ADAPTIVE CONTROL STRATEGIES IN WIND-DIESEL HYBRID SYSTEMS. P. S. Panickar*, S. Rahman**, S. M. Islam**, T. L. Pryor*

Transcription

1 ADAPTIVE CONTROL STRATEGIES IN WIND-DIESEL HYBRID SYSTEMS P. S. Paickar*, S. Rahma**, S. M. Islam**, T. L. Pryor* *Murdoch Uiversity Eergy Research Istitute, Wester Australia ** Departmet of Electrical ad Computer Egieerig, Curti Uiversity, Wester Australia Abstract A wid geeratig system could be better ad more itelligetly operated if the wid availability ca be predicted. These predictios would help i plaig ad cotrollig diesel egies or covetioal geerators to optimize the hybrid system operatio. A statistical method has bee employed to predict wid speeds usig historical wid speed data. This study explais the methodology to compute a oe step ahead predictio of wid speed whe previous values of wid speed is available. A method of predictig more tha oe step ahead values of wid speed has also bee proposed. Results of differet approaches are compared ad assessed for their figures of merit ad applicatio i plaig ad cotrol of the power system operatio. Beefits of shortterm wid speed predictio o diesel dispatch/commitmets are discussed i the cotext of a typical wid-diesel hybrid power system operatioal sceario i Wester Australia. 1. INTRODUCTION Wid-Diesel hybrid power systems are particularly suited for locatios where wid resource availability is high ad the cost of diesel fuel ad geerator operatio cotrol the cost of electrical eergy supplied. As a result of turbie developmets the ecoomics of wid power have ow become competitive with covetioal power source. The ecoomy of operatio of wid turbies is critically depedet o the wid speeds at the site. If the wid turbie is used alog with a diesel egie, the cost of power geeratio could be reduced, i additio to reducig greehouse gas emissio problems. The wid geeratig system could be better ad more itelligetly operated if the wid availability ca be predicted. These predictios would help i plaig ad cotrollig diesel geerators to optimize the system operatio. This will help ot oly i fuel savigs ad greehouse gas reductios but also i extedig the life spa of the diesel egies. Predictio of wid speed is ot a easy task due maily to the fact that wid is the result of a multitude of meteorological factors viz. temperature, pressure variatio, solar radiatio etc. Sice the wid power geerated is a fuctio of the cube of the wid speed, error i predictio ca lead to a major effect o the performace of the system [1]. The curret study aalyses two methods of wid speed predictio ad compares the forecast by estimatig the correlatio coefficiet, Mea Absolute Percetage Error ad Root Mea Square Error of the predicted wid speed values i both the cases. The first method cosiders oly the previous wid speed values to predict the wid speed value of a istat, which is oe sample poit ahead. The secod method predicts the future wid speed values from the previous same hour of a 2-hour data set givig a full day predictio horizo. 2. WIND SPEED PREDICTION MODELING The primary objective of predictig wid speed is to prepare a pla i diesel dispatchig. A twety-four hour predictio horizo would help i a diesel uit commitmet strategy, which i day to day plaig ca save fuel as well as maiteace cost for the power system. A loger-term predictio ad plaig seems ambitious ad impractical due to sudde chages i weather patters iflueced by radom variatios i the behaviour of wid. The quality of predictio would deped o the methodology ad the selectio of appropriate iputs i the computatio model. Ofte o-availability of eough data sets, iflueced by techological ad cost costraits of data collectio equipmet, restricts the umber of iputs to be used, affectig the accuracy of predictio. The popular predictio methods are statistical models, eural etwork or fuzzy logic techiques. Ofte the accuracy of predictio is overshadowed by the system operatioal costraits, viz. miimum operatig load ad ruig time for the diesel egies ad spiig reserve regulatios. Thus for example idlig a diesel egie at the miimum load would be

2 ecoomically more sesible tha shuttig it off ad brigig it o withi a short time spa. Thus operatioal costraits may overwrite the diesel dispatch strategy based o the predicted sceario. A West Australia hybrid power system site is chose for the study for which historical wid speed data is available. The system cosists of two diesel egies ad a wid turbie geerator as illustrated i Figure 1. Itegrated Wid-Diesel Hybrid System Wid Turbie 25 Diesel 1 2 Voltage Regulator Sychroous Geerator Diesel 2 5 Rectifier Sychroous Geerator Voltage Regulator Sychroous Geerator Voltage Regulator Battery Dump Load Ivertor Load The available data were the cotiuous hourly wid speed ad load for oe year. The data was closely aalysed for diural ad seasoal variatios. Figure 2 gives actual load variatio for the sample site for two weeks. Figure 3 gives the actual wid speed durig the same period. 3. WIND TURBINE POWER CURVE The power curve provides the power output spectrum of a wid turbie geerator (WTG) for the correspodig wid speeds. Figure shows the power curve for the wid geerator used i the study. Below the cut-i wid speed (V c ) the geerator does ot produce ay power ad hece the speed variatios are immaterial. Above the rated wid speed (V r ) the turbie produces the same rated power irrespective of the wid speed. Thus the critical rage is betwee the cut-i speed ad the rated speed whe the power output variatios are very sesitive to the wid speed. W r 3 25 Power Curve for a variable speed wid turbie V c wid velocity m/s V r The adaptive cotrol strategies are applicable i this rage of wid speed. Figure 5 shows the power output from a 25 wid turbie for a 2-hour period. Fig. WTG Power Curve for a 25 uit m/s Fig 1. Hybrid system layout used i the study Fig 2. Load variatio over two weeks 2 Wid Speed Wid Power m/s Fig 3. Wid speed for two weeks Fig 5. Wid Speed ad Power Output For a ideal wid-diesel system, dispatch strategy would be K = W + P1+ P2 ie., K W P1 P2 = (1)

3 where K is the system load (icludig losses), W = Wid power, P1= Diesel egie-1 power ad P2 = Diesel egie-2 power. Assumig a load forecastig error withi ±5%, the dispatch strategy would be to predict the wid power iput so that the coditio as give i equatio (1) ca be met by selectig the appropriate loadig o oe or both diesel geerators. Thus the wid forecastig is importat i desigig a predictive cotrol strategy for the hybrid system.. PREDICTION MODELS The cotrol strategy would deped o the predictio horizo, which may vary from oe hour ahead to several hours or days ahead. I the preset study two models are proposed, oe for a oe-hour ahead predictio ad the secod for a oe full day (2 hours) ahead predictio. A diesel dispatch plaig based o a full day plaig is ideal but practical cosideratios as described later i the paper requires a shorter term predictio. Thus a full day predictio ivolvig a adaptive cotrol whereby the ext hourly dispatch is cotiuously updated to adapt the immediate past variatios is probably the best strategy. The output wid speed waveform i the autoregressive (AR) model ca be geerally expressed as a fuctio of the preset ad past wid speed values [2, 3] as follows. y(t) = -a 1 y(t-t) - a 2 y(t-2t)- - a y(t-t) (2) where y(t) is the future (oe step ahead) predictio of the wid speed, a are the liear predictio coefficiets, T is the sample time iterval ad y(t-t) are the past values of the wid speed correspodig to delays of T. The coefficiets are calculated from a selected modellig widow of past observatios. The least square error method is employed for this calculatio..1 Model 1: Oe hour ahead Predictio This model employs the cotiuous hourly wid speed data values. Oe step ahead predictios were computed for a whole year based o the past seve ad fourtee days of iput data. To compute the predictio values for a whole year the program employed a slidig widow techique, ie., the coefficiets were calculated based o the immediate past seve or fourtee days (ie., m=7 or 1) of data. This techique takes care of seasoal variatios of wid speed values. The predictios were made usig three coefficiets (ie. =3 i equatio 2). Various umber of coefficiets were tried ad =3 was foud to be the best choice. A sample oe-hour ahead predictio result for a day is give i Figure 6. The figures of merit used to assess the results of the predictio were Correlatio Coefficiet, Mea Absolute Percetage Error (MAPE) ad Root Mea Wid Speed m/s ( X Y ) Percetage Error, PE = x1 X Square Error (RMSE) betwee the actual ad predicted wid speeds. ( X ( X X )( Y Y ) r = ; where ( X X ) ad ( Y Y ) X ) 2 * ( Y Y ) 2 Fig 6. Sample oe-hour ahead predictio Correlatio Coefficiet, r betwee two variables X ad Y, as described by the product momet formula is are deviatio of X ad Y values from their mea of umber of X ad Y variables. Values of r closer to +1 MAPE = Oe hour ahead predictio: m=3, =7 Corr. Coef. =.6, MAPE = 9.1, RMSE=1.27m/s 2 Actual WS Predicted WS i= 1 i = 1 RMSE = idicate perfect positive correlatio. PE i ( Actual Pr edicted) The predictio was doe for a full year usig a Matlab program with a slidig widow of seve or fourtee days of iput data for coefficiet calculatio. The results were assessed for a whole year as well as for four three-mothly quarter periods of the year. Table 1 gives the cosolidated figures comparig the two cases with =7 ad 1. The predicted results were compared with actual values of wid speed usig Excel spreadsheet software. The results idicate a margially better correlatio coefficiet whe m=1. The results also 2

4 idicate seasoal variatios. The MAPE varied betwee 1.67 ad These figures are very high for the predictio. Lower figures of MAPE idicated better predictio. RMSE idicated the mea error i wid speed values over the desigated period. With oe-hour ahead predictio the dispatch strategy eeds to be cotiuously updated o a hourly basis. A loger-term predictio would be preferred, say for at least oe full day so that a reasoable plaig strategy for the ext 2 hours could be made. Table 1: Compariso of predictio results Figure of Merit for oe poit ahead wid predictio m=3 =7 m=3 =1 FOM 1 Year Qrt 1 Qrt 2 Qrt 3 Qrt Corr-Coeff MAPE RMSE Corr-Coeff MAPE RMSE Table 2: 2-hour predictio results 2 Hour Predictio Results: m=3, =7days FOM 1 Year 1st Qtr 2d Qtr 3rd Qtr th Qtr Corr. Coefficiet (r) MAPE RMSE Model 2: 2 hour predictio It should be metioed here that the System Idetificatio toolbox i Matlab is capable of computig more tha oe sample poit ahead predictio. Usig the previous wid speed values (icludig predicted) as iput variable, the sample 2-hour predictio gave very high MAPE (9.91) ad RMSE (.m/s), which is practically uacceptable. However, other factors like pressure, geographic locatio, humidity, solar irradiatio data ad may others which are directly or idirectly related to the wid speed are ot always available ad it is difficult to establish the exact correlatio betwee them ad the wid speed. A approach ca be adopted to expad the predictio horizo by lookig at the hourly wid speed variatio for each temporal istat (time istat) of the day throughout the year. For example, O clock midight data for the past 7 days may be similarly modelled to predict the ext days O clock wid speed data. Likewise, from the data set of other time istats, the correspodig wid speed values ca be predicted. Thus the predictio horizo could be expaded ad this would perhaps take the other wid speed determiig factors like time of the day effects ito accout. This model also uses the wid speed data of the first model. However, the data was coverted to twety-four sets of hour-wise data. Predictio is made usig the auto-regressive model with m=3 ad =7 ad 1. I choosig the model, a umber of coefficiets have bee tried, ad the correlatio coefficiets were calculated for each set of the predicted ad actual wid speed data values. Fially, the oe, which showed the higher correlatio coefficiet, was chose for the wid speed predictio. Figure 7 gives a sample predictio result for a 2-hour period based o model 2. 2 hour ahead predictio : m=3, =7 Corr. Coeff. =.63, MAPE =.76, RMSE = 1.67m/s 2 Actual WS Predicted WS Wid Speed m/s Fig 7. Twety four hour ahead Predictio

5 The results for the 2-hour predictio (Table 2) idicate that the accuracy of predictio is much lower tha for the oe poit ahead predictio. The advatage here, though, would be to have a larger predictio horizo. But with more iputs i the computatio, the predictio accuracy could be improved, thus helpig i plaig a predictive dispatch strategy. 5. HYBRID POWER SYSTEM CONSTRAINTS The wid eergy cotributio i the hybrid power system ca reduce the level of operatio of the diesel egies. The scope for diesel savigs become sigificat oly i a multi-diesel egie sceario where oe or more of the diesels ca be shut dow depedig o the wid eergy iputs. Due to reasos related to security of power supply or to avoid loss of load, provisio is always made for spiig reserve, which is defied as the capacity of the statio to pickup the immediate loss of the largest uit i service []. The diesel egies are ever allowed to ru below a certai percetage, say %, of its rated power, to avoid critical maiteace problems. The there are the startig ad idlig cost aspects, cotributig to the total system operatioal ecoomics [5]. Thus predictive cotrol of the hybrid system becomes sigificat i the short-term operatio sceario. Wid speed ad load forecastig ca help pla a dispatch strategy for the diesel egies. Wid speed ca be coverted to correspodig power available from the wid turbie usig its power characteristics. is purely based o the differetial load (system load -wid power) to be met at that hour. But whe operatioal aspects like miimum operatig times (usually oe hour) ad startig costs are take ito cosideratio it may be practical to have a differet strategy of operatio as show i figure Diesel 1(2 ) WTG- Differetial Load Hour Diesel2 (5 ) System Load Fig 9. Diesel dispatch strategy o.2 The secod sceario allows the secod, larger diesel to operate durig peak load hours, avoidig itermittet diesel o ad off, thus reducig startig costs ad other maiteace related costs. 6 2 Diesel 1(2 ) WTG- Differetial Load Diesel2 (5 ) System Load 6 Diesel1(2) WTG- Differetial Load Diesel2 (5 ) System Load Hour Fig 1. Diesel dispatch o.3: Low wid 2 Hour Fig. Diesel dispatch strategy o.1 Figure gives a sample diesel dispatch i a 2 hour sceario where there is o predictio of wid ad load, ad the diesel operatioal costraits are limited to % miimum loadig limits. I this sceario it may be oted that the two diesels are forced to operate itermittetly sice their loadig The third sceario (Figure 1) ivolves both the diesels operatig together for two hours durig the peak load hours. A predictive cotrol strategy is relevat i such scearios whereby idividual diesel loadig is plaed ahead to effect a optimal operatio. The idividual diesel loadig is based o the differetial loads for each hour of the day. Figure 11 gives a combied graphical view of the actual, oe-hour ad 2 hour ahead predicted WTG output ad the differetial load to be met by the diesels. This would provide a uderstadig of the effect of predictio o the loadig of the idividual diesel egies. The forecasted sceario ca help i plaig a diesel commitmet strategy for the

6 period. Table 3 gives a compariso of the 2-hour diesel dispatch for the two predictio scearios. Diff. Loadact Diff. Load1pt Diff. Load2pt WTG act WTG 1pt WTG 2pt Fig 11. Diesel Dispatch for 2 hours Table 3. Differetial Load Results for 2 hours Differetial Load MAPE RMSE 1 hour ahead Predictio Hour ahead Predictio uit to the other based o the differetial load requiremet. Sice every diesel egie has a warmup time before it is ready to come olie with the system. Predicted load ad wid resource iformatio would eable the ideal diesel egie to get ready to come o lie. Whe a covetioal dispatch strategy calls for itermittet operatio of the two diesel egies, a forecasted sceario would help i optimisig the diesel commitmet strategy. Figure gives a flowchart for the proposed predictive diesel dispatch strategy. For the oe-hour ahead predictio method, the strategy eeds to be updated every hour so that the ext predictio takes care of the updated value of wid speed ad load. For loger-term predictios, say 2 hours ahead, the diesel uit commitmets could be made for a loger period, provided the predictio errors are withi accepted rage. To icrease the accuracy, artificial eural etwork approach may be used which could probably icrease the accuracy of predictio by itroducig more iput variables ad selectio of appropriate learig algorithm [6]. 7. CONCLUSION This study looked at the implicatios of short-term wid forecastig o a wid-diesel hybrid system operatioal strategy. A predictive cotrol strategy could be beeficial to the operatio of a wid diesel hybrid system to pla ad optimize the system operatio. It was observed that oe-hour ahead predictios give lower errors compared to a logerterm predictio for 2 hours. The predictio errors could be improved usig a differet approach or with more iput variables. The recommeded strategy is to have a 2 hour predictive cotrol strategy combied with a adaptive cotrol to take care of the ext hour predictio i light of the immediate past actual iput values. The predictive cotrol strategy whe applied to a sceario already operated uder other techical ad ecoomic costraits ca lead to efficiet operatio of the diesels. Fig. Predictive Diesel Dispatch Flow chart 6. EFFECT OF PREDICTION ON DIESEL DISPATCH Forecastig of wid ad load, eve oly for a hour ahead, becomes sigificat i plaig a cotrol strategy for the itegrated hybrid power system. It is sigificat durig diesel-chage-over from oe. ACKNOWLEDGEMENTS The authors would like to thak the Bureau of Meteorology ad the Wester Power Corporatio for providig the ecessary data used i the study. The first author gratefully ackowledges the fiacial support from the Australia CRC for Reewable Eergy Program 5.21 o system itegratio. 9. REFERENCES 1. Nfaoui H., Buret J., ad Sayigh A. A. M., Stochastic Simulatio of Hourly Average

7 Wid Speed Sequeces i Tagiers (Morocco), Solar Eergy, Vol. 56, o. 3, 1996, pp Makridakis S., Wheelwright S.C., ad McGee V. E. Forecastig Methods ad Applicatios (2 d ed.), Chapter 6, Joh Wiley & Sos, Ljug L., System Idetificatio Tool Box User s Guide, Maths Works Ic. Mass., Maho, L. L. J., Diesel Geerator Hadbook, Butterworth-Heiema, Oxford, Paickar, P. S., Islam, S. M. ad Nayar, C. V., A New Quasi-Optimal Cotrol Algorithm for a Wid-Diesel Hybrid System, Wid Egieerig, Vol. 22, o. 3, 199, pp Beyer H. G., Deger T., Hausma J., Hoffma M., ad Ruja P., Short Term Predictio of Wid Speed ad Power Output of a Wid Turbie with Neural Networks, 5 th Europea Wid Eergy Associatio Coferece ad Exhibitio (EWEC 9), Thessaloiki - Macedoia, Greece, Vol. 1, October 199, pp