CN102982383A - Energy supply and demand forecasting method based on support vector machine - Google Patents

Abstract

The invention relates to an energy supply and demand forecasting method for a tobacco enterprise. At first, historical data of energy supply and demand is pre-processed, preprocessing comprises rejecting abnormal data, special point analysis and data denoising, and then all data indicators are used to build a support vector machine model, and then the support vector machine model is conducted solving. Tobacco energy supply and demand forecasting historic data is firstly obtained, and then the data is screened, obvious abnormal data is rejected, and the abnormal data mainly refers to energy waste caused by manual work and abnormal tobacco quality caused by abnormal equipment, and then obtained input and output data are utilized to carry out simple preprocessing, and a model is built through the support vector machine and a least square support vector machine. People choose two kinds of kernel functions in a parallel mode, namely a polynomial function and a radial basis function (RBF), wherein the RBF is used as a main body to build the model, and a result that a certain relation (generally named weight) is multiplied by the polynomial function is used as model correction, so that forecasting precision can be improved to a great extent.

Description

A kind of energy supply and demand Forecasting Methodology based on support vector machine

Technical field

The invention belongs to the production of cigarettes technical field, be specifically related to a kind of energy supply and demand Forecasting Methodology based on support vector machine.

Background technology

How China's tobacco enterprise under the prerequisite of guaranteeing stable supplying and safety in production, efficiently utilizes the limited energy, optimizes the supply and demand structure, takes full advantage of secondary energy, and reducing the waste that the energy supply and demand imbalance causes is a very important problem.Take Trend Forecast as the basis, it is most important effectively to dispatch the energy.At present, the energy forecast of most domestic tobacco enterprise is based on the short-term prediction of artificial experience, and is higher to dopester's experience requirement, lacks the support of forecast model.Most of researchs about energy supply and demand all concentrate on the supply and demand prediction of the single energy or two kinds of energy, and the supply and demand prediction that comprises various energy resources is seldom arranged.

System and method about the energy supply and demand prediction mainly contains neural network model and regression equation method at present.Although neural network model has higher non-linear mapping capability, can be with the arbitrary accuracy Nonlinear Function Approximation, but in actual computation, also have some problems: 1. the computation process speed of convergence of backpropagation is slow, generally needs hundreds and thousands of times iterative computation; 2. the minimal value that has energy function; 3. hidden neuron number and connection weight chooses often by experience; 4. the convergence of network is relevant with the structure of network etc.Because the tobacco enterprise energy resource system is complicated, related variety of energy sources is various, and is interrelated between the energy, is not suitable for predicting with regression equation.And use regression equation and estimate when predicting, can only estimate dependent variable by independent variable, do not allow dependent variable to infer independent variable.And use regression equation and estimate when predicting, can only estimate dependent variable by independent variable, do not allow dependent variable to infer independent variable.Therefore, we consider to utilize support vector machine to carry out the energy supply and demand prediction.

Support vector machine (Support Vector Machine, SVM) be that Corinna Cortes and Vapnik8 equal nineteen ninety-five and at first propose, it shows many distinctive advantages in solving small sample, non-linear and higher-dimension pattern-recognition, and can promote the use of in the other machines problem concerning study such as Function Fitting.Support vector machine only needs inputoutput data to a certain extent, and does not need to set up strict reliable mathematical model with regard to inputoutput data.With the demand of the tobacco energy as input, the quality of tobacco of output is as output, set up the supporting vector machine model that only relies on inputoutput data, this model is fit to the small sample data very much, here the small sample of indication is not that index is according to considerably less, but in the situation of very limited sample number, be based upon to a certain degree with scope in extremely reliable mathematical model, namely set up with the mathematical model under the trusted conditions of data-driven.Although this model is simple, still quite high in the confidence level of the prerequisite drag of determining precision.Not only can carry out accurate as far as possible modeling to the tobacco energy supply and demand at the least square method supporting vector machine based on this method, and can also well predict the quality of tobacco and the tobacco culture scheme under set quality.For the reliability that further increases model and to the checking demand of model prediction, we choose different kernel functions and set up model, and be analyzed, the various support vector machine of match are set up result and the statistical relationship structure of model in appropriate circumstances, like this in the situation of large variable fluctuation, set up one rationally, comprehensively and reliably and system model that can Approximate prediction.The statistical model of setting up by support vector machine has been ignored the mutual restriction under many input conditions, connect each other, only from the data of input and output, so no matter to input the data how complicated, association between multifactor is indigestion how, do not need to consider dimensionality reduction classification and the factor relationship analysis of high order system, the complexity that this will greatly simplify the complexity of system and reduce model, and can effectively improve model accuracy.

Summary of the invention

The object of the invention is to for the deficiencies in the prior art, a kind of dimensionality reduction classification and factor relationship analysis that does not need to consider high order system is provided, greatly simplify the complexity of system and the complexity of reduction model, and can effectively improve the energy supply and demand Forecasting Methodology of model accuracy.

The object of the present invention is achieved like this:

A kind of energy supply and demand Forecasting Methodology based on support vector machine is at first carried out pre-service to the historical data of energy supply and demand, and pre-service comprises the rejecting abnormalities data, particular point analysis and data de-noising; Again all data targets are set up supporting vector machine model; Then carrying out supporting vector machine model finds the solution.

Wherein, carry out support vector and find the solution and be, choose objective function and be:

$\left\{\begin{array}{c}\min \frac{1}{2}{\mathrm{\ω}}^T\mathrm{\ω}+\frac{1}{2}\mathrm{\γ}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\ξ}}_i^2\\ s.t.y_i={\mathrm{\ω}}^T\mathrm{\φ}\left(x\right)+b+{\mathrm{\ξ}}_i\end{array}\right.(i=\mathrm{1,2},\mathrm{L\; n}).$

Wherein, described supporting vector machine model can reflect that n relation factor is on the wherein impact of certain factors vary rate, seek in n the relation factor, the relation in the energy supply and demand between certain variable and the other factors, thus analyze the variation of this variable to the impact of cigarette quality.

Wherein, each index of energy supply and demand index of prediction is set up respectively the weights correction model, choosing kernel function is polynomial expression:

K (x _i, x _j)=[(x _iGx _j)+1] ^q, obtain the model of different support vector machine, and the variable that changes set up the inline model of energy supply and demand.

Wherein, use inline model that former predicted value is revised.

Compared with prior art, the present invention has following advantage:

The present invention obtains first the historical data of tobacco energy supply and demand prediction, and carries out the data screening, rejects comparatively significantly abnormal data, and mainly the cigarette quality that causes of the energy dissipation that manually causes of finger and unit exception is unusual.Then after utilizing the inputoutput data of gained to carry out simple pre-service, set up model by support vector machine and least square method supporting vector machine, we walk abreast and choose two kinds of kernel functions therebetween, be polynomial function and radial basis function (RBF), wherein set up model take radial basis function as main body, polynomial function multiply by certain relation (being commonly referred to weights) as the model correction, so just can improve to a great extent precision of prediction.

Description of drawings

Fig. 1 is the basic universal model of support vector machine of the present invention;

Fig. 2 is energy supply and demand prediction module algorithm flow chart of the present invention.

Embodiment

Below in conjunction with drawings and Examples the present invention is made further description, but never in any form the present invention is limited, do to get any change or replacement according to training centre of the present invention, all belong to protection scope of the present invention.

Fig. 1 ~ Fig. 2 is a kind of embodiment of the present invention.

Fig. 1 is the support vector machine parameter configuration process flow diagram of this Forecasting Methodology.The meaning that ultimate principle that this model is set up and parameter are chosen has been described, the method be to a certain extent feasible effectively.

Fig. 2 is energy supply and demand prediction module algorithm flow chart of the present invention.From database, extract the measured data of modeling, and data are carried out pre-service, from algorithms library, extract corresponding forecast model coefficient information, then use the model of pretreated data and extraction to predict and correction result, last Output rusults.

Present embodiment is to as follows to tobacco enterprise energy supply and demand forecasting process:

1. the historical data of tobacco enterprise energy supply and demand is carried out pre-service

The pre-service here comprises the rejecting abnormalities data, particular point analysis and data de-noising.For extensibility and the reusing that strengthens model, select to set up a kind of basically general effective method and carry out the data pre-service.

Step (1) is chosen the energy supply and demand data under the different quality of tobacco from the tobacco historical data base, contain all tobacco leaf kinds when choosing as far as possible;

Step (2) is carried out pre-service to the data of choosing, and rejects obvious abnormal data (generally choosing allowance is 5%), i.e. the tobacco energy demand of one species can not surpass national regulation the energy resource consumption standard 5%, otherwise namely be defined as abnormal data;

Step (3) utilizes all data to numerous Index Establishment supply and demand X mathematical model relevant with cigarette quality Y of tobacco enterprise energy supply and demand;

All data targets are set up following mathematical model:

$\left\{\begin{array}{c}{Y}_1=a_{11}{X}_1+a_{12}{X}_2+L+a_{1m}{X}_m+b_1\\ Y_2=a_{21}{X}_1+a_{22}{X}_2+L+a_{2m}{X}_m+b_2\\ M\\ Y_p=a_{p1}{X}_1+a_{p2}{X}_2+L+a_{\mathrm{pm}}{X}_m+b_p\end{array}\right.$

Use matrix representation:

$\left[\begin{array}{c}{Y}_1\\ Y_2\\ M\\ Y_p\end{array}\right]=\left[\begin{array}{cccc}{a}_{11}& a_{12}& L& a_{1m}\\ a_{21}& a_{22}& L& a_{2m}\\ M& M& & M\\ a_{p1}& a_{p2}& L& a_{\mathrm{pm}}\end{array}\right]\left[\begin{array}{c}{X}_1\\ X_2\\ M\\ X_m\end{array}\right]+\left[\begin{array}{c}{b}_1\\ b_2\\ M\\ b_p\end{array}\right]$

This problem of being about to is converted into matrix of coefficients a and the b that finds the solution among the Y=aX+b, has avoided loaded down with trivial details Mathematical by the supporting vector machine model that all data are set up, and separates above-mentioned matrix equation with the method for similar black box.

2. supporting vector machine model is found the solution

The principle of step (4) supporting vector machine model is that the X territory is mapped to a high-dimensional feature space with a nonlinear function φ, carries out linear regression at high-dimensional feature space again, thereby obtains the effect that former Space Nonlinear returns;

In order to satisfy structural risk minimization principle, choose the objective optimization function and be:

$\left\{\begin{array}{c}\min \frac{1}{2}{\mathrm{\ω}}^T\mathrm{\ω}+\frac{1}{2}\mathrm{\γ}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\ξ}}_i^2\\ s.t.y_i={\mathrm{\ω}}^T\mathrm{\φ}\left(x\right)+b+{\mathrm{\ξ}}_i\end{array}\right.(i=\mathrm{1,2},\mathrm{L\; n})$

Introducing the Lagrange function is:

$L(\mathrm{\ω},b,\mathrm{\ζ},a)=\frac{1}{2}{\mathrm{\ω}}^T\mathrm{\ω}+\mathrm{\γ}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\ζ}}_i^2-\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{a}_i[\mathrm{\ω\φ}\left(x_i\right)+b+{\mathrm{\ζ}}_i-{\mathrm{\γ}}_i]$

In the formula: a _i---the Lagrange multiplier.

Optimum a and b can obtain by following formula:

$\left\{\begin{array}{cc}\frac{\∂L}{\∂\mathrm{\ω}}=0\→\mathrm{\ω}=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_i\mathrm{\φ}\left(x_i\right)& \\ \frac{\∂L}{\∂b}=0\→\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_i=0& \\ \frac{\∂L}{{\∂\mathrm{\ζ}}_i}=0\→{\mathrm{\α}}_i={\mathrm{\λ\ζ}}_i& i=1,L,N\\ \frac{\∂L}{{\∂\mathrm{\α}}_i}=0\→\mathrm{\ω\φ}\left(x_i\right)+b+{\mathrm{\ζ}}_i-{\mathrm{\γ}}_i=0& i=1,L,N\end{array}\right.$

Cancellation ω and ζ _i, optimization problem is converted into finds the solution following equation:

$\left|\begin{array}{cc}0& e^T\\ e& Q+{\mathrm{\γ}}^{-1}I\end{array}\right|\left|\begin{array}{c}b\\ a\end{array}\right|=\left|\begin{array}{c}0\\ y\end{array}\right|$

The result of LSSVM function regression is:

$y\left(x\right)=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{a}_i\mathrm{\ψ}(x_i,x_j)+b$

The radial basis kernel function that kernel function ψ () selects is:

$K(x_i,x_j)=\exp |-\frac{{\left|\right|x_i-x_j\left|\right|}^2}{{\mathrm{\σ}}^2}|$

3. set up forecast model

This model can reflect n relation factor to the wherein impact of certain factors vary rate, and we mainly seek in n the relation factor, the relation in the energy supply and demand between certain variable and the other factors, and analyze the variation of this variable to the impact of cigarette quality.

Step (5) is set up the inline model of energy supply and demand.As S, its dependent variable utilizes above-mentioned all data to set up new S-T model, i.e. the inline model of energy supply and demand as T with the variation variable in the energy supply and demand.According to the method for step (4), obtain inline model T=CS+d;

Step (6) draws whole energy supply and demand according to inline model forecast model carries out the processing of step (5) with the variable S that changes, and draws the X of all energy supply and demands, carries out the processing of step (4) again, just can predict quality of tobacco Y;

4. each index of energy supply and demand index of prediction is set up respectively the weights correction model

Step (7) is for step (4), and choosing kernel function is polynomial expression:

K(x _i,x _j)=[(x _igx _j)+1] ^q

In the formula: q---polynomial exponent number.

Like this, we have just obtained the model of different support vector machine, Y=A ' X+b '.

Step (8) is carried out prediction in step (5), (6) with the resulting supporting vector machine model of polynomial kernel function in the above-mentioned steps (7) to the variable that changes, and the new y ' of gained multiply by weights k the y of former prediction is revised.Following principle is followed in choosing of weights k:

(1) | y-y ' |≤yg δ k=0 δ is the allowance of national Specification;

(2) otherwise, k=δ.

The above only is preferred implementation of the present invention; should be pointed out that for those skilled in the art, under the prerequisite that does not break away from the technology of the present invention principle; can also make some improvement and distortion, these improvement and distortion also should be considered as protection scope of the present invention.

Claims (5)

1. the energy supply and demand Forecasting Methodology based on support vector machine is characterized in that, at first the historical data of energy supply and demand is carried out pre-service, and pre-service comprises the rejecting abnormalities data, particular point analysis and data de-noising; Again all data targets are set up supporting vector machine model; Then carrying out supporting vector machine model finds the solution.

2. as claimed in claim 1 based on the energy supply and demand Forecasting Methodology of support vector machine, it is characterized in that, carry out support vector when finding the solution, choose objective function and be:

3. as claimed in claim 1 based on the energy supply and demand Forecasting Methodology of support vector machine, it is characterized in that, n relation factor of described supporting vector machine model reflection is on the wherein impact of certain factors vary rate, seek in n the relation factor, relation in the energy supply and demand between certain variable and the other factors, thus the variation of this variable analyzed to the impact of cigarette quality.

4. as claimed in claim 1 based on the energy supply and demand Forecasting Methodology of support vector machine, it is characterized in that each index of energy supply and demand index of prediction is set up respectively the weights correction model, and choosing kernel function is polynomial expression:

K (x _i, x _j)=[(x _iGx _j)+1] ^q, obtain the model of different support vector machine, and the variable that changes set up the inline model of energy supply and demand.

5. as claimed in claim 4 based on the energy supply and demand Forecasting Methodology of support vector machine, it is characterized in that, use inline model that former predicted value is revised.

Images