CN117374963A - Ultra-short-term power demand prediction method for iron and steel enterprises based on load characteristics - Google Patents

Ultra-short-term power demand prediction method for iron and steel enterprises based on load characteristics Download PDF

Info

Publication number
CN117374963A
CN117374963A CN202311479020.7A CN202311479020A CN117374963A CN 117374963 A CN117374963 A CN 117374963A CN 202311479020 A CN202311479020 A CN 202311479020A CN 117374963 A CN117374963 A CN 117374963A
Authority
CN
China
Prior art keywords
load
production
prediction
data
model
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
CN202311479020.7A
Other languages
Chinese (zh)
Inventor
何小红
周有为
高忠江
钟雨哲
吴丹
冉薇
秦香玲
范与舟
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.)
Cisdi Electric Technology Co ltd
Original Assignee
Cisdi Electric 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 Cisdi Electric Technology Co ltd filed Critical Cisdi Electric Technology Co ltd
Priority to CN202311479020.7A priority Critical patent/CN117374963A/en
Publication of CN117374963A publication Critical patent/CN117374963A/en
Pending legal-status Critical Current

Links

Classifications

    • H—ELECTRICITY
    • H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
    • H02J—CIRCUIT ARRANGEMENTS OR SYSTEMS FOR SUPPLYING OR DISTRIBUTING ELECTRIC POWER; SYSTEMS FOR STORING ELECTRIC ENERGY
    • H02J3/00—Circuit arrangements for ac mains or ac distribution networks
    • H02J3/004—Generation forecast, e.g. methods or systems for forecasting future energy generation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/217—Validation; Performance evaluation; Active pattern learning techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/27—Regression, e.g. linear or logistic regression
    • 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
    • G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
    • G06Q10/063—Operations research, analysis or management
    • G06Q10/0631—Resource planning, allocation, distributing or scheduling for enterprises or organisations
    • G06Q10/06315—Needs-based resource requirements planning or analysis
    • 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
    • G06Q50/04—Manufacturing
    • 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)
  • Business, Economics & Management (AREA)
  • Theoretical Computer Science (AREA)
  • Human Resources & Organizations (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Physics & Mathematics (AREA)
  • Economics (AREA)
  • Strategic Management (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Tourism & Hospitality (AREA)
  • Evolutionary Biology (AREA)
  • General Business, Economics & Management (AREA)
  • Marketing (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Health & Medical Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Primary Health Care (AREA)
  • General Health & Medical Sciences (AREA)
  • Public Health (AREA)
  • Water Supply & Treatment (AREA)
  • Power Engineering (AREA)
  • Educational Administration (AREA)
  • Manufacturing & Machinery (AREA)
  • Development Economics (AREA)
  • Game Theory and Decision Science (AREA)
  • Quality & Reliability (AREA)
  • Operations Research (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention relates to a load characteristic-based ultra-short-term power demand prediction method for steel enterprises, and belongs to the technical field of power. The method comprises the following steps: s1: constructing and training a prediction model; s2: carrying out engineering application and online automatic correction on the prediction model; the method takes the iron and steel enterprises which pay the basic electric charge according to the mode of maximum demand as a research object, analyzes the load characteristics of the iron and steel enterprises, finishes the ultra-short-term power demand prediction, predicts the maximum demand within 15 minutes in advance, and achieves the aim of saving the basic electric charge. Compared with the prior art, the method improves the extraction capability of the data characteristics and the change rule, and reasonably adopts a traditional time sequence model, a time sequence prediction model based on a neural network and a prediction model based on process characteristics according to the load characteristics.

Description

Ultra-short-term power demand prediction method for iron and steel enterprises based on load characteristics
Technical Field
The invention belongs to the technical field of power, and relates to a load characteristic-based ultra-short-term power demand prediction method for steel enterprises.
Background
Definition:
(1) Two electricity rates: the electricity price is divided into two parts, one part is a basic electricity price, also called capacity (demand) electricity price, and the other part is an electricity price.
The two electricity price making users can pay the basic electricity fee according to one of the following three modes:
(1) basic electricity charge is calculated according to the capacity of the transformer: the calculation is carried out according to the actual running transformer capacity (without capacity of transacted capacity reduction and suspended service) of the user;
(2) basic electricity charge, namely transformer capacity x capacity electricity price is calculated according to the capacity;
(3) and (3) calculating the basic electricity charge according to the mode of maximum demand, wherein the maximum demand value of the user in the current month is the actual maximum demand multiplied by the electricity price of the demand.
(2) Relationship between load and demand:
the load is active power at any moment, namely P;
the required amount is the average value of active power within 15 minutes, and the acquisition interval is calculated according to 1s, namely
The maximum required amount is the maximum value of the average active power value within 15 minutes, i.e
P max ＝max(P max ,P avg )
The power market is currently competing and opening, and not only needs to consider peak Gu Pingsheng production period, but also needs to consider electricity prices fluctuating at any time. In order to realize peak-shifting power consumption, reasonable power charge saving and short-term prediction of power demand are particularly important.
The steel enterprises are large power consumers, steel plants contain a large number of high-power impact loads, so that the power fluctuation of the steel plants connected to a power system is strong, and the stability of a power grid is threatened to a certain extent. The huge power fluctuation of the steel mill damages the system stability on one hand, causes idle standby and capacity waste of the system generator set, forces the power department to newly build a power plant, and aggravates environmental pollution; on the other hand, the cost of the steel mill is greatly increased, and the economic benefit is greatly reduced. The strong volatility of the steel mill can cause great harm to a power grid system and the steel mill, and an optimal control method is required to be sought, so that load fluctuation is stabilized and reduced, and the stability of the power system is improved.
Some researchers have proposed load prediction related algorithms, patent 1-patent 2 are load prediction methods in the steel industry, and patent 3-patent 4 are general load prediction methods:
patent 1: patent document publication number CN 103606018B discloses a system for short-term dynamic prediction of electric load of iron and steel enterprises. The system comprehensively considers the electricity utilization characteristics, the process characteristics, the production plan, the overhaul plan and the production working condition information of each electricity utilization link, establishes a model in a classified manner, and acquires a total load predicted value through superposition of predicted results.
Patent 2: patent document publication No. CN 116417989a discloses a method for predicting active power of an electric power system in a steel plant. According to the daily planned output of the main working procedure, the method calculates the total power consumption by adopting a recursive least square algorithm, takes the average value of the minute level as the daily forecast reference value, and enables the algorithm to adapt to the working condition change. The source of load fluctuation is extracted and analyzed, and a slope-based similarity wave algorithm is adopted to predict the load.
Patent 3: patent publication CN 107578124a discloses a short-term power load prediction method based on a multi-layer improved GRU neural network. The method uses the GRU neural network model to play the role of the cyclic neural network in time sequence data mining, so that the training speed and the training efficiency are improved; a SELU activation function is introduced, so that the problems of gradient disappearance and gradient explosion of the multilayer neural network in the process of parameter optimization by using a gradient descent method are effectively avoided; a multi-layer improved GRU neural network is built, and a Sequence2one training and predicting mode is adopted, so that the data mining capacity and efficiency of the network are further improved.
Patent 4: patent document publication No. CN 111738512B discloses a short-term power load prediction method based on CNN-IPSO-GRU hybrid model. Firstly, collecting data such as historical load, meteorological factors and date information of a power grid, carrying out data normalization processing, dividing a training set and a testing set, extracting multidimensional feature vectors representing load change by using a convolutional neural network technology, and constructing a time sequence as input of a model; then constructing a gating circulation unit network prediction model, optimizing the gating circulation unit network prediction model by utilizing training set data through an improved particle swarm algorithm to obtain two optimal prediction model parameters, and reestablishing the gating circulation unit network model according to the obtained optimal prediction model parameters; and finally, short-term load prediction of the power grid is realized by using the test set data.
The prior art has the following defects:
(1) Because a large number of high-power impact loads exist in the electric power system in the steel industry, the relevance between the loads and seasonal changes, temperature changes and holidays in workdays is not strong, the load prediction algorithm based on the intelligent algorithm in the current electric power industry cannot be directly used, and the process characteristics and actual working conditions in the steel industry need to be fully considered.
(2) Although load prediction algorithms in the steel industry have been proposed by researchers and the power consumption characteristics, process characteristics, production shutdown plans and production working condition information of each power consumption link are considered, the extraction capability of the characteristics and change rules of the data is limited, and the time sequence characteristics of the historical data cannot be fully utilized for prediction.
(3) In the prior art, an automatic correction module is lacking, most of the models are predicted by adopting a fixed algorithm model, the models cannot be corrected according to the latest production data, the self-learning capability is lacking, and the model adaptability is to be improved.
(4) Most of the prior art schemes are predictive loads, and for enterprises adopting the maximum demand charge, the purposes of saving electricity fees and improving economic benefits of enterprises cannot be achieved.
Disclosure of Invention
In view of the above, the invention aims to provide a load characteristic-based ultra-short-term power demand prediction method for steel enterprises.
In order to achieve the above purpose, the present invention provides the following technical solutions:
a method for predicting ultra-short-term power demand of a steel enterprise based on load characteristics comprises the following steps:
s1: constructing and training a prediction model;
s2: carrying out engineering application and online automatic correction on the prediction model;
s11: data collection
Confirming the production mode and the load condition of each production process of a steel plant, and collecting the load data, the production shutdown plan, the production shutdown actual performance and the information of actual production signals of each production process of the steel plant in one year and the load data of 220kV incoming line; the collected load data acquisition interval is 1s;
s12: data cleaning, including load data missing processing and working condition missing processing;
s13: data analysis preprocessing
Each production procedure converts the production mode obtained by investigation and the collected production shutdown plan or actual production signal into a planned production mode, and converts the production mode obtained by investigation and the collected production shutdown actual score into an actual score production mode;
analyzing the load by using an inspection ADF (Augmented Dickey-Fuller Testing), and judging the stability of the load sequence; for the load of non-strong fluctuation, performing autocorrelation analysis ACF, analyzing the correlation of data under different time delays, and observing whether the data has periodicity; the calculation formula is as follows:
where N is the length of the time series X, k is the specific lag of the time series, acf k X represents t And X is t-k Is a correlation of (2);
analyzing whether the load fluctuation of each process has high association degree with the production mode or not, and whether the production mode of each process is changed frequently or not; according to the analysis result, the loads are divided into the following four categories, and the input factors of each category of loads are related to the load characteristics:
(1) more stable load: input factors include load and production mode;
(2) cyclic loading: input factors include load and production mode;
(3) multi-factor load: input factors include load, production mode, steel rolling specification and yield;
(4) strong fluctuating load: input factors include load, production event, and time;
in order to fully consider the load difference of different working conditions of each working procedure, the historical data of each working procedure is intelligently segmented according to different production modes, then the historical data is constructed into a steel enterprise load data set, the data is divided into a training set and a verification set by adopting a leaving method, and the ratio is 7:3, testing by using real data without setting a test set; the consistency of the data distribution of each procedure is maintained, and the proportion of sample types is similar;
s14: predictive algorithm construction
According to the load division of the data analysis preprocessing stage, each procedure is overlapped after the prediction is completed according to the corresponding type, and the total load predicted value of the incoming line is obtained, namely
P pre ＝P 1 +P 2 +...+P n
Wherein P is pre Load predictive value of 220kV incoming line is represented, n represents the number of procedures, and P i A load prediction value indicating a certain process;
the load characteristics of the four types of load prediction methods are considered:
(1) more stable load: using differential autoregressive moving average model ARIMA
The ARIMA model combines three methods of autoregressive AR, differential I and moving average MA to describe the dynamic characteristics of time series data; the ARIMA model is denoted ARIMA (p, d, q), where p represents the autoregressive order, d represents the number of times the time series data is differentiated, and q represents the moving average order;
the AR part represents a linear combination of the current value and the past p values of the time series, expressed as:
X t ＝c+φ 1 X t-1 +φ 2 X t-2 +…+φ p X t-p +ε t
the MA portion represents a linear combination of the current value of the time series and the past q hysteresis errors, expressed as:
X t ＝μ+ε t +θ 1 ε t-1 +θ 2 ε t-2 +…+θ q ε t-q
wherein X is t Representing the current value, epsilon, of the time series t Representing the error, phi, of the current time i And theta i Representing autoregressive coefficients and moving average coefficients, c and μ representing constant terms;
part I eliminates non-stationarity in the time series, which is one of the important components of ARIMA model, and the first-order differential time series is expressed as:
X t ′＝X t -X t-1
after integrating the three parts AR, I, MA, the ARIMA model is expressed as:
X t ′＝c+φ 1 X t-1 ′+…+φ p X t-p ′+ε t +θ 1 ε t-1 +…+θ q ε t-q
solving parameters of the ARIMA model by using a least square method;
(2) cyclic loading: because the fluctuation of the load has obvious periodicity, the aperiodic and periodic factors are comprehensively considered, and a seasonal differential autoregressive moving average model SARIMA is used;
SARIMA is based on ARIMA model taking into account periodicity factors; firstly removing periodicity in the sequence, firstly using ARIMA at periodic intervals to obtain a non-stationary non-periodic time sequence, and then using ARIMA again for analysis on the basis;
(3) multi-factor load: the predicted input variables are more, the relation between the characteristic information and the predicted value is required to be accurately captured, the time characteristic information is fully mined, the current moment is predicted by utilizing short-term and long-term information, a better prediction effect can be obtained, and the time sequence prediction is performed by utilizing a long-short-term memory network LSTM;
LSTM learns which information to memorize and which information to forget through a training process; the cell state is divided into two parts, long term state c (t) And short term state h (t) The method comprises the steps of carrying out a first treatment on the surface of the There are three control gates along the state path: forgetting door f (t) Input gate i (t) And an output gate o (t) ；
Forgetting door f (i) Controlling from a previous long-term state c by a sigmoid activation function (t-1) Is to remove information:
input gate i (i) Controlling information from current output g by sigmoid activation function (t) Added to the current long-term state c (i) In (a) and (b); output door o (i) Using the current long-term state c (t) Is used for controlling the current short-term state h (t) Is formed of (a);
output g (t) In effect is a standard loop layer:
if all control gates have been removed and the long-term and short-term states are combined, the LSTM cell will switch back to the standard loop layer, with output g (t) Equal to output layer z (t) And state layer h (t) The method comprises the steps of carrying out a first treatment on the surface of the In the LSTM cell, output g (t) Only partial transition to current state c (i) And h (t) The method comprises the steps of carrying out a first treatment on the surface of the The control equation output by the final unit, the long-term state and the short-term state are as follows:
(4) strong fluctuating load: according to historical load data, production signals and steel grade information, heating characteristic curves of refining furnaces of all steel grades are constructed, and the characteristic curves are corrected every day according to actual conditions; the prediction is finished in advance according to the production signals or the planning data, and then the real-time load is combined for automatic correction so as to reduce the prediction deviation;
s15: model training
Training the constructed algorithm model by using a training set of each procedure, adjusting the super-parameters of the model and the prediction capability of the preliminary evaluation model by using a verification set, and iterating the prediction model according to the verification result until a good prediction effect can be obtained and no fitting condition exists; and after training and verifying the prediction model of each procedure, obtaining a final algorithm model.
Optionally, in the step S12, the load data missing processing is to check the condition of data missing for the load data of each production process, if the missing time is longer than 1 minute, the processing is not performed, otherwise, the load average value of 30 seconds before and after the missing time is used for filling; the working condition missing treatment is as follows: if the production shutdown plan, the production shutdown actual results or the actual production signals are absent in a part of the time period in each process, the load data of the time period is directly deleted.
Optionally, the step S2 specifically includes the following steps:
s21: acquiring real-time data
The data used for prediction comprises the following parts:
(1) real-time active power collected by IEC-870-5-104 protocol;
(2) a production shutdown plan of each production line provided by a user;
(3) acquiring real-time production signal data from the basic automation system L1 and the process control system L2;
s22: predicting demand
The ultra-short-term power demand prediction interval of the steel industry is 3 seconds, prediction is triggered every three seconds, a load and demand curve of 15 minutes in the future is predicted, and a maximum demand predicted value of 15 minutes in the future is calculated according to the historical load of 15 minutes and the predicted load of 15 minutes in the future;
if the production process belongs to a relatively stable load, a periodic load or a multi-factor load, the prediction interval is 1 hour, namely, the load and demand curve prediction of 1 hour in the future is completed every hour; if the production plan update is received, re-predicting; if the production process belongs to a strong fluctuation load, the prediction interval is 3 seconds; then, when predicting every 3 seconds, carrying out weighted average according to the actual load condition and the completed predicted load, and improving the prediction precision;
s23: automatic correction
The data used for correction comprises the following parts:
(1) historical active power for the previous day;
(2) production shutdown actual results of each production line;
(3) acquiring real-time production signal data from the basic automation system L1 and the process control system L2;
and yesterday historical data is used as a verification set every day, and a prediction model of each production mode under each procedure is automatically iterated, so that the prediction precision is improved.
Alternatively, the differential autoregressive moving average model ARIMA is replaced with an autoregressive moving average model ARMA.
Alternatively, the least square method is replaced by a maximum likelihood estimation method or a moment estimation method.
Alternatively, the SARIMA is replaced with an LSTM or a recurrent neural network RNN.
Alternatively, the LSTM is replaced with a recurrent neural network RNN.
The invention has the beneficial effects that:
the method takes the iron and steel enterprises which pay the basic electric charge according to the maximum demand mode as a research object, analyzes the load characteristics of the iron and steel enterprises, finishes the ultra-short-term power demand prediction, predicts the maximum demand within 15 minutes in advance (namely 15 minutes in the past and 15 minutes in the future), and achieves the aim of saving the basic electric charge.
(1) Compared with the prior art, the method improves the extraction capability of data characteristics and change rules, and reasonably adopts traditional time sequence models (namely ARIMA and SARIMA), a time sequence prediction model (namely LSTM) based on a neural network and a prediction model based on process characteristics according to load characteristics.
(2) Compared with the prior art, the method and the device have the advantages that an automatic correction module is added, and the prediction model is corrected daily according to yesterday data, so that the model adaptability can be improved.
(3) Compared with the prior art, the method further predicts the demand curve and the maximum demand, and can achieve the purposes of saving electricity charge and improving economic benefits of enterprises.
(4) Compared with the prior art, the invention can better predict the load of the refining furnace, and can acquire the real-time production signal data of the refining furnace from the basic automation system (L1) and the process control system (L2), so that a prediction model based on the production signal and the load can be constructed, and the prediction result can be corrected in real time.
Additional advantages, objects, and features of the invention will be set forth in part in the description which follows and in part will become apparent to those having ordinary skill in the art upon examination of the following or may be learned from practice of the invention. The objects and other advantages of the invention may be realized and obtained by means of the instrumentalities and combinations particularly pointed out in the specification.
Drawings
For the purpose of making the objects, technical solutions and advantages of the present invention more apparent, the present invention will be described in the following preferred detail with reference to the accompanying drawings, in which:
FIG. 1 is a flow chart of the present invention;
FIG. 2 is a flow chart of the engineering application and on-line automatic correction of the present invention.
Detailed Description
Other advantages and effects of the present invention will become apparent to those skilled in the art from the following disclosure, which describes the embodiments of the present invention with reference to specific examples. The invention may be practiced or carried out in other embodiments that depart from the specific details, and the details of the present description may be modified or varied from the spirit and scope of the present invention. It should be noted that the illustrations provided in the following embodiments merely illustrate the basic idea of the present invention by way of illustration, and the following embodiments and features in the embodiments may be combined with each other without conflict.
Wherein the drawings are for illustrative purposes only and are shown in schematic, non-physical, and not intended to limit the invention; for the purpose of better illustrating embodiments of the invention, certain elements of the drawings may be omitted, enlarged or reduced and do not represent the size of the actual product; it will be appreciated by those skilled in the art that certain well-known structures in the drawings and descriptions thereof may be omitted.
The same or similar reference numbers in the drawings of embodiments of the invention correspond to the same or similar components; in the description of the present invention, it should be understood that, if there are terms such as "upper", "lower", "left", "right", "front", "rear", etc., that indicate an azimuth or a positional relationship based on the azimuth or the positional relationship shown in the drawings, it is only for convenience of describing the present invention and simplifying the description, but not for indicating or suggesting that the referred device or element must have a specific azimuth, be constructed and operated in a specific azimuth, so that the terms describing the positional relationship in the drawings are merely for exemplary illustration and should not be construed as limiting the present invention, and that the specific meaning of the above terms may be understood by those of ordinary skill in the art according to the specific circumstances.
The method provides a load characteristic-based ultra-short-term power demand prediction method for the steel enterprises, the power utilization characteristics, the production plan and the production working conditions of each branch plant of the steel enterprises are fully considered through collection and cleaning of historical data, classification modeling is carried out according to the analyzed load characteristics, and parameters of a prediction model are automatically corrected every day according to actual conditions so as to continuously improve model adaptability and prediction accuracy. The method has the advantages that the prediction and the warning are advanced, the production rhythm can be adjusted in time by a user, the unexpected out-of-limit of the demand is avoided, and the data support is provided for saving the demand cost.
The steel production process mainly comprises the processes of iron making, steel rolling and the like, and the working conditions and the load conditions of each production process are obviously different, so that the demand prediction of a steel enterprise needs to be carried out on each production process, then the load prediction values of incoming lines are obtained through summarization, the demand prediction values are calculated, and finally the maximum demand prediction value is obtained.
The basic steps of the method for constructing and training the model are shown in fig. 1.
S11: data collection
The data of the method are derived from the real operation data of large-scale steel enterprises.
The method comprises the steps of going deep into an iron and steel enterprise to conduct investigation, firstly, determining the production mode and basic load condition of each production process, and then collecting information such as load data, production shutdown plan, production shutdown actual results, actual production signals and the like of each production process of the steel plant 2022 and 220kV incoming line load data. The load data collection interval was 1s.
S12: data cleansing
(1) Load data missing processing: for the load data of each production process, the condition of data missing is checked, and since only a very small amount of data is missing, if the missing time is longer than 1 minute, the processing is not performed, otherwise, the load average value of 30 seconds before and after the missing time is used for filling.
(2) And (3) working condition missing treatment: if there is a case where a production stop plan, a production stop actual result or an actual production signal is absent in a partial period of each process, load data of the period is directly deleted.
S13: data analysis preprocessing
Each production procedure converts the production mode obtained by investigation and the collected production shutdown plan or actual production signal into a planned production mode, and converts the production mode obtained by investigation and the collected production shutdown actual result into an actual result production mode.
The load was analyzed using ADF test (Augmented Dickey-Fuller Testing) to determine the smoothness of the load sequence. For loads that are not strongly fluctuating, an autocorrelation Analysis (ACF) is performed to analyze the correlation of the data at different time delays and see if there is periodicity. The calculation formula is as follows:
where N is the length of the time series X, k is the specific lag of the time series, acf k X represents t And X is t-k Is a correlation of (3).
And further analyzing whether the load fluctuation of each process has a high correlation with the production pattern, whether the production pattern of each process changes frequently, and other relevant factors. According to the analysis result, the loads can be divided into the following four categories, and the input factors of each category of loads are related to the load characteristics:
(1) more stable load: the production mode is simple, and the load fluctuation of the single production mode is small, such as water treatment, so that input factors comprise load, production mode and the like;
(2) cyclic loading: in a single production mode, load fluctuations are significantly periodically changed, such as oxygen production, so input factors include load, production mode, etc.;
(3) multi-factor load: the load change is not only related to the production mode but also to the rolling specification or yield, such as rolling, so input factors include load, production mode, rolling specification, yield, etc.
(4) Strong fluctuating load: an actual production signal can be received and the load change strongly correlated with the production signal, such as a refining furnace, so input factors include load, production event, time, etc.
In order to fully consider the load difference of different working conditions of each working procedure, the historical data of each working procedure are intelligently segmented according to different production modes, then constructed into a steel enterprise load data set, and the data are divided into a training set and a verification set by adopting a leaving method, wherein the occupation ratio is 7: and 3, directly using real data to test without setting a test set, and detail 'engineering application and online automatic correction' plates. In order to avoid influencing the final result by introducing additional deviation in the data dividing process, the consistency of the data distribution of each process is maintained, and the proportion of sample types is similar.
S14: predictive algorithm construction
According to the load division of the data analysis preprocessing stage, each procedure is overlapped after the prediction is completed according to the corresponding type, and the total load predicted value of the incoming line is obtained, namely
P pre ＝P 1 +P 2 +...+P n
Wherein P is pre Load predictive value of 220kV incoming line is represented, n represents the number of procedures, and P i A load predicted value of a certain process is indicated.
The load characteristics of the four kinds of load prediction methods are fully considered:
(1) more stable load: the load fluctuation is small, the time sequence prediction problem of a single variable is solved, the prediction time overhead and the prediction accuracy are comprehensively considered, and a differential autoregressive moving average model (ARIMA) is used.
The ARIMA model combines three methods, autoregressive (AR), differential (I), and Moving Average (MA), for describing the dynamics of time series data. The general form of the ARIMA model can be expressed as ARIMA (p, d, q), where p represents the autoregressive order, d represents the number of times the time series data is differentiated, and q represents the moving average order.
The AR part represents a linear combination of the current value and the past p values of the time series, which can be expressed as:
X t ＝c+φ 1 X t-1 +φ 2 X t-2 +…+φ p X t-p +ε t
the MA portion represents a linear combination of the current value of the time series and the past q hysteresis errors, which can be expressed as:
X t ＝μ+ε t +θ 1 ε t-1 +θ 2 ε t-2 +…+θ q ε t-q
wherein X is t Representing the current value, epsilon, of the time series t Representing the error, phi, of the current time i And theta i Representing the autoregressive coefficients and the moving average coefficients, c and μ representing constant terms.
The I part can eliminate non-stationarity in the time sequence, which is one of important components of the ARIMA model, and the first-order differential time sequence can be expressed as:
X t ′＝X t -X t-1
after integrating the three parts AR, I, MA, the ARIMA model can be expressed as:
X t ′＝c+φ 1 X t-1 ′+…+φ p X t-p ′+ε t +θ 1 ε t-1 +…+θ q ε t-q
parameters of the ARIMA model are solved using the least squares method.
(2) Cyclic loading: since fluctuations in load are clearly periodic, both aperiodic and periodic factors are considered, using a seasonal differential autoregressive moving average model (SARIMA).
SARIMA is based on ARIMA model taking into account periodicity factors. The method comprises the steps of firstly removing periodicity in the sequence, firstly using ARIMA at periodic intervals to obtain a non-stationary non-periodic time sequence, and then using ARIMA again for analysis on the basis.
(3) Multi-factor load: the predicted input variables are more, the relation between the characteristic information and the predicted value needs to be accurately captured, the time characteristic information is fully mined, and the current moment can be predicted by utilizing short-term and long-term information to obtain a better prediction effect, so that a long-short-term memory network (LSTM) is used for time sequence prediction.
LSTM can learn which information to memorize and which information to forget through the training process. The cell state is divided into two parts, long term state c (t) And short term state h (t) . There are three control gates along the state path: forgetting door f (t) Input gate i (t) And an output gate o (t) 。
Forgetting door f (t) Controlling from a previous long-term state c by a sigmoid activation function (t-1) Is to remove information:
input gate i (t) Controlling information from current output g by sigmoid activation function (t) Added to the current long-term state c (t) Is a kind of medium. Output door o (t) Using the current long-term state c (t) Is used for controlling the current short-term state h (t) Is formed by the steps of (a).
Output g (t) In effect is a standard loop layer:
thus, if all control gates have been removed and the long-term and short-term states are combined, the LSTM cell will transition back to the standard loop layer, with output g (t) Equal to output layer z (t) And state layer h (t) . In LSTM cells, however, the output g (t) Only partial transition to current state c (t) And h (t) . The control equation output by the final unit, the long-term state and the short-term state are as follows:
(4) strong fluctuating load: according to historical load data, production signals and steel grade information, heating characteristic curves of refining furnaces of all steel grades are constructed, and the characteristic curves are corrected every day according to actual conditions. And will be in a high load state over time since the load change of the refining furnace is strongly correlated with the production signal, such as after the ladle is taken. Therefore, prediction can be finished in advance according to production signals or planning data, and then automatic correction is carried out by combining with real-time load so as to reduce prediction deviation.
S15: model training
Training the constructed algorithm model by using a training set of each procedure, adjusting the super parameters of the model and the prediction capability of the preliminary evaluation model by using a verification set, and iterating the prediction model according to the verification result until a good prediction effect can be obtained and no fitting condition exists. And after training and verifying the prediction model of each procedure, obtaining a final algorithm model.
The basic steps of engineering application and on-line automatic correction of the method are shown in fig. 2.
S21: acquiring real-time data
The data used for prediction comprises the following parts:
(1) real-time active power collected by IEC-870-5-104 protocol;
(2) a production shutdown plan of each production line provided by a user;
(3) real-time production signal data is acquired from a basic automation system (L1) and a process control system (L2).
S22: predicting demand
According to the ultra-short-term power demand prediction method for the steel industry, the prediction interval is 3 seconds, the prediction is triggered every three seconds, the load and demand curve of 15 minutes in the future is predicted, and the maximum demand predicted value of 15 minutes in the future is calculated according to the historical load of 15 minutes and the predicted load of 15 minutes in the future.
Because the working procedures of the iron and steel enterprises are more, and most working procedures do not have frequent production mode switching in the same day, if the production working procedures belong to more stable loads, periodic loads or multi-factor loads, the prediction interval is 1 hour, namely, the load and demand curve prediction of 1 hour in the future is completed every hour; if the production plan update is received, re-predicting; if the production process belongs to a strong fluctuation load, the prediction interval is 3 seconds; and then, when predicting every 3 seconds, carrying out weighted average according to the actual load condition and the completed predicted load, and improving the prediction precision.
S23: automatic correction
The data used for correction comprises the following parts:
(1) historical active power for the previous day;
(2) production shutdown actual results of each production line;
(3) real-time production signal data is acquired from a basic automation system (L1) and a process control system (L2).
And (3) taking yesterday historical data as a verification set every day, automatically iterating a prediction model of each production mode under each procedure, and continuously improving the prediction precision.
1. In the more stable load prediction model:
(1) In addition to ARIMA, the prediction method can also use an autoregressive moving average (ARMA) model, wherein the difference is that the ARIMA carries out differential processing on the original data, and if the data is a stable time sequence, the ARMA can be directly used;
(2) The method of parameter estimation may use maximum likelihood estimation and moment estimation in addition to least square method.
2. In the periodic load prediction model:
besides SARIMA, the prediction method can also use an LSTM or a cyclic neural network (RNN), because the LSTM and the RNN are prediction time sequence algorithms based on deep learning, aiming at nonlinear multivariable data with complex structures, particularly unstructured data, the LSTM neural network can obtain good effects, but the periodic load related by the invention belongs to time sequence data with simple structures, so that the SARIMA is directly used for consuming less time, and better effects can be obtained.
3. In the multi-factor load prediction model:
besides LSTM, the prediction method can also use a cyclic neural network (RNN), the difference is that the RNN can only process certain short-term dependence and cannot process long-term dependence, the LSTM can solve the problems of gradient elimination and gradient explosion in the long sequence training process to a certain extent, and the LSTM can better perform in longer sequences, so that the LSTM is directly used in the method.
Finally, it is noted that the above embodiments are only for illustrating the technical solution of the present invention and not for limiting the same, and although the present invention has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that modifications and equivalents may be made thereto without departing from the spirit and scope of the present invention, which is intended to be covered by the claims of the present invention.

Claims (7)

1. A method for predicting ultra-short-term power demand of a steel enterprise based on load characteristics is characterized by comprising the following steps of: the method comprises the following steps:
s1: constructing and training a prediction model;
s2: carrying out engineering application and online automatic correction on the prediction model;
s11: data collection
Confirming the production mode and the load condition of each production process of a steel plant, and collecting the load data, the production shutdown plan, the production shutdown actual performance and the information of actual production signals of each production process of the steel plant in one year and the load data of 220kV incoming line; the collected load data acquisition interval is 1s;
s12: data cleaning, including load data missing processing and working condition missing processing;
s13: data analysis preprocessing
Each production procedure converts the production mode obtained by investigation and the collected production shutdown plan or actual production signal into a planned production mode, and converts the production mode obtained by investigation and the collected production shutdown actual score into an actual score production mode;
analyzing the load by using an inspection ADF (Augmented Dickey-Fuller Testing), and judging the stability of the load sequence; for the load of non-strong fluctuation, performing autocorrelation analysis ACF, analyzing the correlation of data under different time delays, and observing whether the data has periodicity; the calculation formula is as follows:
where N is the length of the time series X, k is the specific lag of the time series, acf k X represents t And X is t-k Is a correlation of (2);
analyzing whether the load fluctuation of each process has high association degree with the production mode or not, and whether the production mode of each process is changed frequently or not; according to the analysis result, the loads are divided into the following four categories, and the input factors of each category of loads are related to the load characteristics:
(1) more stable load: input factors include load and production mode;
(2) cyclic loading: input factors include load and production mode;
(3) multi-factor load: input factors include load, production mode, steel rolling specification and yield;
(4) strong fluctuating load: input factors include load, production event, and time;
in order to fully consider the load difference of different working conditions of each working procedure, the historical data of each working procedure is intelligently segmented according to different production modes, then the historical data is constructed into a steel enterprise load data set, the data is divided into a training set and a verification set by adopting a leaving method, and the ratio is 7:3, testing by using real data without setting a test set; the consistency of the data distribution of each procedure is maintained, and the proportion of sample types is similar;
s14: predictive algorithm construction
According to the load division of the data analysis preprocessing stage, each procedure is overlapped after the prediction is completed according to the corresponding type, and the total load predicted value of the incoming line is obtained, namely
P pre ＝P 1 +P 2 +...+P n
Wherein P is pre Load predictive value of 220kV incoming line is represented, n represents the number of procedures, and P i A load prediction value indicating a certain process;
the load characteristics of the four types of load prediction methods are considered:
(1) more stable load: using differential autoregressive moving average model ARIMA
The ARIMA model combines three methods of autoregressive AR, differential I and moving average MA to describe the dynamic characteristics of time series data; the ARIMA model is denoted ARIMA (p, d, q), where p represents the autoregressive order, d represents the number of times the time series data is differentiated, and q represents the moving average order;
the AR part represents a linear combination of the current value and the past p values of the time series, expressed as:
X t ＝c+φ 1 X t-1 +φ 2 X t-2 +…+φ p X t-p +ε t
the MA portion represents a linear combination of the current value of the time series and the past q hysteresis errors, expressed as:
X t ＝μ+ε t +θ 1 ε t-1 +θ 2 ε t-2 +…+θ q ε t-q
wherein X is t Representing the current value, epsilon, of the time series t Representing the error, phi, of the current time i And theta i Representing autoregressive coefficients and moving average coefficients, c and μ representing constant terms;
part I eliminates non-stationarity in the time series, which is one of the important components of ARIMA model, and the first-order differential time series is expressed as:
X t ′＝X t -X t-1
after integrating the three parts AR, I, MA, the ARIMA model is expressed as:
X t ′＝c+φ 1 X t-1 ′+…+φ p X t-p ′+ε t +θ 1 ε t-1 +…+θ q ε t-q
solving parameters of the ARIMA model by using a least square method;
(2) cyclic loading: because the fluctuation of the load has obvious periodicity, the aperiodic and periodic factors are comprehensively considered, and a seasonal differential autoregressive moving average model SARIMA is used;
SARIMA is based on ARIMA model taking into account periodicity factors; firstly removing periodicity in the sequence, firstly using ARIMA at periodic intervals to obtain a non-stationary non-periodic time sequence, and then using ARIMA again for analysis on the basis;
(3) multi-factor load: the predicted input variables are more, the relation between the characteristic information and the predicted value is required to be accurately captured, the time characteristic information is fully mined, the current moment is predicted by utilizing short-term and long-term information, a better prediction effect can be obtained, and the time sequence prediction is performed by utilizing a long-short-term memory network LSTM;
LSTM learns which information to memorize and which information to forget through a training process; the cell state is divided into two parts, long term state c (t) And short term state h (t) The method comprises the steps of carrying out a first treatment on the surface of the There are three control gates along the state path: forgetting door f (i) Input gate i (t) And an output gate o (t) ；
Forgetting door f (t) Controlling from a previous long-term state c by a sigmoid activation function (t-1) Is to remove information:
input gate i (t) Controlling information from current output g by sigmoid activation function (t) Added to the current long-term state c (t) In (a) and (b); output door o (t) Using the current long-term state c (t) Is used for controlling the current short-term state h (t) Is formed of (a);
output g (t) In effect is a standard loop layer:
if all control gates have been removed and the long-term and short-term states are combined, the LSTM cell will switch back to the standard loop layer, with output g (t) Equal to output layer z (t) And state layer h (t) The method comprises the steps of carrying out a first treatment on the surface of the In the LSTM cell, output g (t) Only partial transition to current state c (t) And h (t) The method comprises the steps of carrying out a first treatment on the surface of the The control equation output by the final unit, the long-term state and the short-term state are as follows:
(4) strong fluctuating load: according to historical load data, production signals and steel grade information, heating characteristic curves of refining furnaces of all steel grades are constructed, and the characteristic curves are corrected every day according to actual conditions; the prediction is finished in advance according to the production signals or the planning data, and then the real-time load is combined for automatic correction so as to reduce the prediction deviation;
s15: model training
Training the constructed algorithm model by using a training set of each procedure, adjusting the super-parameters of the model and the prediction capability of the preliminary evaluation model by using a verification set, and iterating the prediction model according to the verification result until a good prediction effect can be obtained and no fitting condition exists; and after training and verifying the prediction model of each procedure, obtaining a final algorithm model.
2. The load characteristic-based ultra-short-term power demand prediction method for the iron and steel enterprises, as set forth in claim 1, is characterized in that: in the step S12, the load data missing processing is to check the data missing condition for the load data of each production process, if the missing time is longer than 1 minute, the processing is not performed, otherwise, the load average value of 30 seconds before and after the missing time is used for filling; the working condition missing treatment is as follows: if the production shutdown plan, the production shutdown actual results or the actual production signals are absent in a part of the time period in each process, the load data of the time period is directly deleted.
3. The load characteristic-based ultra-short-term power demand prediction method for the iron and steel enterprises, as set forth in claim 1, is characterized in that: the step S2 specifically comprises the following steps:
s21: acquiring real-time data
The data used for prediction comprises the following parts:
(1) real-time active power collected by IEC-870-5-104 protocol;
(2) a production shutdown plan of each production line provided by a user;
(3) acquiring real-time production signal data from the basic automation system L1 and the process control system L2;
s22: predicting demand
The ultra-short-term power demand prediction interval of the steel industry is 3 seconds, prediction is triggered every three seconds, a load and demand curve of 15 minutes in the future is predicted, and a maximum demand predicted value of 15 minutes in the future is calculated according to the historical load of 15 minutes and the predicted load of 15 minutes in the future;
if the production process belongs to a relatively stable load, a periodic load or a multi-factor load, the prediction interval is 1 hour, namely, the load and demand curve prediction of 1 hour in the future is completed every hour; if the production plan update is received, re-predicting; if the production process belongs to a strong fluctuation load, the prediction interval is 3 seconds;
when predicting every 3 seconds, carrying out weighted average according to the actual load condition and the completed predicted load, and improving the prediction precision;
s23: automatic correction
The data used for correction comprises the following parts:
(1) historical active power for the previous day;
(2) production shutdown actual results of each production line;
(3) acquiring real-time production signal data from the basic automation system L1 and the process control system L2;
and yesterday historical data is used as a verification set every day, and a prediction model of each production mode under each procedure is automatically iterated, so that the prediction precision is improved.
4. The method for predicting ultra-short-term power demand of steel enterprises based on load characteristics according to claim 3, wherein the method comprises the following steps: the differential autoregressive moving average model ARIMA is replaced by an autoregressive moving average model ARMA.
5. The method for predicting ultra-short-term power demand of steel enterprises based on load characteristics according to claim 3, wherein the method comprises the following steps: the least square method is replaced by a maximum likelihood estimation method or a moment estimation method.
6. The method for predicting ultra-short-term power demand of steel enterprises based on load characteristics according to claim 3, wherein the method comprises the following steps: the SARIMA is replaced with LSTM or a recurrent neural network RNN.
7. The method for predicting ultra-short-term power demand of steel enterprises based on load characteristics according to claim 3, wherein the method comprises the following steps: the LSTM is replaced with a recurrent neural network RNN.
CN202311479020.7A 2023-11-08 2023-11-08 Ultra-short-term power demand prediction method for iron and steel enterprises based on load characteristics Pending CN117374963A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202311479020.7A CN117374963A (en) 2023-11-08 2023-11-08 Ultra-short-term power demand prediction method for iron and steel enterprises based on load characteristics

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202311479020.7A CN117374963A (en) 2023-11-08 2023-11-08 Ultra-short-term power demand prediction method for iron and steel enterprises based on load characteristics

Publications (1)

Publication Number Publication Date
CN117374963A true CN117374963A (en) 2024-01-09

Family

ID=89396535

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202311479020.7A Pending CN117374963A (en) 2023-11-08 2023-11-08 Ultra-short-term power demand prediction method for iron and steel enterprises based on load characteristics

Country Status (1)

Country Link
CN (1) CN117374963A (en)

Similar Documents

Publication Publication Date Title
CN113962364B (en) Multi-factor power load prediction method based on deep learning
Mamlook et al. A fuzzy inference model for short-term load forecasting
CN110796307B (en) Distributed load prediction method and system for comprehensive energy system
Guo et al. Maximum power demand prediction using fbprophet with adaptive Kalman filtering
Wang et al. Short-term electricity price forecasting based on similarity day screening, two-layer decomposition technique and Bi-LSTM neural network
Sheng et al. Short-term load forecasting based on SARIMAX-LSTM
CN110210670A (en) A kind of prediction technique based on power-system short-term load
CN117595231B (en) Intelligent power grid distribution management system and method thereof
CN109447332A (en) A kind of Middle-long Electric Power Load Forecast method suitable for S type load curve
CN112308270A (en) Long-term electricity load prediction method and device and computer implementation system
CN114140176A (en) Adjustable capacity prediction method and device for load aggregation platform
CN116388150A (en) Matching short-term load prediction method based on digital twin and neural network technology
CN117010946A (en) Thermal power plant production and operation cost accounting system and application method thereof
Ma et al. Scenario-based ultra-short-term rolling optimal operation of a photovoltaic-energy storage system under forecast uncertainty
CN117788082A (en) Power market quotation decision method and system based on electricity price prediction
Zhang et al. A deep learning based real-time load forecasting method in electricity spot market
CN117829374A (en) Prediction method and system for power consumption of iron and steel enterprises
CN118017624A (en) Reactive power optimization method based on multiple time scales of distribution network reconstruction
CN116826745B (en) Layered and partitioned short-term load prediction method and system in power system background
Zhai et al. Prediction of cold rolling gas based on EEMD-LSTM deep learning technology
CN116681173A (en) Data-intensive power load prediction parallel optimization method based on RNN model
CN117374963A (en) Ultra-short-term power demand prediction method for iron and steel enterprises based on load characteristics
Qi et al. Regional electricity sales forecasting research based on big data application service platform
Liu et al. Electricity Consumption Forecasting in Jiangxi Province Based on GRA-PSO-BP
Donghui Short-term Load Forecasting System for Power System Based on Big Data

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