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

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

Intermittent demand forecasting by using Neural Network with simulated data

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 :

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

ME 539, Fall 2008: Learning-Based Control

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

Information-based Feature Selection

Orthogonal Gaussian Filters for Signal Processing

Provläsningsexemplar / Preview TECHNICAL REPORT INTERNATIONAL SPECIAL COMMITTEE ON RADIO INTERFERENCE

Linear Associator Linear Layer

1 Review of Probability & Statistics

The AMSU Observation Bias Correction and Its Application Retrieval Scheme, and Typhoon Analysis

Multilayer perceptrons

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

New Particle Swarm Neural Networks Model Based Long Term Electrical Load Forecasting in Slovakia

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

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

Properties and Hypothesis Testing

Lecture 6 Chi Square Distribution (χ 2 ) and Least Squares Fitting

Chapter 9 - CD companion 1. A Generic Implementation; The Common-Merge Amplifier. 1 τ is. ω ch. τ io

Lecture 6 Chi Square Distribution (χ 2 ) and Least Squares Fitting

10-701/ Machine Learning Mid-term Exam Solution

OPTIMAL ALGORITHMS -- SUPPLEMENTAL NOTES

Chapter 13, Part A Analysis of Variance and Experimental Design

Structuring Element Representation of an Image and Its Applications

REGRESSION (Physics 1210 Notes, Partial Modified Appendix A)

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

A statistical method to determine sample size to estimate characteristic value of soil parameters

Application of Neural Networks in Bridge Health Prediction based on Acceleration and Displacement Data Domain

II. Descriptive Statistics D. Linear Correlation and Regression. 1. Linear Correlation

The Random Walk For Dummies

An Improved Proportionate Normalized Least Mean Square Algorithm with Orthogonal Correction Factors for Echo Cancellation

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

MCT242: Electronic Instrumentation Lecture 2: Instrumentation Definitions

On an Application of Bayesian Estimation

Infinite Sequences and Series

Investigating the Significance of a Correlation Coefficient using Jackknife Estimates

ECE 901 Lecture 12: Complexity Regularization and the Squared Loss

Algebra of Least Squares

THE SYSTEMATIC AND THE RANDOM. ERRORS - DUE TO ELEMENT TOLERANCES OF ELECTRICAL NETWORKS

Massachusetts Institute of Technology

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

Vector Permutation Code Design Algorithm. Danilo SILVA and Weiler A. FINAMORE

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

CEE 522 Autumn Uncertainty Concepts for Geotechnical Engineering

A proposed discrete distribution for the statistical modeling of

GUIDELINES ON REPRESENTATIVE SAMPLING

Modeling and Estimation of a Bivariate Pareto Distribution using the Principle of Maximum Entropy

7. Modern Techniques. Data Encryption Standard (DES)

Run-length & Entropy Coding. Redundancy Removal. Sampling. Quantization. Perform inverse operations at the receiver EEE

1. Linearization of a nonlinear system given in the form of a system of ordinary differential equations

Mechatronics. Time Response & Frequency Response 2 nd -Order Dynamic System 2-Pole, Low-Pass, Active Filter

6.883: Online Methods in Machine Learning Alexander Rakhlin

Measurement uncertainty of the sound absorption

Chapter 10: Power Series

Introduction to Artificial Intelligence CAP 4601 Summer 2013 Midterm Exam

RAINFALL PREDICTION BY WAVELET DECOMPOSITION

Output Analysis (2, Chapters 10 &11 Law)

On a Smarandache problem concerning the prime gaps

CHAPTER 4 BIVARIATE DISTRIBUTION EXTENSION

Sequences A sequence of numbers is a function whose domain is the positive integers. We can see that the sequence

Similarity Solutions to Unsteady Pseudoplastic. Flow Near a Moving Wall

EXPERIMENTAL INVESTIGATION ON LAMINAR HIGHLY CONCENTRATED FLOW MODELED BY A PLASTIC LAW Session 5

ADVANCED SOFTWARE ENGINEERING

WHAT IS THE PROBABILITY FUNCTION FOR LARGE TSUNAMI WAVES? ABSTRACT

The axial dispersion model for tubular reactors at steady state can be described by the following equations: dc dz R n cn = 0 (1) (2) 1 d 2 c.

Microscopic traffic flow modeling

Overdispersion study of poisson and zero-inflated poisson regression for some characteristics of the data on lamda, n, p

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

ARTIFICIAL NEURAL NETWORK MODELS FOR ESTIMATION OF SEDIMENT LOAD IN AN ALLUVIAL RIVER IN INDIA

Invariability of Remainder Based Reversible Watermarking

Goodness-of-Fit Tests and Categorical Data Analysis (Devore Chapter Fourteen)

FIR Filters. Lecture #7 Chapter 5. BME 310 Biomedical Computing - J.Schesser

Modification of Arrhenius Model for Numerical Modelling of Turbulent Flames

5. Fast NLMS-OCF Algorithm

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

A New Multivariate Markov Chain Model with Applications to Sales Demand Forecasting

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

(all terms are scalars).the minimization is clearer in sum notation:

Investigation of artificial neural network models for streamflow forecasting

Overview. p 2. Chapter 9. Pooled Estimate of. q = 1 p. Notation for Two Proportions. Inferences about Two Proportions. Assumptions

STUDY OF ATTRACTOR VARIATION IN THE RECONSTRUCTED PHASE SPACE OF SPEECH SIGNALS

Formation of A Supergain Array and Its Application in Radar

1 Efficient Splice Site Prediction with Context-Sensitive Distance Kernels

Double Stage Shrinkage Estimator of Two Parameters. Generalized Exponential Distribution

Lecture 10: Performance Evaluation of ML Methods

Appendix: The Laplace Transform


A Unified Approach on Fast Training of Feedforward and Recurrent Networks Using EM Algorithm

Finally, we show how to determine the moments of an impulse response based on the example of the dispersion model.

INFINITE SEQUENCES AND SERIES

Using Robust Extreme Learning Machines to Predict Cotton Yarn Strength and Hairiness

This is an introductory course in Analysis of Variance and Design of Experiments.

NUMERICAL METHODS FOR SOLVING EQUATIONS

µ and π p i.e. Point Estimation x And, more generally, the population proportion is approximately equal to a sample proportion

Pixel Recurrent Neural Networks

Varanasi , India. Corresponding author

FACULTY OF MATHEMATICAL STUDIES MATHEMATICS FOR PART I ENGINEERING. Lectures

THE KALMAN FILTER RAUL ROJAS

Transcription:

Iovative Productio Machies ad Systems D.T. Pham, E.E. Eldukhri ad A.J. Soroka (eds) 2008 MEC. Cardiff Uiversity, UK. Forecastig SO 2 air pollutio i Salamaca, Mexico usig a ADALINE. M.G. Cortia a, U.S. Medoza a, J.M. Barró-Adame a, D.Adia b, A. Vega-Coroa a a, Uiversidad de Guaajuato, Facultad de Igeiería Mecáica, Eléctrica y Electróica, México. b Uiversidad Politécica de Madrid, ETSI Telecomuicació, España. Abstract A compariso betwee a liear regressio model ad a No-liear regressio model is preseted i this work for forecastig of pollutio levels due to SO 2 i Salamaca city, Gto. Predictio is performed by meas of a Adaptive Liear Neural Network (ADALINE) ad a Geeralized Regressio Neural Network (GRNN). Predictio experimets are realized for 1, 12 ad 24 hours i advace, ad the results for liear regressio have bee satisfactory. The performace estimatio of both models are determied usig the Root Mea Squared Error (RMSE) ad Mea Absolute Error (MAE). Obtaied results are compared. The fial results idicated that ADALINE outperforms the past approach usig GRNN. Keywords: ADALINE, GRNN, SO 2 cocetratio 1. Itroductio Salamaca city is catalogued as oe of the most polluted cities i Mexico. The mai causes of pollutio i Salamaca are due to fixed emissio sources such as Chemical Idustry, ad Electricity Geeratio, beig the more importat pollutats i Air, Sulphur Dioxide (SO 2 ), measured i part per billio (PPB), ad Particulate Matter less tha 10 Micrometers i diameter (PM 10 ), measured i micrometers i diameter. This article focuses o forecastig SO 2 cocetratio. I a effort to fight pollutio of the zoe, i July 2005, the Program of Evirometal Cotigecy was lauched, the purpose of it beig to protect the health of populatio, especially that of vulerable groups. This program cotemplates the urget ad immediate reductio of SO 2 emissios ad PM 10 whe measuremets of these pollutats register levels above those established by Health Authorities. To accomplish it, 3 phases were established: Pre-cotigecy, Cotigecy Phase I ad Cotigecy Phase II for Sulphur Dioxide, PM 10 particles ad for a combiatio of both [1]. Predictio of pollutat cocetratios i the Atmosphere would allow takig prevetive measures, reducig the emissio of pollutats before reachig levels of a evirometal cotigecy. I this work, the use of a Neural Network ADALINE (ADA) is proposed to predict pollutio levels 1, 12 ad 24 hours i advace for the zoe of Salamaca before a evirometal cotigecy 232

occurs, ad results obtaied are compared with those obtaied with a Geeralized Regressio Neural Network (GRNN) [2]. 2. Methodology Figure 1 shows the flow diagram of the methodology that was followed for the realizatio of this work, which cosists of 3 mai phases: i) Select Traiig ad Test data sets, ii) Neural Network Desig ad iii)simulatio ad Results Evaluatio. system is used i the learig rule. ADA trasfer fuctio is a liear fuctio istead of a hard limit trasfer fuctio of the Perceptro. ADA ad the multiple versio Madalie (MADA) use a learig mechaism kow as Delta Rule of Widrow ad Hoff, also kow as the Least Square Mea Error (LSM) Rule [3], based o the search of the miimum error betwee the desired output ad the liear output obtaied. 2.2.1 Network Structure I geeral terms, the output fuctio of the etwork is give by equatio (1) T a W p (1) where a, is the output vector of the liear euros, W, is the weight matrix, ad p is the iput vector. 2.2.2 Learig Rule ADA is a supervised learig etwork that eeds a priori kowledge of the associated values to each iput, deomiated Widrow-Hoff Rule, also kow as LMS. The Rule of Widrow-Hoff i geeral terms is expressed as idicated i equatio (2). Fig. 1. Flow Diagram for SO 2 2.1. Traiig ad Test Data Sets Data base used for this experimet have bee previously processed as did i [2]. Data base selected to trai the et are those correspodig to the moths of Jauary ad February 2005, ad data selected to make the predictios are those of March 2005. 2.2. Neural Network Desig ADA is a geeralisatio of the perceptro traiig algorith. The mai fuctioal differece with the perceptro traiig rules is the way the output of the W(k + 1) =W(k) + 2 e(k)p T (k) (2) where k represets the curret iteratio of the weights updatig process, W(k+1) is the ext value that vector W is goig to take, ad W(k) is the curret weights vector; e is the vector of curret error, defied as the differece betwee the desired respose ad the etworks output show i equatio (1); is the learig rate. The gai updatig process is give by equatio (3) b(k + 1) = b(k) + 2 e(k) (3) where e is the vector represetig the error, b(k+1) is the Gai updatig vector, ad b(k) is the curret Gai Vector. 233

2.3. Traiig ad Simulatio of the et Iput euros are equal to the umber of observatios. I this work, iput euros are 1344 which is the total umber of observatios correspodig to the traiig period (Jauary ad February 2005). Test group cosists of 672 observatios correspodig to March 2005. Traiig vector have bee made oly to predict SO 2 cocetratios levels, that meas, o other variables were used to perform the forecast. Vector X, is the iput vector at times t=i--1,..., t=i-1, t=i, where i, is the curret hour ad is the umber of forecast hours is doe. Vector Y, is the output vector, whose elemets correspod to the estimatio of SO 2 levels at times t=i+1, t=i+2,..., t=, where i, is the curret hour ad is the umber of forecast hours is doe. Traiig ad simulatio for ADA ad GRNN were performed usig two differet Patter Schemes, sice the scheme used i [2] produces a apparet time-shift i the predictio made by ADA, due to how the patters for the traiig ad test matrices were formed. Due to this situatio, two differet Traiig Schemes were used, the secod oe to correct the apparet timeshift i the forecast for ADALINE etwork. These schemes were amed ADA I ad ADA II. I Traiig Scheme I (ADA I), iput patters x i are formed as idicated i [2], where i time t=i, each patter is X ={x i--1, x i-1, x i }, where x i is the SO 2 cocetratio i the curret hour, ad is the umber of forecast hours is doe. This meas that iput patters are formed with the curret ad past cocetratios. However, the first patter was formed by all zeros, sice for the first data, we had ot apriori iformatio. Output Traiig patters are formed with the ext hours cocetratios, Y={y i+1, y i+2,, y i+ }. I Traiig Scheme II (ADA II), patters were formed as i ADA I, but with the differece that for ADA II x 1 is equal to ADA I x 2, ADA II x 2 is equal to ADA I x 3 ad so o. Aother differece is that ADA I was formed with N patters, ad ADA II with N-1 patters. Due to the structure of the patters it is ecessary that we use N-1 patters. Evaluatio of the forecastig Performace was accomplished usig the Root Mea Squared Error (RMSE) ad the Mea Absolute Error (MAE). Mea-squared error is the most commoly used measure of success of umeric predictio, ad root mea-squared error is the square root of mea-squarederror, take to give it the same dimesios as the predicted values by themselves. This method exaggerates the predictio error - the differece betwee predictio value ad actual value of a test case of test case i which the predictio error is largest tha the others. If this umber is sigificatly greater tha the mea absolute error, it meas that there are test cases i which the predictio error is sigificatly greater tha the average predictio error. Balaguer et al. [4], have used RMSE as a idicator of the relatioship betwee predicted ad observed data. Root Mea Squared Error is computed accordig to equatio (4) RMSE 1 i 1 ( y i y i ) 2 (4) where i, is the predicted value for a determied time t=i, y i, is the real value for the same time ad is the umber of observatios. Mea Absolute Error is the average of the differece i all test cases. Mea Absolute Error (MAE) is computed accordig to equatio (5) 1 MAE i 1 y i y i (5) where i, is the predicted value for a give time t=i, y i, is the real value for the same time, ad is the umber of observatios. 3. Results Table I shows the obtaied results for schemes ADAI, ADA II, ad GRNN, proposed i [2]. GRNN was traied with Scheme I patter. 234

Table 1 Performace of a GRNN agaist a ADA Scheme GRNN ADA I ADA II Hours Ahead (PPB) 1 12 24 RMSE 59,6 58,05 73,94 MAE 34,67 43,72 53,1 RMSE 19,92 38,84 140,00 MAE 16,67 81,53 98,76 RMSE 19,92 58,36 58,36 MAE 16,67 36,13 36,13 The best results were achieved for the predictio of 1- hour ahead SO 2 cocetratios i both GRNN ad ADA etworks, which agrees with results obtaied by Medoza [2] ad Turias [5]. There is a sigificat improvemet usig the ADA II etwork sice both MAE ad RMSE errors are much lower tha those obtaied with GRNN. Results of 1-hr predictio are show i figures 2 ad 3. Fig 3. 1-hour forecastig with a ADA. I figure 5, for the case of SO 2 levels predictio with 12 hours ahead with ADA I, the predictio apparetly presets a time shift, which prevets gettig satisfactory results. This is due to the patters orgaizatio i this scheme. Fig 2. 1-hour forecastig with a GRNN Fig 4. 12-hours forecastig with a GRNN. Results obtaied with ADA II were much better tha those obtaied with ADA I ad GRNN, comparig them i Table 1, MAE ad also RMSE error were reduced, ad ADA II showed o time-shift for the predictio of SO 2. Figures 4, 5 ad 6 show results for 12-hour ahead predictio for the differet etworks 235

that were used. Fig 5. 12-hours forecastig with a ADA I. Fig. 7. 24-hours forecastig with a GRNN. The results for ADA I are show i the figure 8, where it is also time-shifted as results for 12-hr forecastig. Fig 6. 12-hours forecastig with aada II. For the case of 24-hr predictio, agai, ADA II scheme showed a better performace over the GRNN, ad ADAI. Figures 7 ad 9 show results for 24-hr predictio usig GRNN ad ADA II. Fig. 8. 24-hours forecastig with ADA I. 236

I both cases, error icreases as the umber of forecast hours is made icreases. It has bee show that the use of a liear regressio eural etwork improves the SO 2 predictio of cocetratio levels, reducig the error obtaied with a No-liear regressio eural etwork. Refereces [1] Istituto Estatal de Ecología. Calle Aldaa N.12 esquia calle Republica, Pueblito de Rocha, c.p. 36040, Guaajuato, Gto. Fig. 9. 24-hours forecastig with ADA II. Obtaied results i SO 2 forecast cocetratio levels with ADA show that the scheme of patters plays a importat role for obtaiig acceptable results. 4. Coclusios This work shows the compariso of the performace of a Liear Regressio Neural Network (ADALINE) ad a No-Liear Regressio Network (GRNN) to forecast cocetratio levels of SO 2. Oe of the mai differeces is, that a liear regressio etwork eeds less parameters adjustmet tha a Noliear regressio etwork, thus facilitatig its implemetatio, however, to obtai better results with a liear regressio etwork, it is ecessary to search for patter scheme that allows us reduce the error i the SO 2 predictio of cocetratio levels. ADA II outperformed GRNN i all the cases, showig that a appropriate patter Scheme must be used. [2] U.S. Medoza-Camarea, F. Ambriz-Coli, D.M. Arteaga-Jauregui ad A.Vega-Coroa. SO 2 cocetratios forecastig for differet hours i advace for the city of Salamaca, Gto., Mexico. 2005. [3] Mada M.. Gupta, Liag Ji, ad Noriyasu Homma. Static ad Dyamic Neural Networks, 2003. [4] Emili Balguer Ballester, Emilio Soria Olivas, ad Jose Luis Carrasco Rodríguez. Forecastig of surface ozoe cocetratioos 24 hours i advace usig eural etworks. Iteratioal coferece o Neural Networks ad Applicatios, 2001. [5] I. Turias, F.J. Gozalez, ad P.L. Galido. Applicatio of eural techiques to the modellig of time-series of atmospheric pollutio data i the campo de Gibraltar regio. Neural Network Egieerig Experieces, Procc. Of the 8 Iteratioal coferece o applicatio of eural etwork: 9-16, 2003. 237