CN113762387A - Data center station multi-load prediction method based on hybrid model prediction - Google Patents

Data center station multi-load prediction method based on hybrid model prediction Download PDF

Info

Publication number
CN113762387A
CN113762387A CN202111048836.5A CN202111048836A CN113762387A CN 113762387 A CN113762387 A CN 113762387A CN 202111048836 A CN202111048836 A CN 202111048836A CN 113762387 A CN113762387 A CN 113762387A
Authority
CN
China
Prior art keywords
load
data
prediction
model
formula
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.)
Granted
Application number
CN202111048836.5A
Other languages
Chinese (zh)
Other versions
CN113762387B (en
Inventor
李华
丁吉
杨东升
张化光
周博文
李广地
金硕巍
罗艳红
王迎春
闫士杰
杨波
陈乐�
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.)
Northeastern University China
Original Assignee
Northeastern University China
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 Northeastern University China filed Critical Northeastern University China
Priority to CN202111048836.5A priority Critical patent/CN113762387B/en
Publication of CN113762387A publication Critical patent/CN113762387A/en
Application granted granted Critical
Publication of CN113762387B publication Critical patent/CN113762387B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • 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
    • G06F18/23—Clustering techniques
    • G06F18/232—Non-hierarchical techniques
    • G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • 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
    • 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
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
    • Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
    • Y04S10/00—Systems supporting electrical power generation, transmission or distribution
    • Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • Business, Economics & Management (AREA)
  • General Physics & Mathematics (AREA)
  • Economics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Human Resources & Organizations (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Strategic Management (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Marketing (AREA)
  • General Business, Economics & Management (AREA)
  • Tourism & Hospitality (AREA)
  • Health & Medical Sciences (AREA)
  • Probability & Statistics with Applications (AREA)
  • Primary Health Care (AREA)
  • Public Health (AREA)
  • General Health & Medical Sciences (AREA)
  • Water Supply & Treatment (AREA)
  • Development Economics (AREA)
  • Game Theory and Decision Science (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention provides a mixed model prediction-based data center station multi-element load prediction method, and relates to the technical field of automatic control. The invention divides the multivariate data of the data center station into three scenes of spring and autumn, summer and winter, performs multivariate load prediction on the data in various scenes, performs characteristic analysis and normalization on the multivariate load data by adopting a GRA method, inputs the processed data into a QPSO-BP neural network for prediction, adopts the QPSO-BP neural network and a BooXSt model for parallel prediction in the aspect of prediction algorithm, simultaneously applies deep learning and machine learning technologies to the load prediction, effectively combines the two integrated learning modes, fully exerts the advantages of the two models and is beneficial to obtaining a more stable model with stronger generalization capability. The hybrid prediction model can actively enrich input data characteristics with single dimension, avoid the influence of data errors caused by artificial factors in the data acquisition process on the calculation precision, and can realize high-precision load prediction under special conditions of large load fluctuation and the like.

Description

Data center station multi-load prediction method based on hybrid model prediction
Technical Field
The invention relates to the technical field of automatic control, in particular to a data center station multi-element load prediction method based on hybrid model prediction.
Background
In recent years, with the rapid development of internet technology, the scale and the number of data center stations are rapidly expanded, the power consumption of data centers in China also accounts for 1% of the total power consumption of the whole country according to statistics, and the load of the data center stations becomes a considerable power load. Under the requirements of fast and accurate scheduling of a power system and safety and stability of the system, implementation of prediction accuracy on a data center station is important.
The loads of the data center station are mainly divided into two types, one is the load of a server for processing data, the other is the load of storage, illumination, cooling and power distribution for maintaining the normal work of the server, and the load of the data center station is influenced by various factors due to the complexity of electricity consumption of the data center station, and the change of the load is not obvious in regularity. The traditional load prediction method usually selects one factor to carry out single mapping analysis on the load, neglects the influence of other factors, and does not consider the linkage relation among all the influencing factors, so that the analysis of the load characteristic is not accurate enough, the load prediction and the formulation of the power utilization plan are influenced, and the accuracy is low. In addition, the traditional prediction models such as a time series model, a neural network model and an artificial intelligence optimization model have respective advantages and disadvantages, the time series model is simple to assume and calculate and strong in adaptability, but the extrapolation effect is poor, and the prediction range is small; the neural network model has good fitting effect and the capability of processing nonlinear data, but the model is unstable and depends on data characteristics; the artificial intelligence optimization model can be combined with other methods for use, so that the prediction precision is improved, but local optimization is easy to fall into. In addition, the traditional prediction algorithm has typical limitations that errors are insensitive to changes of weight values, error gradient changes are small, adjustment time is long, iteration times are multiple, convergence is slow, a neural network output layer is easy to fall into local minimum, certain defects are caused in the aspects of prediction precision and stability, and the problems provide challenges for accurate load prediction of a data center station.
The traditional prediction method does not fully excavate massive deep sleep historical operation data, often predicts the load in a single scene, neglects the difference of time level load, does not consider the influence of various factors existing in the system on the prediction precision of the system, has long adjustment time, more iteration times, slow convergence, extremely easy falling into local minimum of a neural network output layer, has certain defects in the aspects of prediction precision and stability, and can cause inaccurate system prediction due to the various factors.
Disclosure of Invention
Aiming at the problems in the prior art, the invention provides a data center station multivariate load prediction method based on mixed model prediction.
In order to solve the technical problems, the invention adopts the following technical scheme:
a data center station multivariate load prediction method based on mixed model prediction comprises the following steps:
step 1: collecting and preprocessing data; the method comprises the steps of obtaining historical data of a data center station within preset time, constructing a training set, and preprocessing the data, wherein the historical data comprises cold load, heat load, electric load, light intensity, wind speed, humidity, air pressure and date.
Step 1.1: acquiring historical data of the data center station within preset time, and dividing the load of the data center station into three scenes, namely spring and autumn, summer and winter by adopting a clustering algorithm K-means method to perform scene-based prediction;
step 1.2: selecting three meteorological characteristic factors of solar radiation quantity, temperature and air humidity from historical data, wherein the three meteorological characteristic factors are ranked in a training set as the solar radiation quantity, the temperature and the air humidity, and then collecting the cold, heat and electricity loads and the environmental factors of a data center station to form a training set X as follows:
Figure BDA0003251940450000021
wherein X is a training set; x is the number ofeFor data center electrical loads, xe(i) The ith electric load in the electric load sequence; x is the number ofhFor thermal load, xh(i) Is the ith heat load in the heat load sequence; x is the number ofcFor cold load, xc(i) The ith cold load in the cold load sequence; x is the number ofRIs the amount of solar radiation, xR(i) Is the ith solar radiation amount in the radiation amount sequence; x is the number ofTIs temperature, xT(i) Is the ith temperature quantity in the temperature sequence; x is the number ofMIs the air humidity, xM(i) Is the ith humidity in the humidity sequence; m is the number of sequences in a sequence.
Step 1.3: carrying out importance sorting and feature selection on load prediction feature data by adopting Random Forest (RF) out-of-bag estimation;
the significance is calculated as follows:
Figure BDA0003251940450000022
in the formula, Q is the number of the base learners; errOOBqOut-of-bag error for the qth base learner; errOOB'qAfter noise is added to the qth basis learner, performing importance ranking on load prediction characteristic data and performing characteristic selection by adopting Random Forest (RF) out-of-bag estimation;
step 1.4: calculating the correlation between the data load and the characteristic factors;
analyzing the cold and hot loads and the electric loads of the data center station and the correlation between the multivariate loads and the meteorological influence factors according to three scenes of spring, autumn, summer and winter, establishing a matrix formed by the cold and hot electric loads and the environmental influence factors under the three scenes, and calculating the correlation coefficient and the correlation degree according to the strength, the size and the sequence of the relationship between the cold and hot electric loads of the data center station and the environmental influence factors under the three scenes;
in spring and autumn, the sequence of the cooling, heating and power loads and the environmental influence factors forms the following matrix:
Figure BDA0003251940450000031
in the formula X1The matrix is formed by the cold, heat and electricity loads and environmental influence factors in spring and autumn.
In summer, the sequence of cold load, electrical load and environmental impact factors forms the following matrix:
Figure BDA0003251940450000032
in the formula X2The matrix is formed by summer cold and electricity load and environmental influence factors.
In winter, the data sequence of the thermal load, the electrical load and each environmental influence factor forms the following matrix:
Figure BDA0003251940450000033
in the formula X3A matrix formed for winter thermoelectric loads and environmental influences.
The raw data needs to be normalized, and the formula is as follows:
Figure BDA0003251940450000034
wherein x is selected original data; x is the number ofmaxIs the maximum value of the sample data; x is the number ofminIs the minimum value of the sample data; x' is the value after normalization;
correlation coefficient xijAnd degree of correlation gammajThe calculation formula of (a) is as follows:
Figure BDA0003251940450000035
Figure BDA0003251940450000041
xi in the formulajIs the correlation coefficient, ξ, of the data class jj(k) Its k-th degree of association; gamma rayjThe relevance of the data type j; x is the number of0(k) The kth value of the normalized meteorological factor sequence is obtained; x is the number ofj(k) The kth value of the normalized load sequence; rho is a resolution coefficient, and j represents the type of the normalized data;
step 2, constructing a BP neural network model by adopting a quantum particle swarm algorithm QPSO;
step 2.1: a predicted electrical load calculation model of the data center station is constructed by adopting a BP neural network, and the formula is as follows:
Figure BDA0003251940450000042
wherein l is the number of hidden layer neurons in the model; n is the number of neurons in the input layer, m is the number of sequence quantities, and a is a constant between 1 and 10;
step 2.2: optimizing the neural network model by using a quantum particle swarm algorithm QPSO;
step 2.2.1: calculating the average particle history optimal position as shown in the following formula:
Figure BDA0003251940450000043
in the formula MbestThe historical optimal position of the particle is obtained; s is the scale of the particle swarm; qlocal,iIs the position of the ith particle in the particle iteration; step 2.2.2: updating the particle position as shown in the following formula:
Figure BDA0003251940450000044
in the formula QiThe updated position of the ith particle; alpha is alpha1、α2Is [0,1 ]]A random number in between; qglobalIs the global optimal particle position;
step 2.2.3: and constructing a fitness function by taking the reciprocal of the sum of the calculated value of the electric load and the square of the error of the actual value as an individual fitness function, wherein the fitness function is represented by the following formula:
Figure BDA0003251940450000045
in the formula EiFitness of i populations; the (y) (i) is the actual electrical load represented by the ith population of the data center station; s (i) predicted electrical loads represented by the ith population of the data center; and N is the population number.
After introducing the fitness function, the particle position function is updated as follows:
Figure BDA0003251940450000051
Figure BDA0003251940450000052
in the formula xiIs the position of the ith particle; mu is[0,1]A uniform random number above; the chi is continuously updated along with the increase of the iteration times, and the position of the particle is kept to be optimal; n is a radical ofmaxIs the maximum number of QPSO iterations; n is a radical ofminIs the minimum number of QPSO iterations;
step 3, constructing an XGboost prediction model;
step 3.1: establishing a regularization learning objective function;
for the training set X in the step 1, predicting a predicted value by adopting an additive function equation:
Figure BDA0003251940450000053
in the formula, L is a minimum regularization target function of the model;
Figure BDA0003251940450000054
predicted value for ith target
Figure BDA0003251940450000055
With the actual value yiThe difference between, i.e. the loss function; n is the sample volume, K is the sample feature number, Ω (f)k) Calculating the variable f for the kth iterationkCorresponding to the complexity penalty function of the tree.
Step 3.2: optimizing by using a gradient tree enhancement algorithm;
wherein the second order approximation of the objective function is optimized as:
Figure BDA0003251940450000056
Figure BDA0003251940450000057
Figure BDA0003251940450000058
in the formula (I), the compound is shown in the specification,
Figure BDA0003251940450000059
for the ith predictor, g, in the t iterationiFirst order gradient data in the loss function; h isiFor second order gradient data in the loss function, ft(xi) The variables are calculated for the t-th iteration,
Figure BDA00032519404500000510
is the sign of the gradient;
step 3.3: evaluating the impurity fraction of the decision tree as shown in the following formula:
Figure BDA0003251940450000061
Figure BDA0003251940450000062
in the formula (I), the compound is shown in the specification,
Figure BDA0003251940450000063
the optimal weight for leaf j; l is(t)(q) is the optimum value of the formula structure q. I isjIs the real set of leaves j in the gradient tree; gamma and lambda are self-defined parameters of an XGboost algorithm, wherein gamma is a penalty term of the regularization of a first-order gradient function, and lambda is a penalty term of the regularization of a second-order gradient function; t is the total number of leaf nodes in the gradient tree;
step 4, combining the QPSO-BP neural network model and the XGboost prediction model to construct a hybrid prediction model and calculating the weight of the hybrid prediction model;
calculating the weight values of the output results of the two models, setting an initial value of the weight of the fusion model by combining an average absolute percentage error inverse weight MAPE-RW algorithm with an error index, and searching the optimal weight value by combining the initial value to finally form an optimal load prediction model;
the MAPE-RW algorithm is shown below:
Figure BDA0003251940450000064
in the formula, ωaIs the weight of the predictive model a; sigmaMAPE,a、σMAPE,bMAPE values for predictive models a and b, respectively.
The hybrid predictive model weight calculation is as follows:
fs,x＝wQPSO-BP·fXGBoost,s,x+wXGBoost·fQPSO-BP,s,x (22)
in the formula (f)s,xOutputting a predicted value of the x-th class load of the scene s for the hybrid model; w is aQPSO-BP、wXGBoostWeights of the QPSO-BP neural network and the XGboost model are respectively; f. ofQPSO-BP,s,x、fXGBoost,s,xRespectively predicting values of the QPSO-BP neural network and the XGboost model to the xth load of the scene s; the output scene s values are 1, 2 and 3, which respectively represent three scenes of spring and autumn, summer and winter; the x-type loads include cold loads, heat loads and electric loads;
and 5: and (3) substituting the data preprocessed in the step (1) into a hybrid prediction model for calculation, and finishing the multivariate load prediction of the sub-scene data central station.
The invention has the following beneficial effects:
the invention provides a hybrid model parallel prediction-based multi-load prediction method for a sub-scene data central station, which has the following beneficial effects:
(1) the multivariate data of the data center station are divided into three scenes, namely spring and autumn scenes, summer scenes and winter scenes, and the data residing in the scenes are subjected to multivariate load prediction, so that the prediction time is reduced, and the prediction precision is improved;
(2) in the aspect of characteristic factor processing, various characteristic factors are considered, and the strength, the size and the sequence of the relationship among the factors are described by using the grey correlation degree, so that the prediction error is reduced;
(3) and performing characteristic analysis and normalization on the multivariate load data by adopting a GRA (generalized GRA analysis) method, and inputting the processed data into a QPSO-BP (quick-response-Back propagation) neural network for prediction, so that the learning time of the QPSO-BP neural network on the data can be obviously reduced, and more efficient data mining is realized.
(3) A QPSO-BP neural network is adopted to replace a traditional BP neural network, and the neural network line loss rate calculation model optimized by the genetic algorithm has better nonlinear fitting capability and higher calculation accuracy than a single BP neural network model.
(4) In the aspect of a prediction algorithm, a QPSO-BP neural network and an XGboost model are adopted for parallel prediction, deep learning and machine learning technologies are simultaneously applied to load prediction, two integrated learning modes are effectively combined, the advantages of the two models are fully played, and the model with higher stability and generalization capability is obtained. The hybrid prediction model can actively enrich input data characteristics with single dimension, so that the network learning is more efficient, the influence of data errors caused by artificial factors in the data acquisition process on the calculation precision can be avoided, and high-precision load prediction can be realized under special conditions of large load fluctuation and the like;
(5) the weight of the fusion model is set by applying an MAPE-RW algorithm, the search of the optimal weight is completed, and the error of the fusion model is reduced;
drawings
FIG. 1 is a flowchart illustrating an overall process of multi-load prediction for a data center station according to an embodiment of the present invention;
FIG. 2 is a diagram illustrating the results of a spring-autumn correlation analysis of data center data in an embodiment of the present invention;
FIG. 3 is a diagram illustrating a summer correlation analysis result of data in a data center according to an embodiment of the present invention;
FIG. 4 is a diagram illustrating a result of a winter relevance analysis of data in a data center according to an embodiment of the invention;
FIG. 5 is a diagram illustrating a BP neural network prediction model according to an embodiment of the present invention;
FIG. 6 is a flowchart illustrating a QPSO-BP neural network algorithm according to an embodiment of the present invention;
FIG. 7 is a flow chart of an XGboost neural network algorithm in an embodiment of the present invention;
FIG. 8 is a comparison of the prediction results of the hybrid prediction model of the electric load in spring and autumn and other methods according to the embodiment of the invention;
FIG. 9 is a comparison of the summer cooling load hybrid prediction model with the prediction results of other methods in accordance with an embodiment of the present invention;
FIG. 10 is a comparison of the thermal load hybrid prediction model in winter with the prediction results of other methods according to an embodiment of the present invention.
Detailed Description
The following detailed description of embodiments of the present invention is provided in connection with the accompanying drawings and examples. The following examples are intended to illustrate the invention but are not intended to limit the scope of the invention.
A data center station multivariate load prediction method based on hybrid model prediction is disclosed, as shown in FIG. 1, and comprises the following steps:
step 1: collecting and preprocessing data; the method comprises the steps of obtaining historical data of a data center station within preset time, constructing a training set, and preprocessing the data, wherein the historical data comprises cold load, heat load, electric load, light intensity, wind speed, humidity, air pressure and date.
Step 1.1: historical data of the data center station within a preset time are obtained, and the load of the data center station is divided into three scenes, namely spring and autumn, summer and winter, by adopting a clustering algorithm K-means method to perform scene-by-scene prediction, so that the prediction accuracy is further improved.
Step 1.2: in order to predict the cooling, heating and power loads of the data center station, three meteorological characteristic factors which strongly influence load prediction, namely solar radiation amount, temperature and air humidity, are selected from historical data, the three meteorological characteristic factors are ranked in a training set as the solar radiation amount, the temperature and the air humidity, and then the cooling, heating and power loads and environmental factors of the data center station are collected to form a training set X as follows:
Figure BDA0003251940450000081
wherein X is a training set; x is the number ofeFor data center electrical loads, xe(i) The ith electric load in the electric load sequence; x is the number ofhFor thermal load, xh(i) Is the ith heat load in the heat load sequence; x is the number ofcFor cold load, xc(i) The ith cold load in the cold load sequence; x is the number ofRAs sun radiationShot size, xR(i) Is the ith solar radiation amount in the radiation amount sequence; x is the number ofTIs temperature, xT(i) Is the ith temperature quantity in the temperature sequence; x is the number ofMIs the air humidity, xM(i) Is the ith humidity in the humidity sequence; m is the number of sequences in a sequence.
Step 1.3: carrying out importance sorting and feature selection on load prediction feature data by adopting Random Forest (RF) out-of-bag estimation;
the significance is calculated as follows:
Figure BDA0003251940450000082
in the formula, Q is the number of the base learners; errOOBqOut-of-bag error for the qth base learner; errOOB'qThe out-of-bag error after adding noise for the qth basis learner. In the load prediction of the data center station, historical data used for prediction may include characteristic data related to prediction but redundant, importance ranking and characteristic selection are performed on the load prediction characteristic data by adopting Random Forest (RF) out-of-bag estimation, wherein the more MDA indexes are reduced, the larger the influence of the corresponding characteristics on the prediction result is, and the higher the importance of the corresponding characteristics is.
Step 1.4: calculating the correlation between the data load and the characteristic factors;
analyzing the cold and hot loads and the electric loads of the data center station and the correlation between the multivariate loads and the meteorological influence factors according to three scenes of spring, autumn, summer and winter, establishing a matrix formed by the cold and hot electric loads and the environmental influence factors under the three scenes, and calculating the correlation coefficient and the correlation degree according to the strength, the size and the sequence of the relationship between the cold and hot electric loads of the data center station and the environmental influence factors under the three scenes;
the correlation analysis can further improve the prediction accuracy. The method comprises the following steps that cold and heat supply exist simultaneously in spring and autumn, strong coupling exists between an electric load and a cold and heat load, and correlation among the electric load, the cold and heat load and environmental influence factors of a data center station is analyzed; the method mainly comprises the steps that summer is mainly cooling seasons, the electric load and the cooling load have strong coupling, and the correlation among the electric load, the cooling load and environmental influence factors of a data center station is analyzed; the method is mainly used in heating seasons in winter, the electric load and the heat load have strong coupling, and the correlation among the electric load, the heat load and environmental influence factors of the data center station is analyzed.
In spring and autumn, the sequence of the cooling, heating and power loads and the environmental influence factors forms the following matrix:
Figure BDA0003251940450000091
in the formula X1The matrix is formed by the cold, heat and electricity loads and environmental influence factors in spring and autumn.
In summer, the sequence of cold load, electrical load and environmental impact factors forms the following matrix:
Figure BDA0003251940450000092
in the formula X2The matrix is formed by summer cold and electricity load and environmental influence factors.
In winter, the data sequence of the thermal load, the electrical load and each environmental influence factor forms the following matrix:
Figure BDA0003251940450000093
in the formula X3A matrix formed for winter thermoelectric loads and environmental influences.
Because the cold, heat and power loads and environmental factors of each year have certain periodicity, under the condition that the sample data is less, the trained network model is more accurate by recycling the sample data, and the data is recycled for 2 times. Because the load value of the data center station is large, in order to remove dimension influence and accelerate the network convergence speed, the original data needs to be normalized, and the formula is as follows:
Figure BDA0003251940450000101
wherein x is selected original data; x is the number ofmaxIs the maximum value of the sample data; x is the number ofminIs the minimum value of the sample data; x' is the value after normalization;
the strength, the size and the sequence of the relationship between the cold, heat and electricity loads of the data center station and the environmental influence factors under the three scenes of spring, autumn, summer and winter are judged by adopting a GRA method, the strength of the correlation is judged according to the similarity degree of the curve geometry, namely the correlation degree, and the correlation coefficient xi isjAnd degree of correlation gammajThe calculation formula of (a) is as follows:
Figure BDA0003251940450000102
Figure BDA0003251940450000103
xi in the formulajIs the correlation coefficient, ξ, of the data class jj(k) Its k-th degree of association; gamma rayjThe relevance of the data type j; x is the number of0(k) The kth value of the normalized meteorological factor sequence is obtained; x is the number ofj(k) The kth value of the normalized load sequence; rho is a resolution coefficient, and is usually 0.5; j represents the kind of normalized data;
step 2, constructing a BP neural network model by adopting a quantum particle swarm algorithm QPSO;
a QPSO-BP neural network model is adopted to replace a traditional BP neural network, and the neural network calculation model optimized by the QPSO algorithm has better nonlinear fitting capability and higher calculation accuracy than a single BP neural network model.
Step 2.1: adopting a BP neural network to construct a predicted electrical load calculation model of the data center station;
the determination of the neural network structure mainly determines the network type, the number of layers of the BP neural network, the number of neurons in each layer, the form of an excitation function and the like. The method adopts a BP neural network to construct a predicted electrical load calculation model of the data center station, and selects an S-shaped function as a function between an intermediate layer (hidden layer) and an output layer; the number of input layers is determined by the dimension of the sample data; the output layer is the electrical load of the data center station; the determination of the number of the intermediate layers is not a unified theoretical method at present, and is determined by an empirical formula, fig. 5 is a schematic diagram of a BP neural network prediction model of the patent, and the formula is as follows:
Figure BDA0003251940450000104
wherein l is the number of hidden layer neurons in the model; n is the number of neurons in the input layer, and a is a constant between 1 and 10;
the BP neural network in this patent operates as follows:
(a) opening a Neural network Fitting (Neural Net Fitting) module in a status bar application program in Matlab software;
(b) under the condition of selecting a data page, selecting a standard sample matrix file to be imported, and selecting a target output matrix file to be imported by Targets;
(c) under the verification and Test Data (validity and Test Data) page, 70% of Training Data (Training), 15% of verification Data (validity) and 15% of Test Data (Testing) are selected;
(d) the number of hidden layers is selected to be 4 under the Network Architecture (Network Architecture) page.
Step 2.2: optimizing the neural network model by using a quantum particle swarm algorithm QPSO;
the BP neural network has low learning convergence speed, is easy to fall in local minimum points, has weak nonlinear fitting capability and low calculation accuracy, and a quantum particle swarm algorithm (QPSO) is a global optimization algorithm and optimizes the weight and the threshold of the BP neural network by utilizing the QPSO to obtain an optimal individual. And predicting and calculating the electrical load of the data center station through the optimal weight and the threshold, and avoiding the BP neural network from falling into local optimization, thereby obtaining a more accurate load predicted value.
The position updating operation of the quantum particle swarm algorithm in the patent is as follows:
step 2.2.1: calculating the average particle history optimal position as shown in the following formula:
Figure BDA0003251940450000111
in the formula MbestThe historical optimal position of the particle is obtained; s is the scale of the particle swarm; qlocal,iIs the position of the ith particle in the particle iteration;
step 2.2.2: updating the particle position as shown in the following formula:
according to the particle position updating formula, a traditional random variable is increased into two random variables, so that the convergence speed of the algorithm is increased, the risk is reduced, and meanwhile, the randomness of the algorithm is increased.
Figure BDA0003251940450000112
In the formula QiThe updated position of the ith particle; alpha is alpha1、α2Is [0,1 ]]A random number in between; qglobalIs the global optimal particle position;
step 2.2.3: and constructing a fitness function by taking the reciprocal of the sum of the calculated value of the electric load and the square of the error of the actual value as an individual fitness function, wherein the fitness function is represented by the following formula:
the Fitness (Fitness) represents the degree of goodness and badness of population individuals in the genetic algorithm, a Fitness function can clearly reflect the iterative evolution effect of each particle, QPSO is carried out towards the direction of increasing the Fitness, and the expression is as follows:
Figure BDA0003251940450000113
in the formula EiFitness of i populations; the (y) (i) is the actual electrical load represented by the ith population of the data center station; s (i) is in the dataPredicted electrical loads represented by the ith population of the central station; and N is the population number.
After introducing the fitness function, the particle position function is updated as follows:
Figure BDA0003251940450000121
Figure BDA0003251940450000122
in the formula xiIs the position of the ith particle; mu is [0,1 ]]A uniform random number above; the chi is continuously updated along with the increase of the iteration times, and the position of the particle is kept to be optimal; n is a radical ofmaxIs the maximum number of QPSO iterations; n is a radical ofminIs the minimum number of QPSO iterations;
step 3, constructing an XGboost prediction model as shown in FIG. 7;
the XGboost is essentially a Boosting serial ensemble learning algorithm based on a tree, a base learner has the characteristics of weak prediction models and strong correlation, and the integration mode of the XGboost is formed by continuous serial superposition of the base learner.
Step 3.1: establishing a regularization learning objective function;
for the training set X in the step 1, predicting a predicted value by adopting an additive function equation:
Figure BDA0003251940450000123
in the formula, L is a minimum regularization target function of the model;
Figure BDA0003251940450000124
predicted value for ith target
Figure BDA0003251940450000125
With the actual value yiThe difference between, i.e. the loss function; n is the sample volume, K is the sample feature number, in this patent, the sample feature is a numberAccording to the environmental factors of the central station, the sample characteristic number is 3; omega (f)k) Calculating the variable f for the kth iterationkCorresponding to the complexity penalty function of the tree.
Step 3.2: optimizing by using a gradient tree enhancement algorithm;
wherein the second order approximation of the objective function is optimized as:
Figure BDA0003251940450000126
Figure BDA0003251940450000127
Figure BDA0003251940450000128
in the formula (I), the compound is shown in the specification,
Figure BDA0003251940450000131
for the ith predictor, g, in the t iterationiFirst order gradient data in the loss function; h isiFor second order gradient data in the loss function, ft(xi) The variables are calculated for the t-th iteration,
Figure BDA0003251940450000132
is the sign of the gradient;
step 3.3: evaluating the impurity fraction of the decision tree as shown in the following formula:
Figure BDA0003251940450000133
Figure BDA0003251940450000134
in the formula (I), the compound is shown in the specification,
Figure BDA0003251940450000135
the optimal weight for leaf j; l is(t)(q) is the optimum value of the formula structure q. I isjIs the real set of leaves j in the gradient tree; gamma and lambda are self-defined parameters of an XGboost algorithm, wherein gamma is a penalty term of the regularization of a first-order gradient function, and lambda is a penalty term of the regularization of a second-order gradient function; t is the total number of leaf nodes in the gradient tree;
step 4, combining the QPSO-BP neural network model and the XGboost prediction model to construct a hybrid prediction model and calculating the weight of the hybrid prediction model;
because the QPSO-BP neural network model and the XGboost prediction model have different learning mechanisms and different emphasis points for errors, the prediction results obtained by the two models have certain deviation, and the weight of the output results of the two models needs to be calculated to ensure that the prediction result precision of the hybrid prediction model is higher.
Calculating the weight values of the output results of the two models, setting an initial value of the weight of the fusion model by combining an average absolute percentage error inverse weight MAPE-RW algorithm with an error index, and searching the optimal weight value by combining the initial value to finally form an optimal load prediction model;
the MAPE-RW algorithm is shown below:
Figure BDA0003251940450000136
in the formula, ωaIs the weight of the predictive model a; sigmaMAPE,a、σMAPE,bMAPE values for predictive models a and b, respectively.
The hybrid predictive model weight calculation is as follows:
fs,x＝wQPSO-BP·fXGBoost,s,x+wXGBoost·fQPSO-BP,s,x (22)
in the formula (f)s,xOutputting a predicted value of the x-th class load of the scene s for the hybrid model; w is aQPSO-BP、wXGBoostWeights of the QPSO-BP neural network and the XGboost model are respectively; f. ofQPSO-BP,s,x、fXGBoost,s,xAre respectively asPredicting the x-th load of the scene s by a QPSO-BP neural network and an XGboost model; the output scene s values are 1, 2 and 3, which respectively represent three scenes of spring and autumn, summer and winter; the x-type loads include cold loads, heat loads and electric loads;
and 5: and (3) substituting the data preprocessed in the step (1) into a hybrid prediction model for calculation, and finishing the multivariate load prediction of the sub-scene data central station.
The QPSO-BP integral algorithm is divided into two parts, one part is a QPSO part, the other part is a BP neural network part, and the prediction result output by the hybrid prediction model has smaller error and higher fitting degree than the single QPSO-BP neural network electric load prediction result.
Step 5.1: the calculation of the QPSO-BP neural network model comprises a QPSO algorithm and a BP neural network algorithm as shown in FIG. 5;
the QPSO algorithm comprises the following steps:
step a 1: inputting data and preprocessing the data;
step a 2: randomly creating an initial population, and randomly giving initial values to the positions and the speeds of all particles;
step a 3: calculating the fitness of each particle according to the fitness function, comparing the fitness of each particle, and recording the position of the particle with the optimal fitness and the corresponding fitness value;
step a 4: updating the position of the particle, comparing the current position of the particle with the previous optimal position, and updating the current position into the optimal position if the current position is better than the previous optimal position;
step a 5: and stopping the operation when the iteration number reaches the upper limit, outputting the optimization result of the QPSO, and returning to the operation step a3 to continue the operation if the optimization result does not reach the condition of stopping the operation.
The BP neural network algorithm comprises the following steps:
step b 1: determining a BP neural network topological structure;
step b 2: initial BP neural network weight and threshold length;
step b 3: optimizing optimal weight and threshold value by QPSO algorithm;
step b 4: calculating an error;
step b 5: updating the weight threshold value;
step b 6: and (5) if the output condition is met, outputting the result, otherwise, returning to the operation step (4) to continue the operation.
Step 5.2: the XGboost model algorithm is shown in FIG. 6 and comprises the following steps:
step c 1: inputting a training data set;
step c 2: setting a target loss function;
step c 3: determining a regression tree structure to calculate an independent tree structure q and a leaf weight w;
step c 4: starting XGboost iterative computation, if the iteration times are larger than the set times, outputting a predicted value, otherwise, returning to the step c3 to continue the iteration.
As can be seen from fig. 2, in spring and autumn, the correlation coefficient of the power-saving load and the thermal load is 0.43, the correlation coefficient of the electrical load and the cold load is 0.47, and the coupling relationship between the cold load and the thermal load is not deep; from FIG. 3, it can be seen that the correlation coefficient of the electrical load and the cooling load is as high as 0.87 in summer, and the two are in a strong coupling relationship, which shows that a large part of electricity is used for refrigeration when a data center stands in summer; from fig. 4, it can be seen that the correlation coefficient of the electrical load and the thermal load is as high as 0.67 in winter, which reflects that the western station heating power consumption accounts for a large proportion in the data; and the influence of weather factors on the cooling, heating and power loads can be directly seen from fig. 2-4.
As shown in fig. 8, 9, and 10, the hybrid prediction model proposed in this patent is demonstrated by simulation analysis to be superior, and it can be seen that compared with the BP neural network and the QPSO-BP model, the hybrid prediction model has a prediction output curve more fitting the actual load curve, wherein the QPSO-BP model has higher prediction accuracy than the BP neural network, the prediction accuracy of the hybrid prediction model obtained by performing simulation calculation on the three types of loads, i.e., cooling, heating and power, in fig. 8-10 is above 99.7%, the prediction accuracy of the QPSO-BP model is about 98.12%, and the prediction accuracy of the BP neural network is about 96.52%. The data center station multivariate load prediction method based on the hybrid model prediction provided by the patent has superiority and feasibility.
Finally, it should be noted that: the above examples are only intended to illustrate the technical solution of the present invention, but not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some or all of the technical features may be equivalently replaced; such modifications and substitutions do not depart from the spirit of the corresponding technical solutions and scope of the present invention as defined in the appended claims.

Claims (7)

1. A data center station multivariate load prediction method based on mixed model prediction is characterized by comprising the following steps:
step 1: collecting and preprocessing data; acquiring historical data of the data center station within preset time, constructing a training set, and preprocessing the data, wherein the historical data comprises cold load, heat load, electric load, light intensity, wind speed, humidity, air pressure and date;
step 2, constructing a BP neural network model by adopting a quantum particle swarm algorithm QPSO;
step 3, constructing an XGboost prediction model;
step 4, combining the QPSO-BP neural network model and the XGboost prediction model to construct a hybrid prediction model and calculating the weight of the hybrid prediction model; calculating the weight values of the output results of the two models, setting an initial value of the weight of the fusion model by combining an average absolute percentage error inverse weight MAPE-RW algorithm with an error index, and searching the optimal weight value by combining the initial value to finally form an optimal load prediction model;
and 5: and (3) substituting the data preprocessed in the step (1) into a hybrid prediction model for calculation, and finishing the multivariate load prediction of the sub-scene data central station.
2. The hybrid model prediction-based data center station multivariate load prediction method according to claim 1, wherein the step 1 comprises the following steps:
step 1.1: acquiring historical data of the data center station within preset time, and dividing the load of the data center station into three scenes, namely spring and autumn, summer and winter by adopting a clustering algorithm K-means method to perform scene-based prediction;
step 1.2: selecting three meteorological characteristic factors of solar radiation quantity, temperature and air humidity from historical data, wherein the three meteorological characteristic factors are ranked in a training set as the solar radiation quantity, the temperature and the air humidity, and then collecting the cold, heat and electricity loads and the environmental factors of a data center station to form a training set X as follows:
Figure FDA0003251940440000011
wherein X is a training set; x is the number ofeFor data center electrical loads, xe(i) The ith electric load in the electric load sequence; x is the number ofhFor thermal load, xh(i) Is the ith heat load in the heat load sequence; x is the number ofcFor cold load, xc(i) The ith cold load in the cold load sequence; x is the number ofRIs the amount of solar radiation, xR(i) Is the ith solar radiation amount in the radiation amount sequence; x is the number ofTIs temperature, xT(i) Is the ith temperature quantity in the temperature sequence; x is the number ofMIs the air humidity, xM(i) Is the ith humidity in the humidity sequence; m is the number of sequences in a sequence;
step 1.3: carrying out importance sorting and feature selection on load prediction feature data by adopting Random Forest (RF) out-of-bag estimation;
the significance is calculated as follows:
Figure FDA0003251940440000021
in the formula, Q is the number of the base learners; errOOBqOut-of-bag error for the qth base learner; errOOB'qAfter noise is added to the qth basis learner, performing importance ranking on load prediction characteristic data and performing characteristic selection by adopting Random Forest (RF) out-of-bag estimation;
step 1.4: calculating the correlation between the data load and the characteristic factors;
analyzing the cold and hot loads and the electric loads of the data center station and the correlation between the multivariate loads and the meteorological influence factors according to three scenes of spring, autumn, summer and winter, establishing a matrix formed by the cold and hot electric loads and the environmental influence factors under the three scenes, and calculating the correlation coefficient and the correlation degree according to the strength, the size and the sequence of the relationship between the cold and hot electric loads of the data center station and the environmental influence factors under the three scenes;
in spring and autumn, the sequence of the cooling, heating and power loads and the environmental influence factors forms the following matrix:
Figure FDA0003251940440000022
in the formula X1A matrix formed by the cold, heat and electricity loads and environmental influence factors in spring and autumn;
in summer, the sequence of cold load, electrical load and environmental impact factors forms the following matrix:
Figure FDA0003251940440000023
in the formula X2A matrix formed by summer cold and electricity load and environmental influence factors;
in winter, the data sequence of the thermal load, the electrical load and each environmental influence factor forms the following matrix:
Figure FDA0003251940440000024
in the formula X3A matrix formed for winter thermoelectric loads and environmental impact factors;
the raw data is normalized, and the formula is as follows:
Figure FDA0003251940440000031
wherein x is selected original data; x is the number ofmaxIs the maximum value of the sample data; x is the number ofminIs the minimum value of the sample data; x' is the value after normalization;
correlation coefficient xijAnd degree of correlation gammajThe calculation formula of (a) is as follows:
Figure FDA0003251940440000032
Figure FDA0003251940440000033
xi in the formulajIs the correlation coefficient, ξ, of the data class jj(k) Its k-th degree of association; gamma rayjThe relevance of the data type j; x is the number of0(k) The kth value of the normalized meteorological factor sequence is obtained; x is the number ofj(k) The kth value of the normalized load sequence; ρ is the resolution factor, and j represents the normalized data class.
3. The hybrid model prediction-based data center station multivariate load prediction method according to claim 1, wherein the step 2 comprises the following steps:
step 2.1: a predicted electrical load calculation model of the data center station is constructed by adopting a BP neural network, and the formula is as follows:
Figure FDA0003251940440000034
wherein l is the number of hidden layer neurons in the model; n is the number of neurons in the input layer, m is the number of sequence quantities, and a is a constant between 1 and 10;
step 2.2: and optimizing the neural network model by using a quantum particle swarm algorithm QPSO.
4. The hybrid model prediction-based data center station multivariate load prediction method according to claim 3, wherein the step 2.2 comprises the following steps:
step 2.2.1: calculating the average particle history optimal position as shown in the following formula:
Figure FDA0003251940440000035
in the formula MbestThe historical optimal position of the particle is obtained; s is the scale of the particle swarm; qlocal,iIs the position of the ith particle in the particle iteration;
step 2.2.2: updating the particle position as shown in the following formula:
Figure FDA0003251940440000041
in the formula QiThe updated position of the ith particle; alpha is alpha1、α2Is [0,1 ]]A random number in between; qglobalIs the global optimal particle position;
step 2.2.3: and constructing a fitness function by taking the reciprocal of the sum of the calculated value of the electric load and the square of the error of the actual value as an individual fitness function, wherein the fitness function is represented by the following formula:
Figure FDA0003251940440000042
in the formula EiFitness of i populations; the (y) (i) is the actual electrical load represented by the ith population of the data center station; s (i) predicted electrical loads represented by the ith population of the data center; n is the number of the population;
after introducing the fitness function, the particle position function is updated as follows:
Figure FDA0003251940440000043
Figure FDA0003251940440000044
in the formula xiIs the position of the ith particle; mu is [0,1 ]]A uniform random number above; the chi is continuously updated along with the increase of the iteration times, and the position of the particle is kept to be optimal; n is a radical ofmaxIs the maximum number of QPSO iterations; n is a radical ofminIs the minimum number of QPSO iterations.
5. The hybrid model prediction-based data center station multivariate load prediction method according to claim 1, wherein the step 3 comprises the following steps:
step 3.1: establishing a regularization learning objective function;
for the training set X in the step 1, predicting a predicted value by adopting an additive function equation:
Figure FDA0003251940440000045
in the formula, L is a minimum regularization target function of the model;
Figure FDA0003251940440000046
predicted value for ith target
Figure FDA0003251940440000047
With the actual value yiThe difference between, i.e. the loss function; n is the sample volume, K is the sample feature number, Ω (f)k) Calculating the variable f for the kth iterationkA complexity penalty function corresponding to the tree;
step 3.2: optimizing by using a gradient tree enhancement algorithm;
wherein the second order approximation of the objective function is optimized as:
Figure FDA0003251940440000051
Figure FDA0003251940440000052
Figure FDA0003251940440000053
in the formula (I), the compound is shown in the specification,
Figure FDA0003251940440000054
for the ith predictor, g, in the t iterationiFirst order gradient data in the loss function; h isiFor second order gradient data in the loss function, ft(xi) The variables are calculated for the t-th iteration,
Figure FDA0003251940440000055
is the sign of the gradient;
step 3.3: evaluating the impurity fraction of the decision tree as shown in the following formula:
Figure FDA0003251940440000056
Figure FDA0003251940440000057
in the formula (I), the compound is shown in the specification,
Figure FDA0003251940440000058
the optimal weight for leaf j; l is(t)(q) is the optimum value of the formula structure q, IjIs the real set of leaves j in the gradient tree; gamma and lambda are self-defined parameters of an XGboost algorithm, wherein gamma is a penalty term of the regularization of a first-order gradient function, and lambda is a penalty term of the regularization of a second-order gradient function; t is the total number of leaf nodes in the gradient tree.
6. The hybrid model prediction-based data center station multivariate load prediction method according to claim 1, wherein the MAPE-RW algorithm in the step 4 is shown as follows:
Figure FDA0003251940440000059
in the formula, ωaIs the weight of the predictive model a; sigmaMAPE,a、σMAPE,bMAPE values for predictive models a and b, respectively;
the hybrid predictive model weight calculation is as follows:
fs,x＝wQPSO-BP·fXGBoost,s,x+wXGBoost·fQPSO-BP,s,x (22)
in the formula (f)s,xOutputting a predicted value of the x-th class load of the scene s for the hybrid model; w is aQPSO-BP、wXGBoostWeights of the QPSO-BP neural network and the XGboost model are respectively; f. ofQPSO-BP,s,x、fXGBoost,s,xAnd respectively predicting values of the QPSO-BP neural network and the XGboost model to the xth load of the output scene s.
7. The hybrid model prediction-based data center station multivariate load prediction method according to claim 6, wherein the output scene s has values of 1, 2 and 3, which respectively represent three scenes of spring and autumn, summer and winter; the x-type loads include cold loads, heat loads, and electrical loads.
CN202111048836.5A 2021-09-08 2021-09-08 Multi-element load prediction method for data center station based on hybrid model prediction Active CN113762387B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111048836.5A CN113762387B (en) 2021-09-08 2021-09-08 Multi-element load prediction method for data center station based on hybrid model prediction

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111048836.5A CN113762387B (en) 2021-09-08 2021-09-08 Multi-element load prediction method for data center station based on hybrid model prediction

Publications (2)

Publication Number Publication Date
CN113762387A true CN113762387A (en) 2021-12-07
CN113762387B CN113762387B (en) 2024-02-02

Family

ID=78793770

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111048836.5A Active CN113762387B (en) 2021-09-08 2021-09-08 Multi-element load prediction method for data center station based on hybrid model prediction

Country Status (1)

Country Link
CN (1) CN113762387B (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115660226A (en) * 2022-12-13 2023-01-31 国网冀北电力有限公司 Power load prediction model construction method and construction device based on digital twins
CN116050666A (en) * 2023-03-20 2023-05-02 中国电建集团江西省电力建设有限公司 Photovoltaic power generation power prediction method for irradiation characteristic clustering
CN116227741A (en) * 2023-05-05 2023-06-06 深圳市万物云科技有限公司 Water chilling unit energy saving method and device based on self-adaptive algorithm and related medium

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2015172560A1 (en) * 2014-05-16 2015-11-19 华南理工大学 Central air conditioner cooling load prediction method based on bp neural network
CN110728401A (en) * 2019-10-10 2020-01-24 郑州轻工业学院 Short-term power load prediction method of neural network based on squirrel and weed hybrid algorithm
CN111260116A (en) * 2020-01-10 2020-06-09 河南理工大学 Time-interval refined short-term load prediction method based on BOA-SVR and fuzzy clustering
WO2020135510A1 (en) * 2018-12-29 2020-07-02 中兴通讯股份有限公司 Burst load prediction method and device, storage medium and electronic device

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2015172560A1 (en) * 2014-05-16 2015-11-19 华南理工大学 Central air conditioner cooling load prediction method based on bp neural network
WO2020135510A1 (en) * 2018-12-29 2020-07-02 中兴通讯股份有限公司 Burst load prediction method and device, storage medium and electronic device
CN110728401A (en) * 2019-10-10 2020-01-24 郑州轻工业学院 Short-term power load prediction method of neural network based on squirrel and weed hybrid algorithm
CN111260116A (en) * 2020-01-10 2020-06-09 河南理工大学 Time-interval refined short-term load prediction method based on BOA-SVR and fuzzy clustering

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
CHEN LI等: "Power Load Forecasting Based on the Combined Model of LSTM and XGBoost", 《PRAI\'19 PROCEEDINGS OF THE 2019 THE INTERNATIONAL CONFERENCE ON PATTERN RECOGNITION AND ARTIFICIAL INTELLIGENCE》, pages 46 - 51 *
尹新等: "基于一种混合算法模型的短期电力负荷预测", 《计算机仿真》, vol. 27, no. 10, pages 255 - 258 *

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115660226A (en) * 2022-12-13 2023-01-31 国网冀北电力有限公司 Power load prediction model construction method and construction device based on digital twins
CN116050666A (en) * 2023-03-20 2023-05-02 中国电建集团江西省电力建设有限公司 Photovoltaic power generation power prediction method for irradiation characteristic clustering
CN116227741A (en) * 2023-05-05 2023-06-06 深圳市万物云科技有限公司 Water chilling unit energy saving method and device based on self-adaptive algorithm and related medium

Also Published As

Publication number Publication date
CN113762387B (en) 2024-02-02

Similar Documents

Publication Publication Date Title
CN110705743B (en) New energy consumption electric quantity prediction method based on long-term and short-term memory neural network
CN113762387A (en) Data center station multi-load prediction method based on hybrid model prediction
CN112116144B (en) Regional power distribution network short-term load prediction method
CN108921339B (en) Quantile regression-based photovoltaic power interval prediction method for genetic support vector machine
CN105701572B (en) Photovoltaic short-term output prediction method based on improved Gaussian process regression
CN110598923A (en) Air conditioner load prediction method based on support vector regression optimization and error correction
CN115130741A (en) Multi-model fusion based multi-factor power demand medium and short term prediction method
CN115374995A (en) Distributed photovoltaic and small wind power station power prediction method
CN115481788B (en) Phase change energy storage system load prediction method and system
CN110163444A (en) A kind of water demand prediction method based on GASA-SVR
CN116526473A (en) Particle swarm optimization LSTM-based electrothermal load prediction method
CN115470862A (en) Dynamic self-adaptive load prediction model combination method
CN115759389A (en) Day-ahead photovoltaic power prediction method based on weather type similar day combination strategy
CN113177675A (en) Air conditioner cold load prediction method based on optimization neural network of longicorn group algorithm
CN116205508A (en) Distributed photovoltaic power generation abnormality diagnosis method and system
CN113762591B (en) Short-term electric quantity prediction method and system based on GRU and multi-core SVM countermeasure learning
CN116826710A (en) Peak clipping strategy recommendation method and device based on load prediction and storage medium
CN117081063A (en) Distributed charging load prediction method and system based on GCN-Crossformer model
CN113255223B (en) Short-term prediction method and system for air conditioner load
CN114611757A (en) Electric power system short-term load prediction method based on genetic algorithm and improved depth residual error network
CN116826745B (en) Layered and partitioned short-term load prediction method and system in power system background
CN117522626A (en) Photovoltaic output prediction method based on feature selection and abnormal multi-model fusion
CN117374941A (en) Photovoltaic power generation power prediction method based on neural network
Yong et al. Short-term building load forecasting based on similar day selection and LSTM network
CN116435987A (en) Ultra-short-term wind power prediction method for large-scale wind power clusters

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
GR01 Patent grant
GR01 Patent grant