CN116027672B - Model prediction control method based on neural network - Google Patents
Model prediction control method based on neural network Download PDFInfo
- Publication number
- CN116027672B CN116027672B CN202310310441.0A CN202310310441A CN116027672B CN 116027672 B CN116027672 B CN 116027672B CN 202310310441 A CN202310310441 A CN 202310310441A CN 116027672 B CN116027672 B CN 116027672B
- Authority
- CN
- China
- Prior art keywords
- neural network
- weight coefficient
- transient
- value
- working condition
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/60—Other road transportation technologies with climate change mitigation effect
- Y02T10/72—Electric energy management in electromobility
Abstract
The invention discloses a model prediction control method based on a neural network, which belongs to the technical field of model prediction and comprises the following steps: s1, establishing a state equation of a permanent magnet synchronous motor; s2, discretizing the state equation to obtain a prediction equation; s3, calculating electromagnetic torque T e Effective component i of d-axis stator current wd Total controllable loss P Loss Then calculating a prediction equation of the next moment, and repeating iterative calculation to obtain the prediction value of each variable at the next moment; and S4, designing a weight coefficient by using a neural network. By adopting the model predictive control method based on the neural network, clear connection among a control target, a weight coefficient, a working condition and a performance index is established, the weight coefficient can be automatically set, and the optimality of the weight coefficient is ensured under different working conditions.
Description
Technical Field
The invention relates to a model prediction technology, in particular to a model prediction control method based on a neural network.
Background
Model predictive control is a type of control technology and is widely applied to various control objects in the fields of power systems and power electronics, such as wind farm control, micro-grid control, converter control, motor control and the like. The model predictive control is a new generation control technology in the field of motor driving, has good dynamic performance, is easy to process various constraint conditions, is convenient to implement in a control object of a nonlinear model, and has great development potential in the future.
The conventional model predictive control method has the following disadvantages:
1. no clear link is established between the control target, the weight coefficient, the working condition and the performance index. The weight coefficient needs to be simulated or tested in a large amount, and the setting can be manually completed according to experience, so that the automatic setting can not be realized. When the control object is changed, for example, the control object is changed from the permanent magnet synchronous motor to the grid-connected converter, at this time, the control target, the number of weight coefficients and performance indexes are often changed, and the conventional model prediction control method needs to be subjected to a process of manually setting the weight coefficients again.
2. Most of the method does not consider motor loss and does not bring the system efficiency into performance indexes, so that the universality of the traditional model predictive control method is insufficient in certain occasions needing energy conservation and efficiency improvement. Although a small number of methods consider motor loss and improve system efficiency, there is no consideration of trade-off between system efficiency and other performance indexes such as transient time (in the existing methods, the improvement of system efficiency can bring deterioration of other performance indexes such as transient time, and how to design weight coefficients according to different requirements of users on the performance indexes is a problem of the existing methods).
3. The existing model prediction control method based on the neural network can find out the optimal weight coefficient under a certain specific working condition. But when the operating conditions change, a new neural network is retrained. In the occasion of frequent change of working condition, the existing method is long in time consumption and complicated in weight coefficient setting process.
Disclosure of Invention
In order to solve the problems, the invention provides a model predictive control method based on a neural network, which establishes clear connection among a control target, a weight coefficient, a working condition and a performance index, can automatically set the weight coefficient, and ensures the optimality of the weight coefficient under different working conditions.
In order to achieve the above object, the present invention provides a model predictive control method based on a neural network, comprising the steps of:
s1, establishing a state equation of a permanent magnet synchronous motor;
s2, discretizing the state equation to obtain a prediction equation;
s3, calculating electromagnetic torque T e Effective component i of d-axis stator current wd Total controllable loss P Loss Then calculating a prediction equation of the next moment, and repeating iterative calculation to obtain the prediction value of each variable at the next moment;
and S4, designing a weight coefficient by using a neural network.
wherein i is wd 、i wq The stator current effective components are respectively in a d-axis coordinate system and a q-axis coordinate system; l is the inductance of the surface-mounted permanent magnet synchronous motor; r is R s The stator resistance corresponds to the copper loss of the motor; r is R c The resistor is iron loss resistance and corresponds to the iron loss of the motor; omega e Is the electrical angular velocity, which is equal to the speed ω of motor rotation times the pole pair number N p ;u d 、u q Stator voltages under d-axis and q-axis coordinate systems respectively; psi phi type pm The magnetic flux of the permanent magnet is;
i d 、i q each consisting of two parts, i.e. producing the effective component i of the electromagnetic torque wd 、i wq With iron loss component i cd 、i cq (2)
Wherein i is cd 、i cq The stator current iron loss components under the d-axis coordinate system and the q-axis coordinate system are respectively;
wherein T is e Is the actual value of the electromagnetic torque;
in the psi- d 、ψ q The flux linkages under the d-axis coordinate system and the q-axis coordinate system are respectively d-axis flux linkage and q-axis flux linkage.
Preferably, step S2 specifically includes the following steps;
establishing a stator current effective component prediction equation under a d-axis and q-axis coordinate system:(5)
wherein k is the number of cycles of the current control cycle; i.e wd (k+1)、i wq (k+1) is i respectively wd 、i wq Predicted value at the k+1th control period; i.e wd (k)、i wq (k) I respectively wd 、i wq Predicted values at the kth control period; t (T) s Sampling time;
the formula (6) and the formula (7) are combined to obtain i d 、i q Is the predictive equation of:8)
wherein i is d (k+1)、i q (k+1) is i respectively d 、i q Predicted value at the k+1th control period; u (u) d (k+1)、u q (k+1) is the stator voltage in the d-axis and q-axis coordinate systems of the (k+1) -th control period, respectively.
and (3) adding 1 on the basis of the prediction equation at the moment k+1 obtained in the step S2, so as to obtain the prediction equation at the moment k+2, and repeating iterative calculation to obtain the prediction value of each variable at the moment k+2.
Preferably, the step S4 specifically includes the following design steps:
wherein, a first term in the cost function controls the electromagnetic torque tracking reference value; the second item is controlled by MTPA principleStator current effective component i wd Is of a size of (2); the third term controls the total controllable loss of the motor;、/>for the weight coefficient, the weight coefficient group +.>The cost of different control targets is adjusted, so that each performance index is weighed.
S42, data acquisition and collection
First, the weight coefficient is set、/>Adjusted interval->、/>And discrete step +.>Get about->Is->Individual weight coefficients and about->Is->A plurality of weight coefficients;
and then two are combined to formSets of weight coefficients: />The method comprises the steps of carrying out a first treatment on the surface of the Then establishing a mathematical model of the permanent magnet synchronous motor in simulation software, and specifying the working condition as rated load +.>The motor is accelerated from zero speed to rated speed +.>And hold for a period of time;
finally, only the weight coefficient is changed by utilizing the automatic simulation function, and the operation is repeatedGroup simulation, recording each performance index value obtained by each group simulation;
s43, training an offline neural network under a specific working condition;
s44, training an online neural network for establishing a relation between the working condition and the optimal weight coefficient.
Preferably, the step S43 specifically includes the following steps:
training offline neural networks
Establishing a relation between a weight coefficient and a performance index, wherein an input layer of the offline neural network is a weight coefficient group, an output layer is transient time, torque error, stator current error and system efficiency, training the offline neural network by using a data set under a specific working condition obtained in data acquisition and acquisition, and judging that the offline neural network training is completed when the difference between the performance index value obtained by operating the offline neural network and the performance index value obtained by operating a mathematical model of a permanent magnet synchronous motor is always smaller than a set value;
after training is completed, the time consumption of single operation of the offline neural network is far less than that of a mathematical model of the permanent magnet synchronous motor;
according to the requirements of performance indexes in transient state and steady state, a transient performance requirement function is predefinedSteady state performance demand function->Wherein the transient performance demand function takes torque error, stator current error, system efficiency and transient time into consideration to obtainThe expression is as follows: />(13)
The steady-state performance demand function considers the torque error, the stator current error and the system efficiency to obtainThe expression is as follows:(14)
wherein A, B, C, D, E, F is the weight of each performance index set;
wherein the torque errorWherein->Is the number of sampling points, +.>Is the actual value of the electromagnetic torque of the nth sampling point,/->An electromagnetic torque reference value; stator current error->Wherein->Stator current actual value of nth sampling point, < ->Is the value of the sampling point corresponding to the standard sinusoidal current waveform; system efficiency->Wherein->Is the output power of the motor, ">Is the total controllable loss; transient time->The time taken for the motor to accelerate from zero speed to the rotational speed reference value is defined;
taking one far smaller thanDiscrete step size +.>Repeating the data acquisition and acquisition steps to form a more precise weight coefficient group ++>Traversing ∈10 with trained offline neural network>And respectively finding out the weight coefficient groups which minimize the two performance requirement functions, namely a transient optimal weight coefficient group and a steady optimal weight coefficient group under a specific working condition.
Preferably, the step S44 specifically includes the following steps:
because the user focuses on different performances in transient and steady state periods, respectively training two online neural networks, setting the optimal weight coefficient groups in the transient and steady state periods, respectively marking the optimal weight coefficient groups as a first online neural network and a second online neural network, wherein the input layers of the first online neural network and the second online neural network are rotation speed reference valuesLoad torque->The output layer is the optimal weight coefficient group +.>The input layer dimension is 2, and the output layer dimension is 2;
then obtaining the optimal weight coefficient groups under different working conditions, and setting discrete step sizes for the rotating speed and the load torque respectively、/>Obtain->Sum of rotational speed values->The flux linkage values are combined in pairs to obtain +.>The working conditions are as follows: />The method comprises the steps of carrying out a first treatment on the surface of the Will->And (3) replacing a specific working condition by a working condition, repeating the steps S42 and S43, obtaining the correspondence between one working condition and one transient optimal weight coefficient group and between one steady optimal weight coefficient group once, and finally obtaining +.>Transient optimal weight coefficient sets and steady-state optimal weight coefficient sets corresponding to the working conditions;
by usingIndividual operating conditions and transient optimal weight coefficient set, steady state optimal weight coefficientAnd training the first online neural network and the second online neural network by taking the array as a data set, and when the difference value between the optimal weight coefficient set obtained by operating the online neural network and the optimal weight coefficient set obtained by operating the offline neural network is always smaller than a set value, considering that the online neural network training is completed, and the trained first online neural network and second online neural network are used for selecting the optimal weight coefficient set according to the real-time working condition of the permanent magnet synchronous motor. />
The beneficial effects of the invention are as follows:
1. a clear relation is established among the control target, the weight coefficient, the working condition and the performance index, the weight coefficient can be automatically set, and the optimality of the weight coefficient is ensured under different working conditions.
2. The performance indexes (torque error and stator current error) of the current main stream method are reserved, meanwhile, the system efficiency and transient time are considered, the range of the considered performance indexes is wider, and the adaptability of different occasions is stronger.
3. The relation between the working condition and the optimal weight coefficient is established, and the better weight coefficient can be obtained in a shorter time in the occasion of frequent working condition change, so that the control object shows better performance.
The technical scheme of the invention is further described in detail through the drawings and the embodiments.
Drawings
FIG. 1 is a flow chart of a neural network-based model predictive control method of the present invention;
FIG. 2 is a control block diagram of a neural network-based model predictive control method of the present invention;
fig. 3 is a mathematical model diagram of a neural network-based model predictive control method of the present invention.
Description of the embodiments
The present invention will be further described with reference to the accompanying drawings, and it should be noted that, while the present embodiment provides a detailed implementation and a specific operation process on the premise of the present technical solution, the protection scope of the present invention is not limited to the present embodiment.
The model prediction control method based on the neural network comprises the following steps:
s1, establishing a state equation of a permanent magnet synchronous motor;
wherein i is wd 、i wq The stator current effective components are respectively in a d-axis coordinate system and a q-axis coordinate system; l is the inductance of the surface-mounted permanent magnet synchronous motor; r is R s The stator resistance corresponds to the copper loss of the motor; r is R c The resistor is iron loss resistance and corresponds to the iron loss of the motor; omega e Is the electrical angular velocity, which is equal to the speed ω of motor rotation times the pole pair number N p ;u d 、u q Stator voltages under d-axis and q-axis coordinate systems respectively; psi phi type pm The magnetic flux of the permanent magnet is;
i d 、i q each consisting of two parts, i.e. producing the effective component i of the electromagnetic torque wd 、i wq With iron loss component i cd 、i cq (2)
Wherein i is cd 、i cq The stator current iron loss components under the d-axis coordinate system and the q-axis coordinate system are respectively;
wherein T is e Is the actual value of the electromagnetic torque;
in the psi- d 、ψ q The flux linkages under the d-axis coordinate system and the q-axis coordinate system are respectively d-axis flux linkage and q-axis flux linkage.
S2, discretizing the state equation to obtain a prediction equation;
preferably, step S2 specifically includes the following steps;
establishing a stator current effective component prediction equation under a d-axis and q-axis coordinate system:(5)
wherein k is the number of cycles of the current control cycle; i.e wd (k+1)、i wq (k+1) is i respectively wd 、i wq Predicted value at the k+1th control period; i.e wd (k)、i wq (k) I respectively wd 、i wq Predicted values at the kth control period; t (T) s Sampling time;
the formula (6) and the formula (7) are combined to obtain i d 、i q Is the predictive equation of:(8)
wherein i is d (k+1)、i q (k+1) is i respectively d 、i q Predicted value at the k+1th control period; u (u) d (k+1)、u q (k+1) is the stator voltage in the d-axis and q-axis coordinate systems of the (k+1) -th control period, respectively.
S3, calculating electromagnetic torque T e Effective component i of d-axis stator current wd Total controllable loss P Loss Then calculating a prediction equation of the next moment, and repeating iterative calculation to obtain the prediction value of each variable at the next moment;
iron loss predictive value expression:
(11);
calculating a prediction equation at the next moment, and repeating iterative calculation to obtain a predicted value of each variable at the next moment;
adding 1 to the prediction equation at the moment k+1 obtained in the step S2 to obtain the prediction equation at the moment k+2, repeating iterative calculation to obtain the prediction value of each variable at the moment k+2, namely calculating T on the basis of the formula (3) e (k+2) (electromagnetic torque T e Is a predicted value of (2); calculating the d-axis stator current effective component i on the basis of formulas (5) and (7) wd Is a predicted value of (a).
S4, designing weight coefficients by using a neural network
The step S4 specifically includes the following design steps:
(12)
wherein, a first term in the cost function controls the electromagnetic torque tracking reference value; the second term controls the stator current effective component i based on MTPA (maximum torque to current ratio) principle wd Is of a size of (2); the third term controls the total controllable loss of the motor;、/>for the weight coefficient, the weight coefficient group +.>The cost of different control targets is adjusted, so that each performance index is weighed;
s42, data acquisition and collection
First, the weight coefficient is set、/>Adjusted interval->、/>And discrete step +.>(the above values are empirically set), the result is about +.>Is->Individual weight coefficients and about->A kind of electronic deviceA plurality of weight coefficients;
Then establishing a mathematical model of the permanent magnet synchronous motor in simulation software, and specifying the working condition as rated loadThe motor is accelerated from zero speed to foreheadConstant rotation speed->And hold for a period of time;
finally, only the weight coefficient is changed by utilizing the automatic simulation function, and the operation is repeatedGroup simulation, recording each performance index value obtained by each group simulation;
s43, training an offline neural network under a specific working condition;
preferably, the step S43 specifically includes the following steps:
training offline neural networks
And establishing a relation between the weight coefficient and the performance index, wherein an input layer of the offline neural network is a weight coefficient group, an output layer is transient time, torque error, stator current error and system efficiency, the dimension of the input layer of the offline neural network is 2, and the dimension of the output layer is 4. Training the offline neural network by using a data set under a specific working condition obtained in data acquisition and acquisition by an operation neural network training algorithm (the data set in the embodiment comprises a weight coefficient set and a performance index value), and judging that the offline neural network training is completed when the difference between the performance index value obtained by operating the offline neural network and the performance index value obtained by operating a mathematical model of the permanent magnet synchronous motor is always smaller than a set value;
after training, the time consumption of single operation of the offline neural network is far less than that of a mathematical model of the permanent magnet synchronous motor, so that an optimal weight coefficient set under a specific working condition can be found more accurately and more quickly;
according to the requirements of performance indexes in transient state and steady state, a transient performance requirement function is predefinedSteady state performance demand function->Wherein the transient performance demand function accounts for torque error, stator current error, system efficiency, and transient timeObtain->The expression is as follows: />The steady state performance demand function takes into account torque error, stator current error, system efficiency, and gets +.>The expression is as follows: />Wherein A, B, C, D, E, F is the weight of each performance index set; />
Wherein the torque errorIn (1) the->Is the number of sampling points +.>Is the actual value of the electromagnetic torque of the nth sampling point,/->An electromagnetic torque reference value; stator current error->Wherein->Stator current actual value of nth sampling point, < ->Is the value of the sampling point corresponding to the standard sinusoidal current waveform; system efficiency->Wherein->Is the output power of the motor, ">Is the total controllable loss; transient time->The time taken for the motor to accelerate from zero speed to the rotational speed reference value is defined;
taking one far smaller thanDiscrete step size +.>Repeating the data acquisition and acquisition steps to form a more precise weight coefficient group ++>Traversing ∈10 with trained offline neural network>And respectively finding out the weight coefficient groups which minimize the two performance requirement functions, namely a transient optimal weight coefficient group and a steady optimal weight coefficient group under a specific working condition.
S44, training an online neural network for establishing a relation between the working condition and the optimal weight coefficient.
Preferably, the step S44 specifically includes the following steps:
because the user focuses on different performances in transient and steady state periods, respectively training two online neural networks, setting the optimal weight coefficient groups in the transient and steady state periods, respectively marking the optimal weight coefficient groups as a first online neural network and a second online neural network, wherein the input layers of the first online neural network and the second online neural network are rotation speed reference valuesLoad torque->The output layer is the optimal weightCoefficient set->The input layer dimension is 2, and the output layer dimension is 2;
then obtaining the optimal weight coefficient groups under different working conditions, and setting discrete step sizes for the rotating speed and the load torque respectively、/>Obtain->Sum of rotational speed values->The flux linkage values are combined in pairs to obtain +.>The working conditions are as follows: />;
Running a neural network training algorithm willAnd (3) replacing a specific working condition by a working condition, repeating the steps S42 and S43, obtaining the correspondence between one working condition and one transient optimal weight coefficient group and one steady optimal weight coefficient group each time of operation, and finally obtainingTransient optimal weight coefficient sets and steady-state optimal weight coefficient sets corresponding to the working conditions;
by usingThe optimal weight coefficient set and the steady-state optimal weight coefficient set of each working condition and the transient state thereof are used as data sets, the first online neural network and the second online neural network are trained, and when the online neural network is operated, the optimal weight coefficient set and the optimal weight coefficient set are obtainedWhen the difference value of the optimal weight coefficient set obtained by the off-line neural network is always smaller than a set value, the on-line neural network training is considered to be completed, and the trained first on-line neural network and second on-line neural network are used for selecting the optimal weight coefficient set according to the real-time working condition of the permanent magnet synchronous motor.
The system of the model predictive control method based on the neural network comprises an offline neural network running under a specific working condition, a first online neural network for a transient optimal weight coefficient set and a second online neural network for obtaining a steady optimal weight coefficient set.
The model predictive control method based on the neural network is applied to a real-time physical controller, and a trained online neural network is mounted in the real-time physical controller;
when the method is used, firstly, a mark variable S is set to judge the running state of a control object, when the control object is in a transient period, S=1, a first online neural network is operated, and according to the instant working condition, the first online neural network outputs a transient optimal weight coefficient set, and the transient optimal weight coefficient set is used as an optimal weight coefficient set to be input into a cost function minimization module; when the control object is in a steady state period, S=0, the second online neural network is operated, and according to the instant working condition, the second online neural network outputs a steady state optimal weight coefficient set, and the steady state optimal weight coefficient set is used as the optimal weight coefficient set to be input into the cost function minimizing module.
In summary, the invention adopts the model predictive control method based on the neural network, and mainly adopts the following design:
(1) Selecting a control target: selecting electromagnetic torqueD-axis stator current effective component +.>Total controllable lossAs a control target, a model predictive torque control strategy is constructed, the model predictive control being controlled by a cost function +.>The control target is made to track its reference value.
(2) And (3) designing a weight coefficient: model predictive control determines the behavior of a control object by minimizing a cost function. In the cost function, a plurality of weight coefficients need to be set for different control targets. As described in (1), the control targets of the present invention are 3, and 2 weight coefficients are required to be setAnd->. (3) adaptation of different working conditions: for a permanent magnet synchronous motor, the operating conditions include a rotation speed +.>And load size->Transient and steady state. The designed control method needs to be adjusted according to different working conditions, so that the control object shows the optimal performance under the current working condition.
(4) Testing performance indexes: the effect of the control strategy is measured by the performance index of the control object.
Finally, it should be noted that: the above embodiments are only for illustrating the technical solution of the present invention and not for limiting it, and although the present invention has been described in detail with reference to the preferred embodiments, it will be understood by those skilled in the art that: the technical scheme of the invention can be modified or replaced by the same, and the modified technical scheme cannot deviate from the spirit and scope of the technical scheme of the invention.
Claims (6)
1. A model predictive control method based on a neural network is characterized by comprising the following steps of: the method comprises the following steps:
s1, establishing a state equation of a permanent magnet synchronous motor;
s2, discretizing the state equation to obtain a prediction equation;
s3, calculating electromagnetic torque T e Effective component i of d-axis stator current wd Total controllable loss P Loss Then calculating a prediction equation of the next moment, and repeating iterative calculation to obtain the prediction value of each variable at the next moment;
s4, designing a weight coefficient by using a neural network;
the equation of state expression in step S1 is as follows:(1) Wherein i is wd 、i wq The stator current effective components are respectively in a d-axis coordinate system and a q-axis coordinate system; l is the inductance of the surface-mounted permanent magnet synchronous motor; r is R s The stator resistance corresponds to the copper loss of the motor; r is R c The resistor is iron loss resistance and corresponds to the iron loss of the motor; omega e Is the electrical angular velocity, which is equal to the speed ω of motor rotation times the pole pair number N p ;u d 、u q Stator voltages under d-axis and q-axis coordinate systems respectively; psi phi type pm The magnetic flux of the permanent magnet is;
i d 、i q each consisting of two parts, i.e. producing the effective component i of the electromagnetic torque wd 、i wq With iron loss component i cd 、i cq (2)
Wherein i is cd 、i cq The stator current iron loss components under the d-axis coordinate system and the q-axis coordinate system are respectively;
wherein T is e Is the actual value of the electromagnetic torque;
in the psi- d 、ψ q The flux linkages under the d-axis coordinate system and the q-axis coordinate system are respectively d-axis flux linkage and q-axis flux linkage.
2. The neural network-based model predictive control method according to claim 1, characterized by: step S2 specifically comprises the following steps;
establishing a stator current effective component prediction equation under a d-axis and q-axis coordinate system:(5)
wherein k is the number of cycles of the current control cycle; i.e wd (k+1)、i wq (k+1) is i respectively wd 、i wq Predicted value at the k+1th control period; i.e wd (k)、i wq (k) I respectively wd 、i wq Predicted values at the kth control period; t (T) s Sampling time;
the formula (6) and the formula (7) are combined to obtain i d 、i q Is the predictive equation of:(8)
wherein i is d (k+1)、i q (k+1) is i respectively d 、i q Predicted value at the k+1th control period; u (u) d (k+1)、u q (k+1) is the stator voltage in the d-axis and q-axis coordinate systems of the (k+1) -th control period, respectively.
3. A god-based according to claim 2The model predictive control method through the network is characterized in that: in step S3, the predicted copper loss value expression is expressed by:(9);
and (3) adding 1 on the basis of the prediction equation at the moment k+1 obtained in the step S2, so as to obtain the prediction equation at the moment k+2, and repeating iterative calculation to obtain the prediction value of each variable at the moment k+2.
4. The neural network-based model predictive control method as set forth in claim 3, wherein: the step S4 specifically includes the following design steps:
wherein, a first term in the cost function controls the electromagnetic torque tracking reference value; the second item controls the effective component i of the stator current according to the MTPA principle wd Is of a size of (2); the third term controls the total controllable loss of the motor;、/>the weight coefficients are recorded as weight coefficient groupsThe cost of different control targets is adjusted, so that each performance index is weighed;
s42, data acquisition and collection
First, the weight coefficient is set、/>Adjusted interval->、/>And discrete step +.>Get about->Is->Individual weight coefficients and about->Is->A plurality of weight coefficients; two-by-two combination to form->Sets of weight coefficients: />Then establishing a mathematical model of the permanent magnet synchronous motor in simulation software, and specifying the working condition as rated load +.>The motor is accelerated from zero speed to rated speed +.>And hold for a period of time;
finally, only the weight coefficient is changed by utilizing the automatic simulation function, and the operation is repeatedGroup simulation, recording each performance index value obtained by each group simulation;
s43, training an offline neural network under a specific working condition;
s44, training an online neural network for establishing a relation between the working condition and the optimal weight coefficient.
5. The neural network-based model predictive control method as set forth in claim 4, wherein: the step S43 specifically includes the following steps:
training offline neural networks
Establishing a relation between a weight coefficient and a performance index, wherein an input layer of the offline neural network is a weight coefficient group, an output layer is transient time, torque error, stator current error and system efficiency, training the offline neural network by using a data set under a specific working condition obtained in data acquisition and acquisition, and judging that the offline neural network training is completed when the difference between the performance index value obtained by operating the offline neural network and the performance index value obtained by operating a mathematical model of a permanent magnet synchronous motor is always smaller than a set value;
after training is completed, the time consumption of single operation of the offline neural network is far less than that of a mathematical model of the permanent magnet synchronous motor;
according to the requirements of performance indexes in transient state and steady state, a transient performance requirement function is predefinedSteady state performance demand functionWherein the transient performance demand function accounts for torque error, stator current error, system efficiency, and transient timeObtain->The expression is as follows: />(13)
The steady-state performance demand function considers the torque error, the stator current error and the system efficiency to obtainThe expression is as follows:(14)
wherein A, B, C, D, E, F is the weight of each performance index set;
wherein the torque errorIn (1) the->Is the number of sampling points, +.>Is the actual value of the electromagnetic torque of the nth sampling point,/->An electromagnetic torque reference value; stator current error->In which, in the process,stator current actual value of nth sampling point, < ->Is the value of the sampling point corresponding to the standard sinusoidal current waveform; system efficiencyRate ofWherein->Is the output power of the motor, ">Is the total controllable loss; transient time->The time taken for the motor to accelerate from zero speed to the rotational speed reference value is defined;
taking one far smaller thanDiscrete step size +.>Repeating the data acquisition and acquisition steps to form a more accurate weight coefficient setTraversing ∈10 with trained offline neural network>And respectively finding out the weight coefficient groups which minimize the two performance requirement functions, namely a transient optimal weight coefficient group and a steady optimal weight coefficient group under a specific working condition.
6. The neural network-based model predictive control method of claim 5, wherein: the step S44 specifically includes the following steps:
because the user focuses on different performances in transient and steady state periods, respectively training two online neural networks for setting the optimal weight coefficient groups in the transient and steady state periods, respectively marking as a first online neural network and a second online neural network, and respectively marking the first online neural network and the first online neural network as the first online neural network and the second online neural networkThe input layers of the two online neural networks are all rotation speed reference valuesLoad torque->The output layer is the optimal weight coefficient group +.>The input layer dimension is 2, and the output layer dimension is 2;
then obtaining the optimal weight coefficient groups under different working conditions, and setting discrete step sizes for the rotating speed and the load torque respectively、/>Obtain->Sum of rotational speed values->The flux linkage values are combined in pairs to obtain +.>The working conditions are as follows: />The method comprises the steps of carrying out a first treatment on the surface of the Will->And (3) replacing a specific working condition by a working condition, repeating the steps S42 and S43, obtaining the correspondence between one working condition and one transient optimal weight coefficient group and between one steady optimal weight coefficient group once, and finally obtaining +.>Transient state corresponding to each working conditionAn optimal weight coefficient set, a steady-state optimal weight coefficient set;
by usingAnd training the first online neural network and the second online neural network by taking the individual working condition, the transient optimal weight coefficient group and the steady optimal weight coefficient group thereof as data sets, and when the difference value between the optimal weight coefficient group obtained by operating the online neural network and the optimal weight coefficient group obtained by operating the offline neural network is always smaller than a set value, considering that the online neural network is trained, and the trained first online neural network and second online neural network are used for selecting the optimal weight coefficient group according to the real-time working condition of the permanent magnet synchronous motor. />
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202310310441.0A CN116027672B (en) | 2023-03-28 | 2023-03-28 | Model prediction control method based on neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202310310441.0A CN116027672B (en) | 2023-03-28 | 2023-03-28 | Model prediction control method based on neural network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN116027672A CN116027672A (en) | 2023-04-28 |
CN116027672B true CN116027672B (en) | 2023-06-09 |
Family
ID=86079792
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202310310441.0A Active CN116027672B (en) | 2023-03-28 | 2023-03-28 | Model prediction control method based on neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN116027672B (en) |
Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110266228A (en) * | 2019-07-05 | 2019-09-20 | 长安大学 | Surface permanent magnetic Synchronous Machine Models forecast Control Algorithm based on BP neural network |
Family Cites Families (19)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JPH0746917B2 (en) * | 1987-07-28 | 1995-05-17 | 三菱電機株式会社 | Control device for three-phase converter |
JP4196545B2 (en) * | 2001-04-09 | 2008-12-17 | トヨタ自動車株式会社 | Power output apparatus and electric vehicle equipped with the same |
JP2008228419A (en) * | 2007-03-12 | 2008-09-25 | Mie Univ | Torque control method of motor based on model prediction control |
JP5387614B2 (en) * | 2011-05-16 | 2014-01-15 | 株式会社デンソー | Rotating machine control device |
CN106357191B (en) * | 2016-09-09 | 2018-09-11 | 湖南工业大学 | motor load self-adaptation control method |
CN108038341A (en) * | 2018-01-08 | 2018-05-15 | 山东大学 | A kind of model selection method suitable for interactive multi-model load modeling |
CN109849896B (en) * | 2019-02-28 | 2020-11-03 | 江苏大学 | Hybrid electric vehicle self-adaptive E-H switching coordination control method based on parameter observation |
CN110287638B (en) * | 2019-07-04 | 2022-09-13 | 南京邮电大学 | Method for predicting life of flying saw based on kalman-RNN neural network |
CN110518847B (en) * | 2019-08-30 | 2021-03-30 | 长安大学 | Surface permanent magnet synchronous motor model prediction control method based on BP neural network |
CN110659722B (en) * | 2019-08-30 | 2023-07-18 | 江苏大学 | Electric vehicle lithium ion battery health state estimation method based on AdaBoost-CBP neural network |
CN110775154B (en) * | 2019-10-25 | 2021-02-05 | 南京航空航天大学 | Electric truck steer-by-wire system and network uncertain control method thereof |
CN112740537B (en) * | 2020-12-24 | 2023-02-17 | 深圳市英威腾电气股份有限公司 | MTPA control method, device, system and equipment of permanent magnet synchronous motor |
WO2022204944A1 (en) * | 2021-03-30 | 2022-10-06 | 华为数字能源技术有限公司 | Electric drive system, drivetrain, heating method, and electric vehicle |
CN113659907B (en) * | 2021-09-02 | 2023-09-19 | 大连海事大学 | Direct torque control method of six-phase permanent magnet fault-tolerant motor based on torque model prediction |
CN115840374A (en) * | 2021-09-18 | 2023-03-24 | 中车株洲电力机车研究所有限公司 | Rail transit traction transmission system simulation method and device |
CN114285342B (en) * | 2021-12-22 | 2023-11-24 | 电子科技大学 | Permanent magnet synchronous motor model prediction direct speed synchronous control method |
CN115276419B (en) * | 2022-08-22 | 2023-05-16 | 山东大学 | Sampling noise-resistant dual-active-bridge converter model prediction control method and system |
CN115378324A (en) * | 2022-09-29 | 2022-11-22 | 山东大学 | Permanent magnet synchronous motor control method, system, medium and electronic equipment |
CN115685747A (en) * | 2022-09-30 | 2023-02-03 | 西南交通大学 | Model prediction control method based on residual error neural network optimization |
-
2023
- 2023-03-28 CN CN202310310441.0A patent/CN116027672B/en active Active
Patent Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110266228A (en) * | 2019-07-05 | 2019-09-20 | 长安大学 | Surface permanent magnetic Synchronous Machine Models forecast Control Algorithm based on BP neural network |
Also Published As
Publication number | Publication date |
---|---|
CN116027672A (en) | 2023-04-28 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Li et al. | An adaptive sliding-mode observer for induction motor sensorless speed control | |
CN106788054B (en) | A kind of Speed Sensorless Control Method based on rotation high-frequency signal injection | |
CN102662323B (en) | Adoptive sliding mode control method and adoptive sliding mode control system of wind power generation variable-pitch actuator | |
CN109921707A (en) | A kind of switching magnetic-resistance hub motor is without position prediction control method | |
CN110707981A (en) | Permanent magnet synchronous motor speed controller based on novel extended state observer | |
CN108365785B (en) | Asynchronous motor repeated prediction control method | |
CN102611380B (en) | Online identification method for parameters of double-fed motor | |
CN109787531A (en) | A kind of switching magnetic-resistance hub motor forecast Control Algorithm | |
WO2022134661A1 (en) | Method for selecting magnetization state of adjustable-flux permanent magnet synchronous motor in case of optimal control of full-speed domain efficiency and online control method | |
CN109347382B (en) | Rotor position estimation method of permanent magnet direct-drive wind driven generator | |
CN104393798A (en) | Electric bicycle control method based on integral sliding mode and disturbance observer | |
CN110401393A (en) | Tolerant hierarchical sequence model predictive control method, equipment and storage medium | |
Yan et al. | Torque estimation and control of PMSM based on deep learning | |
Bohari et al. | Speed tracking of indirect field oriented control induction motor using neural network | |
CN111711396A (en) | Method for setting control parameters of speed ring of permanent magnet synchronous motor based on fractional order sliding mode controller | |
CN116027672B (en) | Model prediction control method based on neural network | |
CN106953566A (en) | A kind of method and apparatus of the frequency matching based on fuzzy controller | |
CN104408223B (en) | A kind of tracking optimal control method of Wind turbines | |
CN111835251B (en) | Permanent magnet synchronous motor high-performance control method based on speed-free sensing | |
Feytout et al. | Robust CRONE design for a variable ratio planetary gearing in a variable speed wind turbine | |
CN107689758A (en) | Durface mounted permanent magnet synchronous electric machine torque control method | |
CN110011583B (en) | Permanent magnet synchronous motor sliding mode control system based on singular perturbation theory and modeling method | |
CN114123904B (en) | Predictive current increment control method suitable for operation of permanent magnet synchronous motor in high-speed region | |
CN114006558B (en) | Motor parameter online identification method only for bill parameters | |
Barambones et al. | A robust field oriented control of induction motor with flux observer and speed adaptation |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |