Professional Documents
Culture Documents
Abstract: High performance application of permanent systems have become mature after years’ development.
magnet synchronous motors (PMSM) is increasing. PMSM However, most physical systems show significant
models with accurate parameters are significant for precise nonlinearities as dynamic range is extended in high
control system designs. Acquisition of these parameters during performance equipment. These real world nonlinearities are
motor operations is a challenging task due to the inherent
nonlinearity of motor dynamics. This paper proposes an
generally neglected under some assumptions so that linear
intelligent model parameter identification method using the system theories are applicable to find approximate solutions.
particle swarm optimization (PSO) approach. As an intelligent These solutions are acceptable in general industrial
computational method based on stochastic search, PSO is applications. However, in high performance applications,
shown to be a versatile and efficient tool for this complicated linear system theories are inadequate. There is a need for the
engineering problem. Through both simulation and development of advanced approaches to address the problem
experiment, this paper verifies the effectiveness of the proposed of model identification for nonlinear systems. In this work,
method in identification of PMSM model parameters. we investigate a new parameter identification approach
Specifically, stator resistance and load torque disturbance are using Particle Swarm Optimization (PSO) as applied to
identified in this PMSM application. Though PMSM is
discussed, the method is generally applicable to other types of
PMSM, which is a typical example of highly coupled,
electrical motors, as well as to other dynamic systems with nonlinear systems.
nonlinear model structure. There have been some techniques proposed for the
parameter identification of PMSM from different
Keywords: Parameter identification, Particle Swarm perspectives. Some of them are focused on offline
Optimization, permanent magnet synchronous motors. applications, e.g. [6]-[8]. Online identification of PMSM
electrical parameters is discussed in [9][10], which are based
I. Introduction on the decoupled control of linear systems when the motor
mechanical dynamics are ignored. The basic idea is to
As the high field strength neodymium-iron-boron (NdFeB) transform the original nonlinear system into a linear one.
magnets become commercially available with affordable This method cannot guarantee high performance with
prices, the sinusoidal back EMF permanent magnet sufficient accuracy over the entire operation range and
synchronous motors (PMSM) are receiving increasing decoupling may obscure internal dynamics.
attention due to their high speed, high power density and While the most commonly used identification techniques
high efficiency. These characteristics are very favorable in as mentioned above can not provide satisfactory results in
some special high performance applications, e.g. robotics, this specific application, this work investigates the
aerospace, and electric ship propulsion systems [1][2]. It has possibility of using a numerical approach as a promising
been shown that PMSM can provide significant performance alternative. PSO is a relatively new stochastic optimization
improvement in many variable speed applications [3][4] technique developed in mid 1990s. The algorithm is
To achieve desired system performance, advanced theoretically simple, and computationally efficient. It
control systems usually are required to provide fast and exhibits advantages for many complex engineering
accurate response, quick disturbance recovery, and problems. PMSM model parameter identification problem is
parameter variations insensitivity [5]. Acquiring accurate investigated here utilizing PSO approach.
models for systems under investigation is usually the The rest of this paper is organized as follows. Section 2
fundamental part in advanced control system designs. Model briefly reviews the traditional mathematical model of
identification and parameter estimation techniques for linear PMSM. The parameter identification problem for this
212 Li Liu, Wenxin Liu and David A. Cartes
In this paper, the PSO algorithm, which uses stochastic entire population; vi is the rate of the position change
search, is utilized to track PMSM parameters based on the (velocity) for particle xi. Velocity changes in (5a) comprise
traditional dq-model (1)-(3). The model with these estimated three parts, i.e. the momentum part, the cognitive part, and
parameters should provide improved accuracy in modeling the social part. This combination provides a velocity getting
PMSM dynamics when maintaining its simplicity for closer to pbest and gbest. Every particle’s current position is
controller designs. Specifically, this work focuses on the then evolved according to (5b), which produces a better
identification of the stator resistance Rs and disturbed load position in the solution space.
torque TLd. The definition for TLd is expressed in (4). It is a The implementation procedure of the PSO algorithm can
summation of the actual load torque TL, and the disturbances be illustrated with the flowchart in Figure 2. According to
caused by inertia and frictional coefficient variations. While Figure 2, the optimization process can be divided into six
all the three parameters, TL, J and B in (2) tend to change steps as follows.
with motor operating conditions, their impact to the model is Initialization. During this step, the bounds of each
combined and taken into account by properly defining TLD. element are set for the position and velocity particles; the
1 d ωr 1 population is initialized randomly; the parameters in (5) and
TLd = TL + J +B ωr (4) the terminal condition are all preset; and pbests are set to be
n p dt np this initial population.
In (4), we define J = J − J 0 and B = B − B0 . J0 and B0 Evaluation of the initial population. In this step, the cost
denote the nominal values of motor inertia and viscous for all particles in the initial population are evaluated
friction coefficient listed in PMSM datasheet. In doing so, according to a fitness (cost or objective) function; the global
the number of the parameters to be identified in this problem best particle, gbest, is chosen to be the particle with the best
is also reduced. fitness.
Updating position and velocity. The position and the
velocity particle are updated according to (5a) and (5b)
III. Overview of Particle Swarm Optimization respectively; check if the updated position and velocity
Particle Swarm Optimization (PSO) is a population based particles are beyond their corresponding limits; if it happens,
stochastic search algorithm. It was first introduced by the over limit elements are set to the corresponding limits.
Kennedy and Eberhart in 1995 [14]. Since then, it has been Evaluation of the updated population. Similar to (ii), the
widely used to solve a broad range of optimization updated position particles are evaluated according to their
problems. The algorithm was presented as simulating fitness; the pbest and gbest particles will be updated if
animals’ social activities, e.g. insects, birds, etc. It attempts necessary.
to mimic the natural process of group communication to Check if the terminal condition is satisfied. This is done
share individual knowledge when such swarms flock, according to the preset terminal condition, such as number
migrate or hunt. If one member sees a desirable path to go, of iterations reaching the maximum, quality of the best
the rest of this swarm will follow quickly. In PSO, this solution being satisfactory, etc. If the terminal condition has
behavior of animals is imitated by particles with certain not been satisfied, the updating process (iii) will be
positions and velocities in a searching space, wherein the repeated; otherwise, the optimization process ends.
population is called a swarm, and each member of the Output results. The best solution obtained during the
swarm is called a particle. Starting with a randomly optimization process, gbest, is output in this step.
initialized population, each particle in PSO flies through the From the updating rules and the flow chart of PSO, it can
searching space and remembers the best position it has seen. be seen that PSO is very simple in concept and easy in
Members of a swarm communicate good positions to each realization. Thus, PSO gains popularity in recent years. PSO
other and dynamically adjust their own position and velocity development and its applications for power systems are
based on these good positions. The velocity adjustment is presented in [17]-[19]. Compared to other computational
based upon the historical behaviors of the particles intelligence techniques, PSO has fewer parameters to tune.
themselves as well as their neighbors. In this way, the Since 1995 many researchers from different fields have
particles tend to fly towards better and better searching areas attempted to improve the performance of the original PSO.
over the searching process [14]-[16]. The searching Newer versions and the choices of proper values of the
procedure based on this concept can be described by (5). parameters to improve the performances of PSO are
vik +1 = wi vik + c1rand1 × ( pbesti − xik ) + c2 rand 2 × ( gbest − xik ) (5a) discussed in [20-23]. The modifications of the PSO mainly
consider the swarm size, the number of iterations, and the
xik +1 = xik + vik +1 (5b) velocity to update the search direction and the steps. Both
In (5), c1 and c2 are positive constants, defined as the performance and the execution time of PSO are
acceleration coefficients; w is the inertia weight factor; improved in the modifications.
rand1 and rand2 are two random functions in the range of In this work, the PSO algorithm will be used to solve the
[0,1]; xi represents the ith particle and pbesti the best PMSM model parameter identification problem, which,
previous position of xi; gbest is the best particle among the however, is extendable to other dynamic systems.
Simulation and experiment results demonstrate the
214 Li Liu, Wenxin Liu and David A. Cartes
effectiveness of the proposed approach. states and the output of the model. To evaluate the
parameters to be identified, system output y is compared
Start with model output ŷ . The fitness function can be defined as
Random initialization
the weighted quadratic function shown in (8).
C ( pˆ ) = ∫ ( y − yˆ ) W ( y − yˆ ) dt
T
(8)
Calculate initial cost
t
In (8), W is a positive definite matrix. Obviously, the
Move the particles with PSO velocity fitness function is a function of the estimated parameter
vector, p̂ . The identification error will result in a nonzero
Fitness evaluation for each particle C ( pˆ ) . Thus, the fitness function can be used to guide the
Obtain pbest, gbest
search for better estimation. Now, the identification problem
k=k+1 has been transformed into an optimization problem. Since
Calculate new searching directions traditional algorithms usually have difficulties optimizing
complex nonlinear systems with multiple local optima, the
No intelligent optimization algorithms mentioned before are a
Fitness < optimization error
better choice for this problem.
Yes
The structure of the PSO based parameter identification
Output result
approach can be illustrated using Figure 3. First, the system
input, u, is given to both the system to be identified and the
Stop
system model. Then, the outputs from the system and its
model are input to the performance evaluator, where the
Figure 2 : Flowchart of the PSO algorithm fitness will be calculated. The calculated fitness C ( pˆ ) is
then input to the PSO based identifier to identify the
IV. PSO Approach for PMSM Parameter unknown parameter vector p̂ . For offline applications, the
Identification identification process will end after p̂ has been successfully
identified. For online applications, after p̂ is successfully
For a system with known model structure but unknown
parameters, the parameter identification problem can be identified or the preset maximum iteration number has been
treated as an optimization problem. The basic idea is to reached, the estimated p̂ will be used to update the system
compare the system output with the model output. The model, and then the above process will be repeated.
discrepancy between the system and model outputs is
minimized by optimization based on a fitness function. The System input u Measured output y
fitness function is defined as a measure of how well the System
In (9), the phase currents iabc and the rotor speed ωr are Table 1 : PMSM specifications.
the measurable system outputs; iˆabc and ωˆ r are the estimated
values of iabc and ωr, which are calculated from the system PMSM parameters Nominal values (unit)
model; Rs and TLd are the unknown parameters to be Power rating Pr 19.8 (kw)
Rated speed wr 1700 (rpm)
identified, n is the number of sampled data, and the Current at rated speed Ir 41.56 (Amps RMS)
weighting factor w is defined as w = iabc / ωr . By Torque at rated speed Tr 67.27 (N.m)
introducing the weighting factor w, the impact of the motor Voltage constant ke 1.33 (Volt.sec/rad)
currents and motor speed to the fitness function in (9) is Torque constant kt 1.629 (N.m/Amp)
normalized. With this definition, errors in both the currents Max bus voltage VDC 560 (Volt)
Pole pairs np 4
and the rotor speed are minimized when the fitness function
Stator resistance Rs 0.17 (W)
is optimized. The PSO algorithm in Figure 4 is then q-axis inductance Lq 1.9 (mH)
performed repetitively until reaching the desired error limit. d-axis inductance Ld 1.9 (mH)
In the subsequent sections, both offline and online Static friction Tf 0.1483 (N.m)
simulations are performed to evaluate the performance of Damping coefficient B 0.00115 (N.m.sec/rad)
the PSO based parameter identification. Then, the PSO Moment of inertia J 0.008 (kg.m2)
based approach is also used to identify PMSM parameters
with a sample of experimental data from a practical PMSM During simulations, each swarm generation contains 10
testbed. position particles. Coefficients in (5), w, c1, and c2, are set to
0.8, 2, and 2 respectively. Figure 5 shows the optimization
V. Simulation Results process of the proposed algorithm for a sample dataset. In
this figure, (a) shows the optimization process of the fitness
In this section, a three phase PMSM with nonsalient poles is function (cost), (b) and (c) show the optimization process of
simulated in MATLAB® to demonstrate the performance of the identified parameters, Rs and TLd, respectively.
the proposed PSO approach applied to PMSM model Simulation studies show that the proposed approach
parameter identification. General structure of the simulation converges here within 20 generations. This 20-generation
setup is illustrated in Figure 4. In this diagram, the motor is upper bound was routinely observed. The efficiency and
applied to a variable frequency drive system with cascaded accuracy of the approach make it suitable for applications
PI-controllers. This design is widely used in industrial motor that require fast solutions.
drive applications. Similar structure is used in our Based on the observation on both fitness function and
experimental test bed presented in a later part of this paper. identification errors from offline identification, the PSO
Nominal parameters of the simulated PMSM are listed in based identifier can give good estimation within 20
Table 1. These parameters are chosen to be identical to the iterations in this application to PMSM. To ensure the
test motor datasheet parameters. However, in this estimation accuracy, the maximum number of allowed
simulation, several parameters, i.e. the stator resistance, Rs, iteration is set to 30 for each sample of data. During online
motor inertia, J, and frictional coefficient, B, are defined as simulation, the sampling frequency is set to 100kHz. That
variables of time in order to demonstrate PSO method’s means 3300 pairs of data are sampled within 0.033 second.
capability to track multiple time varying parameters. Then the sample data set is input to the PSO based identifier
According to (4), the time varying inertia and viscous to estimate the stator resistance Rs, and the load torque TLd.
frictional coefficient are interpreted as load torque After PSO completes its searching of 30 iterations, the
disturbance that should be reflected in the identified load identification process for this data set stops. Then, the
torque. All the setting on parameter fluctuations should be estimated parameters will be used as the initial values for
within a range that is likely to occur during realistic motor the estimation of next data set. The above process recurs
operations. during the online application.
Figure 6 and Figure 7 show the online identification
PWM Inverter
ωr* i qr
*
vqr
* results of the proposed algorithm. From these figures, it can
∑ PI ∑ PI be seen that the proposed approach can identify time varying
-
ωr - * * M
PMSM parameters successfully. Figure 8 shows the entire
idr v dr
∑ PI optimization process of the PSO algorithm during online
Encoder
- identification. For each run of the PSO based algorithm, the
r ia previously optimized parameters are used as the initial
i d abc ib
ic value. In doing so, the identification process for online
dq
i qr θr
applications tends to converge faster in terms of less number
of iterations. This is especially true for systems with slowly
time varying parameters. From Figure 8, it can be seen that
Figure 4 : PMSM current control in dq-reference frame the algorithm usually converges within 10 iterations because
of a good initial guess. But from the figure we can also see
216 Li Liu, Wenxin Liu and David A. Cartes
that the initial fitness for the runs of 5, 6, 7, and 8 are not as Note that the convergence speed of the identification for
good as others. This is because the fitness’s sensitivity to one data set is evaluated in terms of the number of PSO
parameter changes is different for different operating points. iterations. The time required is not evaluated strictly during
Sometimes, even a small change in the parameters will the online application. In this online application, the
result in a large change in the fitness. simulation for next period does not begin until the 30
iterations for the previous period are completed. Simulations
in Matlab® and Simulink® show that the identification for
one dataset can be done on the order of minutes. But this
should not be used as a precise evaluation on the required
time. The computational speed is impacted by several
factors, such as implementation hardware setup, the
sampling frequency, and the choice of Ordinary Differential
Equation (ODE) solvers, etc. As just mentioned, for
continuous online applications, tremendous computation
time can be saved due to good guesses of initial values for
each run. Therefore, the computation time should not be a
big issue for online implementations.
It is worth to mention that as a stochastic search
algorithm, PSO cannot guarantee global solutions in every
run. However, the application of this paper is on a
complicated practical problem. The focus is thus more
emphasized on the applicability in real situations rather than
theoretical study on convergence. With the parameter
Figure 5 : PSO optimization process for one dataset (a) settings as mentioned previously, the simulation study does
Fitness function (b) Estimation of stator resistance Rs not encounter the local optima problem. A reason for this is
(c) Estimation of load torque TLd that the PSO optimization is a continuous repetitive process
in this identification problem. Even one solution obtained
0.22 from a dataset has some errors due to a local optima, this
Stator resistance Rs (ohm)
0.16
Acutal resistance as defined VI. Experimental Verification
0 0.05 0.1 0.15 0.2 0.25 0.3
Time (sec) In order to further verify the effectiveness the proposed
identification method, experiments are performed on a
Figure 6 : Identification of the stator resistance Rs testbed PMSM drive system. This section presents the
3.06 experimental results.
3.04
Acutal torque as defined
Identified torque
The experimental data were obtained on a 680-V 19.8-
kw PM servo motor fed from a closed loop drive and
Torque (N.m)
3.02
Acknowledgment
This work is supported by the Office of Naval Research
(ONR) under grant N000140210623.
(a) References
Estimation [1] M. A. Rahman and P. Zhou, “Analysis of brushless
Rotor speed (rad/sec)
130
Measurement permanent magnet synchronous motors”, IEEE
120 Trans. Industrial Electronics, vol. 43, no. 2, 1996,
pp. 256–267.
110 [2] M. Ooshima, A. Chiba, A. Rahman, T. Fukao, “An
0.25 0.3 0.35 0.4
improved control method of buried-type IPM
Time (sec) bearingless motors considering magnetic saturation
(b) and magnetic pull variation”, IEEE Transactions
on Energy Conversion, vol. 19, no. 3, Sept. 2004,
Figure 10 : Experimental results with raw data: estimation pp. 569-575.
vs. Measurement (a) Phase a current ia (b) Rotor speed ωr [3] R. Lateb, N. Takorabet, F. Meibody-Tabar, A.
Mirzaian, J. Enon, and A. Sarribouette,
Instead of comparing the actual and identified “Performances comparison of induction motors and
parameters, the experimental results in Figure 10 compares surface mounted PM motor for POD marine
the performances of the identified model with the original propulsion”, in Conf. Rec. IEEE-IAS Annu.
system. This is important because PMSM’s parameters may Meeting. 2005.
deviate from its nominal value when its operating condition [4] F. Caricchi, F. Crescimbini, and O. Honorati,
changes. It is meaningless to compare the identified “Low-cost compact permanent magnet machine for
parameters to their nominal values. In Figure 10, (a) shows adjustable-speed pump application”, Industry
the stator phase a current ia, and (b) shows the rotor speed Applications, IEEE Trans. Industry Applications,
ωr. Since the performances of the model outputs match the vol. 34, no. 1, Jan.-Feb. 1998, pp. 109 – 116.
measurements quite well, it demonstrates the effectiveness [5] M. A. Rahman, D. M. Vilathgamuwa, M. N. Uddin
of the PSO based model parameter identification approach. and T. King-Jet, “Nonlinear control of interior
In this experiment, raw data from the measurement are permanent magnet synchronous motor”, IEEE
directly used for identification. Therefore, comparing to Trans. Industry Applications, vol. 39, no. 2, 2003,
many other methods, PSO approach presents superior pp. 408–416.
capability at the presence of measurement noises so that [6] S. Weisgerber, A. Proca, and A. Keyhani,
there is no need to process the original measurements in “Estimation of permanent magnet motor
advance, e.g. [24]. parameters,” in Proc. IEEE Ind. Appl. Soc. Annu.
Meeting, New Orleans, LA, Oct. 1997, pp. 29–34.
VII. Conclusions [7] K. M. Rahman, and S. Hiti, “Identification of
machine parameters of a synchronous motor”,
In this paper, a new approach based on particle swarm IEEE Transactions on Industry Applications, vol.
optimization is proposed for parameter identification with 41, no. 2, 2005, pp. 557–565.
nonlinear model structure. The method is applied to the [8] A. Kilthau, and J. Pacas, “Appropriate models for
218 Li Liu, Wenxin Liu and David A. Cartes