CN107944594B - Short-term load prediction method based on spearman grade and RKELM microgrid - Google Patents

Short-term load prediction method based on spearman grade and RKELM microgrid Download PDF

Info

Publication number
CN107944594B
CN107944594B CN201710957155.8A CN201710957155A CN107944594B CN 107944594 B CN107944594 B CN 107944594B CN 201710957155 A CN201710957155 A CN 201710957155A CN 107944594 B CN107944594 B CN 107944594B
Authority
CN
China
Prior art keywords
load
sample
prediction
data
historical
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.)
Active
Application number
CN201710957155.8A
Other languages
Chinese (zh)
Other versions
CN107944594A (en
Inventor
杨荣照
杨苹
陈夏
沈志钧
余伟洲
张云飞
陈亦平
张勇
候君
莫维科
高琴
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.)
South China University of Technology SCUT
Original Assignee
South China University of Technology SCUT
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 South China University of Technology SCUT filed Critical South China University of Technology SCUT
Publication of CN107944594A publication Critical patent/CN107944594A/en
Application granted granted Critical
Publication of CN107944594B publication Critical patent/CN107944594B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • 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
    • G06F18/211—Selection of the most significant subset of features
    • G06F18/2113—Selection of the most significant subset of features by ranking or filtering the set of features, e.g. using a measure of variance or of feature cross-correlation
    • 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)
  • Economics (AREA)
  • Physics & Mathematics (AREA)
  • Human Resources & Organizations (AREA)
  • Strategic Management (AREA)
  • General Physics & Mathematics (AREA)
  • Marketing (AREA)
  • Tourism & Hospitality (AREA)
  • General Business, Economics & Management (AREA)
  • Health & Medical Sciences (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Computation (AREA)
  • Public Health (AREA)
  • Water Supply & Treatment (AREA)
  • General Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Primary Health Care (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (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 short-term load prediction method based on SPSS and RKELM micro-grids, which comprises the following steps: (1) on-line data acquisition and periodic updating of a historical database (2) to preprocess historical data and extract load sample characteristics; (3) constructing an offline load prediction model; (4) screening a historical sample similar to the precursor load of the point to be predicted by adopting a spearman grade correlation method to serve as an online training sample; (5) and calculating a load predicted value at a future moment according to the online training sample and the offline load prediction model. The method uses a rapid simplified kernel function extreme learning machine (RKELM), a chaotic particle swarm algorithm and a spearman grade correlation screening (RKELM) to establish a prediction model containing off-line parameter optimization and on-line load; the timeliness of the algorithm is guaranteed through periodic updating of the model parameters, meanwhile, the complexity of online prediction calculation is reduced, the time-duration data storage amount is reduced, the calculation cost is reduced, and the short-term and ultra-short-term loads of the microgrid can be predicted accurately.

Description

Short-term load prediction method based on spearman grade and RKELM microgrid
Technical Field
The invention belongs to the technical field of microgrid load prediction, and particularly relates to a microgrid short-term load prediction method based on spearman grades and RKELM
Technical Field
The micro-grid is a small power generation and distribution system which is composed of a distributed power supply, an energy storage device, an energy conversion device, a load, a monitoring and protecting device and the like. The grid-connected operation or the independent operation of the microgrid is realized through the controllability of a power supply and a load inside the microgrid. The micro-grid is externally represented as an integral unit, and meanwhile, the micro-grid can be smoothly merged into a main grid for operation, so that the requirements of users on electric energy quality and power supply safety are met. With the rapid development of new energy and new technology, distributed power generation is gradually popularized and applied. The micro-grid can promote the access of distributed clean energy, reduce environmental pollution and power transmission loss, and improve the power supply reliability of users through switching of an isolated/grid-connected mode. Meanwhile, load prediction of the microgrid is an important component of an energy management system of the smart grid, and an effective load prediction model is of great importance to demand side management and development of smart grid technology.
Data analysis shows that the load randomness of the micro-grid is higher, the similarity of the data is not large after the time passes, and the prediction difficulty is higher compared with that of a large power grid. In order to ensure efficient economic operation of the microgrid, accurate load prediction is an important basis for decision-making of microgrid optimization operation and energy management. The method for predicting the short-term load of the power system mainly comprises an extrapolation method, a time series method, an artificial neural network method, a support vector machine method and the like. However, these methods are not well suited for load prediction of the microgrid with higher load randomness or have higher calculation cost. The invention provides a short-term load prediction method based on a spearman grade and an RKELM microgrid.
Disclosure of Invention
The invention provides a short-term load forecasting method based on a spearman grade and an RKELM microgrid, aiming at the problems in the prior art.
The scheme of the invention is as follows: a short-term load prediction method based on a spearman grade and an RKELM microgrid comprises the following steps:
step A: online data collection is periodic and updates the historical database.
And B: and preprocessing the historical data and extracting load sample characteristics.
And C: constructing an offline load prediction model: and performing offline training on the load historical data by using a chaotic particle swarm algorithm and a simplified kernel function extreme learning machine (RKELM) to generate an offline load prediction model.
Step D: and screening a historical sample similar to the historical load of the point to be predicted by adopting a spearman grade correlation method to serve as an online training sample.
Step E: online prediction: and calculating a load predicted value at a future moment according to the online training sample and the offline load prediction model.
Further, for said step a, the online data collection periodically and updates the duration database, including:
and collecting data on line, updating a historical database every day, and storing the data of the previous 30 days in the database.
Further, for the step B, preprocessing the historical data and extracting load sample features, including:
step B1: the historical data is normalized and scaled to a [0,1] interval, and the formula is as follows:
Figure GDA0003216148810000011
wherein xiLoading a power value at the moment i; x is the number ofminThe minimum load active value of the historical data in the database is obtained; x is the number ofmaxThe maximum load active value of the historical data in the database is obtained;
Figure GDA0003216148810000021
and (4) normalizing the load value at the moment i.
Step B2: marking missing data or more than 4 continuous sampling values as problem data PkModified according to the following formula:
Figure GDA0003216148810000022
wherein P isk' As modified data, Pk-96Data for 96 points before time k (i.e., one day ago), Pk-192Data of 192 points before time k, Pk-288The data of 288 points before the time k. The data acquisition period is 15 min.
Step B3: extracting load sample features including sampling time tiWeek information wiAverage load before day daiBefore-day lag load dliBefore-week hysteresis load wli。
For average load d before dayaiThe calculation method is shown in the following formula.
Figure GDA0003216148810000023
xiFor the moment i loaded with a power value, xi-jIs the load active value of the ith-j point; daiEach load point is corresponding to a day-ahead average load; every 15 minutes, the average load before the day is the average load of the first 96 points of the load point; i represents the current load point, i starts at 97 since the data of the first 96 points have no daily average load, and j is 1 to 96; g is the total number of samples of the historical load samples;
for the before-day lag load dliHysteresis load w before cycleliThe calculation method is shown in the following formula.
dli＝xi-96；i＝97…G
wli＝xi-672；i＝673…G
For convenience of description, A is usediAll the load attributes of the ith sample are referred to, and the structure is shown as the following formula.
Ai＝{ti,wi,dai,dli,wli}
Further, for step C, an offline load prediction model is constructed: and performing offline training on the load historical data by adopting a chaotic particle swarm algorithm and a simplified kernel function extreme learning machine to generate an offline load prediction model. The method comprises the following steps:
step C1: using the sample MTParameters to be optimized with the RKELM algorithm (extreme learning machine algorithm with simplified kernel function)
Figure GDA0003216148810000024
And
Figure GDA0003216148810000025
construction of prediction model of T time
Figure GDA0003216148810000026
Reuse of the verification sample VTLoad characteristics of, predict and verify
Figure GDA0003216148810000027
In the formula, g is a prediction model training algebra and corresponds to a chaotic particle swarm optimization algebra;
Figure GDA0003216148810000028
parameters of the RKELM neural network model at the T moment;
Figure GDA0003216148810000029
is a Gaussian kernel function parameter at time T;
Figure GDA00032161488100000210
the RKELM neural network model dimension at the T moment;
Figure GDA0003216148810000031
load prediction values for the validation samples; mTFor training samples at time T, VTA verification sample at time T; a. theiTo verify the characteristics of the sample.
Step C2: solving the optimal model parameters by using a chaotic particle swarm algorithm, wherein the target function of the chaotic particle swarm algorithm (CPSO) is the minimum prediction Error, the Mean Absolute percentage Error (Mean Absolute percentage Error) is used as an evaluation function when the prediction model is constructed, and the evaluation function is shown in the following formula
Figure GDA0003216148810000032
Wherein S in the formula is the total prediction time. Optimizing parameters according to the prediction error and parameters of each group
Figure GDA0003216148810000033
Wherein
Figure GDA0003216148810000034
L∈[6,7,...,12]；Z1,Z2∈[-15,25]。
Step C3: and constructing an offline prediction model by using the optimal parameters C, gamma and L.
Further, for the step D, screening a historical sample similar to the historical load of the point to be predicted by adopting a spearman grade correlation method to serve as an online training sample. The method specifically comprises the following steps:
taking the historical load { x) of the predicted pointi-1,xi-2,xi-3,...,xi-16Is a reference sample, ranked by size
Figure GDA0003216148810000035
Comparing the similarity of the historical load sample at the similar moment of the previous 8 days with the reference sample, wherein the spearman grade correlation coefficient is as follows:
Figure GDA0003216148810000036
wherein x'i-mAre historical load samples. Let the time to be predicted be ToAnd selecting 12 samples with the maximum rho as online training samples.
Further, for said step E, online predicting: and calculating a load predicted value at a future moment according to the online training sample and the offline load prediction model. The method specifically comprises the following steps:
and (3) performing online prediction by using 12 samples screened out by the spearman grade correlation as online training samples and combining an offline load prediction model. Meanwhile, in order to ensure the prediction accuracy, the off-line optimization operation interval is one week, and the latest 1-month historical data is adopted for operation so as to obtain the optimal parameter C at each momentT,γT,LT(T ═ 1,2,.., 96) for online prediction of the next week.
Compared with the prior art, the invention provides a method for optimizing a prokaryotic function model by adopting an RKELM model and screening relevant samples by using a spearman on the basis of analyzing and evaluating the load data of the microgrid, so that the precision is improved and the calculation efficiency is considered.
Drawings
FIG. 1 is a comparison graph of optimal parameters for the week 1 test of building microgrid load prediction in an example.
Fig. 2 is a graph of the results of the weekly building microgrid load predictions of the example.
Fig. 3 is a graph of the load prediction relative error of the building microgrid at week 1.
Fig. 4 is a flowchart of a short-term load prediction method based on spearman rank and RKELM microgrid, which is provided by the invention:
fig. 5 is a schematic implementation diagram of the microgrid load prediction principle.
Detailed Description
The practice of the present invention will be further illustrated with reference to the accompanying drawings and examples, but the practice and protection of the invention is not limited thereto.
Fig. 1 is a short-term load prediction method based on spearman grades and RKELM microgrid. As shown in fig. 1, the method of the present invention comprises the steps of:
step A: the on-line data acquisition periodically updates the historical database, and specifically comprises the following steps:
and the data acquisition device is used for acquiring data on line, updating the historical database every day, and storing the data of the previous 30 days in the database.
And B: preprocessing historical data and extracting load sample characteristics, and the method specifically comprises the following steps:
step B1: the historical data is normalized and scaled to a [0,1] interval, and the formula is as follows:
Figure GDA0003216148810000041
wherein xiLoading a power value at the moment i; x is the number ofminThe minimum load active value of the historical data in the database is obtained; x is the number ofmaxThe maximum load active value of the historical data in the database is obtained;
Figure GDA0003216148810000042
is when i isAnd (4) calibrating the load normalization value.
The collected data is difficult to avoid error data or missing, and therefore, correction of the data is essential.
Step B2: marking missing data or more than 4 continuous sampling values as problem data PkModified according to the following formula:
Figure GDA0003216148810000043
wherein P isk' As modified data, Pk-96Data for 96 points before time k (i.e., one day ago), Pk-192Data of 192 points before time k, Pk-288The data of 288 points before the time k. The data acquisition period is 15 min.
Step B3: extracting load sample features including sampling time tiWeek information wiAverage load before day daiBefore-day lag load dliBefore-week hysteresis load wli。
For average load d before dayaiThe calculation method is shown in the following formula.
Figure GDA0003216148810000044
In the formula: g is the total number of samples of the historical load samples.
For the before-day lag load dliHysteresis load w before cycleliThe calculation method is shown in the following formula.
dli＝xi-96；i＝97…G
wli＝xi-672；i＝673…G
For convenience of description, A is usediAll the load attributes of the ith sample are referred to, and the structure is shown as the following formula.
Ai＝{ti,wi,dai,dli,wli}
And C: constructing an offline load prediction model: and performing offline training on the load historical data by using a chaotic particle swarm algorithm and a simplified kernel function extreme learning machine (RKELM) to generate an offline load prediction model. The method comprises the following steps:
the neural network model of the prokaryote function extreme learning machine (KELM) is shown as follows.
Figure GDA0003216148810000051
Where N is the input layer dimension. K (x)i,xj) Is the selected kernel function. Kernel function K (x)i,xj) It usually takes a gaussian kernel function, as shown below.
ΩELM(xi,xj)＝K(xi,xj)＝exp(-γ||xi-xj||2)
It can be known that KELM is an NxN dimensional neural network, where N is the input layer input data xjN is also the number of training samples. Obviously, under the structure, when N is increased, the complexity of the network is increased, and the calculation amount is increased rapidly, so that the number of training samples is limited, the generalization capability of the kernel function limit learning machine is limited, and the establishment of a prediction model is not facilitated.
According to the literature (Deng Wanyu, on year Soon, Zheng Qinghua. A Fast Reduced Kernel Extreme Learning Machine [ J ]. NEURAL NETWORKS, 2016, 76:29-38.), it is known that the support vector can be randomly selected, and thus the KELM can be simplified to RKELM. Its network can be represented as:
Figure GDA0003216148810000052
in the formula, beta is an output weight connecting the hidden layer and the output layer, y is an RKELM training target value, and L is an RKELM neural network model dimension;
writing in matrix form then:
KN×Lβ＝Y
in the formula KN×L＝k(X,XL) For simplified kernel function momentsIn the matrix, β is an output weight vector of length L.
Minimize output weight β:
Figure GDA0003216148810000053
in the ELM algorithm, the output weight β is usually calculated by taking the least squares solution thereof, and the calculation method is shown as the following formula.
Figure GDA0003216148810000054
Based on the above equation, the neural network function of RKELM can be expressed as:
Figure GDA0003216148810000055
thus, the N × N dimensional neural network can be simplified to an N × L neural network. In comparison to KELM, RKELM requires an optimal solution for L in addition to C and γ.
Step C1: using the sample MTParameters to be optimized with the RKELM algorithm (extreme learning machine algorithm with simplified kernel function)
Figure GDA0003216148810000056
And
Figure GDA0003216148810000057
construction of prediction model of T time
Figure GDA0003216148810000058
Reuse of the verification sample VTLoad characteristics of, predict and verify
Figure GDA0003216148810000059
In the formula, g is a prediction model training algebra and corresponds to a chaotic particle swarm optimization algebra;
Figure GDA00032161488100000510
parameters of the RKELM neural network model at the T moment;
Figure GDA00032161488100000511
is a Gaussian kernel function parameter at time T;
Figure GDA00032161488100000512
the RKELM neural network model dimension at the T moment;
Figure GDA0003216148810000061
load prediction values for the validation samples; mTFor training samples at time T, VTA verification sample at time T; a. theiTo verify the characteristics of the sample.
The selection mode of the verification sample is that parameter optimization is carried out on the load prediction model at the time T, and the verification sample is
VT＝{xi|Ai}；ti＝T
In the formula, VTA verification sample of the load prediction model at the time T; x is the number ofiIs a predicted target value, i.e. an actual value; a. theiTo verify the characteristics of the sample.
For the training sample, the predecessor sample at the T moment and the two-day previous synchronization sample are used as the training sample for the off-line optimization of the T moment, as follows
MT＝{xi|Ai}；ti＝{T-1,T-2,
T-94,…,T-98,T-192,…,T-196}
In the formula MTAnd training samples of the prediction model at the T moment.
Step C2: solving the optimal model parameters by using a chaotic particle swarm algorithm, wherein the target function of the chaotic particle swarm algorithm (CPSO) is the minimum prediction Error, the Mean Absolute percentage Error (Mean Absolute percentage Error) is used as an evaluation function when the prediction model is constructed, and the evaluation function is shown in the following formula
Figure GDA0003216148810000062
Wherein S in the formula is the total prediction time. Optimizing parameters according to the prediction error and parameters of each group
Figure GDA0003216148810000063
Wherein
Figure GDA0003216148810000064
L∈[6,7,...,12]；Z1,Z2∈[-15,25]。
Compared with the particle swarm algorithm, the chaotic particle swarm algorithm has the characteristic of ergodicity, and the probability of obtaining the global optimum is improved by inerting part of particles trapped in the local extreme value, namely introducing chaos. The same part as the particle swarm algorithm is shown in the formula
Figure GDA0003216148810000065
Wherein
Figure GDA0003216148810000066
Represents the g-th generation position of the i-th particle, corresponding to the parameter
Figure GDA0003216148810000067
The g-th change vector of the ith particle, w is the inertia coefficient, c1,c2The degree of trust of the particle to itself and the degree of trust of the group, r1,r2Is [0,1]]Random number between, pbestiThe optimal position of the ith particle is the gbest optimal position of the population.
When the fitness function of the particle satisfies the following formula in 5 successive generations
Figure GDA0003216148810000068
And judging that the particles are trapped in a stagnation state, and entering a chaotic search mode. Where per is the threshold. Randomly generating a 3-dimensional vector space
Y0＝{Y0,1,Y0,2,Y0,3}；Y0,1,Y0,2,Y0,3∈[0,1]
By vector Y0As an iteration initial vector. According to Logistic equation.
Yn+1＝μYn(1-Yn)
Starting chaos sequence iteration to obtain iteration sequence Y0,Y1…Yn-1The Logistic equation has the characteristic of ergodicity, and can iterate a plurality of fields around the local optimal solution. Then by means of a carrier wave, according to
Y′i＝gbest+R(2Yi-1)；i＝0,1,…n-1
And amplifying to a region with the original position of the particle, namely gbest, as the center and the radius of R, wherein R is the radius of the chaotic search. Obtained Y'iEntering a particle swarm optimization process. And when the precision requirement is met or the upper limit of the iteration times is reached to 200 times, the parameter optimization process is terminated.
The off-line optimizing operation interval period is one week, and the latest 1 month historical data is adopted for operation so as to obtain the optimal parameter C at each momentT,γT,LT(T ═ 1,2,.., 96) for online prediction of the next week.
Step C3: using optimum parameters CT,γT,LT(T ═ 1, 2.., 96) an offline prediction model was constructed.
Step D: and screening a historical sample similar to the historical load of the point to be predicted by adopting a spearman grade correlation method to serve as an online training sample. The method specifically comprises the following steps:
taking the historical load { x) of the predicted pointi-1,xi-2,xi-3,…,xi-16Is a reference sample, ranked by size
Figure GDA0003216148810000071
Comparison front 8 days similarSimilarity between the historical load sample and the reference sample at the moment, and a spearman grade correlation coefficient is as follows:
Figure GDA0003216148810000072
wherein x'i-mAre historical load samples. Let the time to be predicted be ToAnd selecting 12 samples with the maximum rho as online training samples.
Step E: online prediction: and calculating a load predicted value at a future moment according to the online training sample and the offline load prediction model. The method specifically comprises the following steps:
and (3) performing online prediction by using 12 samples screened out by the spearman grade correlation as online training samples and combining an offline load prediction model. Meanwhile, in order to ensure the prediction accuracy, the off-line optimization operation interval is one week, and the latest 1-month historical data is adopted for operation so as to obtain the optimal parameter C at each momentT,γT,LT(T ═ 1,2,.., 96) for online prediction of the next week.
Specific examples of predictions:
historical loads of residential building users, which are collected by a certain property company, for 70 days from 1 ten days to 3 months are used as original sampling data; the sampling time interval of the sample is 15min, and the original sampling attribute comprises time and load power.
Before the data is imported into the prediction model, all attributes are normalized and scaled to [0,1], and the load power is taken as an example below
Figure GDA0003216148810000073
Using historical data of previous 40 days as time-sharing training and verification samples for off-line parameter optimization; the load history data of the last 30 days is a test sample of online prediction for testing prediction errors.
The accuracy evaluation criteria adopted by the load prediction are errors: the mean relative error (MAPE), relative error (APE) are used herein
Figure GDA0003216148810000074
Figure GDA0003216148810000081
Wherein S is the number of prediction samples; x is the number ofiIs the actual value of the test sample;
Figure GDA0003216148810000082
predicted values for the test samples.
Off-line parameter optimization results
Each week of the prediction has a set of RKELM parameters, each containing 96 sets of optimal parameter values corresponding to a time point every 15min from 0:00 during the day. Taking week 1 of the prediction test as an example, the values of the parameters are shown in fig. 1.
Load prediction result
After offline parameter optimization, screening a historical sample similar to the historical load of the point to be predicted by adopting a spearman grade correlation method as an online training sample, and obtaining a load prediction result as shown in the following figures 2 and 3:
it can be seen from the prediction results that the load prediction results are more accurate in most time periods, but also because the model is established on the basis of the lagging load data of the similar day, when the similarity between the load prediction target value of a day and the historical data is low, the deviation of the prediction values is larger, as shown in the prediction results of the 6 th day in fig. 3.
TABLE 1 building load prediction error after model period update
Figure GDA0003216148810000083
Table 1 shows the prediction result error after the model optimal parameter period is updated in weeks 2 to 4. Therefore, the prediction model disclosed by the invention has the advantages that the average prediction relative error is not more than 12% after periodic updating, and the requirement that the prediction error is not more than 15% in one month in the industry is met.
Aiming at the characteristics of severe change and large difference of building loads with certain regularity, the kernel function and the particle swarm algorithm are improved, the kernel function is reasonably reduced in dimension, chaotic search is added in the particle swarm algorithm, so that the robustness and the generalization capability of the model are enhanced, similar historical samples are reasonably searched based on the spearman grade to construct an online load prediction model, and the flexibility and the prediction capability of the prediction model are improved.

Claims (4)

1. A short-term load prediction method based on a spearman grade and an RKELM microgrid is characterized by comprising the following steps:
step A: collecting online data and periodically updating a historical database;
and B: preprocessing historical data and extracting load sample characteristics;
and C: constructing an offline load prediction model: the method comprises the following steps of performing offline training on load historical data by using a chaotic particle swarm algorithm and a simplified kernel function extreme learning machine RKELM to generate an offline load prediction model, wherein the offline load prediction model specifically comprises the following steps:
step C1: using the sample MTWith the parameters to be optimized in the RKELM algorithm
Figure FDA0003244899940000011
And
Figure FDA0003244899940000012
construction of prediction model of T time
Figure FDA0003244899940000013
Reuse of the verification sample VTThe load characteristics of (a) are predicted and verified:
Figure FDA0003244899940000014
in the above formula, g is the prediction model training algebraCorresponding to the optimized algebra of the chaotic particle swarm algorithm;
Figure FDA0003244899940000015
parameters of the RKELM neural network model at the T moment;
Figure FDA0003244899940000016
is a Gaussian kernel function parameter at time T;
Figure FDA0003244899940000017
the RKELM neural network model dimension at the T moment;
Figure FDA0003244899940000018
load prediction values for the validation samples; mTFor training samples at time T, VTA verification sample at time T; a. theiTo verify the characteristics of the sample;
step C2: solving the optimal model parameters by using a chaotic particle swarm algorithm, wherein the target function of the chaotic particle swarm algorithm is the minimum prediction error, and the average absolute percentage error is used as an evaluation function in the construction of the prediction model, as shown in the following formula
Figure FDA0003244899940000019
Wherein S in the formula is the total prediction time; optimizing parameters according to the prediction error and parameters of each group
Figure FDA00032448999400000110
Wherein
Figure FDA00032448999400000111
L∈[6,7,...,12]；Z1,Z2∈[-15,25]；
Step C3: an off-line prediction model is constructed by utilizing the optimal parameters C, gamma and L,xifor a moment i loaded with a positive value, MAPEgIn the form of an average absolute percentage error,
Figure FDA00032448999400000112
predicting the value of the test sample;
step D: adopting a spearman grade correlation method to screen a historical sample similar to the historical load of a point to be predicted as an online training sample, and specifically comprising the following steps: taking historical load { x) of point to be predictedi-1,xi-2,xi-3,...,xi-16Is a reference sample, ranked by size
Figure FDA00032448999400000113
Comparing the similarity of the historical load sample at the similar moment of the previous 8 days with the reference sample, wherein the spearman grade correlation coefficient is as follows:
Figure FDA00032448999400000114
wherein x'i-mIs a historical load sample; let the time to be predicted be ToSelecting 12 samples with the maximum rho as online training samples;
step E: online prediction: and calculating a load predicted value at a future moment according to the online training sample and the offline load prediction model.
2. The method according to claim 1, wherein the step A comprises:
and collecting data on line, updating a historical database every day, and storing the data of the previous 30 days in the database.
3. The method according to claim 1, wherein the step B comprises:
step B1: the historical data is normalized and scaled to a [0,1] interval, and the formula is as follows:
Figure FDA0003244899940000021
wherein xiLoading a power value at the moment i; x is the number ofminThe minimum load active value of the historical data in the database is obtained; x is the number ofmaxThe maximum load active value of the historical data in the database is obtained;
Figure FDA0003244899940000022
load normalization value at the moment i;
step B2: marking missing data or more than 4 continuous sampling values as problem data PkModified according to the following formula:
Figure FDA0003244899940000023
wherein P'kFor the corrected data, Pk-96Data for 96 points before time k, Pk-192Data of 192 points before time k, Pk-288288 points before time k; the data acquisition period is 15 min;
step B3: extracting load sample features including sampling time tiWeek information wiAverage load before day daiBefore-day lag load dliBefore-week hysteresis load wli；
For average load d before dayaiThe calculation method is shown as the following formula:
Figure FDA0003244899940000024
xifor the moment i loaded with a power value, xi-jIs the load active value of the ith-j point; daiEach load point is corresponding to a day-ahead average load; one point every 15 minutes, average daily negativeThe load is the average load of the first 96 points of the load point; i represents the current load point, i starts at 97 since the data of the first 96 points have no daily average load, and j is 1 to 96; g is the total number of samples of the historical load samples;
for the before-day lag load dliHysteresis load w before cycleliThe calculation method is shown as the following formula:
dli＝xi-96；i＝97…G
wli＝xi-672；i＝673…G
for convenience of description, A is usediAll load characteristics for the ith sample are expressed as follows:
Ai＝{ti,wi,dai,dli,wli}。
4. the method according to claim 1, wherein in step E:
12 samples screened out by spearman grade correlation are used as online training samples, and online prediction is carried out by combining an offline load prediction model; meanwhile, in order to ensure the prediction accuracy, the off-line optimization operation interval is one week, and the latest 1-month historical data is adopted for operation so as to obtain the optimal parameter C at each momentT,γT,LT(T ═ 1,2,.., 96) for online prediction of the next week.
CN201710957155.8A 2017-09-30 2017-10-16 Short-term load prediction method based on spearman grade and RKELM microgrid Active CN107944594B (en)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN2017109240409 2017-09-30
CN201710924040 2017-09-30

Publications (2)

Publication Number Publication Date
CN107944594A CN107944594A (en) 2018-04-20
CN107944594B true CN107944594B (en) 2021-11-23

Family

ID=61935308

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201710957155.8A Active CN107944594B (en) 2017-09-30 2017-10-16 Short-term load prediction method based on spearman grade and RKELM microgrid

Country Status (1)

Country Link
CN (1) CN107944594B (en)

Families Citing this family (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109918415A (en) * 2019-02-21 2019-06-21 智恒科技股份有限公司 A kind of method and system of the water utilities data prediction of data warehouse technology
CN110648248B (en) * 2019-09-05 2023-04-07 广东电网有限责任公司 Control method, device and equipment for power station
CN111487950B (en) * 2020-04-24 2021-11-16 西安交通大学 'prediction-verification-feedback-optimization' closed-loop system for online early warning and offline diagnosis
CN111581883B (en) * 2020-05-09 2022-09-23 国网上海市电力公司 Method for calculating and predicting load on automation device
CN111652424B (en) * 2020-05-28 2022-06-10 国网甘肃省电力公司经济技术研究院 Hydrogen load prediction method
CN112766585A (en) * 2021-01-25 2021-05-07 三峡大学 Electric power short-term rolling load prediction method, system and terminal based on soft ensemble learning
CN114970938B (en) * 2022-03-11 2024-05-07 武汉大学 Self-adaptive house load prediction method considering user privacy protection

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104008432A (en) * 2014-06-03 2014-08-27 华北电力大学 Micro-grid short-term load forecasting method based on EMD-KELM-EKF
CN104090490A (en) * 2014-07-04 2014-10-08 北京工业大学 Input shaper closed-loop control method based on chaotic particle swarm optimization algorithm
CN105046374A (en) * 2015-08-25 2015-11-11 华北电力大学 Power interval predication method based on nucleus limit learning machine model
CN105354646A (en) * 2015-12-04 2016-02-24 福州大学 Power load forecasting method for hybrid particle swarm optimization and extreme learning machine
CN106570250A (en) * 2016-11-02 2017-04-19 华北电力大学（保定） Power big data oriented microgrid short-period load prediction method

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170082634A1 (en) * 2015-07-21 2017-03-23 The General Hospital Corporation Multiplexed Proteomics and Phosphoproteomics

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104008432A (en) * 2014-06-03 2014-08-27 华北电力大学 Micro-grid short-term load forecasting method based on EMD-KELM-EKF
CN104090490A (en) * 2014-07-04 2014-10-08 北京工业大学 Input shaper closed-loop control method based on chaotic particle swarm optimization algorithm
CN105046374A (en) * 2015-08-25 2015-11-11 华北电力大学 Power interval predication method based on nucleus limit learning machine model
CN105354646A (en) * 2015-12-04 2016-02-24 福州大学 Power load forecasting method for hybrid particle swarm optimization and extreme learning machine
CN106570250A (en) * 2016-11-02 2017-04-19 华北电力大学（保定） Power big data oriented microgrid short-period load prediction method

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
《A Fast Reduced Kernel Extreme Learning Machine》;Wan-Yu Deng等;《NEURAL NETWORKS》;20161231;第29-38页 *
《基于优化组合核极限学习机的网络流量预测》;刘悦等;《计算机技术与发展》;20160630;第26卷(第6期);第73-77页 *
《自适应混沌粒子群算法对极限学习机参数的优化》;陈晓青等;《计算机应用》;20161110;第36卷;第3123-3126页 *

Also Published As

Publication number Publication date
CN107944594A (en) 2018-04-20

Similar Documents

Publication Publication Date Title
CN107944594B (en) Short-term load prediction method based on spearman grade and RKELM microgrid
CN113962364B (en) Multi-factor power load prediction method based on deep learning
Wang et al. Research and application of a hybrid forecasting framework based on multi-objective optimization for electrical power system
Feng et al. A taxonomical review on recent artificial intelligence applications to PV integration into power grids
Bozorg et al. Bayesian bootstrap quantile regression for probabilistic photovoltaic power forecasting
Muzumdar et al. Designing a robust and accurate model for consumer-centric short-term load forecasting in microgrid environment
CN110212524A (en) A kind of region Methods of electric load forecasting
Dong et al. Short-term building cooling load prediction model based on DwdAdam-ILSTM algorithm: A case study of a commercial building
Ullah et al. Deep Learning‐Assisted Short‐Term Power Load Forecasting Using Deep Convolutional LSTM and Stacked GRU
Wang et al. Short-term load forecasting with LSTM based ensemble learning
Buonanno et al. Comprehensive method for modeling uncertainties of solar irradiance for PV power generation in smart grids
Zheng et al. Short-term energy consumption prediction of electric vehicle charging station using attentional feature engineering and multi-sequence stacked Gated Recurrent Unit
Mishra et al. Performance evaluation of prophet and STL-ETS methods for load forecasting
Alam et al. AI-based efficiency analysis technique for photovoltaic renewable energy system
Gaber et al. Hourly electricity price prediction applying deep learning for electricity market management
Liao et al. Short-term load forecasting with temporal fusion transformers for power distribution networks
CN117374920A (en) Ultra-short-term prediction method, device and medium considering environmental factors
CN113537607B (en) Power failure prediction method
CN115759395A (en) Training of photovoltaic detection model, detection method of photovoltaic power generation and related device
Viana et al. Load forecasting benchmark for smart meter data
Kartini et al. Very short term load forecasting based on meteorological with modelling k-NN-feed forward neural network
CN113283638A (en) Load extreme curve prediction method and system based on fusion model
Xia et al. Research on Solar Radiation Estimation based on Singular Spectrum Analysis-Deep Belief Network
Guanoluisa-Pineda et al. Short-Term forecasting of photovoltaic power in an isolated area of Ecuador using deep learning techniques
Zhang Research on Photovoltaic Power Prediction Method Based on TCN-BiLSTM Neural Network

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
OL01 Intention to license declared
OL01 Intention to license declared