Sie sind auf Seite 1von 6

Nonlinear Analysis of Daily Global Solar Radiation Time Series

Departamento de Ingeniera Elctrica y Trmica, Escuela Politcnica Superior, Universidad de Huelva, 21 1!, Huelva, Spain 2 Departamento de "sica #plicada, Universidad de #lmera, #lmera, Spain $ %orresponding #ut&or, ga'riel(lope)*die(u&u(es

G. Lpez1*, F. . !atlles" and .L. !os#$"

Abstra#t In t&is +or, +e analyse a daily glo'al irradiance time series into t&e -rame+or, o- c&aotic dynamic systems in order to e.amine a possi'le underlying nonlinear 'e&aviour( Employed met&ods are 'ased on a p&ase space reconstruction -rom t&e measured data and are devoted to t&e calculation ot&e properties o- an underlying attractor, suc& as t&e /yapunov e.ponents( 0esearc&es on t&ese dynamical system invariants +ill point out t&e presence o- c&aos( 1e also use local lineal models as a test -or nonlinearity( T&e glo'al solar radiation data +ere measured at t&e radiometric station o- t&e University o#lmera 2Spain3 during eig&t years( 0esults &ave s&o+n t&e non4e.istence o- any attractor in t&e p&ase space -or t&e glo'al irradiance time series( 5egative /yapunov e.ponents e.clude a c&aotic 'e&aviour t&at mig&t allo+ a 'etter s&ort term prediction t&an autoregressive models, and t&e idea o- t&e e.istence o- a nonlinear di--erential e6uation system( T&ese results matc& +it& t&ose o'tained -rom applying local linear models -or prediction, o- +&ic& estimations suggest t&at t&e data are 'est descri'ed 'y a linear stoc&astic process( 7ey+ords8 solar radiation, -orecasting, c&aos, nonlinear time series(

1. %ntrod&#tion
In-ormation on t&e availa'ility o- solar radiation is needed in many applications dealing +it& t&e e.ploitation o- solar energy( Particularly, glo'al solar radiation is one o- t&e most important input parameters -or any solar energy system and di--erent tec&ni6ues &ave 'een developed to model and -orecast it( 9nce t&e solar energy system 2li,e concentrated solar po+er plants3 is running, prediction o- po+er load, normally done on an &ourly 'asis +it& a prediction &ori)on 'et+een 1 and 2: &ours, is instrumental -or planning and operation o- t&e total po+er system, e(g( -or 'uying or selling po+er or -or solving t&e unit commitment and dispatc& pro'lems( Studies a'out solar radiation time series and ot&er meteorological varia'les +it& autoregressive or stoc&astic models ;14<= &ave e.perienced an important gro+t& in t&e last years, since synt&etic se6uences statistically indistinguis&a'le -rom t&e original ones are needed to design t&e solar devices properly( Ho+ever, using t&is type o- model, prediction -rom past values is limited due to t&e random c&aracter e.&i'ited 'y t&ese time series( T&e main -actor o- randomness a--ecting glo'al solar radiation data is due to variations in t&e s,y cloud cover, +&ic& ma,e t&e radiation on cloudy days di--icult to predict( #ppearing o- several nonlinear dynamical models developed 'y /oren) ;:=, +it& a -ully irregular and comple. 'e&aviour 2generically named c&aos3, allo+s to study t&e nature o- -luctuations in solar radiation data -rom a ne+ met&odology( Time series -rom many system evolutions, apparently evolving in a random +ay, can no+ 'e predicted +it& &ig&er precision t&an +it& traditional #0># models, at least -or s&ort term predictions( Detection o- c&aos in a time series +ould t&us allo+ a 'etter modelling o- time series against statistically 'ased models( In t&is paper, +e analyse a daily glo'al solar irradiance time series into t&e -rame+or, o- c&aotic dynamic systems in order to e.amine a possi'le underlying nonlinear 'e&aviour( E.istence oc&aos s&ould &elp to improve t&e s&ort term predictions per-ormed 'y means o- t&e traditional statistical tec&ni6ues( Employed met&ods are 'ased on a p&ase space reconstruction -rom t&e

measured data and are devoted to t&e calculation o- t&e properties o- an underlying attractor, suc& as t&e /yapunov e.ponents( 1e also use local lineal models as a test -or nonlinearity(

". '(perimental Data


T&e &ori)ontal glo'al solar radiation time series 2sym'oli)ed 'y ? xt@, 'eing t t&e time step in days3 consisted o- 2!:A daily values( T&ey +ere o'tained 'y integration -rom e.perimental values averaged every ten minutes during t&e years 1!!B41!!2 and every -ive minutes during t&e years 1!!<41!! ( T&e measurements +ere recorded at t&e radiometric station o- t&e University o#lmera 2<C( D 5, 2(::D 13 in sout&4eastern Spain 'y means o- a 7ipp and Eonen %>411 pyranometer( >easurements -rom ot&er t+o pyranometers 2/I4%90, model /I42BB3 +ere also availa'le( T&ey +ere used to detect and replace any anomalous value measured 'y t&e %>411 pyranometer( T&e cali'ration constants o- t&e pyranometers +ere c&ec,ed periodically 'y our researc& group( >easurements 'y t&e /I42BB pyranometers +ere also corrected -ollo+ing ;A=(

). *et$odolo+y and res&lts


).1 ,$ase spa#e re#onstr&#tion T&e -irst step in any analysis o- c&aotic data is to reconstruct an attractor 2de-ined as set o- points in p&ase space visited 'y a signal traFectory a-ter transients are gone3 -rom t&e data( Ta,en ;C= suggested a met&od o- p&ase space reconstruction t&at is ,no+n as t&e method of delays( T&is met&od consists o- em'edding t&e time series into a d4dimensional space, d, +&ic& is e6uivalent to t&e original un,no+n p&ase space composed o- all t&e dynamical varia'les( "igure 1a s&o+s t&e lag space plot -or t&e glo'al solar radiation time series using a lag o- one( Points are distri'uted around a :AD straig&t line denoting a periodic nonc&aotic time series( T&is is also evidenced o'serving t&e correlogram s&o+ed in "ig( 1'( #n annual cycle o- <CA days is presented( T&is is an e.pected result as t&e daily values o- solar radiation are mainly a--ected 'y t&e eart& motion(
<A <B 2A
2

t41

#utocorrelation coe--icient

1(B B(A B(B 4B(A 41(B


'3

3 x t+ 1 2>GHm
2B 1A 1B A B

a3
B A 1B 1A 2B 2A <B <A
2

<CA

xt 2>GHm 3

/ags 2days3

I<B

1B!A

1:CB

"ig( 1( a3 P&ase space structure o- t&e glo'al solar radiation time series using lags o- oneJ '3 %orrelogram -or t&e glo'al solar radiation time series(

In order to remove periodicity and trend, a trans-ormation o- t&e time series is underta,en 'y di--erencing( T&e ne+ data are arit&metic di--erences 'et+een pairs o- o'servations using a lag oone day 2Kxt L xt+1 M xt3( T&e di--erenced time series is denoted as ?K xt@( "igure 2a s&o+s t&e p&ase space plot -or t&e di--erenced time series( T&e periodic pattern s&o+n in t&e a'ove -igure seems to 'e removed( T&is result is corro'orated analysing t&e correlogram displayed in "igure 2'( #utocorrelation coe--icients -all to +it&in t&e random4li,e )one a-ter t&e -irst lag and, -or t&e most part, remain in t&at )one t&erea-ter( It is t&us assumed t&at periodicity is removed( #lt&oug& point distri'ution in "ig( 2a seems to 'e due to some non4random comple. underlying 'e&aviour, t&e presence o- a c&aotic attractor in t&at p&ase space is not clear( To searc& t&e

optimum time delay +e &ave used t&e met&od o- mutual information ;I=( >utual in-ormation, li,e autocorrelation, tries to measure t&e e.tent to +&ic& values K xt+m are related to values o- Kxt, at a given lag( Ho+ever, mutual in-ormation &as t&e advantage o- using pro'a'ilities, rat&er t&an a linear 'asis, to asses t&e correlation and t&us, nonlinear correlations are ta,en into accounts( T&e so-t+are implementation o- t&at algorit&m 2and o- t&ose used &erea-ter3 is -rom t&e TISE#5 pac,age ; =(
2A 2B 1A 1B A B 4A 41B 41A 42B 42A 42A 42B 41A 41B 4A B A 1B 1A 2B 2A
t41

a3

#utocorrelation coe--icient

B(A

'3

x t+ 1 2>GHm 2 3

B(B

4B(A

AB

1BB

1AB

xt 2>GHm 3

/ags 2days3

2BB

2AB

<BB

<AB

:BB

"ig( 2( a3 P&ase space structure o- t&e di--erenced time series using a lag o- oneJ '3 %orrelogram -or t&e di--erenced time series 2das&ed lines are !A percent con-idence limits3(

#s a prescription, a good candidate -or t&e time delay +ill 'e +&ere t&e -irst mar,ed minimum omutual in-ormation occurs( "rom "ig( < it is noted t&at t&e -irst minimum is -or L <( 9nce t&e time delay is c&osen, t&e ne.t step is to select t&e em'edding dimension( 7ennel et al( ;!= introduce t&e met&od called t&e false nearest neighbors, +&ic& calculates t&e minimal em'edding dimension( T+o points in a d4dimensional p&ase space are -alse neig&'ors, in t&e sense o- some distance -unction, +&en t&e distance, NNxid M x-dNN, is small 'ut t&e distance NNxi.1dO1 M x-.1dO1NN in a dO14dimensional p&ase space is not( Piven a distance -unction, t&e Euclidean distance in our case, and some t&res&old si)e, t&e percentage o- -alse near neig&'ors 'ecomes )ero as t&e dimension op&ase space goes to t&e minimal em'edding dimension( "igure : s&o+s t&e percentage o- -alse nearest neig&'ors as a -unction o- t&e em'edding dimension d( T&e main c&aracteristic is t&at t&e -raction o- -alse nearest neig&'ors does not -all to )ero as d increases( T&is implies t&at t&e di--erenced time series &as residual QnoiseR in it, and t&us t&e e.istence o- lo+ dimensional c&aos is to 'e discarded(

B(2B >utual in-ormation B(1A B(1B B(BA B(BB

1B 2B <B Time delay, in days

:B

"ig( <( >utual in-ormation as a -unction o- lag -or t&e di--erenced time series(

1(B
Percentage o- -alse nearest neig&'ors

B( B(C B(: B(2 B(B 1 2 < : A C I ! 1B 11 12

Em'edding dimension, d

"ig( :( Percentage o- -alse nearest neig&'ors as a -unction o- t&e em'edding dimension -or t&e di--erenced time series(

)." Lo#ally linear predi#tion T&e easiest nonlinear met&od o- local prediction +as developed 'y /oren) ;:= and called method of analogies( /et 'e xt a point +it&in t&e d4dimensional p&ase space, t&e predicted value a time T later, xt+T, +ill 'elong to some ,ind o- interpolation 'et+een ,no+n points xt1+T, (((, xtr+T, +&ere xj1, (((, xtr are t&e nearest r points to xt( "armer and Sidoro+ic& ;12= propose an en&anced met&od 'y ma,ing a least4s6uared -it o- a local linear map o- xt into xt+T( Doing T L 1, +e can o'tain t&e predicted value as xt+1 L atxt O bt 'y minimising8
W =

xt Ni

t+ 1

at xt bt

213

in regard to at and bt( Nt is t&e 4neig&'or&ood o- xt( %asdagli ;1B= suggests to use t&ese models as a test -or nonlinearity prediction, o'taining t&e average -orecast error given 'y8

< ; xt xtT =2 > E =< E >= < ; xt < xt > = 2 >


2 2

223

+&ere UVW is t&e arit&metic average( I- E L B t&en t&e prediction is per-ect, and i- E L 1, t&e prediction is no 'etter t&an a constant predictor in t&e time series average( I- E is calculated -or di--erent 4values, occurring t&e optimum -or large neig&'our&ood si)es, t&is +ill indicate t&at t&e data are 'etter descri'ed 'y a linear stoc&astic process in t&is em'edding space( 9n t&e ot&er &and, an optimum at rat&er small si)es +ill point out t&e e.istence o- a nonlinear e6uation o- motion( T&e average -orecast error o- local linear models is plotted as a -unction o- t&e neig&'our&ood si)e in "ig( A( 5ote t&e &ig& value o- E and &o+ t&e minimum value o- E occurs at &ig&er neig&'or&ood si)es( T&is result is associated +it& stoc&astic processes(

1(B

#verage -orecast error, E

B( B(C B(: B(2 B(B

1BBB

1BBBB

5eig&'or&ood si)e,

"ig( A( #verage -orecast error o- t&e local linear modelas a -unction o- t&e neig&'or&ood si)e -ot t&e di--erenced time series(

).) Lyap&no/ e(ponents %&aotic dynamical systems are predicta'le -or -inite time scales, 'ut not -or in-inite ones( /et consider t+o initial states o- t&e system +it& almost identical initial conditions( T&eir traFectories in t&e p&ase space +ill move apart, given a metric, at an e.ponential rate +&ile moving on an attractor( T&is rate is descri'ed 'y t&e largest /yapunov e.ponent in t&e system, ma.( I- any /yapunov e.ponent is positive, t&e system +ill 'e c&aotic 'y de-inition( Xet, s&ould one o- t&em 'e )ero, t&e system can 'e descri'ed 'y a set o- ordinary di--erential e6uations ;1B=( "igure C s&o+s t&e d /yapunov e.ponents -or t&e di--erenced time series as a -unction o- t&e num'er o- samples -ollo+ing t&e suggested algorit&m 'y Sano and Sa+ada ;11=( T&e main interest in t&is test is to -ind /yapunov e.ponents greater or e6ual t&an )ero( T&e t+o /yapunov e.ponents e.&i'ited 'y t&e time series are negative and t&us, t&is time series cannot 'e considered coming -rom a c&aotic dynamical system neit&er 'e descri'ed 'y a set o- di--erential e6uations(

B(A B(B 4B(A

41(B 41(A 42(B B <BB CBB !BB Samples 12BB 1ABB

"ig( C( /yapunov e.ponents against t&e num'er o- samples(

0. 1on#l&sions
T&e glo'al solar radiation data +ere measured at t&e radiometric station o- t&e University o#lmera 2Spain3 during eig&t years( 0esults &ave s&o+n t&e non4e.istence o- any attractor in t&e p&ase space -or t&e glo'al irradiance time series( 5egative /yapunov e.ponents e.clude a c&aotic 'e&aviour t&at mig&t allo+ a 'etter s&ort term prediction t&an autorregresive models, and t&e idea o- t&e e.istence o- a nonlinear di--erential e6uation system( T&ese results matc& +it& t&ose

o'tained -rom applying local linear models -or prediction, o- +&ic& estimations suggest t&at t&e data are 'est descri'ed 'y a linear stoc&astic process(

A#2no3led+ements
T&is +or, +as supported 'y t&e proFect E5E2BBI4CI :!4%B24B2 o- t&e >inisterio de %iencia y Tecnologa o- Spain(

Referen#es
;1= 0( "esta, S( Gain, %("( 0atto, 0ene+a'le Energy 2 21!!23 2<( ;2= G( Soland, Solar Energy AA 21!!A3 <II( ;<= H( >or-, Solar Energy C2 21!! 3 1B1( ;:= E(5( /oren), G( #tmos( Sci( 2B 21!C<3 1<B( ;A= /( #lados4#r'oledas, "(G( Satlles, "(G( 9lmo, Solar Energy A: 21!!A3 1 <( ;C= "( Ta,ens, 21! 13( Dynamical Systems and Turbulence, /ecture 5otes in >at&, Springer4Yerlag, Serlin( ;I= #(>( "raser and H(/( S+inney, P&ys( 0ev( #, << 21! C3 11<:( ; = 0( Hegger, H( 7ant), and T( Sc&rei'er, %&aos, ! 21!!!3 :1<( ;!= >(S( 7ennel, 0( Sro+n, and H(D(I( #'ar'anel, P&ys( 0ev( #, :A 21!!23 <:B<( ;1B= >( %asdagli, G( 0oy( Stat( Soc(, A: 21!!13 <B<( ;11= >( Sano and X( Sa+ada, P&ys( 0ev( /ett(, AA 21! I3 1B 2( ;12= G(D( "armer and G(G( Sidoro+ic&, P&ys( 0ev( /ett(, A! 21! I3 :A(

Das könnte Ihnen auch gefallen