CN103136598A - Monthly electrical load computer forecasting method based on wavelet analysis - Google Patents

Abstract

The invention discloses a monthly power load prediction method based on wavelet analysis, which includes an input module, a wavelet transformation module, a first-type prediction module, a second-type prediction module, a wavelet reconstruction module and an output module. The input module receives the historical load data c ₀ and transmits the historical data c ₀ to the wavelet transform module; the wavelet transform module decomposes the historical data into five components of d ₁ , d ₂ , d ₃ , d ₄ and c ₄ , where d ₁ - d ₄ has the characteristics of annual periodic changes, and c ₄ has a growing trend; the first type of forecasting module predicts the four components of d ₁ - d ₄ ; the second type of forecasting module predicts c ₄ The component is predicted; the wavelet reconstruction module performs wavelet reconstruction on each predicted component to obtain the predicted load c ₀ '; finally, the predicted load is presented through the output module.

Description

Monthly calculation of power load machine Forecasting Methodology based on wavelet analysis

Technical field

The present invention relates to a kind of Methods of electric load forecasting, particularly a kind of monthly calculation of power load machine Forecasting Methodology based on wavelet analysis.

Background technology

Load prediction is the basis of power system development planning, fuel planning and formulation generation schedule.Monthly prediction is take the moon as prediction period, with monthly power consumption as predictive content.By moon load prediction technology, operational plan in mid-term that can the arranged rational electric system reduces operating cost, improves power supply reliability.

Moon load Changing Pattern in each year has annual periodicity, when carrying out moon load prediction, must take full advantage of up-to-date monthly data considering on moon basis of load annual periodicity, embody significantly with the predicted impact of up-to-date Changing Pattern to future.Because the periodicity of monthly power consumption Changing Pattern in each year is similar to the characteristics of Day Load Curve Forecasting to short-term load forecasting, therefore prior art adopts the method that is similar to short-term load forecasting mostly, comprises setting up autoregression-moving average (ARMA) model and artificial neural network method (ANN) etc.

Although said method is widely used, and obtained prediction effect preferably in the technical field of monthly load prediction.Yet above-mentioned several method is only sensitiveer in a short time variation characteristic, to long-term variation characteristic, as the growth feature of exponential type, can not reflect well.

In addition, actual monthly load curve except having the fluctuation characteristic with seasonal variations, along with the raising of socioeconomic development and living standards of the people, also have the variation characteristic of increase year after year.Therefore, adopt the method for regression model and artificial neural network can reflect load curve seasonal fluctuations feature, but can not reflect preferably load curve rising tendency year by year.

In order to reflect simultaneously two kinds of variation characteristics of monthly load curve, need to adopt certain method with the moon load signal curve decompose, then adopt respectively different Forecasting Methodologies to predict to the component of different characteristic.

Summary of the invention

The object of the invention is to according to the deficiencies in the prior art part provide a kind of can effecting reaction for a long time, variation characteristic year by year, more comprehensively based on the monthly calculation of power load machine Forecasting Methodology of wavelet analysis.

The objective of the invention is to realize by following approach:

Based on the monthly calculation of power load machine Forecasting Methodology of wavelet analysis, its main points are, comprise the steps:

The load forecast module is provided, and it comprises load module, wavelet transformation module, first kind prediction module, Equations of The Second Kind prediction module, wavelet reconstruction module and output module;

Load module receives the historical load data c from electric system ₀, and with historical data c ₀Send the wavelet transformation module to, the wavelet transformation module is decomposed historical data, c ₀Resolve into high fdrequency component d ₁With low frequency component c ₁, then with c ₁Further resolve into d ₂And c ₂, c ₂Further resolve into d ₃And c ₃, c ₃Further resolve into d ₄And c ₄Thereby, obtain the d that decomposes ₁, d ₂, d ₃, d ₄, c ₄Five components, wherein d ₁-d ₄Has annual periodically variable feature, c ₄Has ever-increasing variation tendency; Need to carry out curve down-sampledly when adopting wavelet transformation, often carry out wavelet transformation one time, the sampled point of horizontal ordinate will reduce half;

With d ₁, d ₂, d ₃, d ₄Component data send into first kind prediction module, first kind prediction module adopts the data method of the cyclic fluctuation feature can reflect data to d ₁, d ₂, d ₃, d ₄These four components are predicted, obtain d ₁', d ₂', d ₃', d ₄' four transform component; With c ₄Data component is sent into the Equations of The Second Kind prediction module, and the Equations of The Second Kind prediction module can reflect that the method for feature of data sustainable growth is to c ₄Component is predicted, obtains transform component c ₄', need to carry out rising sampling to curve when adopting wavelet reconstruction, often carry out wavelet reconstruction one time, the sampled point of horizontal ordinate will increase half.

Send above-mentioned transform component to the wavelet reconstruction module, adopt the Mallat restructing algorithm, at first with component d ₄' and c ₄' be reconstructed and obtain c ₃', then by that analogy, with d ₃' and c ₃' further reconstruct obtains c ₂', d ₂' and c ₂' further reconstruct obtains c ₁', d ₁' and c ₁' further reconstruct obtains prediction load c ₀';

With the prediction load c that obtains ₀' store, and export by output module.

Wavelet transformation theory is proposed by Jean Morlet, has a wide range of applications in the signal analysis field.Wavelet transformation can resolve into original signal the component of different scale, thereby can feature and the variation tendency of reflected signal in different frequency domains.Than Fourier transform, the signal that wavelet transformation more can the Adaptive change fierceness is stronger to the descriptive power of signal.

The continuous wavelet transform of signal f (t) can be expressed as:

$W(a,b)=\frac{1}{\sqrt{a}}\∫f\left(t\right)\mathrm{\ψ}\left(\frac{t-b}{a}\right)\mathrm{dt}$

Wherein a is scale parameter, and b is displacement parameter.

Be reconstructed by the component with different scale, reducible signal, the formula of wavelet inverse transformation can be expressed as:

$f\left(t\right)=\frac{1}{C_{\mathrm{\ψ}}}{\∫}_0^{+\∞}{\∫}_{-\∞}^{+\∞}\frac{W(a,b)}{\sqrt{a}{a}^2}\mathrm{\ψ}\left(\frac{t-b}{a}\right)\mathrm{dadb}$

C wherein _ψBe enabled condition.

The people such as Mallat adopt the method for wave filter to realize wavelet transform, and concrete computation process is:

$c_{j+1}\left(m\right)=\underset{m}{\mathrm{\Σ}}h(m-2k)c_j\left(m\right)$

$d_{j+1}\left(m\right)=\underset{m}{\mathrm{\Σ}}g(m-2k)c_j\left(m\right)$

Wherein j is decomposition scale, and k, m are translation coefficient, c _j+1(m) and d _j+1(m) be respectively HFS and the low frequency part of wavelet coefficient, h (m-2k) and g (m-2k) are respectively Hi-pass filter and low-pass filter.By adopting the discrete wavelet analysis, can be with original signal c ₀Resolve into HFS d ₁With low frequency part c ₁, and then with c ₁Do further to decompose, resolve into d ₂And c ₂, by that analogy.By signal decomposition being become the component of different frequency, can reflected signal the feature of details more, thereby can better analyze signal.

Down-sampled owing to need to curve being carried out when adopting wavelet transformation, the yardstick after therefore decomposing each time is all different, so the sampled point of horizontal ordinate will reduce half.Find in experiment, after having passed through four layers of wavelet decomposition, low frequency component c ₄Substantially, aperiodicity fluctuation tendency.This shows that four layers of wavelet transformation can separate the cyclical component of load curve and the component of sustainable growth: wavelet coefficient d ₁-d ₄With scale coefficient c ₄Represented the basic variation tendency of load curve, wherein component d ₁-d ₄Have very strong undulatory property, corresponding each year loads with the variation tendency in season, component c ₄Fluctuate less, and have the trend that rises appreciably, corresponding is the characteristic of year load growth.Therefore can be well two kinds of variation characteristics (with the fluctuation characteristic of seasonal variations and the variation characteristic of increase year after year) of monthly load curve be reflected simultaneously by wavelet transformation, monthly electric load has been carried out comprehensive prediction, for formulating operation of power networks plan in mid-term, reduce operating cost, improving power supply reliability provides data basic.

The present invention can further be specially:

The data method of the described cyclic fluctuation feature that can reflect data or for time series method or be neural network.

First kind prediction module is mainly to process calculating for four strong components of undulatory property, therefore needs the data processing method of the cyclic fluctuation feature of employing data to realize.Described time series method can be AR model, MA model, arma modeling etc.Said method is also non exhaustive, and all can reflect that the data processing method of the cyclic fluctuation feature of data all can adopt.

Describedly can reflect that the method for data sustainable growth feature is trend extrapolation.

The Equations of The Second Kind prediction module is mainly for c ₄This stable cycle data of component is processed calculating, therefore need to adopt the data processing method that can reflect data sustainable growth feature, described trend extrapolation can be the method without inclined to one side gray prediction, gray prediction method is a kind of of trend extrapolation, the long term variations of its energy fitting data, and data are predicted.

The present invention has following beneficial effect:

1. this Forecasting Methodology by the method for wavelet analysis, resolves into former load data the component of different Changing Patterns.Can adopt different Forecasting Methodologies to predict for different components, can carry out more exactly a moon load prediction.

2. this Forecasting Methodology can reflect the cyclical variation feature of monthly load data, and this is mainly to be caused by the cyclical variation of power load throughout the year.

3. this Forecasting Methodology can reflect the variation characteristic of monthly load data sustainable growth, and this is mainly the raising due to socioeconomic development and living standards of the people, and causes power load to increase year by year.

Description of drawings

Fig. 1 is the general procedure process flow diagram of the monthly calculation of power load machine Forecasting Methodology based on wavelet analysis of the present invention.

Fig. 2 is the curve map of each variable in Fig. 1.

Fig. 3 is wavelet transformation process flow diagram of the present invention.

Fig. 4 is wavelet reconstruction process flow diagram of the present invention.

Fig. 5 is RBF neural network structure figure of the present invention.

Embodiment

See figures.1.and.2, the monthly Methods of electric load forecasting based on wavelet analysis of the present invention comprises load module 1, wavelet transformation module 2, first kind prediction module 3,4,5,6, Equations of The Second Kind prediction module 7, wavelet reconstruction module 8, output module 9.Load module 1 receives historical load data c ₀, and with historical data c ₀Send wavelet transformation module 2 to; 2 pairs of historical datas of wavelet transformation module are decomposed, and resolve into successively d ₁, d ₂, d ₃, d ₄, c ₄Five components, wherein d ₁-d ₄Has annual periodically variable feature, c ₄Has ever-increasing variation tendency; First kind prediction module 3,4,5,6 couples of d ₁-d ₄These four components are predicted; 7 couples of c of Equations of The Second Kind prediction module ₄Component is predicted; Each component after 8 pairs of predictions of wavelet reconstruction module carries out wavelet reconstruction, obtains prediction load c ₀'; To predict that by output module 9 load presents at last.

As shown in Figure 3, in the present invention, wavelet transformation module 2 adopts the Mallat decomposition algorithm, at first with former month load data c ₀Resolve into high fdrequency component d ₁With low frequency component c ₁, then by that analogy, with c ₁Further resolve into d ₂And c ₂, c ₂Further resolve into d ₃And c ₃, c ₃Further resolve into d ₄And c ₄Need to carry out curve down-sampledly when adopting wavelet transformation, the yardstick after decomposing each time is all different, often carries out wavelet transformation one time, and the sampled point of horizontal ordinate will reduce half.

First kind prediction module 3,4,5,6 couples of d ₁-d ₄These four components are predicted, obtain d ₁'-d ₄', selected Forecasting Methodology can reflect the cyclic fluctuation feature of data, specifically comprises time series method (AR model, MA model, arma modeling etc.), neural network etc.The present embodiment is to coefficient d ₁-d ₄When predicting, employing be the RBF neural net prediction method.

The RBF neural network is a kind of three layers of feedforward network, as shown in Figure 5.It comprises input layer X={x ₁, x ₂... x _n, hidden layer and output layer y, hidden layer is comprised of one group of radial basis function, hidden layer node comprises two important parameters: center C _iAnd width cs _i, radial basis function is chosen Gaussian function usually.The training of RBF neural network realizes weights ω between each basis function and output node by minimizing objective function _iAdjusting.Training method comprises clustering procedure, least square method, and gradient method etc. adopts clustering procedure here.When adopting above-mentioned RBF neural network to predict, input vector X is historical data, and output vector Y is single predicted value, in order simultaneously a plurality of points to be predicted, needs to repeat to set up repeatedly above-mentioned RBF neural network.

Adopt the RBF neural network to d ₁-d ₄Predict, result is d as shown in Figure 2 ₁'-d ₄' curve map.

7 couples of c of Equations of The Second Kind prediction module ₄Component is predicted, obtains c ₄', selected Forecasting Methodology can reflect the feature of data sustainable growth, comprises all kinds of trend extrapolations, as moving average, and exponential smoothing, gray prediction etc.

Gray prediction method is a kind of of trend extrapolation, the long term variations of its energy fitting data, and data are predicted.

To coefficient c ₄When predicting, due to more accurate than Traditional GM (1,1) model without inclined to one side GM (1,1) model, what this paper adopted is the nothing method of gray prediction partially.

Suppose that sequence to be predicted is x ⁽⁰⁾={ x ⁽⁰⁾(1), x ⁽⁰⁾(2) ..., x ⁽⁰⁾(n) }.At first to x ⁽⁰⁾Summation formation sequence x adds up ⁽¹⁾={ x ⁽¹⁾(1), x ⁽¹⁾(2) ..., x ⁽¹⁾(n) }, wherein:

$x^{\left(1\right)}\left(i\right)=\underset{j=1}{\overset{i}{\mathrm{\Σ}}}{x}^{\left(0\right)}\left(j\right),i=\mathrm{1,2,3}...$

According to gray system theory, cumulative sequence x ⁽¹⁾Have the exponential increase rule, can think x ⁽¹⁾Satisfy following Differential Equation Model:

$\frac{{\mathrm{dx}}^{\left(1\right)}}{\mathrm{dt}}+{\mathrm{ax}}^{\left(1\right)}=b$

If following formula adopts discrete form to represent, can turn to:

$x^{\left(0\right)}(k+1)+\frac{1}{2}a[x^{\left(1\right)}\left(k\right)+x^{\left(1\right)}(k+1)]=b,k=\mathrm{1,2,3}...$

Being write as matrix form has:

$\left[\begin{array}{c}{x}^{\left(0\right)}\left(2\right)\\ x^{\left(0\right)}\left(3\right)\\ \·\\ \·\\ \·\\ x^{\left(0\right)}\left(n\right)\end{array}\right]=\left[\begin{array}{cc}-\frac{1}{2}[x^{\left(1\right)}\left(1\right)+x^{\left(1\right)}\left(2\right)]& 1\\ -\frac{1}{2}[x^{\left(1\right)}\left(2\right)+x^{\left(1\right)}\left(3\right)]& 1\\ \·& \·\\ \·& \·\\ \·& \·\\ -\frac{1}{2}[x^{\left(1\right)}(n-1)+x^{\left(1\right)}\left(n\right)]& 1\end{array}\right]\left[\begin{array}{c}a\\ b\end{array}\right];$ Brief note is: $Y_n=B\·\left[\begin{array}{c}a\\ b\end{array}\right]$

Wherein:

$Y_n=\left[\begin{array}{c}{x}^{\left(0\right)}\left(2\right)\\ x^{\left(0\right)}\left(3\right)\\ \·\\ \·\\ \·\\ x^{\left(0\right)}\left(n\right)\end{array}\right],$ $B=\left[\begin{array}{cc}-\frac{1}{2}[x^{\left(1\right)}\left(1\right)+x^{\left(1\right)}\left(2\right)]& 1\\ -\frac{1}{2}[x^{\left(1\right)}\left(2\right)+x^{\left(1\right)}\left(3\right)]& 1\\ \·& \·\\ \·& \·\\ \·& \·\\ -\frac{1}{2}[x^{\left(1\right)}(n-1)+x^{\left(1\right)}\left(n\right)]& 1\end{array}\right]$

Above-mentioned equation utilizes least square method can obtain the least square solution of matrix A:

$\left[\begin{array}{c}\hat{a}\\ \hat{b}\end{array}\right]={\left(B^{T}B\right)}^{-1}{B}^T{Y}_n$

According to without the modeling method of GM (1,1) model partially, need with

With Do following calculating:

${\hat{a}}^{\′}=\ln \frac{2-\hat{a}}{2+\hat{a}},$ $A=\frac{2\hat{b}}{2+\hat{a}}$

Sequence to be measured can represent in order to drag:

$x^{\left(0\right)}(k+1)={\mathrm{Ae}}^{{\hat{a}}^{\′}\·k}$

Adopt without inclined to one side GM (1,1) model coefficient c ₄C in the result of predicting such as Fig. 2 ₄' curve map.

As shown in Figure 4, in the present invention, wavelet reconstruction module 8 adopts the Mallat restructing algorithm, at first with component d ₄' and c ₄' be reconstructed and obtain c ₃', then by that analogy, with d ₃' and c ₃' further reconstruct obtains c ₂', d ₂' and c ₂' further reconstruct obtains c ₁', d ₁' and c ₁' further reconstruct obtains prediction load c ₀'.Need to carry out rising sampling to curve when adopting wavelet reconstruction, often carry out wavelet reconstruction one time, the sampled point of horizontal ordinate will increase half.

Load module 1 and load module 9 are realized by software.By load module 1, operating personnel can import a month load historical data c ₀, and be stored in database.Output module 9 can be derived prediction load c ₀', and and then draw the prediction load curve, thereby convenient the observation.

It is same as the prior art that the present invention does not state part.

Claims (3)

1. based on the monthly calculation of power load machine Forecasting Methodology of wavelet analysis, it is characterized in that, comprise the steps:

The load forecast module is provided, and it comprises load module, wavelet transformation module, first kind prediction module, Equations of The Second Kind prediction module, wavelet reconstruction module and output module;

Load module receives the historical load data from electric system c ₀, and with historical data c ₀Send the wavelet transformation module to, the wavelet transformation module is decomposed historical data, c ₀Resolve into high fdrequency component d ₁And low frequency component c ₁, then will c ₁Further resolve into d ₂With c ₂, c ₂Further resolve into d ₃With c ₃, c ₃Further resolve into d ₄With c ₄Thereby, obtain to decompose d ₁, d ₂, d ₃, d ₄, c ₄Five components, wherein d ₁- d ₄Have annual periodically variable feature, c ₄Has ever-increasing variation tendency; Need to carry out curve down-sampledly when adopting wavelet transformation, often carry out wavelet transformation one time, the sampled point of horizontal ordinate will reduce half;

Will d ₁, d ₂, d ₃, d ₄Component data send into first kind prediction module, first kind prediction module adopts the data method pair of the cyclic fluctuation feature can reflect data d ₁, d ₂, d ₃, d ₄These four components are predicted, obtain d ₁', d ₂', d ₃', d ₄' four transform component; Will c ₄Data component is sent into the Equations of The Second Kind prediction module, and the Equations of The Second Kind prediction module can reflect the method pair of the feature of data sustainable growth c ₄Component is predicted, obtains transform component c ₄', need to carry out rising sampling to curve when adopting wavelet reconstruction, often carry out wavelet reconstruction one time, the sampled point of horizontal ordinate will increase half;

Send above-mentioned transform component to the wavelet reconstruction module, adopt the Mallat restructing algorithm, at first with component d ₄' and c ₄' be reconstructed and obtain c ₃', then by that analogy, will d ₃' and c ₃' further reconstruct obtains c ₂', d ₂' and c ₂' further reconstruct obtains c ₁', d ₁' and c ₁' further reconstruct obtains the prediction load c ₀';

With the prediction load that obtains c ₀' store, and export by output module.

2. the monthly calculation of power load machine Forecasting Methodology based on wavelet analysis according to claim 1, is characterized in that, the data method of the described cyclic fluctuation feature that can reflect data or for time series method or be neural network.

3. the monthly calculation of power load machine Forecasting Methodology based on wavelet analysis according to claim 1, is characterized in that, describedly can reflect that the method for data sustainable growth feature is trend extrapolation.

Images