The Research of PMSM RBF Neural Network PID Parameters

Self-tuning in Elevator

Wang Tong-xu, Ma Hong-yan

School of Electrical and Information Engineering, Beijing University of Civil Engineering and Architecture, Beijing 100044
E-mail: wangtongxulixing@163.com

Abstract: To make the elevator operates more smoothly and obtain quicker response speed, the traditional PID
algorithm of elevator traction machine speed system needs to be improved. This paper adopts RBF neural network PID
control algorithm to control the speed system of PMSM .The simulation results verify that the dynamic performances of
PMSM have been improved. In addition, combined with the elevator speed curve, the simulation results verify that the
RBF neural network PID control algorithm applied to the elevator with PMSM control system is feasible .

Key Words: PMSM, RBF neural network , Elevator

invariable. Without Considering the effects of temperature

1 INTRODUCTION and frequency changes on winding resistance.
Elevator is a significant transportation used in The stator voltage equation of PMSM is
high-rise buildings and it is closely related with people's ­ dψ A
lives. Traction machine is the power source of the °uA =RsiA + dt
elevator and the correct choice of traction machine is very °
important. Compared with the traditional traction machine, ° dψ B ˄1˅
® uB =RsiB +
because of simple structure, high reliability and good ° dt
performance , sinusoidal PMSM is more and more widely ° dψ C
used in elevator-driving system . PID control is a kind of ° uC =RsiC + dt
control strategy which have been developed. It was widely ¯
applied to many areas of production, because of its simple According to the principle of coordinate
algorithm, high reliability and easy implement. transformation , the stator voltage equation of PMSM from
The mathematics model of PMSM is a kind of ABC three-phase coordinate transformation to d-q
high-orders, nonlinear and strongly-coupled system, two-phase rotating coordinate is given by:
conventional PID controller can not be utilized with good
performance. ­ dψ sd
To make the elevator operates more smoothly and °°U sd = dt − ωrψ sq + Rs I sd ˄2˅
obtain quicker response speed, the traditional PID ®
algorithm of elevator traction machine speed system °U = dψ sq +ω ψ + R I
s sq
needs to be improved. °̄ sq dt
r sd

2 The Mathematical Model of PMSM The flux equation of PMSM in d-q synchronous
rotating coordinates is
PMSM is a non-linear, strong-coupling and
multi-variable system. To obtain high dynamic ­°ψ sd = Lsd I sd +ψ f
® ˄3˅
performance, dynamic mathematical models must be °̄ψ sq = Lsq I sq
established.
The mathematical model of PMSM is divided into The electromagnetic torque equation in d-q
three parts: voltage equation, flux linkage equation and synchronous rotating coordinates is
>@
electromagnetic torque equation .
To simplify the mathematical model of PMSM, some
>@
Tem = pn ( ȥf I q - ȥf I d )
assumptions are made as follows : The space harmonic is ˄4˅
neglected and three-phase winding is symmetrical. = pn ª¬ȥf I q + ( Ld - Lq )< I d I q º¼
Magnetic saturation and iron core loss are neglected and the
self-inductance and mutual inductance are considered
Whereˈ U sd — d axis stator voltage;
* U sq — q axis stator voltage;
This work is supported by Funding Project for ministry of housing and
Urban-Rural Development 2011 science and technology project,(approve I sd — d axis stator current;
number: 2011-k8-3).

978-1-4799-7016-2/15/$31.00 2015
c IEEE 3350
 I s q — q axis stator current; W = [w1 , w 2 ," , w 6 ]T ˄6˅
ω r —rotor rotation angular speed; The output of neural network is
ψ sd —d axis stator flux ; ym = w1h1 +w2 h2 + " +w6 h6 ˄7˅
ψ sq — q axis stator flux; The approximation error indicator of neural network is
1
L s d — d axis inductance; J = (y(k) − ym (k))2 ˄8˅
L s q — q axis inductance;
2
The weights learning algorithm of neural network
ψ f —rotor flux. could be obtained by gradient descent method.
The isd = 0 control mode has been adopted in the
PMSM speed control and DC motor equivalent model is ­° Δwj (k) = η (y(k) − y m (k)) h j
®
obtained from the further simplification. This model is
>@ °̄ wj (k) = w j (k − 1) + Δ w j (k) + α (w j (k − 1) − wj (k − 2))
shown in Fig.1 .
˄9˅
TL
w In the same way, the following iterative algorithm
usq 1 .5 ⋅ n p L m d I f Te
 formula can be obtained.
R s + Lsq s
­bj (k) = b j (k − 1) + Δ b j (k) + α (b j (k − 1) − bj (k − 2))
Js

°
® & x j − c ji &2
Δ
° jib (k) = η (y(k) − y m (k)) w j h j
np Lmd I f ¯ bj3
Fig.1: DC motor equivalent model ˄10˅
3 RBF neural network PID control system ­c ji (k) = c ji (k − 1) + Δ c ji (k) + α (c ji (k − 1) − c ji (k − 2))
°
RBF neural network is a kind partial approaches ® x j − c ji
network. It has an excellent capacity of nonlinearity ° Δ c ji (k) = η (y(k) − ym (k)) w j b 2
>@
approximation and generalization . In connection with ¯ j

disadvantage of conventional PI control, RBF neural ˄11˅

>@
network PID self-tuning algorithm could be adopted . Whereˈ i=1, 2ˈ
3˗
3.1. The algorithm of RBF neural network
η is the learning rate˗
α is the momentum factor;
RBF neural network model is made up of input
layer, hidden layer and output layer. RBF neural network
η ∈ [0,1] , α ∈ [0,1] .
>@
structure diagram is shown in Fig.2 . 3.2. RBF neural network PID control system
RBF neural network PID self-tuning control system
diagram is shown in Fig.3. The control idea is: the initial
stage adopts the PID control, and then transition to the
RBF neural network control. Its aim is to take full
advantage of RBF neural network and traditional PID .

Fig.2: The structure of RBF neural network

The input vector of RBF neural network

is x = [x1 , x 2, x 3 ]
T
and the output of neural network
hidden layer is h = [ h1 , h2, " , hj ] .
T

& x( k ) − c 2j &
hj = exp( − ) ˈ j = 1, 2, 3," , 6 ˄5˅
2bj2 Fig.3 : RBF neural network PID self-tuning system

Whereˈ c j = [c j1 , c j2 , c j3 ]T ; The control arithmetic of increment PID is

bj = [b1 ," , b m ]T . u (k ) = u (k − 1) + kP (e(k ) − e(k − 1)) + kI e(k )

The weight vector of neural network is : + kD (e(k ) − 2e(k − 1) + e(k − 2)) ˄12˅

2015 27th Chinese Control and Decision Conference (CCDC) 3351

 The control error of increment PID controller is
e(k) = rin (k) − y(k) ˄13˅
The error indicator of neural network weight is :
1
E (k) = e(k) 2 ˄14˅
2
The inputs of PID is ˖
xc (1) = e(k) − e(k − 1) ˄15˅
xc (2) = e(k) ˄16˅
xc (3) = e(k) − 2 e(k − 1) + e(k − 2) ˄17˅

The adjustment quantity of k P ǃ k I ǃ k D can be Fig.5˖RBF neural network PID controller

obtained by gradient descent method.
∂E ∂E ∂y ∂Δu ∂y The parameters of the PMSM used in this simulation
Δkp = −η = −η = ηe(k) xc (1) are given by table 1 and the other parameters are as
∂kp ∂y ∂Δu ∂kp ∂Δu follows: α =0.05ˈη =0.5ˈ β =0.01.
˄18˅ >@
∂E ∂E ∂y ∂Δu ∂y Table 1:The Parameters of The Motor
Δ k I = −η = −η = ηe(k) xc (2)
∂kI ∂y ∂Δu ∂kI ∂Δu Motor parameter parameter values
˄19˅
Motor pole pairs P 1
∂E ∂E ∂y ∂Δu ∂y
Δ k D = −η = −η = ηe(k) xc (3) stator resistance R s 2.875Ω
∂kD ∂y ∂Δu ∂kD ∂Δu Permanent magnet flux
linkage Ψ f 0.175Wb
˄20˅
d axis inductance L d
∂y (k) ∂y (k) 6 c − Δu (k) 0.0085H
Where, ≈ m = ¦ wjhj ji 2 q axis inductance L
∂Δu (k) ∂Δu (k) j =1 bj q 0.0085H
moment of inertia J 0.0008kg ⋅ m 2
4 The Simulation
The motor is started in the condition of no load and
4.1 The simulation research of PMSM speed control
the given signal is 1000r/min ,the simulation time is 0.3s.
speed
Exerted T=2.5 N·m on t=0.1s .The speed of RBF neural
PMSM RBF neural network PID parameters network PID control is shown in Fig.6, and speed of
self-tuning control system is shown in Fig.4. This control ordinary PID control is shown in Fig.7.
system adopts speed and current double closed loop
control. The inner loop is current loop with classic PI
controller. The outer loop is speed loop with BP neural
network PID controller.

Fig. 6 :The speed of the RBF neural network PID control

Fig.4: PMSM RBF neural network PID parameters self-tuning control

system

The model of RBF neural network PID controller is

shown in Fig. 5.

3352 2015 27th Chinese Control and Decision Conference (CCDC)

 Fig.8: The given speed curve of elevator
Fig.7: The speed of the ordinary PID control

The elevator speed curve module designed as the

As shown in Figure 6-7, the dynamic performance
given signal module and the simulation result is shown in
indexes is shown in table 2.
Fig.9.
Table 2 Dynamic performance comparison table
Control algorithm classic PID RBF neural
network PID
Performance index

 tr /s 0.0146 0.008

 t p /s 0.0155 0.009

 ts /s 0.03 0.013

 M P /s 4% 1.9%

It can be seen from table 2 that, by adopting the RBF

neural network PID algorithm, the adjusting time Fig.9: The actual running speed curve of elevator
decreased from 0.03s to 0.013s. At the same time, the
overshoot reduced from 4% to 1.9%. As shown in Fig.8-9, under the RBF neural network
Therefore, regarding the PMSM speed control system, PID control algorithm, the actual speed of the motor is
the RBF neural network PID control can improve control almost consistent with given elevator speed curve. These
accuracy, reduce overshoot and improve response time. results verify that this algorithm can make the motor speed
better run with a given speed signal.
4.2 The simulation research of elevator's velocity
curve 5 Conclusion
The elevator for the significant transportation, need to In the paper, the PMSM speed control system model
>@
meet the demands of comfort, safety and rapidity . Based based on RBF neural network PID self-tuning control
on the operation characteristics of elevator traction algorithm is established for analyzing the dynamic
>@
system , a elevator speed curve function is designed in performances of PMSM. The simulation results show
that the dynamic performances of PMSM can be
formula (21), and the elevator speed curve is shown in Fig.
effectively improved by the RBF neural network PID
8.
self-tuning control algorithm, as it has a quicker response
­ 375t 2 0s ≤ t < 1s and excellent robustness .
°
° 375+750(t-1) 1s ≤ t < 2s The simulation result of motor speed ,which combines
°1500-375(3-t) 2 2s ≤ t < 3s elevator's velocity curve and PMSM speed system, shows
°
(21) that the actual speed curve is almost consistent with the
® 1500 3s ≤ t < 6s given speed curve and the RBF neural network PID
°1500-375(t-5) 2 6s ≤ t < 7s
° self-tuning control as the elevator traction motor control
° 1125-750(t-6) 7s ≤ t < 8s algorithm is feasible. The RBF neural network PID
° self-tuning control can improve the performance of elevator
¯ 375(t-8)
2
8s ≤ t < 9s
 and it has a certain application value.
As can be seen from the simulation waveform, the
speed curve meet the operation characteristics of the
elevator.

2015 27th Chinese Control and Decision Conference (CCDC) 3353

 REFERENCES
[1] Chen Bo-shi, Electric Drive Automatic Control System̅
Motion Control System ,Ver3, Beijing, Mechanical Industry
Press,2012.
[2] Nie Mu-han , Comparative Study on Permanent Magnet
Synchronous Motor Control Systems in Elevator. The 26th
Chinese Control and Decision Conference, 1568 ̅ 1573
,2014
[3] Sun Zeng-xi, Intelligent control theory and technology,Ver2,
Beijing, Tsinghua University press,2011.
[4] Liu Jin-kun,The advanced PID control MATLAB simulation
Ver3. Beijing, electronic industry press,2011.
[5] Liu Jin-kun, RBF neural network adaptive control MATLAB
simulation,Ver1, Beijing, Tsinghua University press,2014
[6] Hong Nai-gang, Power electronics, motor control system
modeling and simulation,Ver1,Beijing, Mechanical industry
press,2010
[7] Ye An-li, The elevator control technology,Ver2, Beijing,
Mechanical Industry Press,2013

3354 2015 27th Chinese Control and Decision Conference (CCDC)