CN104881707A - Sintering energy consumption prediction method based on integrated model - Google Patents

Abstract

The invention relates to a sintering energy consumption prediction method based on an integrated model. The method comprises the following steps of: 1) selecting a model characteristic based on a RReliefF characteristic selection algorithm; 2) according to the historical data of the selected model characteristic, establishing an improvement limit learning machine intelligent prediction model and a regression support vector machine intelligent prediction model; and 3) performing weighting integration on multiple single intelligent prediction models so as to obtain an integrated sintering energy consumption predicted value. Compared with a method in the prior art, the method is based on the integrated prediction model of information entropy, and has good prediction precision, a generalization ability, high time efficiency, and application and popularization value in real production.

Description

A kind of sintering energy consumption Forecasting Methodology based on integrated model

Technical field

The present invention relates to steel enterprise sintering process energy consumption prediction field, especially relate to a kind of sintering energy consumption Forecasting Methodology based on integrated model.

Background technology

Sintering is one of the maximum operation that consumes energy in iron and steel enterprise, although recent year Key Iron And Steel sintering plant revamp reduces year by year, but it also has larger gap with external advanced value, and energy consumption level difference is obvious between domestic each iron and steel enterprise, therefore research sintering circuit is saved energy and reduce the cost significant.To reduce in sintering ratio and process parameter optimizing research that energy consumption and cost are guiding, sintering energy consumption prediction is one of key issue wherein.Sintering plant revamp mainly comprises power consumption and non-electricity consumption, and wherein non-electricity energy consumption comprises solid fuel consumption, the large class of gas consumption in ignition two, accounts for 80% ~ 90% of total energy consumption, is the Main way of research sintering consumption reduction.At present, all there is certain defect in existing sintering energy consumption Forecasting Methodology both at home and abroad: input parameter is difficult to determine, modeling method precision is not enough, all affects precision and the timeliness of sintering energy consumption prediction.The difficult point of sintering energy consumption prediction is: 1) feature is difficult to determine.The influence factor of energy consumption is intricate, is difficult to determine simply by mechanism and experience, needs to select useful information from many factors, eliminate redundancy variable, obtains mode input parameter more accurately.2) single method precision is not enough.In actual production, energy consumption data fluctuation is comparatively large, and within it when mechanism can not be illustrated completely, Individual forecast method is difficult to obtain good precision of prediction, needs the Integrated research carrying out multiple Method and Technology.

By finding the retrieval of prior art, the method proposed for the prediction of sintering art energy consumption is little, and the method proposed for other field Similar Problems can be used for reference.In Chinese patent " a kind of energy resource consumption Forecasting Methodology and device " (grant number: CN103544544A), the people such as Yang Haidong propose a kind of energy consumption Forecasting Methodology based on SVR for enterprise's electricity needs forecast.The method utilizes history energy consumption data to build training sample set, the temporally series arrangement of energy consumption data wherein, by selecting best input exponent number to obtain best forecast model, and can predict future electrical energy requirements through contrast verification the method.But the precision of this Individual forecast model still has much room for improvement viewed from the result.In Chinese patent " a kind of sizing process rate of sizing flexible measurement method based on Bagging " (grant number: CN103018426A), bagging integrated technology is used in the prediction of the rate of sizing by Tian Huixin, propose the integrated predictive model based on SVR, further increase the prediction accuracy of single SVR.But, this invention, when Confirming model input parameter, only adopts the method for Analysis on Mechanism and micro-judgment.For steel enterprise sintering process, the factor complexity wherein affecting energy consumption change is various, is difficult to the notable feature determining to affect energy consumption by Analysis on Mechanism, and only judges by rule of thumb to be theoretically unsound, and thus needs more objective effective parameter selection method.

Summary of the invention

Object of the present invention be exactly provide a kind of to overcome defect that above-mentioned prior art exists and predict the outcome accurately, the sintering energy consumption Forecasting Methodology based on integrated model that time efficiency is high.

Object of the present invention can be achieved through the following technical solutions:

Based on a sintering energy consumption Forecasting Methodology for integrated model, the method comprises the following steps:

1) feature selecting algorithm based on RReliefF carries out aspect of model selection;

2) according to the historical data of the selected aspect of model, set up and improve extreme learning machine intelligent forecast model and Support Vector Machines for Regression intelligent forecast model;

3) be weighted integrated to multiple single intelligent forecast model, obtain integrated after sintering energy consumption predicted value.

RReliefF algorithm is the regression problem of successive value for the treatment of objective attribute target attribute, supposes that the scale of sample space S is m, Stochastic choice sample D _i(i=1,2 ..., m), and calculate its nearest k _nearindividual neighbour's sample S _k, meanwhile, suppose that the predicted value of sample is τ ().According to Bayes' theorem, the weight of characteristic attribute A can be obtained by following formula:

$\begin{array}{c}W\left[A\right]=\frac{P_{\mathrm{diff\τ}|\mathrm{diffA}}{P}_{\mathrm{diffA}}}{P_{\mathrm{diff\τ}}}-\frac{(1-P_{\mathrm{diff\τ}|\mathrm{diffA}})P_{\mathrm{diffA}}}{1-P_{\mathrm{diff\τ}}}\\ =\frac{P_{\mathrm{diff\τ}\&\mathrm{diffA}}}{P_{\mathrm{diff\τ}}}-\frac{P_{\mathrm{diffA}-\mathrm{diff\τ}\&\mathrm{diffA}}}{1-P_{\mathrm{diff\τ}}}\end{array}$

Here P _diffA=(diff (A, D _i, D _j) | D _j∈ S _k) be sample D _iwith its k _nearthe probability of characteristic attribute A difference between individual neighbour's sample; P _{diff τ}=(diff (τ (), D _i, D _j) | D _j∈ S _k) be D _iwith its k _nearthe probability of prediction index value τ () difference between individual neighbour's sample; P _{diff τ | diffA}=(diff (τ (), D _i, D _j) | diff (A, D _i, D _j), D _j∈ S _k) represent at known D _iwith its k _nearthe conditional probability of its prediction index value τ () difference when characteristic attribute A difference between individual neighbour's sample; P _{diff τ & diffA}=(diff (τ (), D _i, D _j) diff (A, D _i, D _j) | D _j∈ S _k) represent D _iwith its k _nearthe probability of characteristic attribute A difference and its prediction index value τ () difference between individual neighbour's sample.Here,

$\mathrm{diff}(A,D_i,D_j)=\frac{\mathrm{value}(A,D_i)-\mathrm{value}(A,D_j)}{\max \left(A\right)-\min \left(A\right)}$

Wherein value (A, D _i), value (A, D _j) be sample D _i, D _jthe value of middle characteristic attribute A; Max (A) and min (A) is D _iwith its k _nearthe minimum and maximum value of attribute A in individual neighbour's sample.Diff (τ (), D _i, D _j) with it in like manner.

Described step 1) be specially:

Step101: parameter initialization, makes i=1, j=1;

Step102: be select sample D in the sample space S of m in scale _i, select from this sample D from a remaining m-1 sample _inearest k _nearindividual sample, composition neighbour sample set S _k;

Step103: from S _kmiddle selection sample D _j, the weight N under the different predicted value of iterative computation _{diff τ}:

N _diffτ＝N _diffτ+diff(τ(·),D _i,D _j)/k _near

Wherein, the τ () predicted value that is sample;

Step104: to each characteristic attribute A, the weight N of iterative computation different characteristic attribute _diffAand the weight N of different predicted value and different characteristic attribute _{diff τ & diffA}:

N _diffA＝N _diffA+diff(A,D _i,D _j)/k _near

N _diffτ&diffA＝N _diffτ&diffA+diff(τ(·),D _i,D _j)·diff(A,D _i,D _j)/k _near

Step105: make i=i+1, j=j+1, judge whether j meets j≤k _near, if so, then return Step103, if not, then perform Step106;

Step106: judge whether i meets i≤m, if so, then returns Step101, if not, then perform Step107;

Step107: for each attribute A, the final weight calculating each characteristic attribute is estimated:

W[A]＝N _diffτ&diffA/N _diffτ-(N _diffτ-N _diffτ&diffA)/(m-N _diffτ)

Finally according to the characteristic attribute of weight preference pattern.

The process of establishing of described improvement extreme learning machine intelligent forecast model is:

Stepa1: the number M determining sub-learning machine;

Stepa2: with Bootstrap method sampling with replacement from training dataset B, obtain the training dataset B of every sub-learning machine _k, k=1 ..., M, and B _kidentical with B scale, be all N;

Stepa3: use B successively _ktrain corresponding sub-learning machine, obtain M training result, and check precision of prediction with unified test data set C respectively;

Stepa4: carry out integrated by the method for average to all sub-learning machine training results, obtain net result, and verification model precision.

The process of establishing of described Support Vector Machines for Regression intelligent forecast model is:

Stepb1: raw data is normalized:

$x_p=\frac{Q_p-Q_{\min }}{Q_{\max }-Q_{\min }}a+d$

Wherein, Q _pfor p value of each factor, p=1 ..., N, Q _max, Q _minbe respectively the maximal value in each factor and minimum value, a, d are parameter, d=(1-a)/2;

Stepb2: adopt RBF kernel function, determines the parameter of SVR model and training network;

Stepb3: by the precision of test data set C test network, obtain neural network forecast output valve, then carry out renormalization:

${\tilde{y}}_q=(Q_{\max }-Q_{\min }){\hat{y}}_q+Q_{\min }$

Wherein, be the network output valve of q test sample book, L is test data set number of samples, be the predicted value after q test sample book renormalization, q=1,2 ..., L;

Stepb4: performance evaluation is carried out to SVR model.

Described step 3) in, adopt information entropy to be weighted integrated to multiple single intelligent forecast model, be specially:

Step301: calculate the degree of variation predicted the outcome that each intelligent forecast model obtains:

$e_{\mathrm{uq}}=|{\tilde{y}}_{\mathrm{uq}}-y_q|/y_q$

Wherein, e _uq, represent relative error magnitudes and the prediction output valve of u model q sample, y _qrepresent q sample desired output, q=1,2 ..., L, L are test data set number of samples, u=1,2, representative improves extreme learning machine intelligent forecast model and Support Vector Machines for Regression intelligent forecast model respectively;

Step302: the entropy calculating each intelligent forecast model:

$E_u=-\frac{1}{\ln L}\underset{q=1}{\overset{L}{\mathrm{\Σ}}}{P}_{\mathrm{ud}}\ln P_{\mathrm{ud}}$

Wherein, P _udbe the Relative Error ratio of u model of q sample,

Step303: the weights calculating each intelligent forecast model:

$z_u=1-\frac{1-E_u}{{\mathrm{\Σ}}_{r=1}^2(1-E_r)}$

Wherein, z _urepresent the weights of u model;

Step304: the integrated output of weighting calculating each model:

${\stackrel{\‾}{y}}_p={\mathrm{\Σ}}_{u=1}^2{z}_u{\tilde{y}}_{\mathrm{uq}}$

。

Compared with prior art, the invention has the advantages that:

1) the present invention uses RReliefF algorithm to extract sintering energy consumption characteristic factor, only determines to have more theoretical foundation with artificial experience as compared with the past, improves the accuracy predicted the outcome;

2) to adopt multiple intelligent prediction algorithms to be weighted integrated in the present invention, effectively can realize the accurate prediction of sintering energy consumption;

3) merge the advantage of multiple forecast model based on the addition Integrated of information entropy, further increase model accuracy, and there is higher time efficiency.

Accompanying drawing explanation

Fig. 1 is principle schematic of the present invention;

Fig. 2 the present invention is based on feature that RReliefF algorithm the obtains characteristic attributes weight ordering chart for each energy consumption index;

Figure (2a) is solid fuel consumption weight sequencing figure; Figure (2b) is coal gas energy consumption weight sequencing figure.

Embodiment

Below in conjunction with the drawings and specific embodiments, the present invention is described in detail.The present embodiment is implemented premised on technical solution of the present invention, give detailed embodiment and concrete operating process, but protection scope of the present invention is not limited to following embodiment.

As shown in Figure 1, the embodiment of the present invention provides a kind of sintering energy consumption Forecasting Methodology based on integrated model, and the method comprises the following steps:

Step 1), carry out aspect of model selection based on the feature selecting algorithm of RReliefF.

In sintering process, the influence factor of energy consumption is numerous, can be summarized as: parameter, 2 of 1) preparing burden) state parameter, 3) operating parameter three class.Batching parameter mainly comprises major ingredient and auxiliary material addition and dielectric dissipation, these parameters before sintering starts just to determine; The production status values such as state parameter mainly comprises box temperature, pressure, igniting space gas flow, they are obtained by spot sensor; Operating parameter comprises that sintering is on-the-spot can the parameter of manual shift, and as machine speed, thickness of feed layer, mixer amount of water etc., they are determined by operating personnel when sintering and starting.By the analysis to sintering process, sintering energy consumption influence factor is summed up as shown in table 1.

Table 1 sintering energy consumption analysis of Influential Factors

The factor affecting sintering energy consumption is varied, and the height of these factor significance levels is only difficult to determine with Analysis on Mechanism and artificial experience, needs to obtain more objective rational conclusion by data analysis.

The present invention adopts a kind of Relief algorithm of improvement (RReliefF algorithm) to carry out aspect of model selection.Relief serial algorithm is a kind of stochastic search methods typically selecting feature according to weight, its main thought is: assess feature according to the separating capacity of feature to closely sample, good feature should make similar sample close, and make inhomogeneous sample away from.

RReliefF algorithm is the regression problem of successive value for the treatment of objective attribute target attribute,

The detailed process that RReliefF algorithm carries out aspect of model selection is as follows:

Step101: be select certain sample D the sample set of m from scale _i, select from this sample D from a remaining m-1 sample _inearest k _nearindividual sample, wherein 1≤i≤m, k _nearvalue generally get 10 ~ 20 and be advisable.Described distance refers to Euclidean distance.

Step102: iterative computation is at described sample D _ioutput-index value P ₀weight sets n under condition _dC:

$n_{\mathrm{dC}}=n_{\mathrm{dC}}+\frac{{\mathrm{\Σ}}_{i=1}^{k_{\mathrm{near}}}|P_0-P_i|}{P_{\max }-P_{\min }}\·\frac{1}{k_{\mathrm{near}}}$

Wherein, P ₀represent described sample D _ithe value of output-index P, P _i(1≤i≤k _near) be described k _nearthe output-index value of i-th sample in individual sample, P _maxand P _minmaximal value and the minimum value of output-index in m sample respectively;

Calculate at described sample D _iweight sets n under input feature vector A condition _dA[A], carry out iterative computation by following formula:

$n_{\mathrm{dA}}\left[A\right]=n_{\mathrm{dA}}\left[A\right]+\frac{{\mathrm{\Σ}}_{i=1}^{k_{\mathrm{near}}}|A_0-A_i|}{A_{\max }-A_{\min }}\·\frac{1}{k_{\mathrm{near}}}$

Wherein, A ₀sample D belonging to representing _ithe value of input feature vector A, A _i(1≤i≤k _near) be described k _nearthe value of input feature vector A in i-th sample in individual sample, A _maxand A _minmaximal value and the minimum value of input feature vector A in m sample respectively;

Calculate at described sample D _ioutput-index value P ₀with the weight sets n under input feature vector A condition _{dC & dA}[A]:

Step103: perform Step101 ~ 102 successively to each sample in sample set, altogether circulation m time, obtains m n _dC, n _dA[A] and n _{dC & dA}[A], so $N_{\mathrm{dC}}=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{n}_{\mathrm{dC}}^i,N_{\mathrm{dA}}\left[A\right]=\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{n}_{\mathrm{dA}}^i\left[A\right],$

Step104: the weighted value W [A] calculating input feature vector A:

Step105: according to weight preference pattern feature.

Step 2), according to the historical data of the selected aspect of model, set up and improve extreme learning machine intelligent forecast model and Support Vector Machines for Regression intelligent forecast model.

Affect on the basis of sintering energy consumption key character in acquisition, adopt the improvement ELM (extreme learning machine based on bagging, Extreme Learning Machine) algorithm improves stability and the generalization ability of former ELM algorithm, adopts ε-SVR (Support Vector Machines for Regression) algorithm to improve Generalization Capability under condition of small sample.

(1) modeling process based on the improvement ELM sintering energy consumption forecast model B-ELM of Bagging is as follows:

Stepa1: the number M determining sub-learning machine.

Stepa2: determine sub-learning machine training dataset.Bagging data set is obtained by data scrubbing and feature extraction each (x _i, y _i) have n to input and 1 output, i.e. x ∈ R ⁿ, y ∈ R _.select 2/3 as training dataset B, all the other are as test data set C.

With Bootstrap method sampling with replacement from training dataset B, obtain the training dataset B of every sub-learning machine _k, k=1 ..., M, and B _kidentical with B scale, be all N, make k=1.

Stepa3: use B _ktrain corresponding sub-learning machine, obtain training result, and check precision of prediction with unified test data set C, k=k+1.

If K is hidden node number, random acquisition initial input weight w _lwith biased b _l, l=1 ..., K.Calculate hidden layer output matrix H{h _nl(n=1 ..., N, l=1 ..., K).Obtain the generalized inverse matrix H of H ⁺, according to β=H ⁺t calculates and exports weights β.

Stepa4: judge whether k meets k≤M, if so, then returns step Stepa3, if not, then carries out integrated by the method for average to all sub-learning machine training results, obtains net result, and verification model precision.

(2) process based on ε-SVR sintering energy consumption prediction modeling is as follows:

Stepb1: raw data is normalized:

$x_p=\frac{Q_p-Q_{\min }}{Q_{\max }-Q_{\min }}a+d$

Wherein, Q _pfor p value of each factor, p=1 ..., N, Q _max, Q _minbe respectively the maximal value in each factor and minimum value, a, d are parameter, d=(1-a)/2;

In the present embodiment, raw data specification is interval to [0.2,0.8], a=0.6.

Stepb2: determine the parameter of SVR model and training network.This model adopts RBF kernel function, and the variance g in penalty factor c and RBF kernel function is obtained by cross-validation method.Concrete steps be c and g is set to-10 ~ 10 respectively between with 0.01 for step-length changes, under the combination of different c and g, calculate the precision of SVR model respectively, find optimum c and g combined value, in this, as the parameter training network one by one of SVR model.

Stepb3: by the precision of test data set C test network, obtain neural network forecast output valve, then carry out renormalization:

${\tilde{y}}_q=(Q_{\max }-Q_{\min }){\hat{y}}_q+Q_{\min }$

Wherein, be the network output valve of q test sample book, L is test data set number of samples, be the predicted value after q test sample book renormalization, q=1,2 ..., L.

Stepb4: performance evaluation is carried out to SVR model, judgement schematics:

$\mathrm{MeanRe}=\frac{1}{L}\underset{q=1}{\overset{L}{\mathrm{\Σ}}}|{\tilde{y}}_q-y_q|/y_q$

Wherein, MeanRe represents average relative error, y _q(q=1,2 ..., L) be the actual value of q test sample book.

Step 3), be weighted integrated to multiple single intelligent forecast model, obtain integrated after sintering energy consumption predicted value.

After establishing B-ELM and the ε-SVR model of sintering energy consumption, entropy assessment is adopted to be weighted integrated to above-mentioned submodel, with accurately predicting sintering energy consumption.Integrated predictive model idiographic flow based on information entropy is:

Step301: calculate the degree of variation predicted the outcome that each intelligent forecast model obtains:

$e_{\mathrm{uq}}=|{\tilde{y}}_{\mathrm{uq}}-y_q|/y_q$

Wherein, e _uq, represent relative error magnitudes and the prediction output valve of u model q sample, y _qrepresent q sample desired output, q=1,2 ..., L, L are test data set number of samples, u=1,2, represent B-ELM and ε-SVR model respectively.

Step302: the entropy calculating each intelligent forecast model:

$E_u=-\frac{1}{\ln L}\underset{q=1}{\overset{L}{\mathrm{\Σ}}}{P}_{\mathrm{ud}}\ln P_{\mathrm{ud}}$

Wherein, P _udbe the Relative Error ratio of u model of q sample,

Step303: the weights calculating each intelligent forecast model:

$z_u=1-\frac{1-E_u}{{\mathrm{\Σ}}_{r=1}^2(1-E_r)}$

Wherein, z _urepresent the weights of u model.

Step304: the integrated output of weighting calculating each model:

${\stackrel{\‾}{y}}_q={\mathrm{\Σ}}_{u=1}^2{z}_u{\tilde{y}}_{\mathrm{uq}}.$

Produce the integrated iron and steel works of 6,500,000 tons of steel scales per year for certain, its in 2 × 380m2 scale sintered production line year 8360000 tons, finished product sintering deposit, operating rate 94%, usage factor is 1.40t/m2h.The 311 groups of SINTERING PRODUCTION historical datas choosing 1 ~ Dec in 2010 carry out analysis modeling.Often organize sample packages containing solid fuel consumption and coal gas energy consumption two energy consumption indexs and 73 features, characteristic attribute is as shown in table 1, parameter 12 of wherein preparing burden, state parameter 50, operating parameter 11.On the basis of carrying out data scrubbing, obtain the weight sequencing of feature for each energy consumption index based on RReliefF algorithm, as shown in Figure 2.

For solid fuel consumption, come the feature of first 5 respectively: coke powder addition, cloud powder addition, mass flow PD-1, ignition temperature T3, blending ore addition; For coal gas burnup, come the feature of first 5 respectively: ignition temperature T3, house steward's air pressure, blending ore addition, mass flow PD-1, coke powder addition.Can find out that these feature orderings and actual conditions meet substantially.

For checking and comparison model precision and performance, the present embodiment chooses following 4 evaluation indexes:

1) average relative error MeanRe:

$\mathrm{MeanRe}=\frac{1}{L}\underset{q=1}{\overset{L}{\mathrm{\Σ}}}|{\tilde{y}}_q-y_q|/y_q$

2) residual error mean value e:

$e=\frac{1}{L}\underset{q=1}{\overset{L}{\mathrm{\Σ}}}|{\tilde{y}}_q-y_q|$

3) maximum error E _max:

$E_{\max }=\max \left\{\right|{\tilde{y}}_q-y_q\left|\right\}$

4) precision P _r:

$P_r=(1-\frac{\mathrm{\σ}}{\frac{1}{m}{\mathrm{\Σ}}_{i=1}^{m}y\left(i\right)})\×100\%$

Wherein, σ is standard deviation, $\mathrm{\σ}=\sqrt{\frac{1}{L}{\mathrm{\Σ}}_{q=1}^L{[{\tilde{y}}_q-y_q]}^2}.$

In order to verify the validity of integrated model proposed comprehensively, first when not carrying out feature selecting, by this model predict the outcome and ELM, B-ELM, ε-SVR three kinds of models compare, as shown in table 2 and table 3.Model accuracy after feature selecting compares as shown in table 4 and table 5.It should be noted that, for each single algorithm (son) model, all by cross-validation method determination optimization model parameter; Meanwhile, often kind of model all runs 5 times, gets the mean value of each index as net result.Bagging submodel number is set to 10, and characteristic factor retains number and is set to 30.

From table 2 and table 3, the average relative error of the integrated predictive model based on information entropy that the present invention proposes is better than other four kinds of models, and other indexs also show advantage in various degree, thus demonstrate the superiority of institute of the present invention Modling model.Here, because integrated model is the normalization weighted sum of 2 submodels, therefore its maximum error is rational between the maximum error of submodel.From table 4 and table 5, the model accuracy after feature selecting comparatively table 2 and table 3 all has a more substantial increase, thus demonstrates the validity of feature selection approach in the present invention.

In addition, because ELM and ε-SVR has without the need to adjusting network weight, the advantage that training speed is fast; RReliefF algorithm, as a kind of heuritic approach, has the feature that time efficiency is high equally, therefore the integrated model that the present invention proposes has good time efficiency.As shown in table 6, integrated predictive model is about 88s consuming time always.

Table 2 compares without the solid fuel consumption prediction index of feature selecting

Table 3 compares without the coal gas energy consumption prediction index of feature selecting

Table 4 has the solid fuel consumption prediction index of feature selecting to compare

Table 5 has the coal gas energy consumption prediction index of feature selecting to compare

The time efficiency of the different model of table 6 compares

In sum, the integrated predictive model based on information entropy that the present invention proposes has good precision of prediction and generalization ability, has higher time efficiency simultaneously, has application and popularization value in actual production.

Claims (6)

1., based on a sintering energy consumption Forecasting Methodology for integrated model, it is characterized in that, the method comprises the following steps:

1) feature selecting algorithm based on RReliefF carries out aspect of model selection;

2) according to the historical data of the selected aspect of model, set up and improve extreme learning machine intelligent forecast model and Support Vector Machines for Regression intelligent forecast model;

3) be weighted integrated to multiple single intelligent forecast model, obtain integrated after sintering energy consumption predicted value.

2. the sintering energy consumption Forecasting Methodology based on integrated model according to claim 1, is characterized in that, described step 1) be specially:

Step101: parameter initialization, makes i=1, j=1;

Step102: be select sample D in the sample space S of m in scale _i, select from this sample D from a remaining m-1 sample _inearest k _nearindividual sample, composition neighbour sample set S _k;

Step103: from S _kmiddle selection sample D _j, calculate the weight N under different predicted value _{diff τ}:

N _diffτ＝N _diffτ+diff(τ(·),D _i,D _j)/k _near

Wherein, the τ () predicted value that is sample;

Step104: to each characteristic attribute A, calculates the weight N of different characteristic attribute _diffAand the weight N of different predicted value and different characteristic attribute _{diff τ & diffA}:

N _diffA＝N _diffA+diff(A,D _i,D _j)/k _near

N _diffτ&diffA＝N _diffτ&diffA+diff(τ(·),D _i,D _j)·diff(A,D _i,D _j)/k _near

Step105: make i=i+1, j=j+1, judge whether j meets j≤k _near, if so, then return Step103, if not, then perform Step106;

Step106: judge whether i meets i≤m, if so, then returns Step101, if not, then perform Step107;

Step107: for each attribute A, the final weight calculating each characteristic attribute is estimated:

W[A]＝N _diffτ&diffA/N _diffτ-(N _diffτ-N _diffτ&diffA)/(m-N _diffτ)

Finally according to the characteristic attribute of weight preference pattern.

3. the sintering energy consumption Forecasting Methodology based on integrated model according to claim 2, is characterized in that, described k _nearvalue be 10 ~ 20.

4. the sintering energy consumption Forecasting Methodology based on integrated model according to claim 1, is characterized in that, the process of establishing of described improvement extreme learning machine intelligent forecast model is:

Stepa1: the number M determining sub-learning machine;

Stepa2: with Bootstrap method sampling with replacement from training dataset B, obtain the training dataset B of every sub-learning machine _k, k=1 ..., M, and B _kidentical with B scale, be all N;

Stepa3: use B successively _ktrain corresponding sub-learning machine, obtain M training result, and check precision of prediction with unified test data set C respectively;

Stepa4: carry out integrated by the method for average to all sub-learning machine training results, obtain net result, and verification model precision.

5. the sintering energy consumption Forecasting Methodology based on integrated model according to claim 1, is characterized in that, the process of establishing of described Support Vector Machines for Regression intelligent forecast model is:

Stepb1: raw data is normalized:

$x_p=\frac{Q_p-Q_{\min }}{Q_{\max }-Q_{\min }}a+d$

Wherein, Q _pfor p value of each factor, p=1 ..., N, Q _max, Q _minbe respectively the maximal value in each factor and minimum value, a, d are parameter, d=(1-a)/2;

Stepb2: adopt RBF kernel function, determines the parameter of SVR model and training network;

Stepb3: by the precision of test data set C test network, obtain neural network forecast output valve, then carry out renormalization:

${\tilde{y}}_d=(Q_{\max }-Q_{\min }){\hat{y}}_q+Q_{\text{min}}$

Wherein, be the network output valve of q test sample book, L is test data set number of samples, be the predicted value after q test sample book renormalization, q=1,2 ..., L;

Stepb4: performance evaluation is carried out to SVR model.

6. the sintering energy consumption Forecasting Methodology based on integrated model according to claim 1, is characterized in that, described step 3) in, adopt information entropy to be weighted integrated to multiple single intelligent forecast model, be specially:

Step301: calculate the degree of variation predicted the outcome that each intelligent forecast model obtains:

$e_{\mathrm{uq}}=|{\tilde{y}}_{\mathrm{uq}}-y_q|/y_q$

Wherein, e _uq, represent relative error magnitudes and the prediction output valve of u model q sample, y _qrepresent q sample desired output, q=1,2 ..., L, L are test data set number of samples, u=1,2, representative improves extreme learning machine intelligent forecast model and Support Vector Machines for Regression intelligent forecast model respectively;

Step302: the entropy calculating each intelligent forecast model:

$E_u=-\frac{1}{\ln L}\underset{q=1}{\overset{L}{\mathrm{\Σ}}}{P}_{\mathrm{ud}}\ln P_{\mathrm{ud}}$

Wherein, P _udbe the Relative Error ratio of u model of q sample,

Step303: the weights calculating each intelligent forecast model:

$z_u=1-\frac{1-E_u}{{\mathrm{\Σ}}_{r=1}^2(1-E_r)}$

Wherein, z _urepresent the weights of u model;

Step304: the integrated output of weighting calculating each model:

${\stackrel{\‾}{y}}_p={\mathrm{\Σ}}_{u=1}^2{z}_u{\tilde{y}}_{\mathrm{uq}}.$