CN113962456A - Medium-and-long-term load prediction method considering industry relevance - Google Patents

Medium-and-long-term load prediction method considering industry relevance Download PDF

Info

Publication number
CN113962456A
CN113962456A CN202111215938.1A CN202111215938A CN113962456A CN 113962456 A CN113962456 A CN 113962456A CN 202111215938 A CN202111215938 A CN 202111215938A CN 113962456 A CN113962456 A CN 113962456A
Authority
CN
China
Prior art keywords
industry
load
data
value
relevance
Prior art date
Legal status (The legal status 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 status listed.)
Pending
Application number
CN202111215938.1A
Other languages
Chinese (zh)
Inventor
姜宇轩
李昆明
马吉科
王江辉
曹卫青
张鸿鸣
张昆明
章天晗
林之岸
杨莉
林振智
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Zhejiang University ZJU
Jiangsu Fangtian Power Technology Co Ltd
Original Assignee
Zhejiang University ZJU
Jiangsu Fangtian Power Technology Co Ltd
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 Zhejiang University ZJU, Jiangsu Fangtian Power Technology Co Ltd filed Critical Zhejiang University ZJU
Priority to CN202111215938.1A priority Critical patent/CN113962456A/en
Publication of CN113962456A publication Critical patent/CN113962456A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION 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
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/18—Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION 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
    • G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/06—Energy or water supply

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Mathematical Physics (AREA)
  • Health & Medical Sciences (AREA)
  • Business, Economics & Management (AREA)
  • General Engineering & Computer Science (AREA)
  • Software Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Mathematical Optimization (AREA)
  • Pure & Applied Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Economics (AREA)
  • Artificial Intelligence (AREA)
  • Computational Mathematics (AREA)
  • Mathematical Analysis (AREA)
  • Molecular Biology (AREA)
  • Human Resources & Organizations (AREA)
  • Strategic Management (AREA)
  • Tourism & Hospitality (AREA)
  • General Business, Economics & Management (AREA)
  • Marketing (AREA)
  • Operations Research (AREA)
  • Algebra (AREA)
  • Databases & Information Systems (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Probability & Statistics with Applications (AREA)
  • Primary Health Care (AREA)
  • Water Supply & Treatment (AREA)
  • Public Health (AREA)
  • Development Economics (AREA)

Abstract

The invention discloses a medium and long term load prediction method considering industry relevance. The method comprises the following steps: the method comprises the steps of firstly, preprocessing industrial load data aiming at the problems of industrial load data loss and large data magnitude difference, providing an industrial monthly load data restoration method based on iterative interpolation, restoring monthly load data of a typical industry, and carrying out normalization processing on the load data by adopting a min-max standardization method. On the basis, the industry relevance is analyzed and quantified based on the Shapiro-Wilk test and the Pearson correlation coefficient, and the strongly relevant industries are screened according to the relevance indexes. And finally, on the basis of the industry load data, the air temperature data and the industry relevance, predicting the Long-Term load in the industry by using a Long Short Term Memory (LSTM) neural network. The method can realize high-precision medium and long term load prediction by introducing industry relevance analysis.

Description

Medium-and-long-term load prediction method considering industry relevance
Technical Field
The invention relates to the technical field of power systems, in particular to a medium-long term load prediction method considering industry relevance.
Background
With the rapid development of electric power industry in China, the scale of an electric power system is increasingly large, a smart power grid is developed vigorously, and electric power loads are influenced by more and more uncertain factors, such as climate, GDP, industrial structures, policies and the like, which provide greater challenges for load prediction of medium-and long-term electric power systems. In addition to these objective factors, an industry located in an industry chain must be affected by the upstream and downstream industries, so the influence of the association between industries on the load prediction of the industry cannot be ignored. The comprehensive consideration of the multiple influence factors to improve the load prediction precision becomes a hot topic in the current power system research field.
The medium and long-term load prediction result has important significance for resource scheduling and planning of the power system. For example, when a power plant is newly built, a construction site, a power generation scale, a construction period, and the like may be selected based on the result of the prediction of the medium-and long-term load in the area. The high-precision load prediction has important value on the operation planning of the power grid, and the development of big data and artificial intelligence technology also provides certain technical support for the high-precision load prediction. Currently, many researches are based on big data and deep learning technology, and the influences of various natural factors and social factors on load prediction are analyzed, so that various load prediction models are proposed. However, most of the research on load prediction of middle and long-term industries is mainly based on multivariate data such as load data, meteorological data, regional data and economic data, and the influence of the relevance among the industries on the load prediction is not considered, so that the precision of the load prediction of the industries is often not high.
Therefore, the influence of the industry relevance on the load prediction is considered, and a prediction method for improving the medium-long term load prediction precision is yet to be researched.
Disclosure of Invention
The invention provides a medium-long term load prediction method considering the industry relevance in order to overcome the defects in the prior art, and the influence of the industry relevance on load prediction can be considered, so that the accuracy of the industry load prediction is improved.
The invention adopts the following technical scheme:
a medium-long term load prediction method considering industry relevance comprises the following steps:
aiming at the problems of industry load data loss and large data magnitude difference, an industry monthly load data restoration method based on iterative interpolation is provided, monthly load loss data are restored, and load data are normalized by adopting a min-max standardization method;
analyzing and quantifying the association between industries based on Shapiro-Wilk test and Pearson correlation coefficient, and screening out strong correlation industries according to the association index (namely Pearson correlation coefficient);
and on the basis of the industry load data, the air temperature data and the industry relevance, predicting the long-term load and the medium-term load of the industry by using the long-term and short-term memory neural network.
In the above technical solution, further, for the problem of industry load data loss and a large difference between data magnitudes, an industry monthly load data restoration method based on iterative interpolation is proposed to restore monthly load data loss, and a min-max standardization method is adopted to normalize the load data, including the following steps:
because the original load data usually has the problems of data loss, large magnitude difference and the like, the original load data needs to be repaired and preprocessed. The method comprises the following steps of collecting middle and long term load data of the industry at intervals of months, and forming a data input matrix which can be expressed as follows:
Figure RE-GDA0003378062050000021
in the formula, yijThe month load value of the j month of the i industry; m is the total number of industries; n is the total number of months of data acquisition; t isjIndicating the meteorological parameters for month j.
Aiming at the problem of load data missing,
and (2) completing the data by adopting an iterative interpolation method, namely, by means of multiple iterations, taking the load data of the industry where the missing data is located as a dependent variable y, taking the rest of related characteristic data as an independent variable x, performing fitting modeling and interpolation by using a linear regression, performing multiple iteration cycles, and outputting complete load data, wherein the calculation of the y value is shown as the following formula:
y＝a1x1+a2x2+...aMxM
the magnitude of the load may vary greatly from industry to industry, which may lead to increased difficulty in model training. The gradient descent can be accelerated by adopting normalization, so that the convergence speed of the model is improved. Meanwhile, the normalized data is limited to be between 0 and 1, so that the evaluation standard can be unified, and the calculation speed can be improved. And (3) carrying out normalization processing on the load data by adopting a min-max standardization method:
Figure RE-GDA0003378062050000031
in the formula, xi' represents the load data of month i after min-max normalization; x is the number ofiRepresenting the original load data of the ith month; x is the number ofminRepresenting the minimum of elements in the original load data; x is the number ofmaxRepresenting the maximum value of the elements in the original load data.
Further, the association between industries is analyzed and quantified based on Shapiro-Wilk test and Pearson correlation coefficient, and strong related industries are screened out according to the association index, and the steps are as follows:
the correlation coefficient is used to evaluate the strength and direction of the linear relationship between two random variables. And analyzing the correlation among random variables by adopting a Pearson correlation coefficient to meet the condition of normal distribution of sample data, and performing normality test on load data of each industry based on a Shapiro-Wilk normal test method. Suppose H0: the sample data obeys normal distribution; h1: the sample data does not follow a normal distribution. And (3) judging the similarity degree of the distribution of the sample to be detected and the normal distribution by calculating the statistic W:
Figure RE-GDA0003378062050000032
wherein W is the statistic returned by Shapiro-Wilk test, and the closer the value of W is to 1, the closer the data is to the standard normal distribution; x is the number ofiRepresenting the ith data in the sample;
Figure RE-GDA0003378062050000033
represents an average value of sample data; a isiThe coefficients for the Shapiro-Wilk test can be obtained by looking up the table.
Based on the statistic W, a significance level alpha is set, and then the quantile W thereof is obtainedα. If W is less than WαThen H is rejected0Otherwise, accept H0. The Shapiro-Wilk test also calculates the P value of the data samples, which is expressed in the original hypothesis H0The probability of observing the sample and more extreme cases occurring under the true premise. If the P value is less than the significance level alpha (statistically, typically 0.05), H is rejected0。
The Pearson correlation coefficient r can be used for analyzing the relevance between the industry X and the industry Y, and for the industry X to be predicted and the related industry Y, the Pearson correlation coefficient calculation formula is as follows:
Figure RE-GDA0003378062050000041
in the formula, r represents a correlation coefficient between industry X and industry Y, and the value of r is between-1 and + 1. XiAnd YiRespectively represent the electricity load data of the ith month of the industry X and the industry Y,
Figure RE-GDA0003378062050000042
and
Figure RE-GDA0003378062050000043
the monthly electrical load means of industry X and Y are shown respectively.
The value range of the correlation coefficient r is between-1 and +1, namely | r | < 1. The closer | r | is to 1, the higher the degree of linear correlation of X with Y. r ═ -1, indicating a completely negative linear correlation between X and Y; r ═ 1, indicating a completely positive linear correlation between X and Y; r is 0, indicating that there is no linear correlation between the two.
Further, on the basis of industry load data, air temperature data and industry relevance, a long-term and short-term memory neural network is used for predicting the industry medium-term and long-term loads, and the method comprises the following steps:
the LSTM Neural Network is a modified Recurrent Neural Network (RNN). Compared with RNN, LSTM has the greatest advantages of solving the long-term dependence problem and enabling the neural network to have the characteristic information of the hidden load of the node with earlier memory time, so that the LSTM has better performance when processing a long-term sequence and is suitable for medium-term and long-term load prediction. The core concept of long-short term memory neural networks is the state memory unit and the "gate" structure. The state memory unit forms a path for transmitting the load prediction information, so that the information can be transmitted in a time sequence; the structure of the gate determines the degree of memory of the load characteristic information in the transmission process. The long-short term memory neural network realizes the information transmission through three basic 'gate' structures, namely a forgetting gate, an input gate and an output gate, and the transmission flow and the updating steps of the industry load characteristic information in the LSTM unit are divided into the following 4 steps:
(1) forgetting door
The role of the forgetting gate in the LSTM is to determine the degree of feature forgetting of the state memory cell. The forgetting gate predicts the load predicted value h from the previous momentt-1And current input feature xtSimultaneously transmitting a sigmoid function and outputting f of the sigmoid functiontAs a forgetting factor of the state memory cell. f. oftCan be expressed as:
ft＝σ(Wf[ht-1,xt]+bf)
in the formula, WfOutput h for forgetting the moment before the doort-1And current input feature xtA multiplied weight matrix; bfBiasing the item for the forgetting gate; sigma represents a sigmoid function; h ist-1A load prediction output representing a previous time; x is the number oftAnd the load prediction characteristic factors representing the current input comprise industry relevance characteristics, air temperature characteristics and the like.
(2) Input gate
The input gate functions to generate the state memory cell intermediate value. First, the input gate outputs the predicted output value h of the previous timet-1And newly input xtThe common input sigmoid function outputs itAs input node state gtThe update coefficient of (2). Input node state gtFrom ht-1And xtAnd (4) jointly determining.
Sigmoid function output i of input gatetThe expression is as follows:
it＝σ(Wi[ht-1,xt]+bi)
in the formula, WiPassing h of sigmoid function into input gatet-1And input feature xtA multiplied weight matrix; biBiasing the term for the input gate sigmoid function.
Input node state gtThe expression is as follows:
gt＝tanh(Wg[ht-1,xt]+bg)
in the formula, WgTransmitting h of tanh function to input gatet-1And input feature xtA multiplied weight matrix; bgThe term is biased for the input gate tanh function.
(3) Refresh state memory cell
The third step of the LSTM workflow is to update the state memory, i.e., C in the mapt. Firstly, the state of the state memory unit at the previous time and the forgetting gate are outputtMultiplying and outputting the selective memory result. Then, the value is added point by point with the intermediate value of the state memory cell to update the state memory cell Ct，CtThe expression is as follows:
Figure RE-GDA0003378062050000051
in the formula (I), the compound is shown in the specification,
Figure RE-GDA0003378062050000052
is the Hadamard product.
(4) Output gate
The output gate is used for determining the load predicted value h at the current momenttFirstly, predict the load h of the previous timet-1And input feature xtTransmitting sigmoid function into the output of ot. Secondly, will get moreNew state memory cell CtPasses to tanh function, and outputs with otMultiplying to obtain the predicted value h of the load at the current momentt. Wherein o istThe expression of (a) is as follows:
ot＝σ(Wo[ht-1,xt]+bo)
in the formula, WoIs ht-1And input feature xtA multiplied weight matrix; boThe gate bias term is output.
Load predicted value h at current momenttThe expression is as follows:
Figure RE-GDA0003378062050000061
。
and training the model by taking the target industry load data, the strongly-related industry load data and the air temperature data as a data set based on the LSTM neural network model. To evaluate the accuracy of the prediction model, the following evaluation indexes were employed:
Figure RE-GDA0003378062050000062
Figure RE-GDA0003378062050000063
in the formula, ykFor real industry load data, yk' load data predicted for model, R1As mean absolute percentage error, R2Is the root mean square error.
And when the evaluation index meets the requirement, completing model training, and if the evaluation index does not meet the requirement, modifying the LSTM model parameter and continuing training.
The invention has the beneficial effects that:
the invention adopts an industry monthly load data restoration method based on iterative interpolation to realize the industry monthly load missing data restoration, the restoration precision can reach more than 95 percent, and the problem of data missing in the existing engineering is effectively solved; the method for quantifying the industry relevance based on the Pearson correlation coefficient is provided, strong relevant industries can be screened out according to quantification results, a load prediction model is built by combining the industry relevance with an LSTM neural network and taking the industry relevance as characteristics, the middle and long term load of the industry is predicted, the precision of the middle and long term load prediction is effectively improved, and the error is kept within 4%.
Drawings
FIG. 1 is a schematic overall flow diagram of the present invention.
Fig. 2 is a schematic diagram of an LSTM network element of the present invention.
FIG. 3 is a graph of the prediction result of the LSTM-based load prediction of the agricultural and sideline food processing industry.
FIG. 4 is a graph of the results of the LSTM-based "food manufacturing" load prediction of the present invention.
FIG. 5 is a graph showing the results of the LSTM-based "alcoholic beverage and purified tea manufacturing" load prediction according to the present invention.
Detailed Description
For better understanding of the objects, technical solutions and effects of the present invention, the present invention will be further explained with reference to the accompanying drawings.
The invention provides a medium-long term load prediction method considering industry relevance, which comprises the following detailed steps:
step 1, an industry monthly load data restoration method based on iterative interpolation is provided, monthly load missing data of a typical industry are restored, and load data are normalized by adopting a min-max standardization method.
Because the original load data usually has the problems of data loss, large magnitude difference and the like, the original load data needs to be repaired and preprocessed. The method comprises the following steps of collecting middle and long term load data of the industry at intervals of months, and forming a data input matrix which can be expressed as follows:
Figure RE-GDA0003378062050000071
in the formula, yijThe month load value of the j month of the i industry; m is the total number of industries; n is the total number of months of data acquisition; t isjIndicating the meteorological parameters for month j.
Aiming at the problem of load data missing,
completing the data by adopting an iterative interpolation method, namely, performing multiple iterations, regarding the load data of the industry where the missing data is located as a dependent variable y, regarding the rest of related characteristic data as an independent variable x, performing fitting modeling and interpolating the missing y value by using a linear regression, performing multiple iteration cycles, and outputting complete load data; the y value is calculated as shown in the following equation:
y＝a1x1+a2x2+...+aMxM
the magnitude of the load may vary greatly from industry to industry, which may lead to increased difficulty in model training. The gradient descent can be accelerated by adopting normalization, so that the convergence speed of the model is improved. Meanwhile, the normalized data is limited to be between 0 and 1, so that the evaluation standard can be unified, and the calculation speed can be improved. And (3) carrying out normalization processing on the load data by adopting a min-max standardization method:
Figure RE-GDA0003378062050000081
in the formula, xi' represents the load data of month i after min-max normalization; x is the number ofiRepresenting the original load data of the ith month; x is the number ofminRepresenting the minimum of elements in the original load data; x is the number ofmaxRepresenting the maximum value of the elements in the original load data.
And 2, analyzing and quantifying the association between industries based on Shapiro-Wilk test and Pearson correlation coefficient, and screening out strongly-related industries according to association indexes.
The correlation coefficient is used to evaluate the strength and direction of the linear relationship between two random variables. Analyzing the correlation among random variables by adopting a Pearson correlation coefficient, wherein the correlation needs to meet the condition of normal distribution of sample data, and correcting the load data of each industry based on a Shapiro-Wilk normal test methodAnd (5) checking attitude. Suppose H0: the sample data obeys normal distribution; h1: the sample data does not follow a normal distribution. And (3) judging the similarity degree of the distribution of the sample to be detected and the normal distribution by calculating the statistic W:
Figure RE-GDA0003378062050000082
wherein W is the statistic returned by Shapiro-Wilk test, and the closer the value of W is to 1, the closer the data is to the standard normal distribution; x is the number ofiRepresenting the ith data in the sample;
Figure RE-GDA0003378062050000083
represents an average value of sample data; a isiThe coefficients for the Shapiro-Wilk test can be obtained by looking up the table.
Based on the statistic W, a significance level alpha is set, and then the quantile W thereof is obtainedα. If W is less than WαThen H is rejected0Otherwise, accept H0. The Shapiro-Wilk test also calculates the P value of the data samples, which is expressed in the original hypothesis H0The probability of observing the sample and more extreme cases occurring under the true premise. If the P value is less than the significance level alpha (statistically, typically 0.05), H is rejected0。
The Pearson correlation coefficient r can be used for analyzing the relevance between the industry X and the industry Y, and for the industry X to be predicted and the related industry Y, the Pearson correlation coefficient calculation formula is as follows:
Figure RE-GDA0003378062050000091
in the formula, r represents a correlation coefficient between industry X and industry Y, and the value of r is between-1 and + 1. XiAnd YiRespectively represent the electricity load data of the ith month of the industry X and the industry Y,
Figure RE-GDA0003378062050000092
and
Figure RE-GDA0003378062050000093
the monthly electrical load means of industry X and Y are shown respectively.
The value range of the correlation coefficient r is between-1 and +1, namely | r | < 1. The closer | r | is to 1, the higher the degree of linear correlation of X with Y. r ═ -1, indicating a completely negative linear correlation between X and Y; r ═ 1, indicating a completely positive linear correlation between X and Y; r is 0, indicating that there is no linear correlation between the two.
Based on the theory of the correlation coefficient, an industry relevance evaluation index is formulated as shown in table 1.
TABLE 1 industry Association evaluation index
Pearson correlation coefficient range | r- Industry relevance evaluation
0.8～1.0 Very strong correlation
0.6～0.8 Strong correlation
0.4～0.6 Moderate correlation
0.2～0.4 Weak correlation
0～0.2 Very weak correlation
And evaluating the industry correlation according to the calculated Pearson correlation coefficient, and screening out the extremely strong related industries.
And 3, predicting the middle-long term load of the industry by using the long-short term memory neural network on the basis of the industry load data, the air temperature data and the industry relevance.
The LSTM Neural Network is a modified Recurrent Neural Network (RNN). Compared with RNN, LSTM has the greatest advantages of solving the long-term dependence problem and enabling the neural network to have the characteristic information of the hidden load of the node with earlier memory time, so that the LSTM has better performance when processing a long-term sequence and is suitable for medium-term and long-term load prediction. The core concept of long-short term memory neural networks is the state memory unit and the "gate" structure. The state memory unit forms a path for transmitting the load prediction information, so that the information can be transmitted in a time sequence; the structure of the gate determines the degree of memory of the load characteristic information in the transmission process. The long-short term memory neural network realizes the information transmission through three basic 'gate' structures, namely a forgetting gate, an input gate and an output gate, as shown in fig. 2. The transmission process and the updating step of the industry load characteristic information in the LSTM unit are divided into the following 4 steps:
(1) forgetting door
The role of the forgetting gate in the LSTM is to determine the degree of feature forgetting of the state memory cell. The forgetting gate predicts the load predicted value h from the previous momentt-1And current input feature xtSimultaneously transmitting a sigmoid function and outputting f of the sigmoid functiontAs a forgetting factor of the state memory cell. f. oftCan be expressed as:
ft＝σ(Wf[ht-1,xt]+bf)
in the formula, WfOutput h for forgetting the moment before the doort-1And current input feature xtA multiplied weight matrix; bfBiasing the item for the forgetting gate; sigma represents a sigmoid function; h ist-1A load prediction output representing a previous time; x is the number oftLoad prediction characteristic factors representing current inputs, including industry association characteristics, gasTemperature characteristics, etc.
(2) Input gate
The input gate functions to generate the state memory cell intermediate value. First, the input gate outputs the predicted output value h of the previous timet-1And newly input xtThe common input sigmoid function outputs itAs input node state gtThe update coefficient of (2). Input node state gtFrom ht-1And xtAnd (4) jointly determining.
Sigmoid function output i of input gatetThe expression is as follows:
it＝σ(Wi[ht-1,xt]+bi)
in the formula, WiPassing h of sigmoid function into input gatet-1And input feature xtA multiplied weight matrix; biBiasing the term for the input gate sigmoid function.
Tan h function output g of input gatetThe expression is as follows:
gt＝tanh(Wg[ht-1,xt]+bg)
in the formula, WgTransmitting h of tanh function to input gatet-1And input feature xtA multiplied weight matrix; bgThe term is biased for the input gate tanh function.
(3) Refresh state memory cell
The third step of the LSTM workflow is to update the state memory, i.e., C in the mapt. Firstly, the state of the state memory unit at the previous time and the forgetting gate are outputtMultiplying and outputting the selective memory result. Then, the value is added point by point with the intermediate value of the state memory cell to update the state memory cell Ct，CtThe expression is as follows:
Figure RE-GDA0003378062050000111
in the formula (I), the compound is shown in the specification,
Figure RE-GDA0003378062050000112
is the Hadamard product.
(4) Output gate
The output gate is used for determining the load predicted value h at the current momenttThe output gate firstly predicts the load predicted value h at the previous momentt-1And input feature xtTransmitting sigmoid function into the output of ot. Next, the output gate will refresh the state memory cell CtPasses to tanh function, and outputs with otMultiplying to obtain the predicted value h of the load at the current momentt. Wherein o istThe expression of (a) is as follows:
ot＝σ(Wo[ht-1,xt]+bo)
in the formula, WoIs ht-1And input feature xtA multiplied weight matrix; boThe gate bias term is output.
Load predicted value h at current momenttThe expression is as follows:
Figure RE-GDA0003378062050000113
in the formula, CtIs an updated state memory cell.
And training the model by taking the target industry load data, the strongly-related industry load data and the air temperature data as a data set based on the LSTM neural network model. To evaluate the accuracy of the prediction model, the following evaluation indexes were employed:
Figure RE-GDA0003378062050000114
Figure RE-GDA0003378062050000121
in the formula, ykFor real industry load data, yk' load data predicted for model, R1As mean absolute percentage error, R2To all areSquare root error.
And when the evaluation index meets the requirement, completing model training, and if the evaluation index does not meet the requirement, modifying the LSTM model parameter and continuing training.
In order to verify the effectiveness and accuracy of the proposed middle-long term load prediction method considering the industry relevance, example simulation is carried out on 2016-year-month 2020 load data of a certain city, wherein 2 major industries and more than 30 related industries are included.
Taking the load data of food manufacturing industry as an example, randomly selecting the load data of 2 months as a data missing point to carry out iterative interpolation, comparing the interpolated load data with the real load data and carrying out error analysis, wherein the result is shown in table 2.
TABLE 2 comparison of interpolated data with actual data
Data missing Point 1 Data missing Point 2
Interpolating data 16566.41 10648.95
Real data 15944.91 10277.74
Relative error 3.90％ 3.61％
As can be seen from the above table, the result of the iterative interpolation matches the real situation well.
Next, taking the food-related industry as an example, the industry load correlation analysis was performed for "agri-sideline food processing industry", "food manufacturing industry", "wine and beverage and purified tea manufacturing industry", and "tobacco manufacturing industry". The above industry load data was first subjected to the Shapiro-Wilk test, and the results are shown in Table 3.
TABLE 3 Normal distribution test results for industry load data
Figure RE-GDA0003378062050000122
Figure RE-GDA0003378062050000131
From Table 3, it can be seen that "agricultural and sideline food processing industry", "food manufacturing industry", "wine and beverage and refined tea manufacturing industry" obeyed normal distribution and could be analyzed by Pearson's correlation coefficient method.
And after verifying that the industry load data meet the normal distribution condition, calculating a correlation coefficient matrix between industries by using a Pearson correlation coefficient method. Correlation analysis is performed in a period of 12 months, 4 correlation coefficients between every two industries in 2016-2019 years are finally obtained, and an average value is taken as a final correlation coefficient matrix element, as shown in table 4.
TABLE 4 industry load data correlation coefficient matrix
Industry Agricultural and sideline food processing industry Food manufacturing industry Production of alcoholic beverages and refined tea
Agricultural and sideline food processing industry 1 0.962799 0.916886
Food manufacturing industry 0.962799 1 0.792809
Production of alcoholic beverages and refined tea 0.916886 0.792809 1
Based on the LSTM neural network prediction method considering the industry relevance, the industry load data is divided into a training set and a testing set, and the model is trained to obtain the industry load prediction curves shown in figures 3 to 5 and the industry load prediction error analysis table shown in table 5.
According to the table 5, the error of the load prediction result of the food related industry is small after the industry relevance is considered, and the prediction result well meets the precision requirement.
TABLE 5 industry load prediction error analysis
Industry R1 R2
Agricultural and sideline food processing industry 3.963％ 839.881
Food manufacturing industry 1.166％ 232.081
Production of alcoholic beverages and refined tea 0.987％ 81.246
In order to evaluate the value of the industry relevance for the long-term load prediction in the industry, the load prediction method without considering the industry relevance is compared with the method on the basis of the load data of the agricultural and sideline food processing industry, the food manufacturing industry and the wine and refined tea manufacturing industry, and the comparison prediction result is shown in table 6.
TABLE 6 comparison of forecasted results without trade associations
Figure RE-GDA0003378062050000141
As can be seen from Table 6, the prediction evaluation indexes considering the industry relevance are all superior to those under the condition that the industry relevance is not considered, and the prediction precision of the long-term load in the industry is effectively improved by considering the industry relevance.
In conclusion, the method realizes monthly load missing data recovery in typical industries, the recovery precision reaches more than 96%, the problem of data missing in actual services is solved, meanwhile, the medium-long term load prediction error is kept within 5%, and the medium-long term load prediction precision is effectively improved.

Claims (4)

1. A medium-long term load prediction method considering industry relevance is characterized by comprising the following steps:
(1) restoring monthly load missing data by an industry monthly load data restoration method based on iterative interpolation, and performing normalization processing on load data by adopting a min-max standardization method;
(2) analyzing and quantifying the association between industries based on Shapiro-Wilk test and Pearson correlation coefficient, and screening out strong related industries according to association indexes;
(3) and on the basis of the industry load data, the air temperature data and the industry relevance, predicting the long-term load and the medium-term load of the industry by using the long-term and short-term memory neural network.
2. The method for predicting the medium-long term load considering the industry relevance as claimed in claim 1, wherein: the step (1) is specifically as follows:
the method comprises the following steps of collecting middle and long term load data of the industry at intervals of months, and forming a data input matrix which can be expressed as follows:
Figure FDA0003310802120000011
in the formula, yijThe month load value of the j month of the i industry; m is the total number of industries; n is the total number of months of data acquisition; t isjWeather parameters representing month j;
aiming at the problem of load data loss, an iterative interpolation method is adopted to complement data, namely, through multiple iterations, the load data of the industry where the loss data is located is regarded as a dependent variable y, the rest relevant characteristic data is regarded as an independent variable x, a linear regression device is used for carrying out fitting modeling and interpolation on the lost y value, multiple iteration cycles are carried out, and complete load data are output; the y value is calculated as shown in the following equation:
y＝a1x1+a2x2+...+aMxM
and (3) carrying out normalization processing on the load data by adopting a min-max standardization method:
Figure FDA0003310802120000021
in the formula, xi' represents the load data of month i after min-max normalization; x is the number ofiRepresenting the original load data of the ith month; x is the number ofminRepresenting the minimum of elements in the original load data; x is the number ofmaxRepresenting the maximum value of the elements in the original load data.
3. The method for predicting the medium-long term load considering the industry relevance as claimed in claim 1, wherein: the step (2) is specifically as follows:
analyzing the correlation among random variables by adopting a Pearson correlation coefficient, wherein the correlation needs to meet the condition of normal distribution of sample data, and carrying out normality test on load data of each industry based on Shapiro-Wilk test; suppose H0: the sample data obeys normal distribution; h1: sample data does not follow normal distribution; and (3) judging the similarity degree of the distribution of the sample to be detected and the normal distribution by calculating the statistic W:
Figure RE-FDA0003378062040000022
wherein W is the statistic returned by Shapiro-Wilk test, and the closer the value of W is to 1, the closer the data is to the standard normal distribution; x is the number ofiRepresenting the ith data in the sample;
Figure RE-FDA0003378062040000026
represents an average value of sample data; a isiThe coefficients for the Shapiro-Wilk test can be obtained by table lookup;
based on the statistic W, a significance level alpha is set, and then the quantile W thereof is obtainedα(ii) a If it isW＜WαThen H is rejected0Otherwise, accept H0(ii) a And calculating the P value of the data sample using the Shapiro-Wilk test, and rejecting H if the P value is less than a significance level alpha0；
The Pearson correlation coefficient r is used for analyzing the relevance between the industry X and the industry Y, and for the industry X to be predicted and the related industry Y, the Pearson correlation coefficient calculation formula is as follows:
Figure RE-FDA0003378062040000023
wherein r represents a correlation coefficient between industry X and industry Y; xiAnd YiRespectively represent the electricity load data of the ith month of the industry X and the industry Y,
Figure RE-FDA0003378062040000024
and
Figure RE-FDA0003378062040000025
respectively representing monthly electric load mean values of industries X and Y;
the value range of the correlation coefficient r is between-1 and +1, namely | r | is less than or equal to 1; the closer the | r | is to 1, the higher the linear correlation degree of X and Y is; r ═ -1, indicating a completely negative linear correlation between X and Y; r ═ 1, indicating a completely positive linear correlation between X and Y; r is 0, indicating that there is no linear correlation between the two.
4. The method for predicting the medium-long term load considering the industry relevance as claimed in claim 1, wherein: the step (3) is specifically as follows:
(1) predicting the load h from the previous momentt-1And current input feature xtSimultaneously transmitting a sigmoid function and outputting f of the sigmoid functiontAs a forgetting coefficient of the state memory unit; f. oftCan be expressed as:
ft＝σ(Wf[ht-1,xt]+bf)
in the formula, WfOutput h for forgetting the moment before the doort-1And current input feature xtA multiplied weight matrix; bfBiasing the item for the forgetting gate; sigma represents a sigmoid function; h ist-1Indicating a predicted load value at a previous time; x is the number oftRepresenting the currently input load prediction characteristic factors, including industry relevance characteristics and air temperature characteristics;
(2) the predicted output value h of the previous moment is outputt-1And newly input xtThe common input sigmoid function outputs itAs input node state gtThe update coefficient of (2);
output itThe expression is as follows:
it＝σ(Wi[ht-1,xt]+bi)
in the formula, WiPassing h of sigmoid function into input gatet-1And input feature xtA multiplied weight matrix; biBiasing the term for the sigmoid function of the input gate;
input node state gtThe expression is as follows:
gt＝tanh(Wg[ht-1,xt]+bg)
in the formula, WgTransmitting h of tanh function to input gatet-1And input feature xtA multiplied weight matrix; bgBiasing the term for the input gate tanh function;
(3) refresh state memory cell Ct: firstly, the state of the state memory unit at the previous time and the forgetting gate are outputtMultiplying and outputting a selective memory result; then, the selective memory result is added to the intermediate value of the state memory cell point by point to update the state memory cell Ct，CtThe expression is as follows:
Figure FDA0003310802120000031
in the formula (I), the compound is shown in the specification,
Figure FDA0003310802120000032
is a Hadamard product;
(4) determining the load predicted value h at the current momenttFirstly, predict the load h of the previous timet-1And input feature xtTransmitting sigmoid function into the output of ot(ii) a Next, the updated state memory cell CtPasses to tanh function, and outputs with otMultiplying to obtain the predicted value h of the load at the current momentt(ii) a Wherein o istThe expression of (a) is as follows:
ot＝σ(Wo[ht-1,xt]+bo)
in the formula, WoFor the load prediction value h of the previous momentt-1And input feature xtA multiplied weight matrix; boAn output gate bias term;
load predicted value h at current momenttThe expression is as follows:
Figure FDA0003310802120000041
CN202111215938.1A 2021-10-19 2021-10-19 Medium-and-long-term load prediction method considering industry relevance Pending CN113962456A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111215938.1A CN113962456A (en) 2021-10-19 2021-10-19 Medium-and-long-term load prediction method considering industry relevance

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111215938.1A CN113962456A (en) 2021-10-19 2021-10-19 Medium-and-long-term load prediction method considering industry relevance

Publications (1)

Publication Number Publication Date
CN113962456A true CN113962456A (en) 2022-01-21

Family

ID=79465323

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111215938.1A Pending CN113962456A (en) 2021-10-19 2021-10-19 Medium-and-long-term load prediction method considering industry relevance

Country Status (1)

Country Link
CN (1) CN113962456A (en)

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106779346A (en) * 2016-12-01 2017-05-31 国家电网公司 A kind of Forecasting Methodology of monthly power consumption
CN108846528A (en) * 2018-09-19 2018-11-20 云南电网有限责任公司信息中心 A kind of big industrial electrical user long term load forecasting method
CN111027775A (en) * 2019-12-12 2020-04-17 中国长江三峡集团有限公司 Step hydropower station generating capacity prediction method based on long-term and short-term memory network
CN111241755A (en) * 2020-02-24 2020-06-05 国网(苏州)城市能源研究院有限责任公司 Power load prediction method
CN112149879A (en) * 2020-09-02 2020-12-29 上海电力大学 New energy medium-and-long-term electric quantity prediction method considering macroscopic volatility classification
CN112348168A (en) * 2020-10-27 2021-02-09 国网四川省电力公司经济技术研究院 Ultra-short-term load prediction method and system considering data loss and characteristic redundancy
CN112508259A (en) * 2020-12-02 2021-03-16 重庆大学 Medium-and-long-term power consumption prediction method and system for independent user categories

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106779346A (en) * 2016-12-01 2017-05-31 国家电网公司 A kind of Forecasting Methodology of monthly power consumption
CN108846528A (en) * 2018-09-19 2018-11-20 云南电网有限责任公司信息中心 A kind of big industrial electrical user long term load forecasting method
CN111027775A (en) * 2019-12-12 2020-04-17 中国长江三峡集团有限公司 Step hydropower station generating capacity prediction method based on long-term and short-term memory network
CN111241755A (en) * 2020-02-24 2020-06-05 国网(苏州)城市能源研究院有限责任公司 Power load prediction method
CN112149879A (en) * 2020-09-02 2020-12-29 上海电力大学 New energy medium-and-long-term electric quantity prediction method considering macroscopic volatility classification
CN112348168A (en) * 2020-10-27 2021-02-09 国网四川省电力公司经济技术研究院 Ultra-short-term load prediction method and system considering data loss and characteristic redundancy
CN112508259A (en) * 2020-12-02 2021-03-16 重庆大学 Medium-and-long-term power consumption prediction method and system for independent user categories

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
SIBTAIN MUHAMMAD ETAL: "" A Multivariate and Multistage Medium- and Long-Term Streamflow Prediction Based on an Ensemble of Signal Decomposition Techniques with a Deep Learning Network"", 《ADVANCES IN METEOROLOGY》, vol. 2020, no. 1, 18 August 2020 (2020-08-18), pages 1032 - 1037 *
傅靖等: "《基于基因表达式编程的中长期电力负荷预测挖掘》", 《电力信息与通信技术》, vol. 18, no. 05, 25 May 2020 (2020-05-25), pages 7 - 12 *

Similar Documents

Publication Publication Date Title
CN106874581B (en) Building air conditioner energy consumption prediction method based on BP neural network model
CN107169628B (en) Power distribution network reliability assessment method based on big data mutual information attribute reduction
CN107315884B (en) Building energy consumption modeling method based on linear regression
CN109766583A (en) Based on no label, unbalanced, initial value uncertain data aero-engine service life prediction technique
CN110245801A (en) A kind of Methods of electric load forecasting and system based on combination mining model
CN104361414B (en) Power transmission line icing prediction method based on correlation vector machine
CN107463993B (en) Medium-and-long-term runoff forecasting method based on mutual information-kernel principal component analysis-Elman network
CN111105104A (en) Short-term power load prediction method based on similar day and RBF neural network
CN111814956B (en) Multi-task learning air quality prediction method based on multi-dimensional secondary feature extraction
CN113205207A (en) XGboost algorithm-based short-term power consumption load fluctuation prediction method and system
CN112036075A (en) Abnormal data judgment method based on environmental monitoring data association relation
CN109948920B (en) Electric power market settlement data risk processing method based on evidence theory
CN116303786B (en) Block chain financial big data management system based on multidimensional data fusion algorithm
CN115358437A (en) Power supply load prediction method based on convolutional neural network
CN117150409A (en) Power consumption abnormality detection method
CN113988415B (en) Medium-and-long-term power load prediction method
CN116826737A (en) Photovoltaic power prediction method, device, storage medium and equipment
CN113591402A (en) Digital power transformer health state fuzzy comprehensive evaluation method
CN113052353B (en) Air quality prediction and prediction model training method and device and storage medium
CN113191526A (en) Short-term wind speed interval multi-objective optimization prediction method and system based on random sensitivity
CN112232570A (en) Forward active total electric quantity prediction method and device and readable storage medium
CN116565850A (en) Wind power ultra-short-term prediction method based on QR-BLSTM
CN114971022B (en) Wind power prediction method based on D-value-LSTM neural network model
CN116151799A (en) BP neural network-based distribution line multi-working-condition fault rate rapid assessment method
CN116613732A (en) Multi-element load prediction method and system based on SHAP value selection strategy

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination