# US 2023/0412098 A1 · Efficiency optimization control method and system for permanent magnet synchronous motor

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2023/0412098 A1
- Document ID: US20230412098A1
- Status: Published application
- Kind: A1
- Application number: 18/239,994
- Filed: 2023-08-30
- Published: 2023-12-21
- Earliest claimed: 2022-05-18
- Assignees: Xiangtan University; Foshan Green Intelligent Manufacturing Research Institute of Xiangtan University; Foshan Shunde Lepuda Motor Co., Ltd
- Applicants: Xiangtan University; Foshan Green Intelligent Manufacturing Research Institute Of Xiangtan University; Foshan Shunde Lepuda Motor Co.,Ltd
- Inventors: Siqi Peng; Weijun Li; Dan Guo; Hongyu Peng; Hui Li; Pingan Tan
- Web page: https://patentgrep.com/patent/US20230412098A1

## CPC

- H02P 21/0003: {Control strategies in general, e.g. linear type, e.g. P, PI, PID, using robust control}
- H02P 21/0014: {using neural networks}
- H02P 21/14: Estimation or adaptation of machine parameters, e.g. flux, current or voltage
- H02P: CONTROL OR REGULATION OF ELECTRIC MOTORS, ELECTRIC GENERATORS OR DYNAMO-ELECTRIC CONVERTERS; CONTROLLING TRANSFORMERS, REACTORS OR CHOKE COILS

## Abstract

An efficiency optimization control method and system for permanent magnet synchronous motor are provided, and the method includes: step 1: obtaining an approximate optimal direct axis (d-axis) current of a permanent magnet synchronous motor by using a loss model algorithm; step 2: performing, by using the approximate optimal d-axis current as an initial value and using a deep reinforcement learning algorithm, an optimizing process on the approximate optimal d-axis current to construct an optimal deep reinforcement learning model; and step 3: inputting currently acquired state data of the permanent magnet synchronous motor into the optimal deep reinforcement learning model to obtain a control parameter value corresponding to an optimal efficiency of the permanent magnet synchronous motor, and controlling the permanent magnet synchronous motor based on the control parameter value. Therefore, real-time target control for the permanent magnet synchronous motor is achieved.

## Claims

### Claim 1

An efficiency optimization control method for permanent magnet synchronous motor, comprising:
step 1: obtaining an approximate optimal direct axis (d-axis) current id of a permanent magnet synchronous motor by using a loss model algorithm;
step 2: performing, by using the approximate optimal d-axis current as an initial value and using a deep reinforcement learning algorithm, an optimizing process on the approximate optimal d-axis current id to construct an optimal deep reinforcement learning model; and
step 3: inputting currently acquired state data of the permanent magnet synchronous motor into the optimal deep reinforcement learning model to obtain a control parameter value corresponding to an optimal efficiency of the permanent magnet synchronous motor, and controlling the permanent magnet synchronous motor based on the control parameter value.

### Claim 2 (depends on 1)

The efficiency optimization control method for permanent magnet synchronous motor as claimed in claim 1, wherein the step 1 comprises:
analyzing a loss of the permanent magnet synchronous motor to construct a loss model of the permanent magnet synchronous motor;
obtaining the approximate optimal d-axis current based on the loss model of the permanent magnet synchronous motor.

### Claim 3 (depends on 2)

The efficiency optimization control method for permanent magnet synchronous motor as claimed in claim 2, wherein the analyzing a loss of the permanent magnet synchronous motor to construct a loss model of the permanent magnet synchronous motor comprises:
obtaining dynamic voltage equations of the permanent magnet synchronous motor based on a d-axis equivalent circuit of the permanent magnet synchronous motor considering a copper loss and an iron loss and a quadrature axis (q-axis) equivalent circuit of the permanent magnet synchronous motor considering the copper loss and the iron loss, wherein the dynamic voltage equations of the permanent magnet synchronous motor are expressed as follows:
u
q
=
R
s
⁢
i
q
+
L
q
⁢
d
dt
⁢
i
w
⁢
q
+
n
p
⁢
ω
r
(
L
d
⁢
i
w
⁢
d
+
ψ
m
)
;
u
d
=
R
s
⁢
i
d
+
L
d
⁢
d
dt
⁢
i
w
⁢
d
-
n
p
⁢
ω
r
⁢
L
q
⁢
i
w
⁢
q
;
wherein ud represents a d-axis armature voltage, and uq represents a q-axis armature voltage; id represents a d-axis armature current, and iq represents a q-axis armature current; iwd represents an active component of the d-axis current, and iwq represents an active component of the q-axis current; Rs represents an armature resistance; Ld represents a d-axis armature inductance, and Lq represents a q-axis armature inductance; Ψm represents a magnetic linkage of a rotor; ωr represents a mechanical angular velocity of the rotor; and np represents a pole pair number of the permanent magnet synchronous motor;
obtaining an iron loss component of the d-axis current and an iron loss component of the q-axis current based on an equivalent circuit voltage equation of the permanent magnet synchronous motor, wherein the iron loss component of the d-axis current and the iron loss component of the q-axis current are expressed as follows:
i
F
⁢
e
⁢
d
=
-
n
p
⁢
ω
r
⁢
L
q
⁢
i
w
⁢
q
R
F
⁢
e
;
i
F
⁢
e
⁢
q
=
n
p
⁢
ω
r
(
L
d
⁢
i
w
⁢
d
+
ψ
m
)
R
F
⁢
e
;
wherein iFed represents the iron loss component of the d-axis current, iFeq represents the iron loss component of the q-axis current, and RFe represents an equivalent iron loss resistance;
obtaining the loss model of the permanent magnet synchronous motor based on the iron loss component of the d-axis current and the iron loss component of the q-axis current, wherein the loss model of the permanent magnet synchronous motor is expressed as follows:
P
loss
=
P
Cu
+
P
F
⁢
e
=
3
2
⁢
R
s
(
i
d
2
+
i
q
2
)
+
3
2
⁢
R
F
⁢
e
(
i
F
⁢
e
⁢
d
2
+
i
F
⁢
e
⁢
q
2
)
=
3
2
⁢
R
s
⁢
{
[
T
e
n
p
⁢
Ψ
m
+
n
p
⁢
ω
r
(
L
d
⁢
i
w
⁢
d
+
Ψ
m
)
R
F
⁢
e
]
2
+
(
i
w
⁢
d
-
ω
r
⁢
L
q
⁢
T
e
Ψ
m
⁢
R
F
⁢
e
)
2
}
+
3
⁢
n
p
2
⁢
ω
r
2
R
F
⁢
e
[
(
L
d
⁢
i
w
⁢
d
+
Ψ
m
)
2
+
(
L
q
⁢
T
e
n
p
⁢
Ψ
m
)
2
]
;
wherein Te represents an electromagnetic torque of the permanent magnet synchronous motor.

### Claim 4 (depends on 3)

The efficiency optimization control method for permanent magnet synchronous motor as claimed in claim 3, wherein the obtaining the approximate optimal d-axis current based on the loss model of the permanent magnet synchronous motor comprises:
obtaining the active component of the d-axis current based on a formula
dP
l
⁢
o
⁢
s
⁢
s
di
w
⁢
d
=
0
,
wherein the active component of the d-axis current is expressed as follows:
i
w
⁢
d
=
-
n
p
2
⁢
ω
r
2
⁢
ψ
m
⁢
L
d
(
R
s
+
R
F
⁢
e
)
R
s
⁢
R
Fe
2
+
(
R
s
+
R
F
⁢
e
)
⁢
n
p
2
⁢
ω
r
2
⁢
L
d
2
;
obtaining a first formula of the approximate optimal d-axis current based on an equivalent circuit node current equation of the permanent magnet synchronous motor, wherein the first formula of the approximate optimal d-axis current is as follows:
id=iwd+iFed;
obtaining a second formula of the approximate optimal d-axis current based on the first formula of the approximate optimal d-axis current, the active component of the d-axis current, and the iron loss component of the d-axis current, wherein the second formula of the approximate optimal d-axis current is expressed as follows:
i
d
=
(
1
+
n
p
2
⁢
ω
r
2
⁢
L
d
⁢
L
q
R
Fe
2
)
[
-
n
p
2
⁢
ω
r
2
⁢
ψ
m
⁢
L
d
(
R
s
+
R
F
⁢
e
)
R
s
⁢
R
Fe
2
+
(
R
s
+
R
F
⁢
e
)
⁢
n
p
2
⁢
ω
r
2
⁢
L
d
2
]
+
n
p
2
⁢
ω
r
2
⁢
L
q
⁢
ψ
m
R
Fe
2
-
n
p
⁢
ω
r
⁢
L
q
R
F
⁢
e
⁢
i
q
.

### Claim 5 (depends on 4)

The efficiency optimization control method for permanent magnet synchronous motor as claimed in claim 4, wherein the step 2 comprises:
step 2.1: constructing a deep reinforcement learning model, wherein the deep reinforcement learning model comprises a value network QV, a target network QT, a state space S, an action space A and a reward function Reward; structures of the value network Qv and the target network QT are the same; each state in the state space S is represented as a four-dimensional vector (T,v,acc,T1), where T represents an electromagnetic torque of the permanent magnet synchronous motor, v represents a rotary speed of the permanent magnet synchronous motor, acc represents an acceleration of the permanent magnet synchronous motor; T1 represents a load torque of the permanent magnet synchronous motor, and each action in the action space A is a value of the approximate optimal d-axis current id;
step 2.2: performing, by using the deep reinforcement learning model, the optimizing process on the approximate optimal d-axis current to obtain a global optimization data pool;
step 2.3: detecting and eliminating abnormal data in the global optimization data pool; and
step 2.4: training the deep reinforcement learning model based on sample data in the global optimization data pool after eliminating the abnormal data, determining whether a termination condition is satisfied, in a situation that the termination condition is satisfied, then stopping training of the deep reinforcement learning model and thereby obtaining the optimal deep reinforcement learning model; otherwise returning to and performing the step 2.2.

### Claim 6 (depends on 5)

The efficiency optimization control method for permanent magnet synchronous motor as claimed in claim 5, wherein the step 2.2 comprises:
step 2.2.1: constructing a data pool D, wherein a capacity of the data pool D is C;
step 2.2.2: measuring a current state s of the permanent magnet synchronous motor, wherein the current state s comprises a current electromagnetic torque, a current rotary speed, a current acceleration and a current load torque of the permanent magnet synchronous motor; recording the approximate optimal d-axis current as an action a, executing the action a, controlling the permanent magnet synchronous motor based on the approximate optimal d-axis current, automatically updating the current state s of the permanent magnet synchronous motor to a next state s′, and calculating a reward value r for executing the action a in the current state s through the reward function; and storing current sample data (s, a, r, s′) in the data pool D;
step 2.2.3: assigning s′ to the current state s, and inputting the current state s into the value network QV, obtaining a prediction -value of each action in the state space under the current state s by using a parameter θ of the value network QV, determining a target action by using an epsilon (ε)-greedy strategy, and recording the target action as a;
step 2.2.4: executing the target action a, automatically updating the current state of the permanent magnet synchronous motor to a next state s′, and calculating the reward value r for executing the action a in the current state s through the reward function; and storing current sample data (s, a, r, s′) in the data pool D; and
step 2.2.5: executing a loop by returning to the step 2.2.3 until the data pool is full to obtain a current data pool, wherein the current data pool is used as the global optimization data pool.

### Claim 7 (depends on 6)

The efficiency optimization control method for permanent magnet synchronous motor as claimed in claim 6, wherein formulas of the reward function are designed as follow:
r
⁡
(
s
,
a
)
=
a
1
⁢
Reward
⁢
T
⁡
(
k
+
1
)
+
a
2
⁢
Reward
⁢
v
⁡
(
k
+
1
)
+
a
3
⁢
Reward
⁢
acc
⁡
(
k
+
1
)
+
a
4
⁢
Reward
⁢
T
⁢
1
⁢
(
k
+
1
)
;
Reward
⁢
T
⁡
(
k
+
1
)
=
-
∂
T
⁡
(
k
)
∂
i
d
+
T
⁡
(
k
)
2
;
Reward
⁢
v
⁡
(
k
+
1
)
=
-
∂
v
⁡
(
k
)
∂
i
d
+
v
⁡
(
k
)
2
;
Reward
⁢
acc
⁡
(
k
+
1
)
=
-
∂
a
⁢
c
⁢
c
⁡
(
k
)
∂
i
d
+
a
⁢
c
⁢
c
⁡
(
k
)
2
;
Reward
⁢
T
⁢
1
⁢
(
k
+
1
)
=
-
∂
T
⁢
1
⁢
(
k
)
∂
i
d
+
T
⁢
1
⁢
(
k
)
2
;
wherein T(k), v(k), acc(k) and T1(k) respectively represent an electromagnetic torque, a rotary speed, an acceleration and a load torque of the permanent magnet synchronous motor before executing the action a; T(k+1), v(k+1), acc(k+1) and T1(k+1) respectively represent an electromagnetic torque, a rotary speed, an acceleration and a load torque of the permanent magnet synchronous motor after executing the action a; a1, a2, a3 and a4 represent weight parameters of four reward factors; and r(s,a) represents the reward value for executing the action a under the current state s={T(k), v(k), acc(k),T1(k)}.

### Claim 8 (depends on 6)

The efficiency optimization control method for permanent magnet synchronous motor as claimed in claim 6, wherein the step 2.3 comprises:
step 2.3.1: calculating a k-th reachable distance of each data point in a k-th distance neighborhood of a data point p for each data point p in the global optimization data pool, wherein a formula of the k-th reachable distance is expressed as follows:
kreach(o,P)=max(dk(o),d(o,p));
wherein dk(o) represents a k-th distance of a neighborhood point o, and d(o,p) represents a distance between the neighborhood point o and the data point p;
step 2.3.2: calculating a k-th local reachable density Preach (P) of each data point p, wherein a formula of the k-th local reachable density Preach (P) is expressed as follows:
p
r
⁢
each
(
p
)
=
1
/
(
∑
o
∈
N
k
(
p
)
k
reach
(
o
,
p
)
❘
"\[LeftBracketingBar]"
N
k
(
p
)
❘
"\[RightBracketingBar]"
)
;
wherein Nk(p) represents the k-th distance neighborhood of the data point p; and |Nk(p)| is a module of Nk(p);
step 2.3.3: calculating a k-th local outlier factor of each data point p, wherein a formula of the k-th local outlier factor of each data point p is expressed as follows:
LOF
k
(
p
)
=
(
∑
o
∈
N
k
(
p
)
p
reach
(
o
)
❘
"\[LeftBracketingBar]"
N
k
(
p
)
❘
"\[RightBracketingBar]"
)
/
p
r
⁢
each
(
p
)
;
step 2.3.4: analyzing the k-th local outlier factor, where in a situation that the k-th local outlier factor is smaller than 1, the data point p is not an outlier; and in a situation that the k-th local outlier factor is larger than 1, the data point p is an outlier; and
step 2.3.5: eliminating the outlier.

### Claim 9 (depends on 8)

The efficiency optimization control method for permanent magnet synchronous motor as claimed in claim 8, wherein the step 2.4 comprises:
step 2.4.1: initializing an iteration number t as t=1; randomly selecting n sample data from the global optimization data pool after eliminating the abnormal data; in a situation that a reward value r of each of the n sample data is larger than a set threshold, then stopping the training of the deep reinforcement learning model and obtaining a value network QV, wherein currently obtained value network QV is the optimal deep reinforcement model; otherwise performing step 2.4.2;
step 2.4.2: using the randomly selected n sample data as inputs of the value network QV and the target network QT;
for each sample data (s, a, r, s′), obtaining a corresponding prediction -value based on the current state s and the action a of the sample data (s, a, r, s′), and the parameter θ of the value network QV, recording a prediction -value corresponding to i-th sample data as Qpredict(i), wherein i=1, 2, . . . , n;
for the each sample data (s, a, r, s′), obtaining a corresponding -value Q (s′, a′, θT) for executing an actions a′ in the state space under the next state s′ based on the next state s′ and a parameter θT of the target network QT to thereby obtain a plurality of -values, and selecting a maximum value
max
a
′
⁢
Q
⁡
(
s
′
,
a
′
,
θ
T
)
of the plurality of -values; and calculating a target -value QTarget(i) corresponding to the i-th sample data based on the maximum value
max
a
′
⁢
Q
⁡
(
s
′
,
a
′
,
θ
T
)
,
wherein a formula of the target -value QTarget(i) is expressed as follows:
Q
Target
⁡
(
i
)
=
r
i
+
γ
max
a
′
Q
⁡
(
s
′
,
a
′
,
θ
T
)
;
wherein γ represents a reward attenuation factor, and γ is in a range of [0,1]; and ri represents a reward value of the i-th sample data;
step 2.4.3: calculating a loss function L by using the prediction Q-value Qpredict(i) and the target -value QTarget(i), wherein a formula of the loss function L is expressed as follows:
L
=
1
n
⁢
∑
i
=
1
n
⁢
(
Q
Target
⁡
(
i
)
-
Q
predict
⁡
(
i
)
)
2
;
step 2.4.4 minimizing the loss function L by using an error reversal method, updating the parameter θ of the value network QV by using a gradient descent method, wherein a formula for updating the parameter θ is expressed as follows:
θ
=
θ
-
α
⁢
∂
L
∂
θ
;
wherein a represents a learning law;
step 2.4.5: determining whether an equation t=K is satisfied, in a situation that the equation t=K is satisfied, updating the parameter θT of the target network QT as θT=θ, and t=1; otherwise t=t+1; wherein K represents an update cycle of the parameter of the target network; and
step 2.4.6: returning to and performing the step 2.2.3.

### Claim 10 (depends on 9)

The efficiency optimization control method for permanent magnet synchronous motor as claimed in claim 9, wherein in the step 2.4.4, a is taken as 0.01.

### Claim 11 (depends on 9)

The efficiency optimization control method for permanent magnet synchronous motor as claimed in claim 9, wherein in the step 2.4.5, the update cycle K of the parameter of the target network is set as three simulation cycles.

### Claim 12 (depends on 9)

The efficiency optimization control method for permanent magnet synchronous motor as claimed in claim 9, wherein in the step 3, the control parameter value comprises an optimal d-axis current and an optimal q-axis current; and the step 3 comprises: predicting the optimal d-axis current making the permanent magnet synchronous motor run with the optimal efficiency based on the optimal deep reinforcement learning model; controlling the permanent magnet synchronous motor based on the optimal d-axis current, compensating a q-axis current of the permanent magnet synchronous motor based on the optimal d-axis current to thereby achieve an optimal control for efficiency of the permanent magnet synchronous motor in a steady state, wherein a formula of a q-axis current variation Δid is expressed as follows:
Δ
⁢
i
q
=
(
L
d
-
L
q
)
⁢
i
d
Ψ
m
+
(
L
d
-
L
q
)
⁢
(
i
d
+
Δ
⁢
i
d
)
⁢
i
q
⁢
Δ
⁢
i
d
;
wherein Ld represents the d-axis armature inductance, and Lq represents the q-axis armature inductance of; Ψm represents the magnetic linkage of the rotor; id represents a d-axis current before adjusting, and Δid is a d-axis current variation after adjusting.

### Claim 13 (depends on 1)

An efficiency optimization control system for permanent magnet synchronous motor, comprising a memory and a processor, wherein the memory is configured to store a computer program, and the computer program is configured to be executed by the processor to implement the efficiency optimization control method for permanent magnet synchronous motor as claimed in claim 1.

### Claim 14 (depends on 13)

The efficiency optimization control system for permanent magnet synchronous motor as claimed in claim 13, wherein the system further comprises: a data acquisition circuit, configured to acquire current state data of the permanent magnet synchronous motor and input the current state data into the processor;
wherein the current state data comprises a current electromagnetic torque, a current rotary speed, a current acceleration and a current load torque of the permanent magnet synchronous motor.
