power quality control of an autonomous wind–diesel power system based on hybrid intelligent controller

1、Power quality control of an autonomous winddiesel power system based on hybrid intelligent controller1440H.-S. Ko et al. / Neural Networks 21 (2021) 14391446and tune, it is dominantly used in industrial systems (Hagan &Menhaj, 1994). However, it is difficult to obtain good performancefrom the PI

2、D controller only because a nonlinearity makes controlwith a PID controller difficult unless gain scheduling is used.Linearizable systems can be controlled by conventional linearcontrollers such as state space method, optimal control, robustcontrol, model predictive control, etc. Neural network, fuz

3、zy logicand genetic algorithm are widely studied to deal with highlynonlinear systems.The feedforward control concept has attractive features ofpractical relevance. Since it is assumed that a stabilizing controlleris available in advance, the experiment to collect a set of trainingdata sets is easil

4、y performed. Another feature is that one canintroduce the feedforward signal gradually. In applications wherean inappropriate control input can cause damage, this can be a softcontrol strategy (Madsen, 1995).The main reasons for using feedback are to stabilize unstablesystems and to reduce the influ

5、ence from possible disturbancesand model inaccuracies. Using feedback to ensure that the systemrapidly follows changes in the reference is not always a goodpractice. A rapid reference tracking obtained with feedbackgenerally has the side effect that the controller becomes highlysensitive to noise wh

6、ich implies the poor robust properties. Toachieve a satisfying reference tracking without feedback, thefeedforward is applied which is governed only by the reference.Moreover, feedforward control is used for regulation where thereference attains constant levels for longer periods of time.To speed up

7、 the tracking of set-point changes, a feedforwardcontroller is typically designed to provide the steady-state value ofthe control signal for minimizing tracking error (Haley, Soloway, &Gold, 1999; Madsen, 1995).In this paper, fuzzyneural hybrid controller is proposed andapplied for pitch control

8、 of wind turbine. Fuzzy logic is appliedfor designing a feedback controller. Neural network inverse modelis designed for a dynamic feedforward controller. Therefore, fastdamping from fuzzy controller and fast reference tracking can beaccomplished.2. System descriptionThe winddiesel autonomous power

9、system consists of thewindturbinehavingtheinductiongenerator(IG),thedieselengine(DE), synchronous generator (SG), superconducting magneticenergy storage (SMES), and the dumpload. When wind generatedpower is sufficient to serve the load, the DE is disconnectedfrom the SG by electromagnetic clutch, an

10、d the synchronousgenerator acts as a synchronous condenser. The main purposeof the dumpload is to regulate the system frequency. The SG(with/without diesel) is used for reactive power control that isachieved by the excitation system to regulate voltage. The SGalso contributes the reactive power to c

11、ompensate the inductiongenerator.SMES is a control unit for a synchronous machine (Tripathy,Kalantar, & Balasubramanian, 1991). When there is a sudden risein the demand of load, the stored energy is immediately releasedthrough the power system. As the governor and pitch controlmechanism start wo

12、rking to set the power system to the newoperating condition, a SMES unit charges back to its initial valueof current. In the case of sudden release of the loads, a SMESimmediately gets charged towards its full value, thus absorbingsome portion of the excess energy in the system, and as the systemret

13、urns to its steady state, the excess energy absorbed is releasedand SMES current attains its normal value.Fig. 1 shows the prototype of a winddiesel autonomous powersystem(Chedidetal.,2000).Generatordynamicsmodelconsistsofa synchronous machine driven by diesel engine through flywheeland connected in

14、 parallel with an induction machine driven by awind turbine.Fig. 1. The prototype of winddiesel autonomous power system.Fig. 2. The basic configuration of WDAPS.Blade pitch control of wind turbine has the potential forproducing the highest level of interaction because of the presenceof both diesel a

15、nd wind-turbine control loops (Tripathy et al.,1991). The pitch control system consists of a power measurementtransducer, a manual power set-point control, a proportional plusintegral feedback function, and hydraulic actuator, which variesthepitchoftheblades.Turbinebladepitchcontrolhasasignificantim

16、pact on the dynamic behavior of the system. Variable pitchturbines operate efficiently over a wider range of wind speedsthan fixed pitch machines. The generator dynamics model consistsof a synchronous generator driven by a diesel engine through aflywheel and connected in parallel with an induction g

17、eneratordriven by a wind turbine. The diesel generator will act as a dummygrid for the wind generator, which is connected in parallel. Whenwind power rises above the power set point and SMES unit is fullycharged, the pitch control system begins to operate to maintain anaverage power equal to the set

18、 point. The study in this paper isfocused on the designing of turbine blade pitch controller basedon fuzzy logic and neural network.The simplified description of Fig. 1 is in Fig. 2 with SMES(Tripathy et al., 1991).The models of the generators are based on the standardParks transformation (Krause, W

19、asynczuk, & Sudhoff, 1986) thattransforms all stator variables to a rotor reference frame describedbyadirectandquadrature(dq)axis.ThesetofSGandIGequationsare based on the dq-axis in accordance with (InternationalElectrotechnicalCommission,1975).TheSMESmodelcanbefoundin Tripathy et al. (1991).The

20、 nonlinear mathematical model of the winddiesel powersystem is given in detail in Appendix A. The following consid-erations are taken into account to identify component models:the electrical system is assumed as a perfectly balanced three-phase system with pure sinusoidal voltage and frequency. High

21、frequency transients in stator variables are neglected, which indi-cates that the stator voltage and currents are allowed to changeinstantly. This is because this paper is focused on the transient pe-riod instead of sub-transient period. Damper-winding models areignored because their effect appears

22、mainly in a grid-connectedH.-S. Ko et al. / Neural Networks 21 (2021) 143914461441Fig. 3. Membership function of error and change in error.system or a system with several synchronous generators runningin parallel. The different component models are of equal level ofcomplexity.3. Fuzzyneural hybrid c

23、ontrol3.1. Feedback controller based on fuzzy logicFuzzy control systems are rule-based systems in which a setof fuzzy rules represents a control decision mechanism to adjustthe effects of certain system conditions. Fuzzy controller is basedon the linguistic relationships or rules that define the co

24、ntrollaws of a process between input and output (Passino, 1997; Yen& Langari, 1999). This feature draws attention toward a fuzzycontroller due to its nonlinear characteristics and no need for anaccurate system modeling. The fuzzy controller consists of rulebase, which represents a fuzzy logic qu

25、antification of the expertslinguistic description of how to achieve good control, fuzzificationof actual input values, fuzzy inference, and defuzzification of fuzzyoutput. When the experts linguistic description is not available,fuzzy controller still can be designed by using the measurement ofreal-

26、time input/output data (Park, Moon, & Lee, 1995, 1996)In this paper, a total of 121 rules are used for the power systemunder study. The general form of the fuzzy rule is given in theifthen form as follows:if x(k) is A and ?x(k) is B, then y(k) is C,(1)wherex,?x: input signals,y: controller outpu

27、t,A,B,C: linguistic variables.The linguistic values extracted from the experimental knowl-edge are NH (negative high), NL (negative large), NB (negative big),NM (negative medium), NS (negative small), ZE (zero), PS (positivesmall), PM (positive medium), PB (positive big), PL (positive large),and PH

28、(positive high).In the power system under study, generator power deviation(?P) is chosen for the input of a fuzzy controller. The linguisticdescriptions provide experimental expressions of the expert fora control decision-making process and each linguistic variable isrepresented as triangular member

29、ship functions shown in Figs. 3and 4. In the fuzzy controller, the input normalization factors arechosen to represent the proper membership quantifications oflinguistic values. In addition, normalization factors can be used toyield the desired response of the fuzzy controller. g1,g2stand fora normal

30、ization factor for input of fuzzy controller and g0standsfor a denormalization factor for output of fuzzy controller. Fig. 3shows the membership function for error and change in error andFig. 4 depicts the membership function for output.In Figs. 3 and 4, the membership functions are overlapped withe

31、ach other to smooth a fuzzy system output and a fuzzy controlleris designed to regulate a system smoothly when an error and achange in error are near zero. The rules are established to controltransient stability problem for all possible cases. Tables 1 and 2show the inference rule table for two inpu

32、t fuzzy variables innegative and positive sides of change in error, respectively.Fig. 4. Membership function of output.Table 1Inference rule table in negative side of change in error.ErrorChange in error?1?0.8?0.6?0.4?0.20?1?1?1?1?1?1?1?0.8?1?1?1?1?1?0.7?0.6?1?1?1?1?0.7?0.4?0.4?1?1?1?0.7?0.4?0.2?0.2

33、?1?1?0.7?0.4?0.2?0.10?1?0.7?0.4?0.2?0.100.2?0.7?0.4?0.2?0.100.10.4?0.4?0.2?0.100.10.20.6?0.2?0.100.10.20.40.8?0.100.10.20.40.7100.10.20.40.71Table 2Inference rule table in positive side of change in error.ErrorChange in error0.20.40.60.81?1?1?1?1?1?1?0.8?0.4?0.2?0.10?0.7?0.6?0.2?0.100.1?0.4?0.4?0.10

34、0.10.2?0.2?0.200.10.20.4?0.100.10.20.40.700.20.20.40.710.10.40.40.7110.20.60.71110.40.811110.7111111Itisrequiredtofindthefuzzyregionfortheoutputforeachrule.Thecentroidorthecenterofgravitydefuzzificationmethod(Yen&Langari, 1999) is used which calculates the most typical crisp valueof the fuzzy se

35、t and y is C in Eq. (1) can be expressed by (2).y =?iA(yi) × yi?iA(yi)(2)where Ais a degree of the membership function.3.2. Feedforward compensator based on neural network inversemodelA neural network can model an input/output relationship ofa dynamic system. A direct or forward model is a mapp

36、ing thatmaps a system input to a system output. An inverse model, onthe other hand, is an inverse mapping that maps a system outputto a system input. In particular, if one sets the output to be thereference,thentheinversemodelcouldgiveadesiredinputfortheoutput to follow the reference or set point. T

37、he concept of inversemodel was used in designing feedforward controls for dynamicsystems (Harnold, Lee, Lee, & Park, 1998; Kawato, Furukawa, &Suzuki, 1987; Nakanishi & Schaal, 2004; Park, Choi, & Lee, 1996).Kawato et al. (1987) applied the concept of inverse-dynamics1442H.-S. Ko et a

38、l. / Neural Networks 21 (2021) 14391446Fig. 5. Training mode of NNIM.Fig. 6. Neural Network Inverse Model (NNIM).model to control a three-joint robotic manipulator represented ina continuous nonlinear kinematics model. In view of the fact thatthe inverse-dynamics model only gives the ideal computed

39、torque,feedback-error-learning scheme was utilized to compensate forthe output error. Nakanishi and Schaal (2004) reformulated thefeedback-learning scheme for a class of nonlinear systems from aviewpoint of the nonlinear adaptive control theory. Park and Choiet al. (1996) and Harnold et al. (1998) a

40、pproached the problemfromtheviewpointofdiscrete-timemodelofthenonlinearsystem,thus avoiding the issues of the invertibility of a nonlinear model.A two layer neural network is applied to obtain a dynamicfeedforward compensator (Haykin, 1998). In general, the outputof a system can be described with a

41、function or a mapping of theplant inputoutput history (Haykin, 1998; Ng, 1997). For a single-input single-output (SISO) discrete-time system, the mapping canbe written in the form of a nonlinear function as follows:y(k + 1) = f(y(k),y(k ? 1),.,y(k ? n),u(k),u(k ? 1),.,u(k ? m).(3)Solving for the con

42、trol, (3) can be represented as the following:u(k) = g(y(k + 1),y(k),y(k ? 1),y(k ? 2),.,y(k ? n),u(k ? 1),u(k ? 2),u(k ? 3),.,u(k ? m),(4)which is a nonlinear inverse mapping of (3). The objective of thecontrol problem is to find a control sequence, which will drive asystemtoanarbitraryreferencetra

43、jectory.Thiscanbeachievedbyreplacingy(k+1)in(4)withreferenceoutputyreforthetemporarytarget yr(k + 1), evaluated byyr(k + 1) = y(k) + (yref? y(k),(5)where is the target ratio constant (0 < 1). The value of describes the rate with which the present output y(k) approachesthe reference output value,

44、and thus has a positive value between0 and 1 (Park et al., 1995; Park & Moon et al., 1996). In Fig. 5,the training mode is introduced, where ? denotes the vector ofdelay sequence data. Fig. 6 shows the neural network inversemodel (NNIM) in training mode. All activation functions in thehiddenlaye

45、raretanh(x)(describedasfjinFig.6)andtheactivationfunction in output layer is x (depicted as Fiin Fig. 6).Fig. 7. The fuzzyneural hybrid control.The output of the NNIM can be represented as? ui(k) = Fi?nh?j=1Wijfj?n?l=1wjl ? + wj0?+ Wi0?,(6)where ? = y(k + 1),y(k),.,y(k ? n),u(k ? 1),.,u(k ? m)Tand ?

46、 = ?1,?2,?3,.,?n?Twjl: weight between input and hidden layers,nh,n?: number of hidden neurons and external input,Wij: weight between hidden and output layers.Theaboveneuralnetworkinversemodelistrainedbasedontheinputoutput data described in Fig. 5. To train the neural networkinversemodel,theLevenberg

47、Marquardtmethodisappliedwhichis fast and robust (Haykin, 1998; Madsen, 1995; Ng, 1997). Thetrained NNIM is used as a feedforward compensator.The total control scheme is indicated in Fig. 7. In the fuzzycontroller, the input normalization factors are chosen to representthe proper membership quantific

48、ations of linguistic values. Inaddition, normalization factors can be used to yield the desiredresponse of the fuzzy controller. The symbol ? denotes the vectorof delay sequence data. The total control input is u(k) = ufb(k) +uff(k). The feedback control ufb(k) is the output of the fuzzycontroller a

49、nd the output of the feedforward controller, uff(k), canbe represented as the following:uff(k) = g(yr(k + 1),yr(k),yr(k ? 1),.,yr(k ? n),ufb(k ? 1),ufb(k ? 2),.,ufb(k ? m).(7)In Fig. 7, once a signal of a feedforward compensator is givento the control system, the fuzzy controller provides a signal t

50、hatminimizes the error between the system output and its set point.Thiscontrolschemecanbeasoftwayofgeneratingacontrolsignalto minimize the tracking error and improve system performancein the sense that the compensating signal is given in advance(Madsen, 1995). This implies the improvement of the exi

51、stingPID-type controller, which is the main purpose of a feedforwardcontroller in a hybrid control scheme.4. SimulationFirst, a fuzzy controller is designed for a feedback controllerand a neural network inverse model is obtained for a feedforwardcompensator. In this paper, is 0.1 and g1,g2,g0 are 5,

52、50, and 5, respectively, determined by trial and error. TheLevenbergMarquardt method is applied to train a neural networkinversemodel.Thesamplingtimeis0.01s.fortheproposedcontrolaction. The training is carried out by giving varying white noisesignals. Firstly, before training, fuzzy control is imple

53、mented withthe plant. Secondly, white noise signal is inserted into the fuzzycontroller and data set is obtained, using noise signal as input andplant output as output. Then, the neural network inverse modelH.-S. Ko et al. / Neural Networks 21 (2021) 143914461443Fig. 8. Comparison of system response

54、 among PI, FC, and FNHC.(NNIM)istrainedbysettingthenoisesignalasoutputandtheplantoutput as input of the NNIM.The proposed fuzzyneural hybrid controller is tested in awinddiesel autonomous power system (WDAPS). Two cases areconsidered: first, the sudden step load increase of 0.01 per-unit(p.u.) when

55、SMES is in discharging mode (rectifier mode); second,when the SMES is fully discharged and there is a sudden step loadincrease. In this case, SMES is in recharging mode (inverter mode).(The (p.u.) stands for per-unit. It is a normalized value withrespect to a base or reference value.)4.1. Case 1: A

56、sudden step load increaseThe load is suddenly increased by 0.01 p.u. The SMESreleases the charged current (2 p.u.). The governor and pitchmechanism start operating for charging current of SMES anddamping of WDAPS. Fig. 8 shows improvement of the systemfrequency oscillations and power deviations, whe

57、re PI stands forconventional proportionalintegral controller and FC stands forfuzzy logic feedback controller.Fig. 9. Comparison of system response among PI, FC, and FNHC.4.2. Case 2: Sudden step load increase with fully discharged SMESInthiscase,theSMESisfullydischarged(0p.u.).Then,theSMESneeds to

58、recharge current to the set point (2 p.u.). The wind powergeneration from the wind turbine is assumed to be not sufficient.Fig. 9 also shows that the FNHC performance is much better thanthe PI and the FC.5. ConclusionIn this paper, the fuzzyneural hybrid controller for elec-tricity quality control o

59、f wind power generation plants is pre-sented. The main idea of hybrid control is that the dynamicfeedforward control can be used for improving the referencetracking while feedback is used for stabilizing the system andfor suppressing disturbances. Feedforward controller is a neu-ral network inverse model (NNIM), which is trained by the Lev-enbergMarquardt method, and feedback controller is a fuzzycontroller.1444H.-S. Ko et