CN109242139A - A kind of electric power day peak load prediction technique - Google Patents
A kind of electric power day peak load prediction technique Download PDFInfo
- Publication number
- CN109242139A CN109242139A CN201810811805.2A CN201810811805A CN109242139A CN 109242139 A CN109242139 A CN 109242139A CN 201810811805 A CN201810811805 A CN 201810811805A CN 109242139 A CN109242139 A CN 109242139A
- Authority
- CN
- China
- Prior art keywords
- wolf
- peak load
- parameter
- signal
- prediction
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/004—Artificial life, i.e. computing arrangements simulating life
- G06N3/006—Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
-
- 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—Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
- G06Q50/06—Electricity, gas or water supply
Abstract
The invention belongs to Techniques for Prediction of Electric Loads fields, a kind of more particularly to electric power day peak load prediction technique, comprising: sample data of the acquisition including history day peak load, max. daily temperature, Daily minimum temperature, mean daily temperature, per day relative humidity, day maximum wind velocity, date type；The original series of day peak load are carried out with the complete polymerization empirical mode decomposition of adaptive white noise, original series are decomposed into the intrinsic mode functions of limited local feature signal comprising different time scales, and the smooth impulse disturbances of adaptive white noise are all added in decomposing each time, obtain multiple IMF components；By introducing population dynamic evolutionary operator and non-linear convergence factor, grey wolf optimization algorithm is improved, the regularization parameter and Radial basis kernel function parameter to support vector machines optimize, the SVM prediction model after establishing optimization；It is predicted respectively with prediction model, obtains final day peak load prediction result.
Description
Technical field
The invention belongs to Techniques for Prediction of Electric Loads field more particularly to a kind of electric power day peak load prediction techniques.
Background technique
The development of modern society be unable to do without the supply of electric power everywhere, and power industry is raw for social and economic construction and the people
The flat raising of running water plays a crucial role.With the rapid development of power industry, electric system is for Electric Load Forecasting
The required precision of survey is higher and higher.An important ring of the day peak load prediction as load forecast, the accuracy of prediction
Great influence is suffered to the formulation of generation schedule, Dispatch of electric net and electric power, power grid operation and power system power supply reliability.
Therefore, suitable model is constructed, realization is the research work being of great significance to the Accurate Prediction of day peak load.
The key problem of load prediction is the method and model of prediction, with the fast development of science and technology, load prediction technology
Also it is deepening constantly, currently, load prediction technology is gradually transitioned into artificial intelligence Predicting Technique from traditional prediction method.It passes
The load forecasting method of system, such as: time series method, regression analysis, grey method, there are certain deficiencies, to fluctuation
Property biggish complicated load sequence prediction accuracy it is to be improved, and artificial intelligence prediction technique in face of complicated load sequence then
Powerful superiority has been shown, good prediction effect is realized.Artificial neural network algorithm originate from the 1940s,
It is the artificial intelligence technology of a simulation human brain bioprocess.BP (Back Propagation) algorithm is also known as error and reversely passes
Algorithm is broadcast, is the learning algorithm of one of artificial neural network supervised, is usually used in load prediction.But BP neural network
That there are convergence rates is slow for algorithm, the training time is long, the disadvantages of easily falling into locally optimal solution.And the prediction of day peak load is vulnerable to the external world
Factor is interfered and fluctuation is stronger, and load sequence includes that noise is more, this brings great difficulty to prediction work.
Summary of the invention
In order to realize the Accurate Prediction to day peak load, herein on the basis of considering meteorologic factor and date type,
Propose a kind of electric power day peak load prediction technique, comprising:
Step 1: acquisition includes history day peak load, max. daily temperature, Daily minimum temperature, mean daily temperature, per day
Sample data including relative humidity, day maximum wind velocity, date type, is then normalized place to the meteorologic factor of data
Reason；
Step 2: the original series of day peak load are carried out with the complete polymerization empirical mode decomposition of adaptive white noise
(CEEMDAN, Complete Ensemble Empirical Mode Decomposition with Adaptive Noise),
Original series are decomposed into the intrinsic mode functions of limited local feature signal comprising different time scales, and are divided each time
The smooth impulse disturbances of adaptive white noise are all added in solution, obtain multiple IMF components；
Step 3: by introducing population dynamic evolutionary operator and non-linear convergence factor, grey wolf optimization algorithm being changed
Into；
Step 4: using improved grey wolf optimization algorithm to the regularization parameter and Radial basis kernel function of support vector machines
Parameter optimizes, the SVM prediction model after establishing optimization；
Step 5: on the basis of considering meteorologic factor and date type, being passed through to by the complete polymerization of adaptive white noise
It tests the IMF component that mode decomposition obtains to be predicted respectively with the SVM prediction model after optimization, and to prediction
As a result it is reconstructed, obtains final day peak load prediction result.
Specific step is as follows for the complete polymerization empirical mode decomposition of the adaptive white noise:
Step 201: in n times calculating, to signal X (t)+pinj(t) it is decomposed, wherein X (t) is original series, nj
It (t) is white Gaussian noise, parameter piThe signal-to-noise ratio of the white Gaussian noise and sequence signal that introduce for i-th, then first IMF divides
AmountAre as follows:
After the 1st introducing white Gaussian noise, obtained j-th of IMF component is decomposed by EMD, the 1st
Residue signal r1(t) are as follows:
Step 202: defining emdi() is that i-th of IMF component after EMD decomposition is carried out to signal, to sequence r1(t)+
p1emd1(nj(t)) it is decomposed, obtains second IMF componentAre as follows:
2nd residue signal r2(t) are as follows:
Step 203: and so on, k-th of residue signal rk(t) are as follows:
+ 1 IMF component of kthAre as follows:
emdk() is that k-th of IMF component after EMD decomposition is carried out to signal；
Step 204: repeating the above process, until residue signal meets stopping criterion for iteration；If there is L IMF component at this time,
Then original series indicate are as follows:
In formula, r (t) is final residue signal.
The step 3 specifically includes:
In GWO algorithm, grey wolf group updates self-position according to the location information of α wolf, β wolf, δ wolf, at this point, we
More new formula is improved, is enabled:
xα=x1±(ub-lb·r+lb) (28)
xβ=x2±(ub-lb·r+lb) (29)
xδ=x3±(ub-lb·r+lb) (30)
In formula, ub and lb are respectively the upper bound and the lower bound of population search space, r be value range between [0,1] with
Machine number, xα、xβ、xδThe respectively location information of α wolf, β wolf, δ wolf, x1、x2、x3The location information of respectively the 1st, 2,3 wolf；
Updated potential optimal solution vector position x (t+1) are as follows:
Non-linear convergence factor is introduced to improve grey wolf optimization algorithm, as shown in formula (32):
Wherein: t is current iteration number, tmaxIt is maximum number of iterations, e is natural constant, improved convergence factor a
As the increase of the number of iterations is in decreases in non-linear, initial stage successively decreases slowly, is convenient for global search, the later period successively decreases rapidly, reinforcing office
Portion's optimizing.
The step 4 specifically includes:
Step 401: be arranged improved grey wolf optimization algorithm relevant parameter and support vector machines regularization parameter c and
The value range of Radial basis kernel function parameter g；
Step 402: random initializtion grey wolf population, and the position vector of each grey wolf individual is enabled to be made of parameter c, g；
Step 403: training set being learnt using the support vector machines after initialization, calculates the suitable of each grey wolf individual
Answer angle value；
Step 404: grade separation being carried out to grey wolf population according to fitness value, determines α wolf, β wolf, δ wolf and ω wolf
Position；
Step 405: updating wolf pack position, generate new population, calculate corresponding fitness value, and suitable with last iteration
It answers angle value to be compared, preferentially retains；
Step 406: judge whether to reach maximum number of iterations, if reaching, training terminates, the optimal value of output parameter c, g,
Otherwise, it gos to step and 404 continues parameter optimization；
Step 407: establishing SVM prediction model using the parameter c, g after optimization, test set is predicted.
Beneficial effects of the present invention:
The present invention carries out noise reduction process to day peak load sequence using CEEMDAN method, and CEEMDAN method can be to letter
Number tranquilization processing is done, sophisticated signal is decomposed into the eigen mode letter of limited local feature signal comprising different time scales
Number, and the smooth impulse disturbances of adaptive white noise are all added in decomposing each time, make the discomposing effect of signal data more
Completely.And propose that a kind of novel day peak load prediction model --- CEEMDAN-MGWO-SVM be (adaptive white noise
Complete polymerization empirical mode decomposition and improved grey wolf optimization algorithm Support Vector Machines Optimized, Complete Ensemble
Empirical Mode Decomposition with Adaptive Noise Support Vector Machine
Optimized by Modified Grey Wolf Optimization Algorithm).The model passes through adaptive white noise
The complete polymerization empirical mode decomposition algorithm of sound is decomposed to obtain multiple subsequences to day peak load sequence, is then used
MGWO-SVM model predicts each subsequence, finally reconstructs forecasting sequence, obtains prediction result.The model is day peak value
Load prediction provides new direction and thinking.
Detailed description of the invention
Fig. 1 is that embodiment uses CEEMDAN decomposition result of the invention.
Fig. 2 a~2g is embodiment component prediction result and residual error.
Fig. 3 is the final prediction result figure of embodiment.
Fig. 4 is embodiment relative error figure.
Specific embodiment
With reference to the accompanying drawing, it elaborates to embodiment.
1. the complete polymerization empirical mode decomposition of adaptive white noise
1.1EMD
Empirical mode decomposition (Empirical Mode Decomposition) is Hilbert-Huang transformation (HHT)
Core algorithm, essence are that tranquilization processing is done to signal, and it includes different time scales that sophisticated signal, which is decomposed into limited,
The intrinsic mode functions (Intrinsic Mode Function, abbreviation IMF) of local feature signal, should meet following two item
Part:
(1) extreme point of signal and zero crossing number it is equal or it is most difference one；
(2) the coenvelope line of signal and the average value of lower envelope line are zero.
For given signal X (t), may be expressed as: after empirical mode decomposition
In formula, imfiIt (t) is the intrinsic mode functions component of the local feature signal comprising different time scales, rnIt (t) is residual
Remaining signal.
Specific step is as follows for EMD algorithm:
(1) all Local Extremums of X (t) are determined, are fitted the coenvelope line f of X (t) respectively with cubic spline functiona(t)
With lower envelope line fb(t)；
(2) the average value f of coenvelope line and lower envelope line is calculatedm(t)；
(3) X (t) and f are calculatedm(t) difference, E (t)=X (t)-fm(t)；
(4) E (t) is repeated into step (1)~(3) as original series, as envelope mean value fm(t) when tending to 0, first is obtained
A IMF component imf1(t)；
(5) X is enabled1(t)=X (t)-imf1(t), by X1(t) it is repeated the above process as original series, until residue signal
rn(t) be constant function or monotonic function when, stop decompose.
1.2 EEMD
EEMD is white Gaussian noise to be introduced, to solve the problems, such as modal overlap on the basis of traditional EMD.Specific steps
It is as follows:
(1) white Gaussian noise n is added to given original signal X (t)j(t), signal X is obtainedI(t):
XI(t)=X (t)+nj(t) (2)
(2) to signal XI(t) EMD decomposition is carried out, IMF component M is obtainedij(t), wherein Mij(t) Gauss is introduced for i-th
After white noise, obtained j-th of IMF component is decomposed by EMD；
(3) step (1), (2) n times are repeated, different white Gaussian noises is added every time；
(4) mean value of resulting IMF component is final IMF after taking n times to decompose, and formula indicates are as follows:
1.3 CEEMDAN
EEMD method can reduce the generation of modal overlap phenomenon in a certain range, but since white noise sequence is newly added
Column, after the average computation of limited times, error can't be completely counterbalanced by, and will affect the accuracy of reproducing sequence.For this purpose,
M.A.Colominas proposes the complete polymerization empirical mode decomposition of adaptive white noise on the Research foundation of EMD
(Complete Ensemble Empirical Mode Decomposition with Adaptive Noise)。CEEMDAN
Method all adds the smooth impulse disturbances of adaptive white noise in decomposing each time, utilizes the spy of white Gaussian noise zero-mean
Property, keep the discomposing effect of signal data more complete, modal overlap phenomenon can be efficiently solved.
Specific step is as follows for the decomposition of CEEMDAN method:
(1) identical as EEMD algorithm, CEEMDAN algorithm is in n times calculating, to signal X (t)+pinj(t) it is decomposed,
In, parameter piThe signal-to-noise ratio of additional noise and original signal is controlled, then first IMF component are as follows:
Residue signal at this time are as follows:
(2) emd is definedi() is that k-th of IMF component after EMD decomposition is carried out to signal, to sequence r1(t)+p1emd1(nj
(t)) it is decomposed, second IMF component can be obtained are as follows:
Residue signal at this time are as follows:
(3) and so on, k-th of residue signal are as follows:
+ 1 IMF component of kth are as follows:
(4) it repeats the above process, until residue signal meets stopping criterion for iteration.It is if there is L IMF component at this time, then former
Beginning sequence may be expressed as:
In formula, r (t) is final residue signal.
2. improved grey wolf optimization algorithm Support Vector Machines Optimized
2.1.SVM
Support vector machines is built upon theoretical one on Structural risk minization basis of VC dimension of Statistical Learning Theory
Kind machine learning method, it seeks best compromise between the complexity and learning ability of model according to limited sample information,
In the hope of obtaining best Generalization Ability.
When carrying out prediction modeling using SVM, pass through nonlinear mapping functionInput sample is mapped to the spy of higher-dimension
Space H is levied, and carries out linear regression, regression function of the SVM in high-dimensional feature space in H are as follows:
In formula: ω is high-dimensional feature space weight vector, ω ∈ Rk；B is offset constant, b ∈ R.
According to institutional risk minimization principle, formula (11) can be converted are as follows:
In formula: | | ω | |2The complexity of Controlling model, c are regularization parameter, and ε is insensitive coefficient, ξi,For relaxation
The factor.
Lagrange multiplier is introduced, converts convex double optimization problem for the above problem:
In formula,It is Lagrange multiplier, and
In order to accelerate solving speed, formula (13) is changed into dual form, then is had:
Using kernel function K (xi,xj) instead of the inner product of vectors in higher dimensional spaceTo avoid dimension disaster, then
The regression function of SVM are as follows:
Radial basis kernel function is selected to be defined as follows as the kernel function of SVM herein:
In formula: σ is the width parameter of Radial basis kernel function.
2.2.MGWO
2.2.1.GWO
GWO algorithm is the novel heuristic calculation proposed by the leader's hierarchy and predation of imitation grey wolf population
Method.Grey wolf is social animal, and there are more than ten grey wolves in usual each group, constructs stringent grey wolf pyramid hierarchy.
α wolf is located at pyramid top layer, is the head of wolf pack, and grade highest is mainly responsible for predation, decision etc., all wolf packs
Its commander must be accepted.
β wolf is located at the pyramid second layer, is mainly responsible for auxiliary α wolf and makes decision, status is only second to α wolf.β wolf is in wolf pack
Other individuals there is the right of eminent domain, while the trend of other wolves can be fed back to β wolf.
δ wolf is located at pyramidal third layer, is mainly responsible for the decision for executing α wolf and β wolf, and grade ratio ω wolf is high.
ω wolf is located at the pyramidal bottom, and quantity is more, and main function is to assist predation.
In GWO algorithm, behavior of chasing is executed by α, β, δ wolf, ω wolf follows former three to carry out prey tracking and encircles and suppresses, finally
Complete predation task.
The specific practice that GWO algorithm imitates grey wolf predation process is a group grey wolf to be first first randomly generated in search space, so
Afterwards by the position of α, β, δ wolf estimation prey, other wolves calculate separately the distance of itself and α, β, δ wolf to estimate itself and prey
Distance, and, encirclement close to prey, finally successfully bag the game.
The process of mathematical modeling of GWO algorithm is as follows: assuming that there is M grey wolf in a grey wolf population, search space is k dimension,
Then the position of i-th grey wolf is represented by xi=(xi1, xi2..., xik), the behavior that grey wolf gradually surrounds prey may be expressed as:
D=| cxp(t)-x(t)| (17)
X (t+1)=xp(t)-b×d (18)
Wherein, t represents current the number of iterations, position when x (t) is the t times iteration of grey wolf, xp(t) prey position is represented.
Vector b and d can be obtained by formula (19) and formula (20):
B=2ar1-a (19)
C=2r2 (20)
Wherein, r1And r2It is the random vector between [0,1].With the increase of the number of iterations, a is linearly reduced to 0 from 2.
It is assumed that α wolf, β wolf, δ wolf can then estimate the position of prey near prey by the position of this three in wolf pack
It sets.Grey wolf group updates process such as formula (21)-formula (27) of self-position according to the location information of α wolf, β wolf, δ wolf
dα=| c1·xα-x| (21)
dβ=| c2·xβ-x| (22)
dδ=| c3·xδ-x| (23)
x1=xα-b1×dα (24)
x2=xβ-b2×dβ (25)
x3=xδ-b3×dδ (26)
2.2.2 MGWO
For complicated optimization problem, GWO algorithm easily falls into locally optimal solution, for this problem, herein by introducing
Population dynamic evolutionary operator and non-linear convergence factor, improve GWO algorithm, avoid falling into local optimum.
Population dynamic evolutionary operator is introduced, it can be by the search range of wolf pack in GWO algorithm in each algorithm iteration process
In expand to entire solution space, thus increase its obtain globally optimal solution probability, the specific method is as follows:
In GWO algorithm, grey wolf group updates self-position according to the location information of α wolf, β wolf, δ wolf, at this point, we
More new formula is improved, is enabled:
xα=x1±(ub-lb·r+lb) (28)
xβ=x2±(ub-lb·r+lb) (29)
xδ=x3±(ub-lb·r+lb) (30)
In formula, ub and lb are respectively the upper bound and the lower bound of population search space, r be value range between [0,1] with
Machine number.
Updated potential optimal solution vector position are as follows:
Since the convergence factor a in GWO algorithm is the increase with the number of iterations from 2 linear decreases to 0, but algorithm exists
Constantly it is convergent during be not it is linear, convergence factor a linear decrease cannot embody actual Optimizing Search mistake completely
Therefore journey introduces non-linear convergence factor herein and improves to GWO algorithm, as shown in formula (32):
Wherein: t is current iteration number, tmaxIt is maximum number of iterations, e is natural constant.
Improved convergence factor a is in decreases in non-linear with the increase of the number of iterations, and initial stage successively decreases slowly, convenient for the overall situation
Search, later period successively decrease rapidly, strengthen local optimal searching.
2.3 MGWO-SVM
The value of regularization parameter c and Radial basis kernel function parameter g are pre- for the SVM algorithm using Radial basis kernel function
The precision for surveying model has direct influence, selects improved grey wolf optimization algorithm (MGWO) to carry out the parameter of SVM herein excellent
Change, on the basis of grey wolf algorithm, by introducing population dynamic evolutionary operator and non-linear convergence factor, reinforces grey wolf optimization
The ability of searching optimum of algorithm avoids falling into local optimum, and then improves the accuracy of SVM algorithm prediction.MGWO-SVM is specific
Steps are as follows:
Step1: the value of setting MGWO relevant parameter and SVM algorithm regularization parameter c and Radial basis kernel function parameter g
Range.
Step2: random initializtion grey wolf population, and enable the position vector of each grey wolf individual by c, g parameter is constituted.
Step3: learning training set using the SVM after initialization, calculates the fitness value of each grey wolf individual.
Step4: grade separation is carried out to grey wolf population according to fitness value, determines the position of α wolf, β wolf, δ wolf and ω wolf
It sets.
Step5: updating wolf pack position, generates new population, calculates corresponding fitness value, and the adaptation with last iteration
Angle value is compared, and is preferentially retained.
Step6: judging whether to reach maximum number of iterations, if reaching, training terminates, output c, the optimal value of g, otherwise,
It jumps to Step4 and continues parameter optimization.
Step7: using the parameter c after optimization, g establishes SVM prediction model, predicts test set.
3 prediction models based on CEEMDAN-MGWO-SVM
The prediction accuracy of day peak load will receive the influence of factors, in order to realize to the accurate of day peak load
Prediction proposes the prediction mould based on CEEMDAN-MGWO-SVM herein on the basis of considering meteorologic factor and date type
Type is used for the prediction of day peak load.The prediction steps of the model are as follows:
(1) data acquisition and pretreatment
Collecting sample data, including history day peak load, max. daily temperature, Daily minimum temperature, mean daily temperature, day are flat
The data such as equal relative humidity, day maximum wind velocity, date type.Then data prediction is carried out, meteorologic factor is normalized
It handles, festivals or holidays are indicated with 1 in date type, and work daily 0 expression.
(2) the sequence noise reduction based on CEEMDAN
The complete polymerization empirical mode decomposition that adaptive white noise is carried out to original day peak load sequence, obtains multiple
IMF component.
(3) the day peak load prediction based on MGWO-SVM
On the basis of considering meteorologic factor and date type, to the complete polymerization empirical modal by adaptive white noise
It decomposes obtained IMF component to be predicted respectively with MGWO-SVM model, and the result of prediction is reconstructed, obtain final
Day peak load prediction result.
The day peak load that this patent chooses S power grid in March, 2017 to May is research object, acquires day peak load, day
The data such as maximum temperature, Daily minimum temperature, mean daily temperature, per day relative humidity, day maximum wind velocity, date type, by March
To April data as training set sample, the data in May are as test set sample.The input of model includes: prediction day first three days
Day peak load, predict the last week day average day peak load, prediction max. daily temperature, prediction Daily minimum temperature, prediction
Mean daily temperature, the per day relative humidity of prediction, prediction day maximum wind velocity, prediction day date type, model output are prediction day
Peak load.
The complete polymerization empirical mode decomposition that adaptive white noise is carried out to original day peak load sequence, by S power grid
In March, 2017 to May day peak load as signal sequence input CEEMDAN model, obtain six IMF components and one be residual
Remaining signal, as shown in Figure 1；
Obtained IMF component and residue signal is predicted respectively with MGWO-SVM model, sets grey wolf population
Quantity is 20, and the search range of regularization parameter is [0.1,200], the search range of Radial basis kernel function parameter be [0.01,
20], maximum number of iterations 200, prediction result is as shown in Fig. 2 a~2g；
The prediction result of IMF component and residue signal is reconstructed, the day peak value for obtaining S power grid in May, 2017 is negative
Lotus prediction result, as shown in Figure 3；
From figure 3, it can be seen that predicted with day peak load of the CEEMDAN-MGWO-SVM model to S power grid May,
The fitting degree of prediction curve and actual curve is preferable, and prediction effect is significant.
In order to which more accurately the estimated performance of evaluation CEEMDAN-MGWO-SVM model, this patent use relative error
(RE), average absolute percentage error (MAPE) and nonlinear function approximation goodness (R2) etc. three indexs carry out the pre- of evaluation model
Survey precision.Shown in the calculating of index such as formula (1)-formula (3):
Be obtained by calculation CEEMDAN-MGWO-SVM model prediction result average absolute percentage error (MAPE) and
Nonlinear function approximation goodness (R2) it is respectively as follows: 0.196%, 99.77%, and CEEMDAN-MGWO-SVM model prediction result
Relative error is as shown in table 1 and Fig. 4；
1 prediction result relative error of table
According to index calculated result and above-mentioned chart, with CEEMDAN-MGWO-SVM model to S power grid May
Day, peak load was predicted that precision of prediction is very high, and the relative error of each future position is no more than 0.5%.
This embodiment is merely preferred embodiments of the present invention, but scope of protection of the present invention is not limited thereto,
In the technical scope disclosed by the present invention, any changes or substitutions that can be easily thought of by anyone skilled in the art,
It should be covered by the protection scope of the present invention.Therefore, protection scope of the present invention should be with scope of protection of the claims
Subject to.
Claims (4)
1. a kind of electric power day peak load prediction technique characterized by comprising
Step 1: acquisition includes history day peak load, max. daily temperature, Daily minimum temperature, mean daily temperature, per day opposite
Sample data including humidity, day maximum wind velocity, date type, is then normalized the meteorologic factor of data；
Step 2: the original series of day peak load are carried out with the complete polymerization empirical mode decomposition of adaptive white noise, it will be original
Sequence is decomposed into the intrinsic mode functions of limited local feature signal comprising different time scales, and in decomposing each time all
The smooth impulse disturbances of adaptive white noise are added, multiple IMF components are obtained；
Step 3: by introducing population dynamic evolutionary operator and non-linear convergence factor, grey wolf optimization algorithm being improved；
Step 4: using improved grey wolf optimization algorithm to the regularization parameter and Radial basis kernel function parameter of support vector machines
It optimizes, the SVM prediction model after establishing optimization；
Step 5: on the basis of considering meteorologic factor and date type, to the complete polymerization Empirical Mode by adaptive white noise
The IMF component that state is decomposed uses the SVM prediction model after optimization to be predicted respectively, and to the result of prediction
It is reconstructed, obtains final day peak load prediction result.
2. method according to claim 1, which is characterized in that the complete polymerization empirical mode decomposition of the adaptive white noise
Specific step is as follows:
Step 201: in n times calculating, to signal X (t)+pinj(t) it is decomposed, wherein X (t) is original series, nj(t) it is
White Gaussian noise, parameter piThe signal-to-noise ratio of the white Gaussian noise and sequence signal that are introduced for i-th, then first IMF componentAre as follows:
IMF1 j(t) after for the 1st introducing white Gaussian noise, obtained j-th of IMF component, the 1st remnants are decomposed by EMD
Signal r1(t) are as follows:
Step 202: defining emdi() is that i-th of IMF component after EMD decomposition is carried out to signal, to sequence r1(t)+p1emd1(nj
(t)) it is decomposed, obtains second IMF componentAre as follows:
2nd residue signal r2(t) are as follows:
Step 203: and so on, k-th of residue signal rk(t) are as follows:
+ 1 IMF component of kthAre as follows:
emdk() is that k-th of IMF component after EMD decomposition is carried out to signal；
Step 204: repeating the above process, until residue signal meets stopping criterion for iteration；It is if there is L IMF component at this time, then former
Beginning sequence is expressed as:
In formula, r (t) is final residue signal.
3. method according to claim 1, which is characterized in that the step 3 specifically includes:
In GWO algorithm, grey wolf group updates self-position according to the location information of α wolf, β wolf, δ wolf, at this point, we are to more
New formula improves, and enables:
xα=x1±(ub-lb·r+lb) (28)
xβ=x2±(ub-lb·r+lb) (29)
xδ=x3±(ub-lb·r+lb) (30)
In formula, ub and lb are respectively the upper bound and the lower bound of population search space, and r is random number of the value range between [0,1],
xα、xβ、xδThe respectively location information of α wolf, β wolf, δ wolf, x1、x2、x3The location information of respectively the 1st, 2,3 wolf；
Updated potential optimal solution vector position x (t+1) are as follows:
Non-linear convergence factor is introduced to improve grey wolf optimization algorithm, as shown in formula (32):
Wherein: t is current iteration number, tmaxMaximum number of iterations, e is natural constant, improved convergence factor a with
The increase of the number of iterations is in decreases in non-linear, and initial stage successively decreases slowly, is convenient for global search, and the later period successively decreases rapidly, strengthens part and seeks
It is excellent.
4. according to claim 1 or 2 or 3 the methods, which is characterized in that the step 4 specifically includes:
Step 401: the regularization parameter c and radial direction of improved grey wolf optimization algorithm relevant parameter and support vector machines are set
The value range of base kernel functional parameter g；
Step 402: random initializtion grey wolf population, and the position vector of each grey wolf individual is enabled to be made of parameter c, g；
Step 403: training set being learnt using the support vector machines after initialization, calculates the fitness of each grey wolf individual
Value；
Step 404: grade separation being carried out to grey wolf population according to fitness value, determines the position of α wolf, β wolf, δ wolf and ω wolf；
Step 405: updating wolf pack position, generate new population, calculate corresponding fitness value, and the fitness with last iteration
Value is compared, and is preferentially retained；
Step 406: judging whether to reach maximum number of iterations, if reaching, training terminates, and the optimal value of output parameter c, g are no
Then, it gos to step and 404 continues parameter optimization；
Step 407: establishing SVM prediction model using the parameter c, g after optimization, test set is predicted.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810811805.2A CN109242139A (en) | 2018-07-23 | 2018-07-23 | A kind of electric power day peak load prediction technique |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810811805.2A CN109242139A (en) | 2018-07-23 | 2018-07-23 | A kind of electric power day peak load prediction technique |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109242139A true CN109242139A (en) | 2019-01-18 |
Family
ID=65072901
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810811805.2A Pending CN109242139A (en) | 2018-07-23 | 2018-07-23 | A kind of electric power day peak load prediction technique |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109242139A (en) |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111428418A (en) * | 2020-02-28 | 2020-07-17 | 贵州大学 | Bearing fault diagnosis method and device, computer equipment and storage medium |
| CN111457927A (en) * | 2020-04-26 | 2020-07-28 | 北京工商大学 | Unmanned cruise ship multi-target path planning method under dynamic barrier |
| CN111625816A (en) * | 2020-04-21 | 2020-09-04 | 江西理工大学 | Intrusion detection method and device |
| CN111857876A (en) * | 2020-07-20 | 2020-10-30 | 北京字节跳动网络技术有限公司 | Service processing method and device, electronic equipment and computer readable medium |
| CN111882106A (en) * | 2020-06-16 | 2020-11-03 | 江苏大学 | Short-term power load prediction method based on comprehensive factors and CEEMD-IGWO-GRNN |
| CN111985692A (en) * | 2020-07-22 | 2020-11-24 | 河海大学 | CEEMDAN-based power load prediction method |
| CN112163710A (en) * | 2020-10-09 | 2021-01-01 | 武汉理工大学 | Short-term reactive load prediction method and device based on ensemble empirical mode |
| CN112541622A (en) * | 2020-12-01 | 2021-03-23 | 国网浙江省电力有限公司经济技术研究院 | Method and device for predicting short-term power consumption and electronic equipment |
| CN113239624A (en) * | 2021-05-21 | 2021-08-10 | 长沙理工大学 | Short-term load prediction method, equipment and medium based on neural network combination model |
| CN113496315A (en) * | 2021-09-08 | 2021-10-12 | 广东电网有限责任公司湛江供电局 | Load interval prediction method and system based on quantile gradient lifting decision tree |
| CN113627639A (en) * | 2020-05-07 | 2021-11-09 | 中国石油化工股份有限公司 | Well testing productivity prediction method and system for carbonate fracture-cave reservoir |
-
2018
- 2018-07-23 CN CN201810811805.2A patent/CN109242139A/en active Pending
Non-Patent Citations (1)
| Title |
|---|
| DAI, S. 等: ""Daily Peak Load Forecasting Based on Complete Ensemble Empirical Mode Decomposition with Adaptive Noise and Support Vector Machine Optimized by Modified Grey Wolf Optimization Algorithm"", 《ENERGIES》 * |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111428418A (en) * | 2020-02-28 | 2020-07-17 | 贵州大学 | Bearing fault diagnosis method and device, computer equipment and storage medium |
| CN111625816A (en) * | 2020-04-21 | 2020-09-04 | 江西理工大学 | Intrusion detection method and device |
| CN111457927A (en) * | 2020-04-26 | 2020-07-28 | 北京工商大学 | Unmanned cruise ship multi-target path planning method under dynamic barrier |
| CN113627639A (en) * | 2020-05-07 | 2021-11-09 | 中国石油化工股份有限公司 | Well testing productivity prediction method and system for carbonate fracture-cave reservoir |
| CN111882106A (en) * | 2020-06-16 | 2020-11-03 | 江苏大学 | Short-term power load prediction method based on comprehensive factors and CEEMD-IGWO-GRNN |
| CN111857876A (en) * | 2020-07-20 | 2020-10-30 | 北京字节跳动网络技术有限公司 | Service processing method and device, electronic equipment and computer readable medium |
| CN111985692A (en) * | 2020-07-22 | 2020-11-24 | 河海大学 | CEEMDAN-based power load prediction method |
| CN112163710A (en) * | 2020-10-09 | 2021-01-01 | 武汉理工大学 | Short-term reactive load prediction method and device based on ensemble empirical mode |
| CN112541622A (en) * | 2020-12-01 | 2021-03-23 | 国网浙江省电力有限公司经济技术研究院 | Method and device for predicting short-term power consumption and electronic equipment |
| CN113239624A (en) * | 2021-05-21 | 2021-08-10 | 长沙理工大学 | Short-term load prediction method, equipment and medium based on neural network combination model |
| CN113496315A (en) * | 2021-09-08 | 2021-10-12 | 广东电网有限责任公司湛江供电局 | Load interval prediction method and system based on quantile gradient lifting decision tree |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109242139A (en) | A kind of electric power day peak load prediction technique | |
| Lin et al. | An improved moth-flame optimization algorithm for support vector machine prediction of photovoltaic power generation | |
| Liang et al. | A novel wind speed prediction strategy based on Bi-LSTM, MOOFADA and transfer learning for centralized control centers | |
| CN113962364B (en) | Multi-factor power load prediction method based on deep learning | |
| Mahmoud et al. | An advanced approach for optimal wind power generation prediction intervals by using self-adaptive evolutionary extreme learning machine | |
| Hu et al. | Development and application of an evolutionary deep learning framework of LSTM based on improved grasshopper optimization algorithm for short-term load forecasting | |
| CN107622329A (en) | The Methods of electric load forecasting of Memory Neural Networks in short-term is grown based on Multiple Time Scales | |
| Peng et al. | Accelerating minibatch stochastic gradient descent using typicality sampling | |
| Bacanin et al. | Multivariate energy forecasting via metaheuristic tuned long-short term memory and gated recurrent unit neural networks | |
| CN109242212A (en) | A kind of wind-powered electricity generation prediction technique based on change Mode Decomposition and length memory network | |
| Niu et al. | Research on short-term power load time series forecasting model based on BP neural network | |
| CN109934422A (en) | Neural network wind speed prediction method based on time series data analysis | |
| Wang et al. | A novel decomposition-ensemble forecasting system for dynamic dispatching of smart grid with sub-model selection and intelligent optimization | |
| Rani R et al. | Training radial basis function networks for wind speed prediction using PSO enhanced differential search optimizer | |
| Tian et al. | An adaptive ensemble predictive strategy for multiple scale electrical energy usages forecasting | |
| CN110222908A (en) | A kind of short-term electricity demand forecasting method based on data-driven | |
| Akpinar et al. | Forecasting natural gas consumption with hybrid neural networks—Artificial bee colony | |
| CN109858665A (en) | Photovoltaic short term power prediction technique based on Feature Selection and ANFIS-PSO | |
| Niu et al. | Short-term wind speed hybrid forecasting model based on bias correcting study and its application | |
| Wang et al. | A power load forecasting model based on FA-CSSA-ELM | |
| Sen | Time Series Prediction based on Improved Deep Learning | |
| Lu et al. | An adaptive neural architecture search design for collaborative edge-cloud computing | |
| Jena et al. | Electricity consumption & prediction using machine learning models | |
| Eren et al. | A comprehensive review on deep learning approaches for short-term load forecasting | |
| Li et al. | Radial basis function neural network model for dissolved oxygen concentration prediction based on an enhanced clustering algorithm and Adam |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190118 |
|
| RJ01 | Rejection of invention patent application after publication |