CN105426989A - EEMD and combined kernel RVM-based photovoltaic power short-term prediction method - Google Patents

Abstract

The invention discloses an EEMD (Ensemble Empirical Mode Decomposition) and combined kernel RVM (Relevance Vector Machines)-based photovoltaic power short-term prediction method. Through the adopted EEMD, a problem that the possibility of mode mixing during empirical mode decomposition is high is avoided, so that higher resolution and very strong non-linear processing capacity are provided, the data complexity is reduced, the effect is excellent, the result is accurate, the prediction accuracy is effectively improved, and the method can be applied to the data preprocessing of the photovoltaic output power well; through the adoption of an RVM method for short-term photovoltaic power prediction, the method has the advantages of being sparse in model height, less in kernel parameters to be optimized, flexible in kernel function selection, and high in model generalization ability; and by use of a combined kernel function, the prediction accuracy of photovoltaic power through a model is further improved when the weather changes suddenly, so that the universal adaptability and the generalization performance of the model are improved.

Description

Based on the photovoltaic power short term prediction method of EEMD and combination core RVM

Technical field

The invention belongs to the technical field of generation of electricity by new energy and intelligent grid, be specifically related to a kind of photovoltaic power short term prediction method based on EEMD and combination core RVM.

Background technology

After 20 century 70s, along with industrialized development, fossil fuel faces exhaustion, and environmental problem also becomes increasingly conspicuous.In order to solve this difficult problem, the mankind start to pay close attention to regenerative resource, and wherein sun power becomes the focus that everybody pays close attention to.Reach 14GW to China's photovoltaic plant installed capacity end of the year in 2014, estimate that the year two thousand thirty photovoltaic installed capacity will reach 100-200GW.But photovoltaic generation is subject to the interference of many climatic factors, the safe and stable operation of electrical network when the disturbance of grid-connected rear power is serious, may be affected, therefore the forecasting research of photovoltaic output power just be seemed particularly necessary.

At present, the method for photovoltaic output power prediction roughly can be divided into two classes: a class is indirect predictions, and another kind of is direct prediction.Indirect predictions is the predicted value estimation photovoltaic output power utilizing solar radiation amount.Indirect predictions needs detailed weather data as support, but China only has 98 solar radiation observation websites at present, supports data less; Meanwhile, the accuracy of weather forecast is limited, causes the poor effect of indirect predictions; Direct prediction predicts the photovoltaic output power in following a period of time according to photovoltaic history output power data and weather effect factor, and demand data is relatively less, is the main stream approach of domestic photovoltaic power prediction, and what the present invention adopted is exactly directly prediction.

Consider the non-stationary of actual photovoltaic output power, directly larger to its error predicted.Improving one's methods of main flow is decomposition by raw data, reduce data complexity, wherein compare typical method and have wavelet analysis, EMD (EmpiricalModeDecomposition, empirical mode decomposition), LMD (LocalMeanDecomposition, local average decompose) etc.

But the information processing method such as wavelet analysis, EMD and LMD Problems existing is: artificially need set, subjectivity is strong, empirical mode decomposition easily occurs modal overlap.

Traditional load forecasting method has a lot, such as time series, ANN (ArtificialNeuralNetworks, artificial neural network), SVM (SupportVectorMachines, support vector machine) etc. obtained and used widely.But in actual application, time series is owing to putting aside that extraneous factor affects, and when larger change occurs external environment, predicated error is often larger; ANN method easily causes learning problem that is not enough or over-fitting in training;

Though the machine learning algorithms such as SVM can effectively avoid the risk being absorbed in Local Minimum, can realize predicting comparatively accurately, but still have the following disadvantages: 1. kernel function must meet Mercer condition, and optional kernel function is less; 2. parameter is more, and support vector along with the increase of training sample linear increase, calculated amount is larger; 3., when inputting influence factor and being more, forecast model structure will be caused too complicated, and training effectiveness is low.

Summary of the invention

Goal of the invention: the present invention is directed to the defect existed in prior art, there is provided a kind of based on EEMD (EnsembleEmpiricalModeDecomposition, set empirical mode decomposition) and combine core RVM (RelevanceVectorMachines, Method Using Relevance Vector Machine) photovoltaic power short term prediction method.

Technical scheme: a kind of photovoltaic power short term prediction method based on EEMD and combination core RVM, comprises the following steps:

S1: photovoltaic power data are divided into fine day, cloudy day, rainy day and broken sky 4 type according to weather conditions, and modeling respectively;

S2: adopt EEMD by the photovoltaic power data decomposition of non-stationary be a series of preliminary steadily and there is surplus and the IMF component of different characteristic yardstick;

S3: select the surplus of first 5 days of weather pattern identical with day to be predicted per moment photovoltaic powers and IMF component to construct sample input respectively, analyze and choose the historical data of influence factor of photovoltaic power prediction and predicted data inputs as a supplement, using the surplus of corresponding day to be predicted and IMF component data as output, construct training sample and forecast sample and carry out samples normalization;

The iterative initial value of S4: setting combination core RVM forecast model and model parameter hunting zone, in the process using training sample to train combination core RVM forecast model, adopt grid search Optimal Parameters, replaced by the iteration of training error and obtain the wide and combination core weight parameter of optimum core;

S5: the input of forecast sample is imported the combination core RVM model trained, model exports and is the photovoltaic power surplus of day to be predicted and predicting the outcome of IMF component;

S6: predicting the outcome of surplus and each IMF component is carried out superposition summation, obtains the predicted value of day to be predicted photovoltaic output power.

Preferably, described step S2 comprises following sub-step:

S2.1: amplitude k and the total degree M carrying out EMD decomposition of setting white noise;

S2.2: add white Gaussian noise in photovoltaic power data sequence;

S2.3: the data sequence obtained by step S2.2 according to EMD decomposition process is carried out decomposition and obtained a series of surplus and IMF component;

S2.4: repeat M step S2.2 to S2.3, the different white noise sequences for identical amplitude at every turn added during repeating said steps S2.2, decompose to M EMD each IMF and residual components computation of mean values of obtaining

${\stackrel{\‾}{c}}_i\left(t\right)=\underset{m=1}{\overset{M}{\Σ}}{c}_{i.m}\left(t\right)/M,{\stackrel{\‾}{r}}_n\left(t\right)=\underset{m=1}{\overset{M}{\Σ}}{r}_{n.m}\left(t\right)/M,$

Wherein, c _i,mt () is that the m time EMD decomposes i-th the IMF component obtained; r _n,mt () is that the m time EMD decomposes the n-th surplus obtained; T represents t data;

S2.5: export with the IMF component decomposed respectively as EEMD and residual components.

Preferably, the EMD decomposition process described in step S2.3 comprises following sub-step:

S2.3.1: loop initialization variable i=1, x ₁(t)=x (t), wherein x (t) is original data sequence to be decomposed;

S2.3.2: loop initialization variable j=1, y ₁(t)=x ₁(t);

S2.3.3: find out sequences y _jt all local maximums in () also fit to coenvelope line u _jt (), finds out y _jt all local minimums in () also fit to lower envelope line v _jt (), makes u _j(t) and v _jt data point that () envelope is all; Try to achieve u _j(t) and v _jthe mean value of (t) the difference h of original signal and envelope average _j(t)=y _j(t)-m _j(t);

S2.3.4: judge h _jt whether () meet two conditions of IMF component, do not meet, then j=j+1, y _j(t)=h _j-1t (), returns step S2.3.3; Meet, then can obtain i-th IMF component c _i(t)=h _j(t), residual components r _i(t)=x _i(t)-c _i(t);

S2.3.5: judge r _it whether () meet end condition, do not meet, then x _i+1(t)=r _i(t), i=i+1, repeats step S2.3.2 to S2.3.4; Meet, then decompose end; Decomposable asymmetric choice net goes out n IMF component c altogether thus _i(t) and a residual components r _n(t), the decomposable process of EMD to x (t) terminates; X (t) is expressed as

Preferably, the influence factor of the photovoltaic power prediction described in step S3 comprises temperature and light according to intensity.

Preferably, the structure sample input described in step S3 is specially:

The input vector of sample is: X (i, t)=[L (i-1, t), L (i-2, t), L (i-3, t), L (i-4, t), L (i-5, t), T (i, t), T (i-1, t), T (i-2, t), T (i-3, t), T (i-4, t), T (i-5, t), S (i, t), S (i-1, t), S (i-2, t), S (i-3, t), S (i-4, t), S (i-5, t)], i > 5, j=1 ..., M, output vector is y (i, t)=L (i, t);

Wherein, X (i, t) represents that i-th sample inputs the influence factor of t; Y (i, t) represents the output of i-th corresponding t of sample input; L (i-1, t) represents the photovoltaic data in prediction moment the previous day i-th day; L (i-2, t) represents the photovoltaic data in prediction a few days ago the i-th moment day; L (i-3, t) represents the photovoltaic data in prediction moment day first three day i-th; L (i-4, t) represents the photovoltaic data in four days a few days ago the i-th moment of prediction; L (i-5, t) represents the photovoltaic data in prediction moment day the first five day i-th; T (i, t) represents the temperature forecast data in prediction moment day i-th; T (i-1, t) represents the temperature data in prediction moment the previous day i-th day; T (i-2, t) represents the temperature data in prediction a few days ago the i-th moment day; T (i-3, t) represents the temperature data in prediction moment day first three day i-th; T (i-4, t) represents the temperature data predicting four days the i-th moment a few days ago; T (i-5, t) represents the temperature data in prediction moment day the first five day i-th; S (i, t) represents the illumination forecast data in prediction moment day i-th; S (i-1, t) represents the photometric data in prediction moment the previous day i-th day; S (i-2, t) represents the photometric data in prediction a few days ago the i-th moment day; S (i-3, t) represents the photometric data in prediction moment day first three day i-th; S (i-4, t) represents the photometric data predicting four days the i-th moment a few days ago; S (i-5, t) represents the photometric data in prediction moment day the first five day i-th.

Preferably, the samples normalization described in step S3 is specially:

Wherein, for the data value after normalization; X (i) is raw data; x _max, x _minbe respectively the maximal value in raw data and minimum value.

Preferably, described step S4 comprises following sub-step:

The iterative initial value of S4.1: setting combination core RVM forecast model and model parameter hunting zone;

The kernel function of S4.2: calculation combination core RVM model, the wide parameter of the core of grid search to model and combination core weight parameter is adopted to be optimized, whether the parameter that inspection current iteration obtains meets the demands, and meets, then this parameter is combination core RVM forecast model optimized parameter; Do not meet, then undated parameter, require until meet iteration or reach maximum iteration time.

Preferably, the kernel function of the calculation combination core RVM model described in step S4.2 adopts the combination of karyomerite-gaussian kernel and overall core-polynomial kernel, is specially: K (x, x _i)=wG (x, x _i)+(1-w) G (x, x _i), G (x, x _i)=exp (-‖ x-x _i‖ ²/ σ ²), P (x, x _i)=[(xx _i)+1] ²,

Wherein, K (x, x _i) represent model combination after kernel function; G (x, x _i) represent gaussian kernel; G (x, x _i) representative polynomial core; X represents associated vector; x _irepresent the input parameter after normalization; σ represents that the core of gaussian kernel is wide; W represents the kernel function weight coefficient of combination core.

Beneficial effect:

1, the EEMD that the present invention adopts overcomes the deficiency of wavelet analysis, EMD and LMD, avoid the problem that modal overlap easily appears in empirical mode decomposition, there is higher resolution and very strong Nonlinear Processing ability, well can apply to the data prediction of photovoltaic output power; EEMD reduces the complexity of data, excellent effect, and result is accurate, effectively improves precision of prediction;

2, the present invention adopts and carries out the prediction of short-term photovoltaic power based on the RVM method of bayesian theory, not only well remain the outstanding predictive ability of SVM, further improve the weak point of ANN, SVM, there is the advantage that sparse, the to be optimized parameter of model height is few, Selection of kernel function is flexible, generalization ability is strong, precision of prediction is high;

3, use compound kernel function to improve the precision of model to photovoltaic power prediction under sudden change weather further, improve general adaptability and the Generalization Capability of model.

Accompanying drawing explanation

Fig. 1 is method flow diagram of the present invention;

Fig. 2 is set empirical mode decomposition flow chart of steps;

Fig. 3 is set empirical mode decomposition effect schematic diagram;

Fig. 4 is fine day photovoltaic output power short-term forecasting result;

Fig. 5 is broken sky photovoltaic output power short-term forecasting result;

Fig. 6 is cloudy photovoltaic output power short-term forecasting result;

Fig. 7 is rainy day photovoltaic output power short-term forecasting result;

Embodiment

Below in conjunction with accompanying drawing and instantiation, technical scheme of the present invention is described in detail.It should be noted that, this example is only not used in for illustration of the present invention and limits the scope of the invention, and the amendment of those skilled in the art to the various equivalent form of value of the present invention all falls within the application's claims limited range.

As shown in Figure 1, for the precision of prediction influential system sacurity dispatching of photovoltaic output power and the problem of stable operation, the present invention proposes a kind of photovoltaic power short term prediction method based on EEMD and combination core RVM.First according to weather conditions, photovoltaic power data are divided into fine day, cloudy day, rainy day and broken sky Four types, and modeling respectively, ensure that the consistance of sample data; And adopt EEMD by above four kinds of photovoltaic power data decomposition be a series of preliminary steadily but there is IMF (IntrinsicModeFunction, intrinsic mode function) component and the surplus of different characteristic yardstick, reduce the complexity of data; Then adopt the combination of gaussian kernel and polynomial kernel to form the kernel function of RVM, adopt combination core RVM to set up forecast model respectively to surplus and each IMF component, improve the estimated performance of forecast model; Finally predicting the outcome of surplus and IMF component is carried out superposition summation and obtained overall predicting the outcome.The inventive method processes from data prediction and model optimization two aspects, effectively raises precision of prediction and the generalization ability of model, can apply to engineering problem preferably.

First, consider that the sunrise in each season, sunset time are different, for ensureing that data all have value, only get every day 8:00-17:00 the data of totally 10 hours carry out training and predicting.Secondly, there is larger difference in the intensity of illumination under different weather type, photovoltaic output power also exists significant difference.Therefore in order to predict photovoltaic power more accurately, photovoltaic power data are divided into these four kinds of situations of fine day, cloudy day, broken sky and rainy day according to weather pattern, and modeling respectively.

Moreover for avoiding accidental data or bad data on the impact predicted the outcome, carry out pre-service to photovoltaic power data, use the longitudinal pairing comparision of data in the present invention, its computing formula is as follows:

If meet $\left\{\begin{array}{c}|L(i,t)-L(i-1,t)|>\mathrm{\α}\left(t\right)\\ |L(i,t)-L(i+1,t)|>\mathrm{\β}\left(t\right)\end{array}\right.,$ Then: $L(i,t)=\frac{L(i-1,t)+L(i+1,t)}{2}$

Wherein, L (i, t) is the photovoltaic power data of i-th day t; L (i-1, t) is the photovoltaic power data of the i-th-1 day t, namely L (i, t) the previous day photovoltaic power data in the same time; L (i+1, t) is the photovoltaic power data of the i-th+1 day t, i.e. L (i, t) photovoltaic power data in the same time one day after; α (t) and two threshold values of β (t) for presetting.

Carry out EEMD to the original photovoltaic power data completing bad data rejecting, concrete steps are as follows:

1. the amplitude k setting white Gaussian noise and the total degree M carrying out EMD decomposition (general M get 100, k get 0.05 ~ 0.5 times be advisable), make m=1;

2. given photovoltaic hour level power data L (t), calculates number of data points x, and calculates the number N=log of component accordingly ₂x-1; Again by component by high frequency to low frequency sequence number be decided to be i (wherein i=1,2 ..., N).

When 3. carrying out the m time EMD decomposition, photovoltaic power data L (t) adds white Gaussian noise n _mt (), this white Gaussian noise obtains random white Gaussian noise by the randn function in MATLAB, and the follow-up noise amplitude added when repeatedly repeating this step is the identical amplitude k set in advance.Obtain pending adding din-light voltage certificate:

L _m(t)＝L(t)+kn _m(t)

4. to L _mt () carries out EMD decomposition, obtain N number of IMF component c _i,m(t) and a residual components r _n,m(t);

If 5. 3. m < M, m=m+1, return, otherwise continue to perform downwards;

6. each IMF component of obtaining is decomposed to M EMD and residual components computation of mean values is:

${\mathrm{IMF}}_i={\stackrel{\&OverBar;}{c}}_i\left(t\right)=\underset{m=1}{\overset{M}{\&Sigma;}}{c}_{i.m}\left(t\right)/M,R_n={\stackrel{\&OverBar;}{r}}_n\left(t\right)=\underset{m=1}{\overset{M}{\&Sigma;}}{r}_{n.m}\left(t\right)/M,i=1,...,N$

7. export EEMD and decompose the N number of IMF component and residue surplus R that obtain _n, decompose and terminate.The step block diagram Sum decomposition effect of EEMD respectively as shown in Figure 2 and Figure 3.

Wherein, described EMD decomposes and comprises following sub-step:

(1) loop initialization variable i=1, x ₁(t)=x (t), wherein x (t) is original data sequence to be decomposed;

(2) loop initialization variable j=1, y ₁(t)=x ₁(t);

(4) sequences y is found out _jt all local maximums in () also fit to coenvelope line u _jt (), finds out y _jt all local minimums in () also fit to lower envelope line v _jt (), makes u _j(t) and v _jt data point that () envelope is all; Try to achieve u _j(t) and v _jthe mean value of (t) the difference h of original signal and envelope average _j(t)=y _j(t)-m _j(t);

(5) h is judged _jt whether () meet two conditions of IMF component, do not meet, then j=j+1, y _j(t)=h _j-1t (), returns step S2.3.3; Meet, then can obtain i-th IMF component c _i(t)=h _j(t), residual components r _i(t)=x _i(t)-c _i(t);

(6) r is judged _it whether () meet end condition, do not meet, then x _i+1(t)=r _i(t), i=i+1, repeats step S2.3.2 to S2.3.4; Meet, then decompose end; Decomposable asymmetric choice net goes out n IMF component c altogether thus _i(t) and a residual components r _n(t), the decomposable process of EMD to x (t) terminates; X (t) is expressed as

Select day to construct sample input with weather pattern first 5 days per moment photovoltaic powers with prediction, same procedure is adopted to process to surplus and each IMF component, analyze and choose the historical data of influence factor (temperature, intensity of illumination etc.) and the predicted data input variable as a supplement of photovoltaic power prediction, using the corresponding prediction surplus of day and IMF component data as output, construct training sample and forecast sample and normalization; The formation of sample input influence factor is as shown in table 1.

Table 1 sample input influence factor is formed

The input vector of sample is expressed as thus:

X(i,t)＝[L(i-1,t),L(i-2,t),L(i-3,t),L(i-4,t),L(i-5,t),T(i,t),T(i-1,t),

T (i-2, t), T (i-3, t), T (i-4, t), T (i-5, t), S (i, t), S (i-1, t), S (i-2, t), S (i-3, t), S (i-4, t), S (i-5, t)], i > 5, j=1 ... M, output vector is y (i, t)=L (i, t); Normalization formula is:

$\tilde{x}\left(i\right)=\left(x\right(i)-x_{min})/(x_{max}-x_{min})$

Combination core RVM model is adopted to set up forecast model respectively to the component that EEMD obtains.

For given training sample input set with the output collection of correspondence rVM regression model may be defined as:

$t_i=\underset{i=1}{\overset{N}{\&Sigma;}}{w}_{i}K(x,x_i)+w_0+\mathrm{\&epsiv;}$

Wherein ε is for obeying N (0, σ ²) each independent sample error of distributing, w _ifor weight coefficient, K (x, x _i) be kernel function, N is sample size.

For separate output collection, the likelihood function of whole sample is:

$p\left(t\right|w,{\mathrm{\&sigma;}}^2)=\underset{i=1}{\overset{N}{\&Pi;}}N\left(t_i\right|y\left(x_i;w\right),{\mathrm{\&sigma;}}^2)={\left(2{\mathrm{\&pi;\&sigma;}}^2\right)}^{-N/2}\exp (-||t-\mathrm{\&Phi;}(x\left)w\right|{|}^2/2{\mathrm{\&sigma;}}^2$

Wherein: t=(t ₁, t ₂..., t _n), w=[w ₀, w ₁..., w _n] ^t,

According to probabilistic forecasting formula, required conditional probability is:

$P\left(t_{*}\right|t)=\&Integral;p\left(t_{*}\right|w,{\mathrm{\&sigma;}}^2)p(w,{\mathrm{\&sigma;}}^2|t){\mathrm{dwd\&sigma;}}^2$

If directly use the method for maximum likelihood to solve w and σ ², result can cause serious crossing to adapt to usually, for avoiding this phenomenon, adds condition precedent to w.According to bayesian theory, w be distributed as zero standardized normal distribution, introduce hyper parameter α=[α simultaneously ₀, α ₁, α ₂..., α _n] ^t, can obtain

$p\left(w\right|\mathrm{\&alpha;})=\underset{i=0}{\overset{N}{\&Pi;}}N\left(w_i\right|0,{\mathrm{\&alpha;}}_i^{-1})$

Therefore, probabilistic forecasting formula changes into:

$p\left(t_{*}\right|t)=\&Integral;p\left(t_{*}\right|w,{\mathrm{\&sigma;}}^2)p(w,\mathrm{\&alpha;},{\mathrm{\&sigma;}}^2|t){\mathrm{dwd\&alpha;d\&sigma;}}^2$

Each weights being limited to the method for condition precedent, is a key character of RVM.α is hyper parameter corresponding to weight w, meets gamma distribution.After enough update times, most of α _imeeting convergence is infinitely great, and the weights of its correspondence are tending towards 0, and other α _imeeting stably convergence finite value.And x corresponding with it _ibe referred to as associated vector, realize RVM sparse characteristic.

After defining prior probability distribution and likelihood distribution, according to Bayes principle, can be in the hope of the Posterior probability distribution of all unknown parameters just:

$p\left(w\right|t,\mathrm{\&alpha;},{\mathrm{\&sigma;}}^2)={\left(2\mathrm{\&pi;}\right)}^{-(N+1)/2}{\left|\mathrm{\&Psi;}\right|}^{-1/2}\&CenterDot;\exp \{-\frac{1}{2}{(w-\mathrm{\&mu;})}^T{\mathrm{\&Psi;}}^T(w-\mathrm{\&mu;})\}$

Wherein, posteriority covariance matrix is:

Ψ＝(σ ^-2Φ ^TΦ+A) ^-1

μ＝σ ^-2ΨΦ ^Tt

A＝diag(α ₀,α ₁,…,α _N)

In order to Confirming model weights, first need the optimum value obtaining hyper parameter, can be tried to achieve by iterative algorithm, namely

${{\mathrm{\&alpha;}}_i}^{new}=\frac{1-{\mathrm{\&alpha;}}_i{\mathrm{\&Psi;}}_{i,i}}{{\mathrm{\&mu;}}_i^2}$

${\left({\mathrm{\&sigma;}}^2\right)}^{new}=\frac{\left|\right|t-\mathrm{\&Phi;}\mathrm{\&mu;}|{|}^2}{N-\underset{i=0}{\overset{N}{\&Sigma;}}(1-{\mathrm{\&alpha;}}_i{\mathrm{\&Psi;}}_{i,i})}$

In formula: μ _ibe that i-th posteriority is on average weighed, Ψ _i,ifor the diagonal element of i-th in posteriority covariance matrix, N is sample data number.

If given new input value x ^*, then corresponding output probability distribution Gaussian distributed, its corresponding predicted value is:

For the quality of evaluation and foreca effect, relative error (relativeerror, RE) and these 2 error criterions of average absolute value percentage error (meanabsolutepercenterror, MAPE) are adopted to carry out error analysis to predicting the outcome.Its value is less, and effect is better.The expression formula of index is:

$RE=\frac{\left|y_{true}-y_{pre}\right|}{y_{true}}\&times;100\%$

$MAPE=\frac{1}{N}\underset{i=1}{\overset{N}{\&Sigma;}}\frac{\left|y_{true}-y_{pre}\right|}{y_{true}}\&times;100\%$

In formula, y _truerepresent actual value, y _prerepresent predicted value, N represents data count.

Forecast model block diagram of the present invention as indicated with 2.

In order to prediction effect of the present invention is described, respectively short-term forecasting has been carried out to the photovoltaic output power under four kinds of weather patterns, its prediction effect as Figure 4-Figure 7, predict the outcome and index analysis as shown in table 2-table 5.

Table 2 fine day predicated error is added up

Table 3 broken sky predicated error statistics

The cloudy predicated error statistics of table 2

Table 2 rainy day predicated error statistics

Can obviously be found by the index analysis shown in the prediction effect shown in Fig. 4-Fig. 7 and table 2-table 5, the trend that the inventive method predicts the outcome is closing to reality photovoltaic power more, and predicated error is obviously less.Therefore, what the photovoltaic power short term prediction method based on EEMD and combination core RVM that the present invention adopts can be higher realizes the prediction of short-term photovoltaic power, has higher engineer applied and is worth.

Claims (8)

1., based on a photovoltaic power short term prediction method of EEMD and combination core RVM, it is characterized in that, comprise the following steps:

S1: photovoltaic power data are divided into fine day, cloudy day, rainy day and broken sky 4 type according to weather conditions, and modeling respectively;

S2: adopt EEMD by the photovoltaic power data decomposition of non-stationary be a series of preliminary steadily and there is surplus and the IMF component of different characteristic yardstick;

S3: select the surplus of first 5 days of weather pattern identical with day to be predicted per moment photovoltaic powers and IMF component to construct sample input respectively, analyze and choose the historical data of influence factor of photovoltaic power prediction and predicted data inputs as a supplement, using the surplus of corresponding day to be predicted and IMF component data as output, construct training sample and forecast sample and carry out samples normalization;

The iterative initial value of S4: setting combination core RVM forecast model and model parameter hunting zone, in the process using training sample to train combination core RVM forecast model, adopt grid search Optimal Parameters, replaced by the iteration of training error and obtain the wide and combination core weight parameter of optimum core;

S5: the input of forecast sample is imported the combination core RVM model trained, model exports and is the photovoltaic power surplus of day to be predicted and predicting the outcome of IMF component;

S6: predicting the outcome of surplus and each IMF component is carried out superposition summation, obtains the predicted value of day to be predicted photovoltaic output power.

2. the photovoltaic power short term prediction method based on EEMD and combination core RVM according to claim 1, is characterized in that: described step S2 comprises following sub-step:

S2.1: amplitude k and the total degree M carrying out EMD decomposition of setting white noise;

S2.2: add white Gaussian noise in photovoltaic power data sequence;

S2.3: the data sequence obtained by step S2.2 according to EMD decomposition process is carried out decomposition and obtained a series of surplus and IMF component;

S2.4: repeat M step S2.2 to S2.3, the different Gaussian sequences for identical amplitude at every turn added during repeating said steps S2.2, decompose to M EMD each IMF and residual components computation of mean values of obtaining

${\stackrel{\&OverBar;}{c}}_i\left(t\right)=\underset{m=1}{\overset{M}{\&Sigma;}}{c}_{i.m}\left(t\right)/M,{\stackrel{\&OverBar;}{r}}_n\left(t\right)=\underset{m=1}{\overset{M}{\&Sigma;}}{r}_{n.m}\left(t\right)/M,$

Wherein, c _i,mt () is that the m time EMD decomposes i-th the IMF component obtained; r _n,mt () is that the m time EMD decomposes the n-th surplus obtained; T represents t data;

S2.5: export with the IMF component decomposed respectively as EEMD and residual components.

3. the photovoltaic power short term prediction method based on EEMD and combination core RVM according to claim 2, is characterized in that: the EMD decomposition process described in step S2.3 comprises following sub-step:

S2.3.1: loop initialization variable i=1, x ₁(t)=x (t), wherein x (t) is original data sequence to be decomposed;

S2.3.2: loop initialization variable j=1, y ₁(t)=x ₁(t);

S2.3.3: find out sequences y _jt all local maximums in () also fit to coenvelope line u _jt (), finds out y _jt all local minimums in () also fit to lower envelope line v _jt (), makes u _j(t) and v _jt data point that () envelope is all; Try to achieve u _j(t) and v _jthe mean value of (t) the difference h of original signal and envelope average _j(t)=y _j(t)-m _j(t);

S2.3.4: judge h _jt whether () meet two conditions of IMF component, do not meet, then j=j+1, y _j(t)=h _j-1t (), returns step S2.3.3; Meet, then can obtain i-th IMF component c _i(t)=h _j(t), residual components r _i(t)=x _i(t)-c _i(t);

S2.3.5: judge r _it whether () meet end condition, do not meet, then x _i+1(t)=r _i(t), i=i+1, repeats step S2.3.2 to S2.3.4; Meet, then decompose end; Decomposable asymmetric choice net goes out n IMF component c altogether thus _i(t) and a residual components r _n(t), the decomposable process of EMD to x (t) terminates; X (t) is expressed as

4. the photovoltaic power short term prediction method based on EEMD and combination core RVM according to claim 1, is characterized in that: the influence factor of the photovoltaic power prediction described in step S3 comprises temperature and light according to intensity.

5. the photovoltaic power short term prediction method based on EEMD and combination core RVM according to claim 1, is characterized in that: the structure sample input described in step S3 is specially:

The input vector of sample is: X (i, t)=[L (i-1, t), L (i-2, t), L (i-3, t), L (i-4, t), L (i-5, t), T (i, t), T (i-1, t), T (i-2, t), T (i-3, t), T (i-4, t), T (i-5, t), S (i, t), S (i-1, t), S (i-2, t), S (i-3, t), S (i-4, t), S (i-5, t)], i > 5, j=1 ..., M, output vector is y (i, t)=L (i, t);

Wherein, X (i, t) represents that i-th sample inputs the influence factor of t; Y (i, t) represents the output of i-th corresponding t of sample input; L (i-1, t) represents the photovoltaic data in prediction moment the previous day i-th day; L (i-2, t) represents the photovoltaic data in prediction a few days ago the i-th moment day; L (i-3, t) represents the photovoltaic data in prediction moment day first three day i-th; L (i-4, t) represents the photovoltaic data in four days a few days ago the i-th moment of prediction; L (i-5, t) represents the photovoltaic data in prediction moment day the first five day i-th; T (i, t) represents the temperature forecast data in prediction moment day i-th; T (i-1, t) represents the temperature data in prediction moment the previous day i-th day; T (i-2, t) represents the temperature data in prediction a few days ago the i-th moment day; T (i-3, t) represents the temperature data in prediction moment day first three day i-th; T (i-4, t) represents the temperature data predicting four days the i-th moment a few days ago; T (i-5, t) represents the temperature data in prediction moment day the first five day i-th; S (i, t) represents the illumination forecast data in prediction moment day i-th; S (i-1, t) represents the photometric data in prediction moment the previous day i-th day; S (i-2, t) represents the photometric data in prediction a few days ago the i-th moment day; S (i-3, t) represents the photometric data in prediction moment day first three day i-th; S (i-4, t) represents the photometric data predicting four days the i-th moment a few days ago; S (i-5, t) represents the photometric data in prediction moment day the first five day i-th.

6. the photovoltaic power short term prediction method based on EEMD and combination core RVM according to claim 1, is characterized in that: the samples normalization described in step S3 is specially:

Wherein, for the data value after normalization; X (i) is raw data; x _max, x _minbe respectively the maximal value in raw data and minimum value.

7. the photovoltaic power short term prediction method based on EEMD and combination core RVM according to claim 1, is characterized in that: described step S4 comprises following sub-step:

The iterative initial value of S4.1: setting combination core RVM forecast model and model parameter hunting zone;

The kernel function of S4.2: calculation combination core RVM model, the wide parameter of the core of grid search to model and combination core weight parameter is adopted to be optimized, whether the parameter that inspection current iteration obtains meets the demands, and meets, then this parameter is combination core RVM forecast model optimized parameter; Do not meet, then undated parameter, require until meet iteration or reach maximum iteration time.

8. the photovoltaic power short term prediction method based on EEMD and combination core RVM according to claim 7, it is characterized in that: the kernel function of the calculation combination core RVM model described in step S4.2 adopts the combination of karyomerite-gaussian kernel and overall core-polynomial kernel, be specially: K (x, x _i)=wG (x, x _i)+(1-w) G (x, x _i), G (x, x _i)=exp (-‖ x-x _i‖ ²/ σ ²), P (x, x _i)=[(xx _i)+1] ²,

Wherein, K (x, x _i) represent model combination after kernel function; G (x, x _i) represent gaussian kernel; G (x, x _i) representative polynomial core; X represents associated vector; x _irepresent the input parameter after normalization; σ represents that the core of gaussian kernel is wide; _wrepresent the kernel function weight coefficient of combination core.