Professional Documents
Culture Documents
Jan H. Fisch Yun Li, P.C. Kjaer, J.J. Gribble, T.J.E. Miller
Institut für Regelungstechnik Department of Electronics and Electrical Engineering
Fachgebiet Regelsystemtechnik und University of Glasgow
Prozeßautomatisierung Rankine Building
Technical University of Darmstadt Oakfield Avenue
D-64283 Darmstadt, Landgraf-Georg-Str. 4 Glasgow G12 8LT
GERMANY UNITED KINGDOM
Abstract - Research into integrated control of the process can, however, be automated in a more
severely nonlinear switched reluctance motor is in its computerised way. In nature, the process of evolution
infancy. This paper reports an application of genetic provides continuously improving solutions in terms
algorithms to this area, aiming at providing motor and of adapting to the environment. This process is
drive engineers with a helpful method and data for emulated in evolutionary optimisation methods such
commissioning. Using the genetic algorithm method, as genetic algorithms (GAs). These algorithms
optimal firing angles are obtained for maximal torque combine the Darwinian-Wallace principle of
control under multiple operating conditions. For ‘survival-of-the-fittest’ with ‘intelligent’ information
‘minimum commitment design’ at the CAD stage, exchange. They have found to be highly efficient and
Pareto-optimal firing angles are also evolved for both robust in searching for multi-optimal solutions in a
efficiency and torque maximisation, which have not difficult multi-dimensional solution space [6], [7].
been successful in the past due to methodological Obviously, GAs have potential to offer in SRM
limitations. The outcome should be of immediate use control. Examples of this will be reported in this
in inverse model based optimal operation and paper. Following setting the scene of the control
integrated manufacturing of switched reluctance problem in Section II, existing methods and the GA
motors. based method will be discussed in Section III. Firing
angles for various operating conditions will be
optimised with respect to maximum torque
I. Introduction
production and maximum efficiency in Sections IV
In recent years, there has been an increasing interest and V. Conclusions are drawn in Section VI.
in switched reluctance motors (SRMs). Since high
voltage and current switching devices for SRM
II. Firing Angle Optimisation in Inverse
control have been available at relatively low cost, it is
Model Based Control of SRMs
now practical to make use of the advantages of SRMs,
such as simple design, rigidity, low manufacturing For an SRM, a number of variables may need to be
cost, reliability, robustness, and efficiency [1]. Most controlled, such as torque and speed. This paper
of these properties are due to the simple geometric focuses on inverse model based speed control. The
construction of the SRM, whose rotor consists of a basic idea of inverse model control is to cascade a
simple iron core without a single winding. nonlinear filter in front of the system element that
On contrast, control of an SRM appears not tends to cancel out the nonlinear behaviour of the
simple. The behaviour of the SRM and its drive are so motor such that the resulting transfer from the
nonlinear that modern control methods can hardly be demand to actual speeds is unity. Such a control
applied in the known manner for integrated technique is mainly used for steady-state
manufacturing [1]-[5]. One of the issues is to find performance at different operating points for
optimal firing angles in current control. Despite propulsion systems rather than servo systems.
extensive research over the past two decades, there is To proceed, study the torque produced by a single
still no complete knowledge about how control motor phase for a single pole pair
∂W '(i ,θ ) ∂
parameters should be set for desirable operation. i
∂θ ∫0
Largely due to these reasons, Miller [1] stated that Tk (i, θ ) = = ψ (i1 ,θ ) di1 (1)
the only method available to find a global mapping for ∂θ
optimal angles appeared to be conducting exhaustive where Tk denotes the torque for phase k, W’ the
search, either by dynamometer tests or by numerical coenergy converted, θ the rotor angle, ψ the flux
simulations or by both extensively, if one does not linkage, i1 the instantaneous phase current and i the
wish to carry out ‘intelligent’ manual search by trial- current level. By assumption of superposition, the
and-error. sum of all phase torque given by Tk yields the total
Such a manual trial-and-error based search motor torque, where k = 1, 2, …, m for an m-phase
91
SRM. By Faraday’s law, ψ is given by Here IRMS represents the root mean-square ampere in
θ2 one phase. Based on these equations, the simulation
dθ
ψ = ∫ (v − Ri ) + ψ0 (2a) subroutine SRM_model() to be used in the
θ1
ω optimisation work reported here delivers torque and
where the voltage at the phase terminals, v, is efficiency as a function of the turn-on angle, turn-off
obtained from the fixed supply voltage, VDC, by angle, reference current and angular velocity in the
chopping: form:
T
v = VDC sgn(i ref - i) . (2b)
[
η = SRM _ model θon θoff i ref ω VDC R . (5) ]
Here iref is the reference current, which is usually
constant. In (2a), R is the phase resistance, ω the Given the torque, inverse model speed control can
angular velocity of the rotor (assumed constant) and be derived. Fig. 1 shows a block diagram representing
[θ1, θ2] the range of the shaft movement in an SRM and its load with an inverse model in front.
mechanical degrees. Here, T is controlled by iref, the major control
Speed control is achieved through torque control. variable, with θon and θoff chosen from a lookup table.
In current driven torque control, positive torque The table is pre-set at distinct values of iref and ω [1],
generated by a pole pair in a single phase is [4], [5]. The following discusses how iref can be
determined by the amplitude and duration of an derived in order to obtain a desired ω.
idealised rectangular wave of current running through
the phase windings of the SRM during the period of 1 Tdes 1 iref T ω
ωdes SRM load
rising inductance, i.e., while the rotor poles are load SRM
attracted to the stator poles. Supply voltage is
θon θoff
chopped to achieve such a waveform. Hence, the
control variables are: lookup
table
1. Reference current iref: amplitude of the rectangle,
2. Turn-on angle θon: beginning of the rectangle,
3. Turn-off angle θoff: end of the rectangle. Fig.1. Inverse model speed control of an SRM.
By (1) and (2), the steady-state torque also Desired torque Tdes may be calculated from
depends on the angular velocity and thus optimal desired angular speed ωdes via the inverse dynamics of
torque control is derived at a given ω. The average the SRM and its load using the lookup table
torque of the SRM may be calculated according to (1) containing the firing angles. Such inverse model
and (2) within an iteration loop (in time steps), control technique offers a simple and effective way
starting at the time when the rotor passes θon. For for motor engineers practically to handle the
this, the hard chopping controller is simulated first, nonlinearity of SRMs as is well-known [1], [4], [5].
where the phase voltage is alternated between the full Model inaccuracy could be tackled by adding a
supply voltage +VDC and the full reversed supply conventional feedback control loop that ensures zero
voltage -VDC depending on whether the phase current steady-state errors or, if desired, by adding a robust
exceeds iref or not. The new rotor position is nonlinear controller such as the sliding mode
computed from ωt. The actual flux is calculated by controller [8]. However, a crucial problem remains,
adding ∆t(v - Ri) to the former value according to (2). which is how exactly the firing angles should be
The phase current corresponding to the rotor angle θ chosen to appear in the lookup table that are needed
and flux ψ can be derived from the individual to optimise the performance of the SRM.
magnetisation curves of an SRM. The work done, W, To obtain maximum torque, θon should be set near
is then updated using (1) by a contour integration the minimum inductance position to start positive
from θon to θext , where θext is the rotor position at torque production and θoff must be about the
which the phase current reaches zero (i.e. where the maximum inductance position. To prevent negative
energy conversion terminates). For an m phase and Nr torque, flux should be allowed to decay before the
rotor pole SRM, the total torque averaged over one rotor reaches the area of negative torque production.
revolution, T, is thus given by This implies that the supply may be turned off early,
m Nr but if θoff is not carefully chosen it may result in
T= W (3) possible loss of positive torque production. Thus an
2π
optimisation problem arises concerning both θon and
which leads to the efficiency
θoff for a range of iref and ω settings. In this paper, the
Tω
η= . (4) optimisation problem is studied for a 3 horse power
T ω + m R I 2RMS 42-frame 4-phase 8/6 SRM, driven by a 300 V / 60 A
92
DC power supply via a 4-phase insulated-gate bipolar usually only applicable to uni-modal, single-objective
transistor inverter with a fully digital controller. The and, preferably, single-parameter optimisation [6],
first tasks can be characterised as follow: [7]. Largely due to these difficulties, practical
methods for finding a global mapping for optimal
• To find average-torque maximising firing angles angles have been by way of exhaustive search [1].
θon and θoff at The exhaustive search or enumeration technique
• a fixed supply voltage VDC = 270 V with belongs to a-posteriori based guidance methods. As
• constant phase resistance R = 0.3 Ω, the terminology implies, the best solution can only be
• reference current iref in the interval [5, 40] A at found after all possible trials have been conducted,
steps of 5 A and since there is no ‘intelligent’ means incorporated in
• angular velocity ω in the interval [50, 300] rad/s at the search. An example of such action was reported
steps of 25 rad/s. by Torrey and Lang [9], who found firing angles for
optimised efficiency in the course of iterative
simulations similar to hill-climbing. An example of
III. A-Posteriori Guided Search Method applying such a technique to finding an optimal θoff
A. Existing Techniques (in electrical degrees) for maximum torque control
under a pre-determined operating condition is shown
As discussed in the previous sections, the problem of
in Fig. 2. However, in practice, this technique may
designing an SRM controller is an optimisation
only be applied to single parameter optimisation [1],
problem. For this, Lovatt and Stevenson [2]
since the search time will increase exponentially with
maximised the torque output per RMS ampere by
the number of parameters involved. The design cycle
varying the current waveform via the derivative of an
may be shortened by manually incorporating trial-
objective function. The obtained waveform can result
and-error based adjustment. This was attempted by
in optimised torque, but can also lead to considerable
Stiebler and Ge [10], who found firing angles that
torque ripple and place a high demand on the current
deliver certain maximal efficiency, maximal
control loop. Less effort is involved in fixing the
effectiveness and minimal noise. Clearly, manual
current waveform and changing the firing angles only.
trial-and-error based search can be painstakingly long
For this, Chiba and Fukao [3] derived phase angles
and room may still be left for further improvement.
that promise maximum power output per RMS
These existing difficulties in optimal designs of such
ampere by a calculus means. In order to be able to
nonlinear control systems have thus led to the
derive these angles analytically, however, a very
conclusion that no perfect methods are yet known for
restrictive assumption must be made, where the
finding optimal firing angles in SRM control [1].
inductance between the rotor and the stator is
approximated by a sinusoidal function of the rotor 1.8
antecedent generation, but individuals will have to IV. Torque Maximising Firing Angles
prove their fitness in comparison within their own
A. Software Interfacing
generation.
In an evolutionary algorithm, candidate solutions The 8/6 motor data are included in the SRM_model()
can be encoded in numerical or symbolic strings that module. The performance of the integrated on-board
mirror genetic chromosomes. A GA encourages controller plus the SRM system is evaluated by the
information exchange and diversifies the search by module as part of a computer-aided SRM design
structured creativity via ‘genetic’ crossover and package written in C. The use of efficient C compiled
mutation of the ‘chromosomes’, whose usefulness executable code is to speed up the simulation. As
can only be judged a-posteriori. Thus a GA requires shown in Fig. 4, this module is linked to the Matlab
no derivatives of objective functions. Such algorithms environment by a mex-file, link-SRMt(), which passes
have been found highly robust and efficient in a non- θon and θoff for evaluation with various iref and ω. The
deterministic polynomial (NP) manner in searching evaluated fitness is used by anther m-file, SRM-
for multiple optimal solutions in a multi-dimensional taut(), to call the GA toolbox to optimise iref and ω
and multi-modal space [6]-[8]. Obviously genetic for inverse model based control.
algorithms have potential to offer in SRM control.
The GA software used for optimal control settings Matlab environment
reported in this paper is a commercially available
call
Matlab toolbox, FTGA® [7]. The system parameters to FTGA
toolbox f (Τ, η )
SRM
taut
be optimised can be encoded by binary, floating-point
θ on θ off
or logarithmic strings. The toolbox provides a i ref
ω
graphical user interface and user flexibility in setting link
R
SRMt
the GA operators, selection mechanism and niching VDC
for up to four objectives. Pareto optimality is enabled
for multi-objectives, which acknowledges only
superior solutions that leave at least one goal better θon
off and not a single goal worse off. The operation of a θ off η
iref
decimal-coded GA can be summarised in Fig. 3 [8]. ω S R M _ m o d e l (in C)
VDC T
R
Initial/random Final optimised
designs coded designs
Selection
f(P 3) Fig. 4. Matlab FTGA® interfacing with C simulation.
f(P1: 1 2 0 9 0 2 1 7)=5%
f(P2: 4 0 0 3 0 1 6 1)=60% f (P1)
f(P 3: 0 1 6 4 1 8 0 1)=35% f( P2)
B. Results Obtained for Controller
New Integration
Decoding, generation
simulation, The optimisation was run with a candidate population
evaluation
Mutation
Crossover
P 2: 4 0 0 3 0 1 6 1
size of 39, 1 of which was used for “steady-state” GA
P2: 4 0 0 3 0 1 6 1 P 2: 4 0 0 3 0 1 6 1 inheritance. A size 2 tournament selection scheme
P2’: 4 0 1 3 0 8 0 1
P3’: 0 1 6 4 1 1 6 1 P 3: 0 1 6 4 1 8 0 1 and efficient integer coding at a resolution of 1°
Conventional CAD
(electrical) for θon ∈ [140°, 300°] and θoff ∈ [200°,
360°] were used. Such resolution is deemed
Fig.3. GA optimisation coupled CAD simulation. sufficient by motor engineers. The selection of the
turn-on and turn-off ranges is based on the argument
In the figure, the conventional CAD package made in Section II. Should exhaustive search be used,
simulates the performance of candidate solutions. the number of function evaluations would be (300-
The evaluated performance is then fed back to the GA 140) × (360-200) = 25600 per set of (iref, ω).
to guide the evolutionary selection a-posteriori, Using the GA toolbox, the probability of
which in turn alters candidates by crossover and crossover was selected as 0.77 and that of mutation
mutation. Then new candidates are further tested by as 0.077. It was observed that one GA optimisation
the CAD program. Such a process continues until the run typically involved some 780 function evaluations
candidates may not be improved meaningfully. for each operating point and it took approximately
Clearly, the human intelligence used in a manual trial- four hours on a 100 MHz Pentium processor to
and-error based search process is borrowed in the GA complete the torque optimisation for entire 88
to automate the trial and search. Owing to the NP operating conditions. Note that the time taken would
characteristics, a GA requires an exponentially be some 780/25600 = 3% of that needed by
reduced time compared with exhaustive search. exhaustive search. At the end of the torque
optimisation process, multiple optimal turn-on angles
were revealed by the terminating generation of the
94
towards at least one goal and does not harm any other. Fig. 6. Maximal torque offered by Pareto-optimal
The price of gaining true multi-objective optimisation angles given by Tables IV and V for both maximal
is, however, that the solutions may not be unique. torque and maximal efficiency.
η
Usually, various Pareto-optimal solutions exist which 0.9
are not Pareto dominated by other solutions. So the
designer can have the final decision on which solution 0.8
θoff | Tmax
3 5 0
iref
1 5 0
θon | Tmax, ηmax
1 0 0
5 359 358 357 355 353 351 348 347 346 344 342
5 0 Min θon | Tmax
10 358 357 355 352 349 346 344 343 342 340 338
0
15 356 354 352 351 347 346 344 344 342 341 337
477
716
955
1194
1432
1671
1910
2149
2387
2620
2865
20 355 354 352 348 347 343 341 341 341 341 337 ω (rpm)
25 355 354 352 347 347 343 341 341 341 339 336
30 355 353 350 347 346 341 341 340 338 336 335
35 354 352 349 347 345 341 340 338 336 334 332
Fig. 8. GA recommended single-objective and Pareto-
40 354 352 349 346 344 341 338 336 334 332 330
optimal angles versus steady-sate ω for iref = 10 A.
θon [deg]
T [Nm]
200
30
180
20
10 160
0
ω [rad/s] ω [rad/s]
100
iref [A]
100
200 35 iref [A]
25
35
300 15 200
25
5
15
300 5
96
Acknowledgements
Fig. 9. Pareto-optimal turn-on angles in Tables IV.
The authors are grateful to their colleague, Dr. G.J.
θoff [deg]
360
Gray, for his generous and enduring help on linking
between the C and Matlab code and to Lucas
350 Aerospace and MagnaPhysics Corp for motors. It is
340
gratefully acknowledged the EPSRC funding for
GR/J06238 (Control Techniques for Switched
ω [rad/s]
Reluctance Drives) and GR/K24987 (Evolutionary
100 Programming for Nonlinear Control) that provided
iref [A] the environment in which this work was embedded.
200 35
25
15 References
300 5
[1] T. J. E. Miller, Switched Reluctance Motors and
Fig. 10. Pareto-optimal turn-off angles in Tables V. Their Control, Hillsboro, OH: Magna Physics and
Oxford: Clarendon Press, 1993.
[2] H. C. Lovatt and J.M. Stephenson, “Computer-
VI. Conclusions and Further Work
optimised current waveforms for switched
Applications of genetic algorithms to the infant reluctance motors,” IEE Proc. - Electr. Power
research area of nonlinear integrated control of Appl., vol. 141, no. 2, pp. 45-51, 1994.
switched reluctance motors have been investigated. It [3] A. Chiba and T. Fukao, “An analysis and an
has been shown that a GA can be used easily to find operating method of switched reluctance motors
optimal firing angles efficiently. Torque maximising based on a simple inductance representation,”
firing angles for 88 operating conditions have been Proc. IAS Annual Meeting, 1995, pp. 419-426.
obtained. The multiple solutions offered by the GA [4] R. Orthmann and H. P. Schoener, “Turn-off angle
reveal the merit of these optimal angles in terms of control of switched reluctance motors for
sensitivity and robustness. This has led to the search optimum torque output,” EPE Proc., 1993, pp.
for optimal bounds of the turn-on angles, within 20-25.
which any setting may be used for torque maximising [5] P. C. Kjaer, C. Cossar, J. J. Gribble, T. J. E.
control. Hence, within these ranges, room is left for Miller, and Y. Li, “Minimization of reactive
choosing turn-on angles for another objective, such power flow in switched reluctance generators,”
as efficiency, speed or current limit. Proc. IPEC, Yokohama, Japan, 1995, pp. 1022-
The GA has also been applied to Pareto multi- 1027.
objective optimisation for ‘minimum commitment [6] D. E. Goldberg, Genetic Algorithms in Search,
design’ at the CAD stage, resulting in Pareto-optimal Optimization, and Machine Learning, Reading,
firing angles for both torque and efficiency MA: Addison Wesley, 1989.
maximisation. This avoids pre-commitment on [7] FlexTool(GA) User Manual, Tuscaloosa, AL:
weighting the design objectives, which has not been Flexible Intelligence Group LLC, 1995.
successful in the past due to methodological [8] Y. Li, K. C. Ng, D. J. Murray-Smith, G. J. Gray,
limitations. The method and outcome should be of and K. C. Sharman, “Genetic algorithm
immediate use by design and commissioning automated approach to design of sliding mode
engineers for optimal operation and integrated control systems,” Int. J. Control, vol. 63, no. 4,
manufacturing of switched reluctance motors. 1996, pp. 721-739.
Note that, however, the usefulness of the [9] D. A. Torrey and J. H. Lang, “Optimal-efficiency
recommended optimal turn-on and turn-off angles excitation of variable-reluctance motor drives”,
depends upon the accuracy of the motor and drive IEE Proc - Pt. B, vol. 138, no. 1, 1991, pp. 1-14.
CAD package used, but the recommendations can at [10] M. Stiebler and J. Ge, “A low voltage switched
least act as a guide for fine adjustments via on-line reluctance motor with experimentally optimised
dynamometer tests. Further work related to this control”, ICEM Proc., 1992, pp. 532-536.
would thus be to establish a highly accurate model of [11] X. Guan and K. J. MacCallum, “Adopting a
the severely nonlinear SRM and its drive for highly minimum commitment principle for computer
accurate CAD. The model may be described by a aided geometric design systems,” in Artificial
‘grey-box’, representing the global dynamics Intelligence in Design, Gero, J. S. and
determined by physical laws in a clear structure, Sudweeks, F., eds, Kluwer Academic Publishers,
whilst accommodating unmeasurable dynamics in 1996, pp. 623-639.
local black-boxes.