Digital Implementation of Predictive Mutschler, Peter (Member, IEEE); sct Control Algorithms for Induction Motors Flach, Erich Institute of Power Electronics and Control of Drives Darmstadt University of Technology Landgraf-GeorgSr. 4 1D-64283 Darmstadt, Germany Abstract ~ To achicve fast control action, dest control rethods should be wied, Two new types of thse are presented In this paper: 1) "Direct Mean Torque Control" (DMC) combines the 00d dynamite performance of Direct Torque Control with fdvantages of inherently constant switching frequency a time equidistant control for Implementation ina digital ig procesor. 11). Presently, speed control of electrical drives is preferably done by eascaded control loops. To improve. dynamic performance of speed controlled drives, « "Direct SPecd Control (OSPC) method is proposed. The switching events of, the inverter are directly calculated in such a way, that in steady state the speed of the drive is kept in an (extremely Small) hysteresis band and a disturbed system will return (0 Stendy sate as fst as posible, Le. time-opins [erropucrion In the last years, methods to control direcly torque and ‘ux of induction motors gained increasing interest, since they promise an excellent dynamic performance, Based on ‘works of Depenbrock [1] and Takahashi (2), several ypes of Direct Torque Control (DTC) evolved. Time-continuous (analogue) implementations of the control algorithm have ‘ben realized first. For many reasons, a digital realization of direct torque control methods is desirable (3). In this paper, two direct control methods and their digital implementation are presented. Both methods use sophisti cated algorithms to predict the next two switching actions of the inverter. To do this, the state equations (2) of the ‘machine are numerically integrated in a digital signal processor (DSP). This integration is done a small amount of time (€.g. 30..100us) ahead of realtime. This means, that the computed torque (3) is earlier available for the contol algorithm than it wll be produced by the real machine, Both methods use the derivation dT/dt of (3) to determine the next switching actions inthe inverter. In part I of the paper, the Direct Mean Torque Control (MTC) method and its implementation are discussed. In contrast to [2] and (3, this method has the advantage of an inherently constant ‘switching frequency. Experimental results show the validity of the control scheme. (0-7803-4943-1/98/S10,00 © 1998 IEEE In part Il, a new speed control method, the Direct SPsed Control (DSPC) method, is presented. By this method, the switching actions of the inverter are caiculated in such a way, thatthe spood of the drive is directly controlled in a time optimal way. Mopetisa ‘The classical DTC scheme uses a simple machine model based on the measured stator currents and known stator ‘voltages. It consists on the integration Jeu-Ryipae o To avoid a significant delay while determining switching state, both control schemes call for the prediction of the system states. The currents used in (1) are no available from the future. Thus, a complete machine model in the stator frame of reference (a,P) with the fluxes as state variables, and the stator voltages as inpats is used according to: ae vie] fia vee" ] 0 va} Lo ® Equation 2) is numerically integrated using Runge-Kutta sccond order integration. The widih of integration steps is chosen in such a way, that the boundaries of the integration intervals coincide with the switching of the inverter. The stator voltages are predicted from the de-link voltage and the Aesired switching state. Measured currents may be taken ‘nto account afterwards to correct the estimated sates. The correction of the model, thus the observer, is no subject of this paper. Using this made, the torgue is given by 2), 4 PTL TD (Wie Was = Wie Wap) 8 Gh-) " Likewise to (2), we determine the sector of the flux by comparing its components, yup with thre digital com paralrs YY The comparator outputs allow to classifyFig. I: Sectors, Voltage and Zero Vectors the voltage vectors (VV) of the converter to flux and torque increasing and decreasing vector, illustrated in fig. ‘The VVant0 according to sector m (ie. (1) is according to Sect) isto be used as a mainly ux increasing VV. The ‘Wmé is ux and torque increasing, VVin+2 decreases the ‘lux and increases the torque. Similarly, VVi-1,2 decrease the torque Pat: Dimecr Mean Torque ConTRoL oF aX INDUCTION MoroR “The main application area of [1] ate high power motors with considerably large leakage inductance. Duc to this, currents and forque change relatively slow. This eases the digital implementation of the control algorithm using. a DSP. With a large leakage inductance in the motor, a quasi- ‘analogue implementation of direct torque contol is straight forward, (On the other hand, up to now the main application area ‘of [2] are medium and low power drives, having smaller leakage inductance. Especially servo drives show a small leakage inductance in order to achieve high dynamics, Thus, currents and torque can change very quickly. To keep the torque within the desired hysteresis band, the switching states of the inverter must be changed in short intervals ‘This means, for a quasi-analogue implementation, a very high repetition rate of the contol algorithm is necessary For a servo drive, the torque may travel through the 44s Inysteresis band in less than Sys. In this case, a quasi analogue implementation is not satisfactory because the ‘computational effort would be too high. Using a predictive algorithm, fast switching soquences can be calculated in advance. The computation time may be equally distributed within two successive switching intervals ofthe inverter, MTC inherently ses constant switching frequency and determines two switching events in advance fora fixed cycle ts. The aim isto place the switching events directly in a way ‘thatthe mean torque over the eycle is equal to the desired value. In most cases, alternate switching ofa voliage vector (VV) and a zero voltage vector (ZV) satisfies the demand. Fig. 2 shows the contol structure of DMC. Fig. 3 shows a typical cycle of operation. Applying a VV first, the torque increases atthe beginning, Then, applying @ ZN, the torque decreases. In [4 the timing of the switching events was determined by equalizing the different hatched areas in fg. 3. In that case steady state will not be reached at the end of the cycle, if Tt) # Tet) and Tbr) # Twa In this paper we mainly use another algorithm (3), which reaches steady state atthe end ofthe cycle, Converter 2 Exceptio Data Acquisition ‘Choice of WW and ZV Check alternative vv O (Wim20 of Wine2, fg. 3 | | wiv 9m Inverter Modo aw 4s sector J Jostermination— wt] et r ms Fig, 2: Contol structure of DMITCFigure 3: ‘Typical operation yele of DMC Preliminary steps In steady sat, the forque T(4) atthe beginning of a eye should be equal to its value Ty) at the end. Instead of imply equalizing the different hatched areas for the cycle m, it should also be possible to switch the VV to reach directly the value of Ts) according to steady state, ehus Teas) = Tent © Tre ~ Y-eT, ® where eT is the "virtual hysteresis width", Sustaining Aus and torque as well, Vint is normally the mest taken WV. Hence we assume eT as the "virtual hysteresis width while using Vail. eT ean be expressed as fevcher=teG-t) ‘Tvy can be calculated by deriving (3). Since the state of the machine a ty snot krown Jet its the easiest way 10 fet Toy tom the last ele Te)= Td © wv ‘hy With Tyy =‘Tvvi, this equation is used in (4) to determine Mitosis terme nae ere sees oe ae, face Based on fig 3, a preselection of the VV is done. If the aspired value of Te) (Le. Tys~ ¥-€T) is greater than the value which would be reached when applying the ZV all the qyele time (ie TM) + Tev'te, then Vmi0,1.2- are presclected, Thus, the counter clockwise rotation (Le ‘WWan0,12) is chosen it Ter ~ Yar eT > Te) + fay ts 6 Otherwise, we use VVi-01,2. Tai Tae le Figure 4: Poculiarty of novel method for eT 2 Tay ts Torque Control For given tw =‘Twymia of Vme0,1.2, we search its on- time hy to attain exactly Toa The torque T(g) at the end ofthe cycle can be expressed as: Teas = T+ yy -hyy + Fey ts —hyy) ® Solving tis equation for hyy leads with (4) finally to: Typ My) 4-2 =t5 Tey by = (ao) ‘Twat Ordinarily, a VV should only be applied, if the result of (10) is within liga S bv S t-hm. In this 386, the torque will stay in the “viral hysteresis band", Since &T hha been calculated for Tyvi, Vim] (= hyyi) sities the ‘demand. Fr lower angular speed ofthe flux vector, VVm2 fs well suited, t00, whereas VWm0 has to be checked out Depending on the state of the machine, VVmi9 may increase, or even decrease the torque. A special problem crops up ifeT 2 ty-ts, ie. applying a ZV, the torque decreases very slowly, as ilustrated in fig 4 This is typical for low speed operation. In this case, calculating hyy from (10) may lead to hyy<14 hon, ich, ‘would be rounded to zero. In consequence, no switching ‘would occur att, The same situation would appear att ‘The torque would follow the doted line in fig. 4, which causes a negative torque offset. ‘To avoid this, in cases where (10) leads to yy Yi» rounded 10 ii, thus tracking the continuous line in fig” 4. By this, the ‘torque is kept as close a possible at Ty Flux Control Since high dynamic torque control is our main goal, the ‘ux controller should not interfer with it, Similarly to [2], the stator flux is kept as close as possible to a circular trajectory. After the preselection of the VVini0,1,2, the fox controller uses the flux propagation and some supplementary rules to choose the appropriate VV. The length ofthe fax vector and its derivative ae given by vielen aa om) w For the VV, siagy Feuls from (2), whereas fr the ZV, ‘iuevcan simply be taken from the last cycle similar to (6): War (lta) —W(E)) ytd as, Starting from (10), resp. 12, ofthe torque controller, and ‘js, we ean predic the evolution of the flux vector over the next eyele foreach VV. At higher speed, the Mux controller chooses between VVaet and VVim#? to influence the fu. ‘A. common problem of DTC methods is the flux Iaintenance at lower speed, because the onctimes ofthe VV ecrease. First it is posible to check ifthe V0 is able to control either flax and torque as well. Is on-time hyo is Timited t0 Tyyanas by the maximum fox increase which seems tobe tolerable (using ay from (18a) ra Wa tpg) | VE AV +C) Wat Viwe vvomae =Cr co) Empirical tests showed acceptable results with C,=1.$ and Cx=2% If 10) leads to Yara $ vv © home an then VVimt0 is adequate to contol the torque. If this is not the ease, and the flux cannot be preserved over several yels, itis necessary to apply a fux-supporting VV (i.e ‘Wi0), although the torque control may be affected I the fux-supporting VV is switched too late, flux decrease has ‘become important, hence it requires a longer onsime. The idea is to detec the necessity of fluxsupporting VV as soon. as posible o diminish its influence aver the torque. A basic mention was to find expressions which check out 487 Fig. 5: Flow char of flux contr relationships between the influences ofthe switching states ‘over the fux. Avoiding to compare with absolute quantities has the advantage of uncomplicated parameterization for different machines. The algorithm consists of a set of ‘decision rules. We define its inputs as: By" YC) = Yet) = Ct) (isa) Ayew=dy—Y4-bwvo reve (a8) yyy Heh uver ase) Aya =Ay—hwvathiyve=(le— how) Wav asa) Ay ay = W(le)~ Wlba) (Be) Fig. 5 shows the flow chant of the lux controller, which is performed in fig 2, block ©. There are three principle operation modes to enter: ‘+ "Normal Mode" where to chose between fox increasing, Vat or lx decreasing VWms2. "Consider Plux-Supporting VW* where VVims0 may only be switched without torque disruption, VVmel otherwise. + "Force Flux Supporting VV" where VVm40 must be switched, although Texvwo Tae —'4-eT. In ,FluxSupporting VW" mode, decision G, if two ocreases the torque error (Le. Tre Titer) Ye, ‘Wine is switched for hyve 10 Keep the torque as good as possible. Otherwise, hyyo is reduced and the switching ‘order is inverted, This means, in order to minimize theAorgue error over the time, the ZV is preserved from the last cycle, and after it the WV. The resulting torque eror over such a cycle cam be taken into account by adding it t0 Tye just or the next ele. ‘The decision rules A-G in fig $ are mainly (© condition is satisfied, > otherwise) Ar Ay 2 Cooyer (he Co=3%) AND yilt)-vited 50GB Br dvi <0 +C -®D aye! $ Lavoe > Wine Vm a Er fay! S lvl > VVime0 > Wwmel Fi dy > -0.75-te-2v AND hhvvi siuvart (s-buv) Wi2y $0 3G Wmal Gs sign(Tyve) = signCtww) > Wint0 first, then ZV, byvorbyvans > Wino, inverted switching order, aywurhvvenin For the branches A + G, resp. P+ G, hvvonin i 10 he ee asa Wav avaniag = CoB LUs hina) ey a Wawo Since branch condition A (fig. 5) checks for a somewhat critical flux decrease, (19a) results in longer on-times. Setting C,="4, (198) assumes that, inthe succeeding cycle, ‘Win will provide a certain fux increas, too. Some additional constraints are generally included for the ‘beach conditions to avoid suspicious switching patterns: ‘When entering "Force Flux-Supporting VV" mode, a flag is set. After this cycle, we never use flux decreasing Wmne2, “Force Flux-Supporting VV" mode is never entered after a cyele with VVm#2, Implementation in DSP based System ‘The conte algorithm induces some constraints, In con- ‘rast to quasi-analogue implementations, itis not possible to ‘estimate the sate of the machine based on actual measured ‘quantities, because the currents ofthe machine change too fast. The delay between measurement and selection of the appropriate VV would be too lage. 488 | — | elt ifs Hal i WS | BARE de [ek ! flitaladte fell] [i ie ai : Fig. 6: Implementation f DMTC in DSP TMS320C30 ‘A quasi-analogue implementation has always to observe the sate of the machine in order to match the switching criterion (i.e. the hysteresis boundary) as good as possible, ‘A simple machine model allows to Increase the sompling frequency, resulting in considerable lower switching frequency, In our case, a fixed switching frequency of about 667KHZ is used. The average switching frequency per transistor is with DMTC usualy half of the controllers ‘itching frequency, thus 3.33 kHz. The cycle time of 130 us is long enough to calculate a full machine model given by 2). The integration with a Runge-Kutta algorithm of second order permits to predict the suites reached in the real machine at the beginning of the nest cycle. The according stator currents inp ean only be taken into account afterwards, Assuming an approximately constant diference between internal and measured values, the observer becomes a classical Luenberger approach. Although its design is no subject of this paper, itis to be mentioned that a Matrix Riocati approach seems tobe well suited Fig. 6 illustrates the consecuion ofthe conto algorithm. ‘A well defined time ahead of the switching instant ty a timer interrupt (TINTO) is generated. The control algorithm compares fux and torque of the model predicted for tox with its setpoint values. Then the control scheme described above is processed. The second timer (TINTI) is programmed for the ontime of the VV. At tthe VV Decomes active and the counter for TINTI slats. The different branches of the ux controller may cause varying,delays forthe computation whereas the model requires exact, consistency tot. This problem can be solved by polling the timer of TINTO, waiting forty After switching the WV, the sate forty cam be predicted. Two Runge-Kutta steps, one ‘with the step-width hiv applying the VV to the model and. ‘one with the step-width hy and zero voltage, ae caleulated Experimental results for DMTC ‘The control scheme has been implemented on a DSP. -TMS320C30 at 40 MEIz, mainly programmed in high level programming language C. To analyse the inner torque control loop without further influences, we propose a brie start-up with constant torque (i.e. 50% of rated torque Tuas), followed by a deceleration of about 10% Tras Fig. 7 shows thatthe start-up with consiant torque works well. A fast change of the measured phase current can be noticed During acceleration, the "viral hysteresis" enlarges due to increasing [Tzy|. Requesting lover torque at lower angular speed leads to short on-times hvy which yields to significant imerference of the flux controller, but torque ouliers are rather insignificant. Slight perturbations in the speed signal are dve to & mechanical resonance of approx, 300 Hz. PART: DIRECT SPEED CONTROL (DSPC) Presently, speed control of electrical drives is preferably ddone by cascaded control loops. In an inner control lop, the torque of the machine, or some equivalent quantity, and its flux are controlled, whereas in the outer contol loop, the speed ofthe drive is controled In this section, a new method is proposed, which controls the speed directly in a time-optimal way. As many drives, ¢.g. in machine tools, use a superimposed pesition contol Joop, the inner speed control loop is desired to be as fast a5, possible, This is achieved by the proposed method, Basic idea ‘The switching events in the inverter are put into effect in such a way, tht in steady late the speed ofthe drive is kept in an (extremely small-) hysteresis band and a disturbed system will return to steady state as fast a8 possible, i. time-optimal [6]. The new method is based on the following approach: One imagines, the machine would consist of a ‘subsystem which produces the slope of the torque T, and a further integrator, which delivers the torque T. This is shown in fig. 8 Block 1 represents the control, the inverter and the subeystem of the machine. Integrator 2 produces the torque and integrator 3 models the total inertia for a drive with a stiffly coupled loa. 449 z Ss reas ba SNe, Ty Fig: Esp. start-up, Ten Fora short time interval, the torques of machine and load are approximated by their tangents in the actual operating, point, i.e. T= const. Then we get fom fig. 8 0 ~Oq¢ = Aaldl= der At ~ dort We further assume doog/dt=0 and get eld do / dt a eo en Dividing 21) by 20): a at i wT awed a2at dt de Tea J id eae Inxegratng 22) e23ap) Fig. 8: Block diagramm ofthe driveFora short time interval, during which ‘ is constant, (23) defines ase of parabola, which are shown in ig. 9 Is assumed, the initial state of the system to be ey in fig. 9. In this state, a ,torque increasing voltage vector (VV) has to be produced by the inverter. By this, the state travels along the dtted parabola P, till the point @/ay is reached. This isthe intersection wich another parabola Po; which is valid fora torque decreasing. ZV, passing exactly the point +H. So, in e/a the inver- ter changes toa ZV. Then the state travels along the parabo- 4a Ps till point eof i reached, This isthe intersection with parabola Py, which is valid for @ torque increasing ‘WW and passes through point -Hy. In steady state, the sytem moves along the path *Hy..ys/te2.H..4s/tus..ABY. 1k should be noted, that in all converter fed drives, the ‘machine's torque has a ripple, and therefore speed shows periodical deviations Inthe proposed DSPC method, speed ‘deviation is kept in a defined band (-Hy.+HYy). As the width ‘of this band is reduced, switching frequency increases. In ‘ig. 9 intially the system was not in steady state. But the luajectores chosen by the algorithm guarantee the fastest possible way fo proceed into steady state, Thus, this is a {ime-optimal speed conto, DSPC algorithm Simitar to part 1, torque derivatives for diferent switching states have to be calculated. Here we neod the slope Tax for the actually used switching state and a prediction ofthe slope Toes fo the next switching state, To Calculate the intersection of the actual andthe next parabola, the constant C in og. (23) is determined by the actual point c/a in the state plane: Gra ate ow ‘The ordinate ay. ofthe intersection with the next parabola (see fig. 9s 450 ‘The time necessary to travel from the actual point to the intersection point withthe next parabola is: de Te yd Tra {@x41-%) 26) Ta A time interval of this duration may be created by using ‘some timer / counter hardware. One of several possibilities isto load the computed value of ts in a compare register associated with a timer, When the timer count becomes ‘equal fits compare register, the timer is reseted, the pulse pattem for the new switching state is transferred to the inverter, and the control algorithm starts again, But some modifications of this basic scheme are neoessary in special cases. When tin proves 10 be large, 6, texei> 120ps, the assumption ofa constant Tat in the ‘whole interval may be violated. Then a reduced value is Toaded into the compare register, e.g Kaos 60us. Using this value, the point e.od/as in fig. 9 may be reached where the control algorithm is started again and will compute the intersection with the next parabola more precisely. Torque limit To protect the equipment, the torque has tobe limited. The inverters current is indirectly limited by limiting the torque, ‘When the torque is going to exceed its limit, the "torque Timiting mode" is entered, which is similar to the method described in part I with Tyr Ty Observer for speed and acceleration The actual point e/a in the state plane can not directly ‘be measured. As the stationary speed ripple (Hy..#Hy) is, ‘extremely small, especially for drives with @ high foal iner- tia (), a direct measurement ofthe speed o will not provide sulficient resolution. Additionally, a direct measurement of the acceleration "a fora rotating machine is difficult, noisy, and expensive. Therefore, these quantities are calculated as floating point variables using the a/a-observer shown in fig. 10. For areal drive, only the position "x* may be measured, using a high resolution optical encoder with sinveos- output signals and appropriate evaluation ofthese signals. As with conventional control methods, the quality of the position signal "xis crucial for the whole system. If "x" is noisy, then the observer's dynamics have tobe slow in order to get smooth steady state behaviour. But in this case, the suppression of load disturbances Tim is very poor. AnotherFig. 10: Proposed signal processing crucial point is the parameter Iau, which is used by the ‘control algorithm. If is much larger than the real inertia J, then even instability may occu. Therefore, in applications with heavily varying inertia J, some on-line parameter adaptation for J wil be necessary ‘Simulation results The system in fig. 10 has been simulated with the data of ‘an asynchronous servo motor. The torque-limit was set t0 thre times rated torque. With the selected hysteresis widths (iy), in steady state an average switching frequency of 3kitz ‘was observed. In fig. 11 att=2ms. step change in speed reference from ‘6004pm to 606rpm was applied. This causes a torque increa- sing voltage vector, which is applied til the torque limit is reached. AAS the time 10 reach the torque limit is about 400ns, this time is divided into several intervals of 60ys, and in each ofthese intervals, the algorithm is started again. Then, due to the torque limit, the state in the e/eplane travels along a limit of acceleration. At (=2.7ms, a torque decreasing vector is applied, which causes the speed devia- tion to become approx. zero at =3ms. So, within approx. ‘ms the speed has changed from 60pm to 606xpm. Further work When selecting a voltage vector it is also taken into ‘account whether the fax has to increase or to decrease. This simple fx control method gives no satisfactory results at low speeds. In part I of this paper, a sophisticated Mux ‘control is descrbod, which will have tobe adapted in some ‘way to the proposed method of direct speed control (DSPC). Till now, the DSPC has only been simulated. Next, an ‘experimental set-up will be build 10 implement the DSPC algorithm on suitable hardware. 451 ertitt fia Fig. 1 Simulation of step ‘change in the speed reference valve REFERENCES. [1] Depenbrock, M.: Direkte Selostregelung (DSR) fur hhochdynamische Drehfeldantricbe mit Stromrichterspeisung, ezArchiv, Bd. 7, 1985, Heft ?, 211-218 ‘Takahashi, L, Noguchi, T.: A New Quick-Response and High-Efficiency Control Strategy ofan Induction Motor", IEEE Transactions of Industry Applications, ‘Vol. 14:22, No. 5, Sept/Oct. 1986) Aaltonen, M, Titinen, P., Lal, J, Heikki Direkte Drehimomentregeiung von Drchstromantrieben, ABB Technik 3/1995, p, 19-24 Flach, E, Hoffman, R, Mutschler, P. Direct Mean ‘Torque Control ofan Induction Motor, EPE TrondheinvNorway 1997, Vol. 3, p. 672-677 Flach, E: Improved Algorithm for Direct Mean ‘Torque Control ofan Induction Motor, PCIM'98 Nuremberg/Germany 1998, Ineligent Motion p. 261-267 ‘Mutachle, P.: A New Speed Control Method for Induction Motors, PCIM'98 Nuremberg/ Germany 1998, Inelligent Motion p. eB 6) 1) 6 @ 36