The prediction of cement energy demand based on support vector machine Xin Zhang a, *, Shaohong Jing b

Similar documents
Electricity consumption forecasting method based on MPSO-BP neural network model Youshan Zhang 1, 2,a, Liangdong Guo2, b,qi Li 3, c and Junhui Li2, d

Study on Coal Consumption Curve Fitting of the Thermal Power Based on Genetic Algorithm

Support vector machine revisited

Research Article Research on Application of Regression Least Squares Support Vector Machine on Performance Prediction of Hydraulic Excavator

Research on real time compensation of thermal errors of CNC lathe based on linear regression theory Qiu Yongliang

New Exponential Strengthening Buffer Operators and Numerical Simulation

10-701/ Machine Learning Mid-term Exam Solution

Grey Correlation Analysis of China's Electricity Imports and Its Influence Factors Hongjing Zhang 1, a, Feng Wang 1, b and Zhenkun Tian 1, c

Evapotranspiration Estimation Using Support Vector Machines and Hargreaves-Samani Equation for St. Johns, FL, USA

Research on Dependable level in Network Computing System Yongxia Li 1, a, Guangxia Xu 2,b and Shuangyan Liu 3,c

Chapter 7. Support Vector Machine

ECE 8527: Introduction to Machine Learning and Pattern Recognition Midterm # 1. Vaishali Amin Fall, 2015

Linear Support Vector Machines

REGRESSION (Physics 1210 Notes, Partial Modified Appendix A)

Construction and test of forecasting push-pull demand model of urban. emergency supplies

Reliability model of organization management chain of South-to-North Water Diversion Project during construction period

Direction of Arrival Estimation Method in Underdetermined Condition Zhang Youzhi a, Li Weibo b, Wang Hanli c

Prediction of Compressive Strength of Concrete from Early Age Test Result Using Design of Experiments (RSM)

Massachusetts Institute of Technology

6.867 Machine learning

Linear Classifiers III

A collocation method for singular integral equations with cosecant kernel via Semi-trigonometric interpolation

Week 1, Lecture 2. Neural Network Basics. Announcements: HW 1 Due on 10/8 Data sets for HW 1 are online Project selection 10/11. Suggested reading :

Measurement uncertainty of the sound absorption

Classification of fragile states based on machine learning

ME 539, Fall 2008: Learning-Based Control

New Correlation for Calculating Critical Pressure of Petroleum Fractions

Topics Machine learning: lecture 2. Review: the learning problem. Hypotheses and estimation. Estimation criterion cont d. Estimation criterion

Research Article An Optimized Classification Algorithm by Neural Network Ensemble Based on PLS and OLS

Intermittent demand forecasting by using Neural Network with simulated data

OBJECTIVES. Chapter 1 INTRODUCTION TO INSTRUMENTATION FUNCTION AND ADVANTAGES INTRODUCTION. At the end of this chapter, students should be able to:

DERIVING THE 12-LEAD ECG FROM EASI ELECTRODES VIA NONLINEAR REGRESSION

GUIDELINES ON REPRESENTATIVE SAMPLING

Machine Learning Regression I Hamid R. Rabiee [Slides are based on Bishop Book] Spring

The Input-Output Analysis of Construction Industry in West China. Ke Bin Zhang 2, b

Design of distribution transformer loss meter considering the harmonic loss based on one side measurement method

MCT242: Electronic Instrumentation Lecture 2: Instrumentation Definitions

1 Review of Probability & Statistics

AN OPEN-PLUS-CLOSED-LOOP APPROACH TO SYNCHRONIZATION OF CHAOTIC AND HYPERCHAOTIC MAPS

As metioed earlier, directly forecastig o idividual product demads usually result i a far-off forecast that ot oly impairs the quality of subsequet ma

Machine Learning Theory Tübingen University, WS 2016/2017 Lecture 11

A Study on Travel Time Value in Railway Transportation Corridors

The Choquet Integral with Respect to Fuzzy-Valued Set Functions

Final Examination Solutions 17/6/2010

Runoff Prediction with a Combined Artificial Neural Network and Support Vector Regression

BIOINF 585: Machine Learning for Systems Biology & Clinical Informatics

specials, or gone to the university bookstore and found that the texts for your course are on backorder? 204 / 481

Linear Regression Models

ENGI 4421 Confidence Intervals (Two Samples) Page 12-01

THE KALMAN FILTER RAUL ROJAS

Boosting. Professor Ameet Talwalkar. Professor Ameet Talwalkar CS260 Machine Learning Algorithms March 1, / 32

Information-based Feature Selection

IP Reference guide for integer programming formulations.

Introduction to Machine Learning DIS10

Approximating the ruin probability of finite-time surplus process with Adaptive Moving Total Exponential Least Square

The DOA Estimation of Multiple Signals based on Weighting MUSIC Algorithm

Stability Evaluation of Surfactant-Grafted Polyacrylamide Used in Oilfields by Evolution of Ammonia

10/2/ , 5.9, Jacob Hays Amit Pillay James DeFelice

Estimation of Population Mean Using Co-Efficient of Variation and Median of an Auxiliary Variable

Higher-order iterative methods by using Householder's method for solving certain nonlinear equations

Forecasting SO 2 air pollution in Salamanca, Mexico using an ADALINE.

Pattern recognition systems Laboratory 10 Linear Classifiers and the Perceptron Algorithm

A Study of the Collaborative Degree between Economic Growth and Carbon Reduction Targets under International Comparative Perspective

Short Term Load Forecasting Using Artificial Neural Network And Imperialist Competitive Algorithm

Machine Learning. Ilya Narsky, Caltech

6.867 Machine learning, lecture 7 (Jaakkola) 1

LOSS-MINIMIZATION CONTROL OF SCALAR- CONTROLLED INDUCTION MOTOR DRIVES

Classification of problem & problem solving strategies. classification of time complexities (linear, logarithmic etc)

The Random Walk For Dummies

8. СОВЕТУВАЊЕ. Охрид, септември ANALYSIS OF NO LOAD APPARENT POWER AND FREQUENCY SPECTRUM OF MAGNETIZING CURRENT FOR DIFFERENT CORE TYPES

Solution of Final Exam : / Machine Learning

Prediction of returns in WEEE reverse logistics based on spatial correlation

Investigating the Significance of a Correlation Coefficient using Jackknife Estimates

Hypothesis Testing. Evaluation of Performance of Learned h. Issues. Trade-off Between Bias and Variance

Open Access Research of Strip Flatness Control Based on Legendre Polynomial Decomposition

Comparison of Minimum Initial Capital with Investment and Non-investment Discrete Time Surplus Processes

To the use of Sellmeier formula

Mathematical Modeling of Optimum 3 Step Stress Accelerated Life Testing for Generalized Pareto Distribution

Presenting A Framework To Study Linkages Among Tqm Practices, Supply Chain Management Practices, And Performance Using Dematel Technique

The improvement of the volume ratio measurement method in static expansion vacuum system

REGRESSION WITH QUADRATIC LOSS

Lecture 4. Hw 1 and 2 will be reoped after class for every body. New deadline 4/20 Hw 3 and 4 online (Nima is lead)

Monte Carlo Optimization to Solve a Two-Dimensional Inverse Heat Conduction Problem

Polynomial Multiplication and Fast Fourier Transform

Machine Learning Theory Tübingen University, WS 2016/2017 Lecture 12

Machine Learning Brett Bernstein

Power Weighted Quantile Regression and Its Application

NEW IDENTIFICATION AND CONTROL METHODS OF SINE-FUNCTION JULIA SETS

Topics Machine learning: lecture 3. Linear regression. Linear regression. Linear regression. Linear regression

MBACATÓLICA. Quantitative Methods. Faculdade de Ciências Económicas e Empresariais UNIVERSIDADE CATÓLICA PORTUGUESA 9. SAMPLING DISTRIBUTIONS

STA Learning Objectives. Population Proportions. Module 10 Comparing Two Proportions. Upon completing this module, you should be able to:

Multi-variable weakening buffer operator and its application

Interval Intuitionistic Trapezoidal Fuzzy Prioritized Aggregating Operators and their Application to Multiple Attribute Decision Making

Machine Learning Assignment-1

RAINFALL PREDICTION BY WAVELET DECOMPOSITION

Bayesian Methods: Introduction to Multi-parameter Models

Binary classification, Part 1

Molecular Mechanisms of Gas Diffusion in CO 2 Hydrates

Analytic Theory of Probabilities

Chapter 6 Sampling Distributions

Transcription:

4th Iteratioal Coferece o Computer, Mechatroics, Cotrol ad Electroic Egieerig (ICCMCEE 05) The predictio of cemet eergy demad based o support vector machie Xi Zhag a,, Shaohog Jig b School of Electrical Egieerig Uiversity of Jia, Jia 500, Chia a 85436693@63.com, b cse_jsh@uj.edu.c Keywords:support vector regressio, cemet eergy forecast. Abstract. Aimig at the eergy waste problems resultig from the cotradictio betwee eergy demad ad supply i cemet productio, a eergy demad predictio model based o support vector regressio is built. Ad eergy cosumptio of a cemet eterprises i Shaxi cemet productio lie, for example, to validate the model simulatios show that the model has good predictive results.. Itroductio The cemet idustry has always bee oe of high eergy-cosumig idustries,whose eergy cosumptio accouts for about 8% of atioal eergy cosumptio,while it accouts for more tha 40% of the total cemet productio ad cosumptio. Now eergy shortages ad risig eergy prices have bee restrictig the developmet of the cemet idustry, so the cemet compaies must reduce eergy cosumptio to the miimum rage as much as possible, so as to reduce the cost of cemet productio to make the cemet have a price advatage i moder idustry, ad promote the rapid developmet of the cemet idustry. It is imperative to predict the eergy cosumptio of the cemet productio ad the cotrol it accordig to the results. This article establishes SVM model o the basis of the daily electric eergy cosumptio i April 04 i Shaxi Provice # cemet lie, ad predicts accordig to the model results, makes simulatio experimets usig MATLAB software, ad compare the performace with the gray predictio GM(,), ad make a coclusio that SVM ca better predict the eergy cosumptio i cemet productio.. Aalysis of Eergy Cosumptio i cemet compaies At preset cemet productio adopts basically ew dry productio techology. It has largely replaced the out processes, as wet process or kils, ad achieved certai results i terms of icreasig the productio of cemet compaies. But, the higher yield also brigs higher eergy cosumptio []. Depedig o the scale of cemet productio lie, there may be differeces i cemet techology, but basically there will be crushig ad pre-homogeizatio of raw materials, raw material preparatio, coal preparatio, cliker, cemet gridig, cemet packig ad delivery ad so o. Withi a certai rage, each step of the cemet busiess i eergy cosumptio ca directly reflect the scee of the eergy cosumptio of cemet. So forecastig the eergy cosumptio of every cemet productio lie ca reflect the eergy cosumptio of cemet each process to a certai extet ad reflect the eergy cosumed i cemet productio. The cemet compay maagers ca predict the results which based o cemet cosumptios to fid the cause of abormal cosumptio ad cotrol it. So they ca achieve effective developmet of cemet productio. The eergy cosumptio i the cemet productio process is a power ad coal. The two parts accouts for more tha 99% of the total eergy cosumptio of the etire productio. This oly predicts the electricity cosumptio of a sigle step i the cemet productio process. Every cemet productio eeds electricity. I the preset ew dry cemet productio, the electricity that really used for material for gridig is less tha 0% i the total electricity cosumptio of the cemet productio process, while the sum of the key facilities of eergy cosumptio o the total cemet productio process accouts for more tha 80% of the cemet productio process. Research o the eergy 05. The authors - Published by Atlatis Press 6

cosumptio of the sigle cemet productio process eables maagers systematically of cemet compaies uderstad utilizatio of eergy, i order to fid the cause of wasteful eergy usage. The fial result is that cemet compaies ca provide the data support i the improvig the eergy efficiecy of cemet. 3. Support Vector Machies Basic Itroductio Support vector machie (Support Vector Machie, SVM) is a ew ad efficiet machie learig based o the statistics theory. It has may advatages compared to eural etworks, for example, strog geeralizatio ability, easy to trai, o local miimum value ad there is o eed to pre-determie the etwork topology. SVM is a ew techology of data miig ad is primarily by meas of optimizatio methods. So it ca successfully deal with regressio (Support vector Machie for regressio, SVR) ad patter recogitio (support vector classificatio, SVC), ad may other questio. Besides, the SVM ca be exteded to the predictio field. At preset, the theory research ad applied research related to SVM have received extesive attetio domestically ad iteratioally. The predictio uses the oe of the SVM (Support Vector Machie for Regressio, SVR) []. There is a little differece of the basic idea betwee the SVR ad SVM. SVM is becomig to look for a optimal classificatio aspect, so that there is a error miimum from all traiig samples to the optimal classificatio aspect. The idea of realizig SVR is to use a o-liear mappig to map data x ito a high-dimesioal feature space, ad the implemet liear regressio i the space. 3. The establishmet of eergy demad predictio model. d Give traiig set{( xi, yi), i =,,3,..., }, which represets the quatity of samples, xi( xi R ) represets the iput colum vector ofi traiig sample, ad y i Rfor the correspodig iput values. SVR regressio fuctio expressios are as follows: f( x) = ωϕ ( x) + b () I the expressio, ϕ ( x) represets a oliear mappig fuctio, ω, b are the parameters that eed to be idetified i the system model [3]. Usig the structural risk miimizatio priciple, the parameters that eed to beidetified i the last formula is hadled as followig: R( ω) = Ce ( i ) + ω () i= I () formula above, R( ω ) is called empirical risk, ω called the cofidece risk, Ce ( i ) called the loss fuctio. Accordig the SVR priciple, i order to solve to () is equal to optimizatio of (3): mi ω + C ( xi + xi ) i= yi ωϕ( xi) b ε + xi (3) st.. yi ωϕ( xi) b ε + xi xi 0, xi 0, i =,,3,..., I which ε represets for precisio parameter of regressio fuctio, C for pealty factor, ad ξ i, ξi the relaxatio factor. max ( ai ai )( aj aj) K( xi, xj) + ( ai ai ) yi ε ( ai + ai ) ai, ai i, j= i= i= (4) ( ai ai ) = 0 st.. i= ai, ai [0, C] 63

I the above formula, K( xi, x j) is the core of fuctio. Accordig to the Mercer coditio, kerel is fuctio of models choose a Gaussia radial basis kerel fuctios. The kerel fuctio is put ito the formula, the: x i xj ( ) ( i i )exp( ) i= σ f x = a a + b (5) Calculatig regressio fuctio f( x) comes actually dow to the calculatio a i ad a i. The optimal solutio of ai ad a i ca be determied by miimizig the coditios. 3. Steps of Cocretig Cemet Eergy Predictio Model Based o Support Vector Machie. Costructio iput ad output data.use of SVM models eed to select the traiig samples ad test samples. Ad sample data comes from the database of cemet productio, ad the database stores the electricity that cosumed i the cemet productio process. Selectig the origial data from the database, icludig the eergy cosumptio data i a differet time (days, moths, ad years) ca be used for predictig eergy cosumptio of day, moth ad year. The build SVR iput ad output accordig to these data. Data preprocess.the raw data of cemet field has a big differece about the value. It may be a big chage if directly operated. This will reduce the accuracy of the predictio.so the raw data of database eed to be preprocessed. It meas data ormalizatio preprocessig usig (6). xi xmi x = (6) xmax xmi Selectio SVR model parameters ad performace evaluatio of models.svr predictio model eed to determie several parameters before the establishmet, amely a Gaussia kerel fuctio parametersσ ad two free parameters C adε.now there is o detailed descriptio of how to select these three parameters i the referece. But they all make the fial predictio result deviatio as model performace evaluatio stadard. Select the appropriate parameter values from experiece, make repeated cross-validatio ad the fializec = 4, σ = 0.35, soε = 0.0.Meawhile, i order to verify the effectiveess of cemet predictio model, daily electric eergy cosumptio of a cemet mill lie of a cemet factory i Shaxi Provice i April 04 is predicted accordig to the SVR predictio model established. Gettig the Fial SVR Model ad Predict Eergy Cosumptio of Cemet Compaies.After repeated experiece, we obtaied f( x ).Whe determie the eergy cosumptio of cemet scee i differet days, moths, or years, usig cemet eergy cosumptio predictio model, you ca obtai the correspodig cemet eergy cosumptio predictio (day, moth, ad year). 4. Simulatio Experimet Select a cemet mill lie from the database of a Shaxi cemet factory, ad make the power cosumptio data withi 30 days i April 04 as the simulatio data. Meawhile, combied with the MATLAB software, we ca realize cemet eergy predictio based o SVM. The cemet eergy data are as follows: The above data, a total of 30 data poits, ca be composed of 8 samples pairs of iput ad output. The iput sample set is as follows: { x(04.4.), x(04.4.), x(04.4.3)},{ x(04.4.), x(04.4.3), x(04.4.4)},{ x(04.4.3), x(04.4.4), x(04.4.5)},{ x(04.4.4), x(04.4.5), x(04.4.6)},...,{ x(04.4.8), x(04.4.9), x (04.4.30)} The output sample set is as follows: { x(04.4.4), x(04.4.5), x(04.4.6), x(04.4.7), x(04.4.8),..., x (04.4.30)}. I order to forecast cemet mill eergy with SVM, we select 5% data sample set. That is, 8 samples set are test sample set, ad 0 samples sets are the traiig sets. 64

Table Electricity Cosumptio i Shaxi # cemet mill lie from 04.4 Date Electricity Cosumptio (te Electricity Cosumptio (te Date thousad KWH) thousad KWH) 04.4. 9. 04.4.6 9.4 04.4. 3.5 04.4.7.57 04.4.3 34.8 04.4.8 6.4 04.4.4 6.34 04.4.9 38.5 04.4.5 8.7 04.4.0 46.0 04.4.6 46.8 04.4. 50.3 04.4.7 59.65 04.4. 5.6 04.4.8 48.68 04.4.3 48. 04.4.9 46.59 04.4.4 46.76 04.4.0 44.5 04.4.5 4.87 04.4. 4.67 04.4.6 44.34 04.4. 34.54 04.4.7 45.3 04.4.3 8.45 04.4.8 46.5 04.4.4 38.03 04.4.9 4.37 04.4.5 4.38 04.4.30 46.6 Fig SVR traiig sample set results Fig SVR predictio sample set results It s cemet eergy predictio resultwhich is show i Table below: Table Cemet eergy predictio result Date actual value Predictive value Relative error Absolute error 04.4.3 48. 46.78659 -.334.7% 04.4.4 46.7 45.9665-0.77335.65% 04.4.5 4.87 43.80643 0.93643.% 04.4.6 44.34 44.6588 0.388 0.7% 04.4.7 45.3 45.3808 0.0808 0.7% 04.4.8 46.5 45.9405-0.595 0.48% 04.4.9 4.37 43.56337.9337.8% 04.4.30 46.5 46.05074-0.996 0.43% 65

From the predictio results above, the errors that forecastig cemet eergy with SVM makes are withi the allowable rage. However, the absolute error of traiig sample is more 0.5078 tha the predictio sample. It is oly.9 ad withi the rage of allowable error. 4. Performace Compare. I order to verify the effectiveess ad feasibility of SVM forecastig the cemet eergy, we choose the gray GM (,) predictio method that has the greater advatage of small samples ad the compare the predictio results [4]. Figure 3 is a gray GM (,) predictio result: Fig 3 GM (,) Predictio Result The results that cemet eergy data get with GM (,) predictio method is show i Table 3: Table3 GM (,) predictioresults Date Actual value Predictive value Relative error Absolute error 04.4.3 48. 4.583-5.847.7% 04.4.4 46.7 4.690-4.0098 8.65% 04.4.5 4.87 43.66 0.566 0.6% 04.4.6 44.34 43.5674 0.776.74% 04.4.7 45.3 44.07 -.873.84% 04.4.8 46.5 44.466-0.595 0.48% 04.4.9 4.37 44.970.9337.8% 04.4.30 46.5 45.376-0.8739.89% You ca lear from the compariso of the data i Table ad Table 3 that, predictio accuracy of support vector machie SVR is better tha GM (,) predictio accuracy. It is proved that for the cemet eergy cosumptio predictio, SVM has more advatages. 5. Summary This article is maily aimed at eergy cosumptio i the cemet productio, we build the cemet eergy demad predictio model. Ad based o the historical data of cemet eergy cosumptio, we make the simulatio experimet for the daily electric eergy cosumptio of # cemet mill lie i a Shaxi Provice i April 04.The results proved that: SVR cemet eergy demad predictio model this article used has the good usability. The model overcomes the shortage of traditioal predictio method. It has better predictio ability tha the gray predictio method, ad the predictio accuracy is higher. It is a good referece to reasoably arrage for eergy cosumptio i the cemet productio process. Thus, i solvig cemet eergy demad predictio, SVR model has may advatages ad is very suitable for the research for the eergy cosumptio predictio i cemet productio process. Ackowledgemets This work was fiacially supported bymajor projectsofshadog Provice idepedet iovatio achievemets(04cgzh060) ad Chia-EU SMEs Cooperatio Fud for eergy coservatio research project (SQ03ZOC600003). 66

Refereces []Ceg Xuemi. Cemet idustry eergy cosumptio status ad eergy savig potetial [J]. Chia cemet, 006, (3): DOI:0.3969/j.iss.67-83.006.03.006. 6-. [] Yag Were. Research ad implemetatio of eergy maagemet system based o eergy cosumptio forecastig model [J]. South Chia Uiversity of Techology, 03 [3]Su Ha, Yag Purog, Cheg Jihua. Eergy demad forecastig model based o Matlab support vector regressio model [J]. system egieerig theory ad practice, 0, 3 (0): 00-007. [4] Zhou Ruipig. GM (,) model of grey predictio method to predict urba populatio [J]. Joural of Ier Mogolia Normal Uiversity: Natural Sciece Editio, 005, 34 (): 8-83. 67