitive load power.
  • the meteorologically sensitive load power curve for supervised training in the fully connected layer is calculated using the following steps:
  • Step 10 Perform data processing on the total load power and meteorological data of a certain area or a substation, and reorder them to obtain "vertical" data samples composed of total load power and meteorological data on the same day and at the same time in different months;
  • Step 20 Establish a load-meteor nonlinear correlation model between the total load power, the meteorologically sensitive load power, and multiple types of meteorological information, and identify the model parameters using the gradient method;
  • Step 30 Substituting the identified model parameters, longitudinal historical meteorological data, and total load power data into the correlation model, calculating the longitudinal meteorological sensitive load power curve, and arranging the historical daily meteorological sensitive load power curve in a normal time sequence.
  • the step of performing data processing on the total load power and the meteorological data in step 10 includes data cleaning, removal of long-term growth of the base load power, and correction calculations of various types of meteorological factors considering temperature accumulation, hysteresis effects, and body temperature and humidity.
  • the original temperature is corrected based on the cumulative temperature and hysteresis effects.
  • the correction formula is:
  • T DayMod (T day1 ⁇ day1 + T day2 ⁇ day2 ) / ( ⁇ day1 + ⁇ day2 )
  • T DayMod is the correction temperature after considering the temperature accumulation effect; T day1 is the original temperature of the day; T day2 is the correction temperature of the previous day; ⁇ day1 is the correction coefficient of the day and ⁇ day2 is the correction coefficient of the previous day.
  • T DayMod is the corrected temperature after considering the temperature accumulation effect; H is the relative humidity; H T is the correction value of the humidity factor.
  • step 20 the load-meteor nonlinear correlation model is:
  • step 20 when using the gradient method to identify parameters in the load-meteor nonlinear correlation model, the objective function is established as the maximum correlation coefficient, that is:
  • m is the total number of samples
  • i is the i-th sample
  • step 30 the identified model parameters, “vertical” modified meteorological data and power data are substituted into the correlation model, and the meteorologically sensitive load curve Y at the time corresponding to the “vertical” sample is calculated; the maximum and minimum of the meteorologically sensitive load curve Y is performed.
  • the values are normalized to obtain the proportion of meteorologically sensitive load power in the total load power:
  • the actual meteorologically sensitive load power estimate is:
  • P max and P min are the maximum and minimum values of the total load power of the sample
  • ⁇ weather is the proportion of the weather-sensitive load power of the sample at a sampling time.
  • the "longitudinal" meteorologically sensitive load power calculated by the correlation model is arranged in a normal time sequence, and the "horizontal" meteorologically sensitive load power curve is obtained.
  • the estimation model proposed in this application can directly obtain the meteorologically sensitive load power curve from the daily load curve, for example, it is applicable to the situation where the meteorological data is often missing in practical applications.
  • SAE can extract the dimensionality reduction features of the daily load curve without supervision, greatly reducing the number of input neurons in the fully connected layer, thereby greatly reducing the network parameters of the fully connected layer, and significantly reducing the difficulty of model training.
  • Figure 1 Structure of a weather-sensitive load power estimation model based on a stack autoencoder
  • the SAE estimation model uses SAE's unsupervised learning to extract the dimensionality reduction features of the daily load curve; adding multiple layers of fully connected layers at the output of the SAE, taking the dimensionality reduction features as the input of the fully connected layers, and using the correlation model or traditional method calculation results as Fully connected layer output labels, training fully connected layers.
  • the estimation model can directly obtain the power curve of meteorologically sensitive load according to the daily load curve, thereby significantly improving the practicability of the method.
  • the structure of a weather-sensitive load estimation model based on SAE is shown in FIG.
  • the model consists of SAE and fully connected layers.
  • the input of SAE is the daily load curve, and the input dimension is 144, which is the number of sampling points of the daily load curve.
  • the output dimension of SAE and the number of encoding and decoding layers are hyperparameters, which need to be determined by tuning during model training and testing.
  • the fully connected layer is located at the output of the SAE.
  • the input dimension of the fully connected layer is consistent with the output dimension of the SAE.
  • the output is 144 points of daily weather-sensitive load power, thereby forming the deep characteristics of the daily load curve extracted by SAE to weather-sensitive Mapping of load power curve.
  • the model training and weather sensitive load estimation steps are as follows.
  • Step 10 Take all daily load curves from April to October in a year as a sample to perform unsupervised training on the SAE, thereby reducing the dimensionality of the daily load curve and extracting deep features of the daily load curve as an example.
  • the input of the first layer of SAE is x i , and the output of the coding layer of the first AE is calculated:
  • W 1 and b 1 are a weight matrix and a bias matrix, respectively, and s f is an activation function.
  • W 1 ′ and b 1 ′ are respectively a weight matrix and a bias matrix during reconstruction, and s g is an activation function during reconstruction.
  • N is the number of training samples.
  • the input of the next AE is h (1) i , and so on.
  • the final SAE is stacked by multiple AEs.
  • Step 20 Use the calculation result of the weather-sensitive load power curve calculation method to train the fully connected layer with labeled samples.
  • I and O are the input and output vectors of the layer
  • W and b are the weight matrix and bias matrix of the fully connected layer, respectively
  • R is the activation function of the fully connected layer.
  • the deep features of the daily load curve after SAE reduction on a certain day are used as the input of the fully connected layer, and the corresponding day, the weather-sensitive load power curve is used as the label of the fully connected layer for training, and the optimal fully connected layer parameter ⁇ is calculated ′ * :
  • O i is the output of the last fully connected layer of the i-th sample
  • I the meteorologically sensitive load power of the i-th sample
  • N ′ is the number of days when the meteorologically sensitive load power can be calculated.
  • Step 30 After the estimation model is trained, the daily load curve of the day to be estimated is used as an input, and the output of the model is the meteorologically sensitive load power curve to be estimated.
  • a 220 kV (kV) substation in a prefecture-level city is selected as the research object for the description of the implementation mode.
  • the substation includes industrial, commercial, residential and traction loads, and the load types are comprehensive.
  • the data collected are the load power of the station in 2015 (sampling interval: 5 minutes), and temperature and humidity data (sampling interval: 10 minutes).
  • Step 10 Prepare the sample data.
  • the load-meteor nonlinear correlation model was used to calculate the daily meteorologically sensitive load power curve arranged in normal time sequence.
  • Step 20 Normalize the sample data.
  • Range normalization was performed for each sample of 70 daily meteorologically sensitive load power data and 145 daily load data. which is:
  • x i is the ith data of a sample
  • x min and x max are the minimum and maximum values of the sample.
  • Step 30 Take all the daily load curve data from April to October as samples to perform unsupervised training on the SAE, thereby reducing the dimensionality of the daily load curve data and extracting the deep features of the daily load curve data.
  • MAPE minimum mean absolute error
  • x i is the actual daily load power
  • x i ′ is the output value of the decoder
  • n is the total number of sampling points.
  • SAE hyperparameters were finally selected: SAE encoding and decoding layers are 4 layers, that is, 4 self-encoding processes are performed, and the daily load data of 144 points is finally reduced to 5 deep feature parameters.
  • SAE encoding and decoding layers are 4 layers, that is, 4 self-encoding processes are performed, and the daily load data of 144 points is finally reduced to 5 deep feature parameters.
  • dimension reduction the input dimension (5 dimensions) and the number of neurons of the fully connected layer are greatly reduced, that is, the weights and bias parameters to be determined are greatly reduced, which effectively reduces the training difficulty of the fully connected layer.
  • Step 40 Train the fully connected layer using the calculation result of the association model as the labeled sample.
  • the output label of the fully connected layer is the daily meteorological sensitive load power curve calculated by the correlation model.
  • the input samples are the deep features of the daily load power curve after SAE reduction.
  • the training objective function is the minimum MAPE.
  • the activation function of the first layer is a Linear Rectification Function (ReLU) function
  • the second layer is a tanh function. So in fact, the samples with the weather-sensitive load power curve label need only train two fully connected layers.
  • ReLU Linear Rectification Function
  • Step 50 Restore the normalized calculation results of each sample output by the complete model:
  • Step 60 Test the SAE training results.
  • the power curve of the SAE codec is highly consistent with the actual load curve, which shows that the SAE dimension reduction and extraction of deep features (144-dimensional input reduced to 5-dimensional SAE output) can be more fully reflected Enter the curve information.
  • Step 70 Test the fully connected layer training results.
  • a test sample from July 27-30 is selected, and the weather sensitivity load power curve output by the model and the correlation model result are shown in FIG. 3.
  • the two curves are generally close. Taking into account the small training sample, the test results in the figure are good, indicating that the SAE estimation model can approximate the calculation results of the correlation model, so when the weather data is missing and it is difficult to use the correlation model, the SAE estimation model can be used to directly obtain the daily weather sensitivity Load power curve.
  • Step 80 After the model is trained and tested, the daily load curve of the date on which the meteorologically sensitive load power curve needs to be estimated is taken as the total input of the model, and the final output of the model is the meteorologically sensitive load power curve to be estimated.

Landscapes

  • Engineering & Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Economics (AREA)
  • General Physics & Mathematics (AREA)
  • Human Resources & Organizations (AREA)
  • Strategic Management (AREA)
  • Health & Medical Sciences (AREA)
  • Marketing (AREA)
  • General Business, Economics & Management (AREA)
  • Tourism & Hospitality (AREA)
  • General Health & Medical Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Water Supply & Treatment (AREA)
  • Power Engineering (AREA)
  • Environmental & Geological Engineering (AREA)
  • Evolutionary Computation (AREA)
  • Primary Health Care (AREA)
  • Public Health (AREA)
  • Development Economics (AREA)
  • Game Theory and Decision Science (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Molecular Biology (AREA)
  • Computer Hardware Design (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • Mathematical Physics (AREA)
  • Geometry (AREA)
  • Atmospheric Sciences (AREA)
  • Biodiversity & Conservation Biology (AREA)
  • Ecology (AREA)

Abstract

本申请公开了一种气象敏感负荷功率估算方法及装置。包括:获取气象敏感负荷功率估算模型;向气象敏感负荷功率估算模型输入待估算日的日负荷曲线,以提取待估算日的日负荷曲线降维特征;根据待估算日的日负荷曲线降维特征,以及日负荷曲线的降维特征到气象敏感负荷功率间的映射,输出气象敏感负荷功率。

Description

一种气象敏感负荷功率估算方法及装置
本申请要求在2018年06月13日提交中国专利局、申请号为201810606900.9的中国专利申请以及在2019年05月22日提交中国专利局、申请号为201910430705.X的中国专利申请的优先权,该申请的全部内容通过引用结合在本申请中。
技术领域
本申请属于电力系统负荷预测与负荷功率模型领域,例如涉及一种气象敏感负荷功率估算方法及装置。
背景技术
随着全球变暖趋势愈演愈烈,国民生活水平的不断提高,以空调为主的气象敏感负荷的用电功率逐年攀升,2017年苏州等部分地区的夏季空调耗电导致该地区负荷异常增长。研究气象敏感负荷功率的估算问题不仅能提高负荷功率模型的准确性,为夏季电网的安全稳定运行提供调控依据,也能为需求侧响应能力评估提供依据,具有重要的研究意义。
专利申请号为201810607600.2的专利提出了一种基于负荷-气象非线性关联模型的气象敏感负荷功率估算方法,但该模型对负荷功率及气象样本数据的完整性要求高。实际情况下,气象数据,尤其是10分钟采样间隔的气象因子变化曲线,容易存在数据缺失的情况。如果当日的气象数据缺失较多,则无法利用负荷-气象非线性关联模型估算该日的气象敏感负荷。
发明内容
针对以上问题,本申请提出了一种气象敏感负荷功率估算方法及装置,能够由日负荷曲线直接获得气象敏感负荷功率曲线,尤其适用于实际应用时气象数据经常缺失的情况。
本申请采用的技术方案如下:
第一方面,本申请实施例提供一种气象敏感负荷功率估算方法,包括:
获取气象敏感负荷功率估算模型;
向气象敏感负荷功率估算模型输入待估算日的日负荷曲线,以提取待估算日的日负荷曲线降维特征;
根据待估算日的日负荷曲线降维特征,以及日负荷曲线的降维特征到气象敏感负荷功率间的映射,输出气象敏感负荷功率。
在一实施例中,获取气象敏感负荷功率估算模型包括:
训练获得气象敏感负荷功率估算模型,并对气象敏感负荷功率估算模型进行测试。
在一实施例中,气象敏感负荷功率估算模型包括堆栈自编码器(stacked auto-encoder,SAE)模型和全连接层;
训练获得气象敏感负荷功率估算模型包括:
训练SAE模型,以及训练全连接层;
向气象敏感负荷功率估算模型输入待估算日的日负荷曲线,以提取待估算日的日负荷曲线降维特征包括:
向SAE模型输入待估算日的日负荷曲线,以提取待估算日的日负荷曲线降维特征;
根据待估算日的日负荷曲线降维特征,以及日负荷曲线的降维特征到气象敏感负荷功率间的映射,输出气象敏感负荷功率包括:
全连接层根据SAE模型提取到的待估算日的日负荷曲线降维特征,以及日负荷曲线的降维特征到气象敏感负荷功率间的映射,输出气象敏感负荷功率。
在一实施例中,训练SAE模型包括:
以历史数据样本为SAE模型的输入及输出的标签,对SAE模型的第一个自编码器(auto-encoder,AE)进行训练;
以第一个AE的编码层的输出为输入的标签,对SAE模型的下一个AE进行训练,直至SAE模型的所有AE训练完毕;
其中,训练的目标函数为SAE模型的输出与相应历史数据样本的日负荷曲线的相对百分误差绝对值的平均值MAPE最小,
Figure PCTCN2019091142-appb-000001
x i为实际的日负荷功率,x i′为SAE模型的输出,n为总的采样点数。
在一实施例中,对SAE模型的第一个AE进行训练满足如下公式:
h(1) i=s f(W 1x i+b 1);x i为SAE模型的第一个AE的输入,h(1) i为第一个AE的编码层的输出,W 1、b 1分别为权值矩阵和偏置矩阵,s f为激活函数;
Figure PCTCN2019091142-appb-000002
为所述SAE模型的第一个AE的输出,W 1′、b 1′分 别为重构时的权值矩阵和偏置矩阵,s g为重构时的激活函数;
Figure PCTCN2019091142-appb-000003
与x i的均方误差最小,θ *为第一个AE的编码层及解码层的最优全连接层参数,N为历史数据样本数。
在一实施例中,训练全连接层包括:
以历史数据样本的日负荷曲线降维特征为全连接层的输入的标签、气象敏感负荷功率曲线为全连接层的输出的标签,对全连接层进行训练,得到最优全连接层参数θ′ *,其中,历史数据样本的日负荷曲线降维特征与气象敏感负荷功率曲线的对应日期相同;
Figure PCTCN2019091142-appb-000004
O i为第i个样本的最后一层全连接层的输出,P W i为第i个样本的气象敏感负荷功率,N′为全连接层训练样本的日期数。
在一实施例中,全连接层的计算公式满足O=R(WI+b);
其中,I、O分别为全连接层的输入和输出向量,W、b分别为全连接层的权值矩阵和偏置矩阵,R为全连接层的激活函数。
在一实施例中,还包括:
在训练SAE模型前,对历史数据样本进行归一化处理;
在训练全连接层后,对全连接层输出的每个样本的归一化计算结果进行还原。
第二方面,本申请实施例还提供一种气象敏感负荷功率估算装置,包括:堆栈自编码器SAE模型和全连接层;其中,
SAE模型,设置为输入待估算日的日负荷曲线,提取待估算日的日负荷曲线降维特征,并将待估算日的日负荷曲线降维特征输入全连接层;
全连接层与SAE模型的输出端相连,设置为根据待估算日的日负荷曲线降维特征,以及日负荷曲线的降维特征到气象敏感负荷功率间的映射,输出气象敏感负荷功率。
在一实施例中,待估算日的日负荷曲线的维数为待估算日的日负荷曲线的采样点数;待估算的气象敏感负荷功率的维数为待估算日的日负荷曲线的采样点数。
在一实施例中,SAE模型由多个自编码器AE堆叠而成,每个AE均包括编码层和解码层;
全连接层的层数为至少一层。
第三方面,本申请实施例提供一种气象敏感负荷功率估算方法,在堆栈自编码器SAE模型的输出端增加多层全连接层,建立基于SAE的气象敏感负荷功率估算模型;
利用SAE的无监督训练方法提取日负荷曲线的降维特征,利用气象敏感负荷功率曲线作为有标签样本训练全连接层,从而在全连接层形成由日负荷曲线降维特征到气象敏感负荷功率间的映射。
估算模型由两部分构成:第一部分为传统的堆栈自编码器SAE,第二部分为在SAE输出端上叠加的多个全连接层。
在一实施例中,估算模型的输入为日负荷曲线,输入维数为日负荷曲线的采样点数;输出为气象敏感负荷功率,输出维数为日负荷曲线采样点数。
在一实施例中,SAE的前向传播计算公式如下:
SAE的第一层输入为x i,计算第一个自编码器AE的编码层的输出:
h(1) i=s f(W 1x i+b 1)
其中,W 1、b 1分别为权值矩阵和偏置矩阵,s f为激活函数;
由AE的编码层输出,按下式再通过解码层重构输入向量:
Figure PCTCN2019091142-appb-000005
其中,W 1′、b 1′分别为重构时的权值矩阵和偏置矩阵,s g为重构时的激活函数。
在一实施例中,SAE的无监督训练方法如下:
采用历史每日的日负荷曲线数据样本为SAE的输入及输出的标签进行训练,按SAE计算所得的
Figure PCTCN2019091142-appb-000006
与SAE输出的标签x i均方误差最小,计算AE的编码层及解码层的最优全连接层参数θ *
保留h(1) i,以h(1) i为下一AE的输入及其输出的标签,采用以上方式继续下一AE训练,下一个AE的输入即为h(1) i,以此类推,最终的SAE由多个AE堆叠而成。
AE的编码层及解码层的最优全连接层参数θ *计算公式如下:
Figure PCTCN2019091142-appb-000007
其中,N为训练样本数。
在一实施例中,全连接层的前向传播计算公式如下:
O=R(WI+b);
式中,I、O分别为全连接层的输入输出向量,W、b分别为全连接层的权值矩阵和偏置矩阵,R为全连接层的激活函数。
在一实施例中,全连接层的有监督训练方法如下:
以某日的日负荷曲线经SAE降维后的深层特征为全连接层的输入,以对应日期下,气象敏感负荷功率曲线为全连接层输出的标签进行训练,计算最优全连接层参数θ′ *
Figure PCTCN2019091142-appb-000008
式中,O i为第i个样本的最后一层全连接层的输出,P W i为第i个样本的气象敏感负荷功率,N′为气象敏感负荷功率的日期数。
在一实施例中,用于全连接层有监督训练的气象敏感负荷功率曲线采用以下步骤进行计算:
步骤10:对某一地区或某一变电站的总负荷功率及气象数据进行数据处理,并重新排序,获得由不同月的相同日相同时刻的总负荷功率、气象数据组成的“纵向”数据样本;
步骤20:建立总负荷功率、气象敏感负荷功率及多类气象信息之间的负荷-气象非线性关联模型,并利用梯度法辨识模型参数;
步骤30:将所辨识的模型参数、纵向的历史气象数据及总负荷功率数据代入关联模型,计算纵向气象敏感负荷功率曲线,并按正常时序排列得到历史每日的气象敏感负荷功率曲线。
步骤10中对总负荷功率及气象数据进行数据处理的步骤,包括数据清洗、去除基础负荷功率的长期增长量、考虑温度积累、迟滞效应和体感温湿度的多类气象因子修正计算。
对温度气象因子进行修正的步骤为:
基于温度的累积、迟滞效应对原始温度进行修正,修正公式为:
T DayMod=(T day1λ day1+T day2λ day2)/(λ day1day2)
λ day1=1-exp[-exp(T day1-26/6)]
λ day2=1-exp[-exp(T day2-26/6)]
其中,T DayMod为考虑温度积累效应后的修正温度;T day1为当天的原始温度;T day2为前一天的修正温度;λ day1为当天的修正系数、λ day2为前一天的修正系数。
对体感湿度气象因子进行修正的公式为:
H T=T DayModH
其中,T DayMod为考虑温度积累效应后的修正温度;H为相对湿度;H T为湿度因子的修正值。
步骤20中,所述的负荷-气象非线性关联模型为:
Figure PCTCN2019091142-appb-000009
Figure PCTCN2019091142-appb-000010
其中,r XY为总负荷功率与气象敏感负荷之间的相关系数;X为经步骤1处理并归一化后的总负荷功率;Y为对应X相同时刻下的气象敏感负荷,
Figure PCTCN2019091142-appb-000011
为某一样本的X、Y曲线均值;i为采样点序号,X i、Y i为某样本第i个采样点经步骤1处理并归一化后的总负荷功率,和相同时刻下的气象敏感负荷。n为单个样本的采样点个数;a 1,a 2,b 1,b 2,w 1,w 2为负荷-气象关联模型的待辨识参数。T DayMod为考虑温度积累效应后的修正温度,H T为相对湿度的修正值;气象敏感负荷Y与修正温度T DayMod及修正湿度之间的关系为扩展的Sigmoid函数。
步骤20中,采用梯度法辨识负荷-气象非线性关联模型中的参数时,建立目标函数为相关系数最大值,即:
Figure PCTCN2019091142-appb-000012
式中,m表示总样本个数,i表示第i号样本,
Figure PCTCN2019091142-appb-000013
表示第i号样本总负荷功率与气象敏感负荷之间的相关系数。
步骤30中,将所辨识的模型参数、“纵向”的修正气象数据及功率数据代入关联模型,计算该“纵向”样本对应时间下的气象敏感负荷曲线Y;对气象敏感负荷曲线Y进行最大最小值归一化,获得气象敏感负荷功率在总负荷功率中的占比:
Figure PCTCN2019091142-appb-000014
式中,
Figure PCTCN2019091142-appb-000015
为某一样本中第j个数据的气象敏感负荷功率占比,Y (j)为某一样本的第j个气象敏感负荷功率估计值Y,Y max、Y min分别为气象敏感负荷的最大、最小值;
则实际的气象敏感负荷功率估算值为:
P weather=ρ weather·(P max-P min)
其中,P max、P min为该样本的总负荷功率最大、最小值,ρ weather为该样本某采样时刻的气象敏感负荷功率占比。
将关联模型计算获得的“纵向”气象敏感负荷功率按照正常时序排列,获得“横向”排列的日气象敏感负荷功率曲线。
本申请所达到的有益效果:
本申请提出的估算模型可以由日负荷曲线直接获得气象敏感负荷功率曲线,例如适用于实际应用时气象数据经常缺失的情况。SAE可以无监督提取日负荷曲线的降维特征,大幅减少了全连接层的输入神经元个数,从而大幅减少了全连接层的网络参数,显著降低了模型训练难度。
附图说明
图1基于堆栈自编码器的气象敏感负荷功率估算模型结构;
图2算例测试集中7月27-30日总负荷实际值与SAE计算结果比较;
图3算例测试集中7月27-30日本申请方法与关联模型计算结果比较。
具体实施方式
下面结合附图和具体的实施例对本申请技术方案进行描述,以使本领域的技术人员可以更好的理解本申请并能予以实施,但所举实施例不作为对本申请的限定。
SAE估算模型利用SAE的无监督学习提取日负荷曲线的降维特征;在SAE的输出端增加多层全连接层,以降维特征为全连接层的输入,并以关联模型或传统方法计算结果作为全连接层输出的标签,训练全连接层。在实际应用中,该估算模型可以根据日负荷曲线直接获得气象敏感负荷功率曲线,从而显著提升了方法的实用性。
基于SAE的气象敏感负荷估算模型结构附图1所示。模型包括SAE和全连接层2个部分。SAE的输入为日负荷曲线,输入维数为144,也即日负荷曲线的 采样点数。SAE的输出维数以及编码、解码层数为超参数,需要在模型训练、测试时调优确定。全连接层位于SAE的输出端,全连接层的输入维数与SAE的输出维数一致,输出为144点的日气象敏感负荷功率,从而形成由SAE提取的日负荷曲线深层特征到气象敏感负荷功率曲线的映射。
模型的训练及气象敏感负荷的估算步骤如下。
步骤10.以某一年中4-10月所有的日负荷曲线作为样本对SAE进行无监督训练,从而对日负荷曲线进行降维并提取日负荷曲线的深层特征为例。
SAE的第一层输入为x i,计算第一个AE的编码层的输出:
h(1) i=s f(W 1x i+b 1)
其中,W 1、b 1分别为权值矩阵和偏置矩阵,s f为激活函数。
由AE的编码层输出,按下式再通过解码层重构输入向量:
Figure PCTCN2019091142-appb-000016
其中,W 1′、b 1′分别为重构时的权值矩阵和偏置矩阵,s g为重构时的激活函数。
采用历史每日日负荷曲线数据样本进行训练,按
Figure PCTCN2019091142-appb-000017
与x i均方误差最小,寻求AE的编码层及解码层的最优全连接层参数θ *。其计算公式如下。
Figure PCTCN2019091142-appb-000018
其中,N为训练样本数。
保留h(1) i,采用以上方式继续下一AE训练,下一个AE的输入即为h(1) i,以此类推,最终的SAE由多个AE堆叠而成。
步骤20.以气象敏感负荷功率曲线的计算方法的计算结果为有标签样本训练全连接层。
全连接层的计算公式为:
O=R(WI+b)
其中,I、O分别为该层的输入、输出向量,W、b分别为全连接层的权值矩阵和偏置矩阵,R为全连接层的激活函数。
以某日的日负荷曲线经SAE降维后的深层特征为全连接层的输入,以对应日期下,气象敏感负荷功率曲线为全连接层输出的标签进行训练,计算最优全连接层参数θ′ *
Figure PCTCN2019091142-appb-000019
式中,O i为第i个样本的最后一层全连接层的输出,
Figure PCTCN2019091142-appb-000020
为第i个样本的气象敏感负荷功率,N′为可计算出气象敏感负荷功率的日期数。
步骤30.估算模型训练好后,以待估算日的日负荷曲线作为输入,则模型的输出即为需估算的气象敏感负荷功率曲线。
实施例1
选取某地级市内某220千伏(kV)变电站为研究对象进行实施方式说明。该变电站下包含了工业、商业、居民及牵引负荷,负荷类型全面。采集的数据为该站2015年全年负荷功率(采样间隔5分钟),以及温度、湿度数据(采样间隔10分钟)。
步骤10:准备样本数据。
由于气象数据的残缺,计算出按正常时间顺序排列的日气象敏感负荷功率曲线数据共70条(4-10月,每个月10天),其中65条数据作为训练SAE模型全连接层的带标签样本,另5条作测试样本。再取2015年4-10月(共214 天,其中包含节假日69天)所有工作日该变电站日负荷曲线数据140条作为无标签样本训练SAE多层,另5条作测试样本。
由负荷-气象非线性关联模型计算出按正常时间顺序排列的日气象敏感负荷功率曲线。
步骤20:归一化处理样本数据。
对70条日气象敏感负荷功率数据及145条日负荷数据的每个样本进行极差归一化。即:
Figure PCTCN2019091142-appb-000021
式中,x i为某一样本的第i个数据,x min、x max为该样本的最小值、最大值。
步骤30:以4—10月所有的日负荷曲线数据作为样本对SAE进行无监督训练,从而对日负荷曲线数据进行降维并提取日负荷曲线数据的深层特征。以解码器输出与相应日负荷曲线数据的相对百分误差绝对值的平均值(mean absolute percentage error,MAPE)最小作为训练的目标函数。MAPE的计算公式为:
Figure PCTCN2019091142-appb-000022
其中,x i为实际的日负荷功率,x i′为解码器输出值,n为总的采样点数。
实际测试后,最终选定SAE的超参数为:SAE编码、解码层均4层,即进行4次自编码过程,最终将144点的日负荷数据降维至5个深层特征参数。经过降维,全连接层的输入维数(5维)及神经元个数大大减少,也即需确定的权重及偏置参数大量减少,有效降低了全连接层的训练难度。
步骤40:以关联模型计算结果作为有标签样本训练全连接层。全连接层的输出标签为关联模型计算得到的日气象敏感负荷功率曲线,输入样本为日负荷功率曲线经SAE降维后的深层特征,训练的目标函数为MAPE最小。
经过实际测试,最终设置2层的全连接层,分别含25、144个神经元。其第1层的激活函数取线性整流函数(Rectified Linear Unit,ReLU)函数,第2层取tanh函数。故实际上带气象敏感负荷功率曲线为标签的样本仅需训练两层全连接层即可。
步骤50:对完整模型输出的每个样本的归一化计算结果进行还原:
y i=y′ i·(x max-x min)+x min
式中,y′ i为模型输出的归一化气象敏感负荷功率值,x max、x min为模型输入的日负荷曲线样本的最大实际值及最小实际值。
步骤60:对SAE训练结果进行测试。
取测试集中7月27-30日总负荷功率实际值与SAE编解码后输出曲线进行比较,如附图2。分别计算每天两曲线间的MAPE值,估算测试误差,如下表所示。
表1
Figure PCTCN2019091142-appb-000023
由上表及附图2可见,SAE编解码后的功率曲线与实际负荷曲线高度吻合,说明了SAE降维提取深层特征(输入的144维降至SAE输出的5维)时能较完整的反映输入曲线信息。
步骤70:对全连接层训练结果进行测试。
选取7月27-30日的测试样本,模型输出的气象敏感负荷功率曲线与关联模型结果对比如附图3所示。
由附图3可以看出,两曲线总体接近。考虑到训练样本较小这一因素,图 中测试结果良好,说明SAE估算模型可逼近关联模型计算结果,故可在气象数据缺失较多而难以使用关联模型时采用SAE估算模型直接获得日气象敏感负荷功率曲线。
步骤80:估算模型训练、测试后,以所需估算气象敏感负荷功率曲线的日期的日负荷曲线作为模型的总输入,则模型的最终输出即为需估算的气象敏感负荷功率曲线。

Claims (19)

  1. 一种气象敏感负荷功率估算方法,包括:
    获取气象敏感负荷功率估算模型;
    向所述气象敏感负荷功率估算模型输入待估算日的日负荷曲线,以提取待估算日的日负荷曲线降维特征;
    根据所述待估算日的日负荷曲线降维特征,以及日负荷曲线的降维特征到气象敏感负荷功率间的映射,输出气象敏感负荷功率。
  2. 根据权利要求1所述的气象敏感负荷功率估算方法,其中,所述获取气象敏感负荷功率估算模型包括:
    训练获得所述气象敏感负荷功率估算模型,并对所述气象敏感负荷功率估算模型进行测试。
  3. 根据权利要求2所述的气象敏感负荷功率估算方法,其中,所述气象敏感负荷功率估算模型包括堆栈自编码器SAE模型和全连接层;
    所述训练获得所述气象敏感负荷功率估算模型包括:
    训练所述SAE模型,以及训练全连接层;
    所述向所述气象敏感负荷功率估算模型输入待估算日的日负荷曲线,以提取待估算日的日负荷曲线降维特征包括:
    向所述SAE模型输入待估算日的日负荷曲线,以提取待估算日的日负荷曲线降维特征;
    所述根据所述待估算日的日负荷曲线降维特征,以及日负荷曲线的降维特征到气象敏感负荷功率间的映射,输出气象敏感负荷功率包括:
    所述全连接层根据所述SAE模型提取到的所述待估算日的日负荷曲线降维特征,以及日负荷曲线的降维特征到气象敏感负荷功率间的映射,输出气象敏感负荷功率。
  4. 根据权利要求3所述的气象敏感负荷功率估算方法,其中,所述训练SAE模型包括:
    以历史数据样本为所述SAE模型的输入及输出的标签,对所述SAE模型的第一个AE进行训练;
    以所述第一个自编码器AE的编码层的输出为输入的标签,对所述SAE模型的下一个AE进行训练,直至所述SAE模型的所有AE训练完毕;
    其中,训练的目标函数为所述SAE模型的输出与相应历史数据样本的日负 荷曲线的相对百分误差绝对值的平均值MAPE最小,
    Figure PCTCN2019091142-appb-100001
    x i为实际的日负荷功率,x i′为所述SAE模型的输出,n为总的采样点数。
  5. 根据权利要求4所述的气象敏感负荷功率估算方法,其中,所述对所述SAE模型的第一个AE进行训练满足如下公式:
    h(1) i=s f(W 1x i+b 1);x i为所述SAE模型的第一个AE的输入,h(1) i为第一个AE的编码层的输出,W 1、b 1分别为权值矩阵和偏置矩阵,s f为激活函数;
    Figure PCTCN2019091142-appb-100002
    Figure PCTCN2019091142-appb-100003
    为所述SAE模型的第一个AE的输出,W 1′、b 1′分别为重构时的权值矩阵和偏置矩阵,s g为重构时的激活函数;
    Figure PCTCN2019091142-appb-100004
    Figure PCTCN2019091142-appb-100005
    与x i的均方误差最小,θ *为第一个AE的编码层及解码层的最优全连接层参数,N为历史数据样本数。
  6. 根据权利要求3所述的气象敏感负荷功率估算方法,其中,所述训练所述全连接层包括:
    以历史数据样本的日负荷曲线降维特征为所述全连接层的输入的标签、气象敏感负荷功率曲线为所述全连接层的输出的标签,对所述全连接层进行训练,得到最优全连接层参数θ′ *,其中,所述历史数据样本的日负荷曲线降维特征与所述气象敏感负荷功率曲线的对应日期相同;
    Figure PCTCN2019091142-appb-100006
    O i为第i个样本的最后一层全连接层的输出,P W i为第i个样本的气象敏感负荷功率,N′为全连接层训练样本的日期数。
  7. 根据权利要求6所述的气象敏感负荷功率估算方法,其中,所述全连接层的计算公式满足O=R(WI+b);
    其中,I、O分别为所述全连接层的输入和输出向量,W、b分别为所述全连接层的权值矩阵和偏置矩阵,R为所述全连接层的激活函数。
  8. 根据权利要求3所述的气象敏感负荷功率估算方法,还包括:
    在训练所述SAE模型前,对历史数据样本进行归一化处理;
    在训练所述全连接层后,对所述全连接层输出的每个样本的归一化计算结果进行还原。
  9. 一种气象敏感负荷功率估算装置,包括:堆栈自编码器SAE模型和全连接层;其中,
    所述SAE模型,设置为输入待估算日的日负荷曲线,提取待估算日的日负荷曲线降维特征,并将所述待估算日的日负荷曲线降维特征输入所述全连接层;
    所述全连接层与所述SAE模型的输出端相连,设置为根据所述待估算日的日负荷曲线降维特征,以及日负荷曲线的降维特征到气象敏感负荷功率间的映射,输出气象敏感负荷功率。
  10. 根据权利要求9所述的气象敏感负荷功率估算装置,其中,所述待估算日的日负荷曲线的维数为待估算日的日负荷曲线的采样点数;所述待估算的气象敏感负荷功率的维数为所述待估算日的日负荷曲线的采样点数。
  11. 根据权利要求9所述的气象敏感负荷功率估算装置,其中,
    所述SAE模型由多个自编码器AE堆叠而成,每个所述AE均包括编码层和解码层;
    所述全连接层的层数为至少一层。
  12. 一种气象敏感负荷功率估算方法,在堆栈自编码器SAE模型的输出端增加多层全连接层,建立基于SAE的气象敏感负荷功率估算模型;
    利用SAE的无监督训练方法提取日负荷曲线的降维特征,利用气象敏感负荷功率曲线作为有标签样本训练全连接层,从而在全连接层形成由日负荷曲线降维特征到气象敏感负荷功率间的映射。
  13. 根据权利要求12所述的气象敏感负荷功率估算方法,其中,估算模型的输入为日负荷曲线,输入维数为日负荷曲线的采样点数;输出为气象敏感负荷功率,输出维数为日负荷曲线采样点数。
  14. 根据权利要求12所述的气象敏感负荷功率估算方法,其中,SAE的前向传播计算公式如下:
    SAE的第一层输入为x i,计算第一个自编码器AE的编码层的输出:
    h(1) i=s f(W 1x i+b 1)
    其中,W 1、b 1分别为权值矩阵和偏置矩阵,s f为激活函数;
    由AE的编码层输出,按下式再通过解码层重构输入向量:
    Figure PCTCN2019091142-appb-100007
    其中,W 1′、b 1′分别为重构时的权值矩阵和偏置矩阵,s g为重构时的激活函数,h(1) i为第一个AE的编码层的输出。
  15. 根据权利要求14所述的气象敏感负荷功率估算方法,其中,所述SAE的无监督训练方法如下:
    采用历史每日的日负荷曲线数据样本为SAE的输入及输出的标签进行训练,按SAE计算所得的
    Figure PCTCN2019091142-appb-100008
    与SAE输出的标签x i均方误差最小,计算AE的编码层及解码层的最优全连接层参数θ *
    保留h(1) i,以h(1) i为下一AE的输入及其输出的标签,采用以上方式继续下一AE训练,下一个AE的输入即为h(1) i,以此类推,最终的SAE由多个AE堆叠而成。
  16. 根据权利要求15所述的气象敏感负荷功率估算方法,其中,AE的编码层及解码层的最优全连接层参数θ *计算公式如下:
    Figure PCTCN2019091142-appb-100009
    其中,N为训练样本数。
  17. 根据权利要求12所述的气象敏感负荷功率估算方法,其中,全连接层的前向传播计算公式如下:
    O=R(WI+b);
    式中,I、O分别为全连接层的输入输出向量,W、b分别为全连接层的权值矩阵和偏置矩阵,R为全连接层的激活函数。
  18. 根据权利要求17所述的气象敏感负荷功率估算方法,其中,全连接层的有监督训练方法如下:
    以某日的日负荷曲线经SAE降维后的深层特征为全连接层的输入,以对应日期下,气象敏感负荷功率曲线为全连接层输出的标签进行训练,计算最优全 连接层参数θ′ *
    Figure PCTCN2019091142-appb-100010
    式中,O i为第i个样本的最后一层全连接层的输出,P W i为第i个样本的气象敏感负荷功率,N′为气象敏感负荷功率的日期数。
  19. 根据权利要求17所述的气象敏感负荷功率估算方法,其中,用于全连接层有监督训练的气象敏感负荷功率曲线采用以下方式计算:
    对某一地区或某一变电站的总负荷功率及气象数据进行数据处理,并重新排序,获得由不同月的相同日相同时刻的总负荷功率、气象数据组成的“纵向”数据样本;
    建立总负荷功率、气象敏感负荷功率及多类气象信息之间的负荷-气象非线性关联模型,并利用梯度法辨识模型参数;
    将所辨识的模型参数、纵向的历史气象数据及总负荷功率数据代入关联模型,计算纵向气象敏感负荷功率曲线,并按正常时序排列得到历史每日的气象敏感负荷功率曲线。
PCT/CN2019/091142 2018-06-13 2019-06-13 一种气象敏感负荷功率估算方法及装置 WO2019238096A1 (zh)

Applications Claiming Priority (4)

Application Number Priority Date Filing Date Title
CN201810606900.9 2018-06-13
CN201810606900.9A CN108763820A (zh) 2018-06-13 2018-06-13 一种基于堆栈自编码器的气象敏感负荷功率估算方法
CN201910430705.XA CN109992930A (zh) 2018-06-13 2019-05-22 一种气象敏感负荷功率估算方法及装置
CN201910430705.X 2019-05-22

Publications (1)

Publication Number Publication Date
WO2019238096A1 true WO2019238096A1 (zh) 2019-12-19

Family

ID=64022472

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/091142 WO2019238096A1 (zh) 2018-06-13 2019-06-13 一种气象敏感负荷功率估算方法及装置

Country Status (3)

Country Link
US (1) US20190384879A1 (zh)
CN (2) CN108763820A (zh)
WO (1) WO2019238096A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114511058A (zh) * 2022-01-27 2022-05-17 国网江苏省电力有限公司泰州供电分公司 一种用于电力用户画像的负荷元件构建方法及装置
CN117076990A (zh) * 2023-10-13 2023-11-17 国网浙江省电力有限公司 基于曲线降维和聚类的负荷曲线辨识方法、装置及介质

Families Citing this family (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108763820A (zh) * 2018-06-13 2018-11-06 国网江苏省电力有限公司电力科学研究院 一种基于堆栈自编码器的气象敏感负荷功率估算方法
CN111597650B (zh) * 2020-04-28 2022-09-02 国网电力科学研究院武汉南瑞有限责任公司 换流变压器阀侧套管出线装置绝缘结构的优化方法及系统
CN111598328A (zh) * 2020-05-14 2020-08-28 国网河北省电力有限公司电力科学研究院 一种计及疫情事件的电力负荷预测方法
CN113011045B (zh) * 2021-04-13 2022-09-30 河海大学 夏季用电负荷中空调功率获取方法及计算机可读介质
CN113434493A (zh) * 2021-06-28 2021-09-24 湘潭大学 一种基于Transformer的非侵入式负荷分解方法
CN113642676B (zh) * 2021-10-12 2022-02-22 华北电力大学 基于异构气象数据融合的区域电网负荷预测方法和装置
CN115130899B (zh) * 2022-07-18 2023-04-18 东南大学溧阳研究院 一种基于Kmeas—GM的空调负荷日前响应能力评估方法
CN117318049B (zh) * 2023-11-27 2024-02-13 国网山东省电力公司枣庄供电公司 一种基于系数校正的配电网柔性负荷建模方法及系统

Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108763820A (zh) * 2018-06-13 2018-11-06 国网江苏省电力有限公司电力科学研究院 一种基于堆栈自编码器的气象敏感负荷功率估算方法

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108763820A (zh) * 2018-06-13 2018-11-06 国网江苏省电力有限公司电力科学研究院 一种基于堆栈自编码器的气象敏感负荷功率估算方法

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
CHEN, YANXIANG ET AL.: "Estimation Method of Meteorological Sensitive Load Power Based on Correlation Analysis and Stacked Auto-encoder", ELECTRIC POWER AUTOMATION EQUIPMENT, vol. 38, no. 5, 4 May 2018 (2018-05-04), pages 222 - 223 *
ZHU, QIAOMU ET AL.: "Transient Stability Assessment Based on Stacked Autoencoder", PROCEEDINGS OF THE CSE E, vol. 38, no. 10, 20 May 2018 (2018-05-20), pages 2939 *

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114511058A (zh) * 2022-01-27 2022-05-17 国网江苏省电力有限公司泰州供电分公司 一种用于电力用户画像的负荷元件构建方法及装置
CN114511058B (zh) * 2022-01-27 2023-06-02 国网江苏省电力有限公司泰州供电分公司 一种用于电力用户画像的负荷元件构建方法及装置
CN117076990A (zh) * 2023-10-13 2023-11-17 国网浙江省电力有限公司 基于曲线降维和聚类的负荷曲线辨识方法、装置及介质
CN117076990B (zh) * 2023-10-13 2024-02-27 国网浙江省电力有限公司 基于曲线降维和聚类的负荷曲线辨识方法、装置及介质

Also Published As

Publication number Publication date
CN109992930A (zh) 2019-07-09
US20190384879A1 (en) 2019-12-19
CN108763820A (zh) 2018-11-06

Similar Documents

Publication Publication Date Title
WO2019238096A1 (zh) 一种气象敏感负荷功率估算方法及装置
US20190325334A1 (en) Method for predicting air quality with aid of machine learning models
Ohtsuka et al. Forecasting electricity demand in Japan: A Bayesian spatial autoregressive ARMA approach
CN111080032A (zh) 一种基于Transformer结构的负荷预测方法
CN110210993B (zh) 基于循环神经网络模型的城市短期燃气负荷预测方法
CN108549960A (zh) 一种24小时电力负荷预测方法
CN110047291A (zh) 一种考虑扩散过程的短时交通流预测方法
CN106600037B (zh) 一种基于主成分分析的多参量辅助负荷预测方法
CN103020485B (zh) 基于beta噪声核岭回归技术的短期风速预报方法
CN109636063A (zh) 一种短期负荷预测的方法
CN112803398A (zh) 基于经验模态分解和深度神经网络的负荷预测方法及系统
CN115760213A (zh) 短期电价预测模型的构建方法、短期电价预测方法、计算机可读介质、电子设备
CN110222910B (zh) 一种有源配电网态势预测方法及预测系统
CN108830415A (zh) 一种基于非线性关联模型的气象敏感负荷功率估算方法
CN114091758A (zh) 一种中短期负荷预测方法、系统以及存储介质
Li et al. A new grey forecasting model based on BP neural network and Markov chain
CN111724072A (zh) 一种点拍网的基于图神经网络gat的专利估值方法及系统
CN113191069B (zh) 基于双分支深度学习模型的空调负荷估算方法与系统
CN110991723A (zh) 一种人工智能在季节性负荷预测中的应用方法
CN114862032B (zh) 一种基于XGBoost-LSTM的电网负荷预测方法及装置
CN115641498A (zh) 基于空间多尺度卷积神经网络的中期降水预报后处理订正方法
Pan et al. Modeling optimization method based on Gamma test and NSGA II for forecast of PV power output
CN114971081A (zh) 一种基于时间序列分析与日统计的辐照预测方法
Lu et al. A New Power Load Forecasting Model (SIndRNN): independently recurrent neural network based on softmax kernel function
CN113112085A (zh) 一种基于bp神经网络的新能源场站发电负荷预测方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19819267

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19819267

Country of ref document: EP

Kind code of ref document: A1