CN116954072A - Prediction control weight factor dynamic optimization method based on depth residual error network - Google Patents
Prediction control weight factor dynamic optimization method based on depth residual error network Download PDFInfo
- Publication number
- CN116954072A CN116954072A CN202310830632.XA CN202310830632A CN116954072A CN 116954072 A CN116954072 A CN 116954072A CN 202310830632 A CN202310830632 A CN 202310830632A CN 116954072 A CN116954072 A CN 116954072A
- Authority
- CN
- China
- Prior art keywords
- network
- representing
- current
- training
- weight factors
- 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.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 43
- 238000005457 optimization Methods 0.000 title claims abstract description 38
- 238000012549 training Methods 0.000 claims abstract description 42
- 238000012545 processing Methods 0.000 claims abstract description 8
- 238000013178 mathematical model Methods 0.000 claims abstract description 7
- 238000010606 normalization Methods 0.000 claims abstract description 6
- 238000012795 verification Methods 0.000 claims abstract description 4
- 230000006870 function Effects 0.000 claims description 75
- 238000005070 sampling Methods 0.000 claims description 35
- 238000004422 calculation algorithm Methods 0.000 claims description 21
- 230000007935 neutral effect Effects 0.000 claims description 17
- 239000003990 capacitor Substances 0.000 claims description 13
- 238000013528 artificial neural network Methods 0.000 claims description 12
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 9
- 238000004088 simulation Methods 0.000 claims description 9
- 238000013461 design Methods 0.000 claims description 8
- 239000013598 vector Substances 0.000 claims description 8
- 238000011156 evaluation Methods 0.000 claims description 7
- 238000013135 deep learning Methods 0.000 claims description 6
- 238000002474 experimental method Methods 0.000 claims description 6
- 230000004913 activation Effects 0.000 claims description 4
- 238000002790 cross-validation Methods 0.000 claims description 4
- 230000008569 process Effects 0.000 claims description 4
- 230000004907 flux Effects 0.000 claims description 3
- 238000010801 machine learning Methods 0.000 claims description 3
- 238000013139 quantization Methods 0.000 claims description 3
- 238000004804 winding Methods 0.000 claims description 3
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 description 6
- 238000009825 accumulation Methods 0.000 description 5
- 238000010586 diagram Methods 0.000 description 5
- 230000000694 effects Effects 0.000 description 4
- 238000004364 calculation method Methods 0.000 description 3
- 230000000295 complement effect Effects 0.000 description 3
- 230000008034 disappearance Effects 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 210000002569 neuron Anatomy 0.000 description 3
- 238000013507 mapping Methods 0.000 description 2
- 230000011218 segmentation Effects 0.000 description 2
- 238000013459 approach Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008033 biological extinction Effects 0.000 description 1
- 230000015556 catabolic process Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000012937 correction Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000006731 degradation reaction Methods 0.000 description 1
- 238000003058 natural language processing Methods 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- G—PHYSICS
- G05—CONTROLLING; REGULATING
- G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
- G05B13/00—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion
- G05B13/02—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric
- G05B13/04—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric involving the use of models or simulators
- G05B13/042—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric involving the use of models or simulators in which a parameter or coefficient is automatically adjusted to optimise the performance
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- H—ELECTRICITY
- H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
- H02M—APPARATUS FOR CONVERSION BETWEEN AC AND AC, BETWEEN AC AND DC, OR BETWEEN DC AND DC, AND FOR USE WITH MAINS OR SIMILAR POWER SUPPLY SYSTEMS; CONVERSION OF DC OR AC INPUT POWER INTO SURGE OUTPUT POWER; CONTROL OR REGULATION THEREOF
- H02M7/00—Conversion of ac power input into dc power output; Conversion of dc power input into ac power output
- H02M7/42—Conversion of dc power input into ac power output without possibility of reversal
- H02M7/44—Conversion of dc power input into ac power output without possibility of reversal by static converters
- H02M7/48—Conversion of dc power input into ac power output without possibility of reversal by static converters using discharge tubes with control electrode or semiconductor devices with control electrode
- H02M7/483—Converters with outputs that each can have more than two voltages levels
- H02M7/487—Neutral point clamped inverters
-
- H—ELECTRICITY
- H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
- H02M—APPARATUS FOR CONVERSION BETWEEN AC AND AC, BETWEEN AC AND DC, OR BETWEEN DC AND DC, AND FOR USE WITH MAINS OR SIMILAR POWER SUPPLY SYSTEMS; CONVERSION OF DC OR AC INPUT POWER INTO SURGE OUTPUT POWER; CONTROL OR REGULATION THEREOF
- H02M7/00—Conversion of ac power input into dc power output; Conversion of dc power input into ac power output
- H02M7/42—Conversion of dc power input into ac power output without possibility of reversal
- H02M7/44—Conversion of dc power input into ac power output without possibility of reversal by static converters
- H02M7/48—Conversion of dc power input into ac power output without possibility of reversal by static converters using discharge tubes with control electrode or semiconductor devices with control electrode
- H02M7/53—Conversion of dc power input into ac power output without possibility of reversal by static converters using discharge tubes with control electrode or semiconductor devices with control electrode using devices of a triode or transistor type requiring continuous application of a control signal
- H02M7/537—Conversion of dc power input into ac power output without possibility of reversal by static converters using discharge tubes with control electrode or semiconductor devices with control electrode using devices of a triode or transistor type requiring continuous application of a control signal using semiconductor devices only, e.g. single switched pulse inverters
- H02M7/5387—Conversion of dc power input into ac power output without possibility of reversal by static converters using discharge tubes with control electrode or semiconductor devices with control electrode using devices of a triode or transistor type requiring continuous application of a control signal using semiconductor devices only, e.g. single switched pulse inverters in a bridge configuration
- H02M7/53871—Conversion of dc power input into ac power output without possibility of reversal by static converters using discharge tubes with control electrode or semiconductor devices with control electrode using devices of a triode or transistor type requiring continuous application of a control signal using semiconductor devices only, e.g. single switched pulse inverters in a bridge configuration with automatic control of output voltage or current
- H02M7/53875—Conversion of dc power input into ac power output without possibility of reversal by static converters using discharge tubes with control electrode or semiconductor devices with control electrode using devices of a triode or transistor type requiring continuous application of a control signal using semiconductor devices only, e.g. single switched pulse inverters in a bridge configuration with automatic control of output voltage or current with analogue control of three-phase output
- H02M7/53876—Conversion of dc power input into ac power output without possibility of reversal by static converters using discharge tubes with control electrode or semiconductor devices with control electrode using devices of a triode or transistor type requiring continuous application of a control signal using semiconductor devices only, e.g. single switched pulse inverters in a bridge configuration with automatic control of output voltage or current with analogue control of three-phase output based on synthesising a desired voltage vector via the selection of appropriate fundamental voltage vectors, and corresponding dwelling times
-
- H—ELECTRICITY
- H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
- H02P—CONTROL OR REGULATION OF ELECTRIC MOTORS, ELECTRIC GENERATORS OR DYNAMO-ELECTRIC CONVERTERS; CONTROLLING TRANSFORMERS, REACTORS OR CHOKE COILS
- H02P21/00—Arrangements or methods for the control of electric machines by vector control, e.g. by control of field orientation
- H02P21/0003—Control strategies in general, e.g. linear type, e.g. P, PI, PID, using robust control
- H02P21/0014—Control strategies in general, e.g. linear type, e.g. P, PI, PID, using robust control using neural networks
-
- H—ELECTRICITY
- H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
- H02P—CONTROL OR REGULATION OF ELECTRIC MOTORS, ELECTRIC GENERATORS OR DYNAMO-ELECTRIC CONVERTERS; CONTROLLING TRANSFORMERS, REACTORS OR CHOKE COILS
- H02P21/00—Arrangements or methods for the control of electric machines by vector control, e.g. by control of field orientation
- H02P21/0003—Control strategies in general, e.g. linear type, e.g. P, PI, PID, using robust control
- H02P21/0017—Model reference adaptation, e.g. MRAS or MRAC, useful for control or parameter estimation
-
- H—ELECTRICITY
- H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
- H02P—CONTROL OR REGULATION OF ELECTRIC MOTORS, ELECTRIC GENERATORS OR DYNAMO-ELECTRIC CONVERTERS; CONTROLLING TRANSFORMERS, REACTORS OR CHOKE COILS
- H02P21/00—Arrangements or methods for the control of electric machines by vector control, e.g. by control of field orientation
- H02P21/14—Estimation or adaptation of machine parameters, e.g. flux, current or voltage
-
- H—ELECTRICITY
- H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
- H02P—CONTROL OR REGULATION OF ELECTRIC MOTORS, ELECTRIC GENERATORS OR DYNAMO-ELECTRIC CONVERTERS; CONTROLLING TRANSFORMERS, REACTORS OR CHOKE COILS
- H02P27/00—Arrangements or methods for the control of AC motors characterised by the kind of supply voltage
- H02P27/04—Arrangements or methods for the control of AC motors characterised by the kind of supply voltage using variable-frequency supply voltage, e.g. inverter or converter supply voltage
- H02P27/06—Arrangements or methods for the control of AC motors characterised by the kind of supply voltage using variable-frequency supply voltage, e.g. inverter or converter supply voltage using dc to ac converters or inverters
- H02P27/08—Arrangements or methods for the control of AC motors characterised by the kind of supply voltage using variable-frequency supply voltage, e.g. inverter or converter supply voltage using dc to ac converters or inverters with pulse width modulation
- H02P27/12—Arrangements or methods for the control of AC motors characterised by the kind of supply voltage using variable-frequency supply voltage, e.g. inverter or converter supply voltage using dc to ac converters or inverters with pulse width modulation pulsing by guiding the flux vector, current vector or voltage vector on a circle or a closed curve, e.g. for direct torque control
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Power Engineering (AREA)
- Theoretical Computer Science (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Software Systems (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Health & Medical Sciences (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Medical Informatics (AREA)
- Automation & Control Theory (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Feedback Control In General (AREA)
Abstract
The invention discloses a predictive control weight factor dynamic optimization method based on a depth residual error network, which specifically comprises the following steps: establishing a predictive mathematical model of a permanent magnet motor driving system fed by a three-level inverter, designing a cost function, and obtaining a data sample; defining an objective function, carrying out normalization processing on the data set, and dividing the data set into a training set and a cross verification set; constructing a depth residual error network, taking the weight factors as an input layer of the network, determining super parameters of the residual error network by minimizing root mean square error, and selecting a prediction index with the minimum objective function as an optimal weight factor combination; and establishing a database by utilizing the optimal weight factors obtained by prediction, and realizing dynamic optimization of the model predictive control weight factors by a table look-up method. The invention can efficiently and accurately predict the optimal weight factor under the complex operation condition of the motor, and can carry out on-line adjustment on the predicted weight factor, thereby realizing the dynamic optimization of the model predictive control performance.
Description
Technical Field
The invention belongs to the field of model predictive control, and particularly relates to a predictive control weight factor dynamic optimization method based on a depth residual error network.
Background
The multi-level inverter has the advantages of good electromagnetic compatibility, low current harmonic, high efficiency and the like, and is widely applied to medium-voltage high-power occasions. Compared with a diode neutral point clamped inverter, an active neutral point clamped inverter (ANPC) has the advantages of better wear leveling capability, flexible control, strong fault tolerance performance and the like, and is paid attention to. However, control of the ANPC inverter requires consideration of multiple control targets at the same time, including tracking current, neutral-point potential balancing, switching frequency adjustment, and the like.
In recent years, model Predictive Control (MPC) can effectively solve the problem of multi-objective optimization compared with conventional vector control. MPC typically uses a specific cost function to achieve multi-objective optimization, with a weight factor assigned to each control objective. The control performance of the MPC is closely related to the design of the cost function and the weight factor, and different working states affect the value of the weight factor, which means that the weight factor is not unique, and should be dynamically adjusted according to the operation state, and the control targets may conflict with each other, so that the design of the optimal weight factor becomes a big research hot spot in recent years.
Several methods for adjusting weight factors have been proposed by scholars at home and abroad, including a method without weight factors and a method based on observation functions. The former can be roughly divided into a segmentation cost function and a reduction cost function. The segmentation cost function is optimized by an independent cost function, which will increase the complexity of the MPC. Simplifying the cost function uses additional constraints depending on the specific optimization objective, however, the computational burden is a significant challenge. Another type of method monitors the performance of the MPC through an observer, adjusts the control state of the MPC on line, minimizes the cost function, and further realizes on-line adjustment and optimization of the weight factors. However, this approach to on-line optimization adds to the control complexity and computational burden on the MPC, which is more pronounced when applied to multi-level converter systems. Therefore, a method for realizing dynamic optimization of the weight factor without increasing the computational burden has been studied. With the wide application of Artificial Neural Networks (ANNs) in pattern recognition, natural language processing, biotechnology, traffic and the like, ANNs have good application advantages in optimizing predictive control weight factors.
Disclosure of Invention
In order to overcome the defects, the invention provides a predictive control weight factor dynamic optimization method based on a depth residual error network.
The invention discloses a predictive control weight factor dynamic optimization method based on a depth residual error network, which comprises the following steps:
step 1: establishing a three-level inverter feed permanent magnet motor driving system prediction mathematical model, predicting a cost function of a control algorithm according to a control target design model, and obtaining corresponding motor stator current harmonic total distortion THD under different weight factor combinations through MATLAB/Simulink simulation c Neutral point voltage deviation V of DC side bus ag And average switching frequency f sw Data samples are composed.
Step 2: based on the control target of the permanent magnet motor driving system, an objective function is defined to evaluate the control performance, then a data set is established and normalized, and meanwhile, the data set is divided into a training set and a cross verification set in proportion and is respectively used for network design, training and network performance evaluation.
Step 3: constructing a depth residual error network, taking the weight factors as an input layer of the network, and THD c 、V ag and fsw As an output layer of the network, training the network according to the data set and Adam algorithm, determining the hyper-parameters of the residual network by minimizing the root mean square error RMSE, and selecting the prediction index of the minimized objective function as the optimal weight factor combination of the model predictive control algorithm.
Step 4: and constructing a data set by using the optimal weight factors under different working conditions and the corresponding working conditions, and simultaneously realizing dynamic optimization of the model predictive control weight factors under different working conditions by using a table look-up method.
Further, the establishment of the predictive mathematical model and the cost function in the step 1 is specifically:
the method comprises the steps of taking a permanent magnet motor fed by a three-level active neutral point clamped inverter as a control object, performing discretization processing on a motor stator voltage equation under a dq rotating coordinate system by adopting a first-order forward Euler method, and establishing a discretization stator current prediction model of a permanent magnet motor driving system under the rotating coordinate system, wherein the discretization stator current prediction model is expressed as follows:
wherein k represents the sampling time, T s For sampling period, i d (k+1)、i q (k+1) represents the dq-axis stator current prediction value at the sampling time k+1, i d (k)、i q (k) Representing the dq-axis stator current sampling value at the current sampling time k, u d (k)、u q (k) Representing the dq-axis stator voltage value at the current sampling time k, L d 、L q Representing dq-axis stator inductance in rotating coordinate system, R s Representing the resistance value, ω, of the stator winding r Representing the electrical angular velocity of a permanent magnet motor, ψ f Representing the permanent magnet flux linkage.
According to the voltage values of the upper capacitor and the lower capacitor at the side of the direct current bus, a discretization prediction model of the voltage difference value of the capacitor at the side of the direct current bus is established, wherein the prediction model of the potential of the neutral point is expressed as follows:
wherein ,vc1 and vc2 Representing the upper and lower capacitance voltages, v n Represents neutral point potential, C dc Representing the capacitance of the DC side capacitor, u x Representing normalized output phase voltage, i x Representing phase current, u x (k) Representing the normalized phase voltage, i, at the current sampling instant k x (k) Represents the phase current, deltav, at the current sampling instant k n(k) and Δvn (k+1) represents the upper and lower capacitor voltage deviations at the current sampling time k and the sampling time k+1, respectively.
The number of on and off switching times of the inverter switching tube is expressed as:
Δu x (k+1)＝||u x (k+1)-u x (k)||
wherein ,ux (k+1) represents the normalized phase voltage at the next sampling time k+1.
To achieve comprehensive optimization of multiple objectives, a cost function is defined by considering stator current tracking, midpoint potential balancing, and switching adjustment as:
wherein ,Ji Cost function representing tracking current error, J dc Cost function, J, representing midpoint voltage deviation sw Cost function, i, representing switching frequency tracking error sref Representing the reference phase current, i s (k+1)＝[i d (k+1)，i q (k+1)] T The stator current predicted value in the rotation coordinate system at the next sampling time dq is indicated.
The method comprises the steps of establishing a total cost function according to the established cost functions of tracking current errors, midpoint voltage deviations and switching frequency tracking errors, wherein the total cost function is expressed as:
J＝J i +λ dc J dc +λ sw J sw
wherein J is the total cost function, lambda dc 、λ sw Weight factors respectively representing neutral point potential balance and switching frequency adjustment; and enumerating 27 space voltage vectors, calculating a cost function J, and selecting the voltage vector with the smallest J as the optimal output of the current control period.
Further, step 2 defines an objective function, establishes a data set, and divides the data set for network training and evaluation specifically as follows:
to facilitate evaluation of control performance, the depth residual network objective function is defined as:
wherein , and />Respectively predicting, controlling and normalizing the current harmonic total distortion, midpoint voltage deviation and average switching frequency of the model; alpha 1 、α 2 And alpha is 3 For weights predefined in the objective function, alpha 1 The larger the value of (2), the larger the weight of the optimized current harmonic total distortion in the objective function; alpha 2 The larger the value of (2), the larger the weight of the control midpoint voltage deviation in the objective function; alpha 3 The larger the value of (2), the larger the weight of the average switching frequency at the objective function.
For simplicity of work, the range of weights was determined to be 0-5 by simulation, and the step sizes were 0.25. Then corresponding current harmonic distortion THD is acquired c Voltage deviation V of midpoint ag And a switching frequency f sw Forming a data set and entering the data setAnd carrying out normalization processing to eliminate interference of different units and amplitudes, thereby improving the training speed of the residual error network.
The data sets were divided into training sets for training the designed network and cross-validation sets for evaluating network performance at 70% and 30% duty cycles.
Further, the depth residual error network constructed in the step 3 specifically comprises:
the depth residual network consists of a Full Connection (FC) layer and an activation function (ReLU), expressed as:
f(x)＝max(0,x)
the training process of the residual network is to update parameters by using gradient descent of a back propagation algorithm, and is expressed as follows:
wherein ω 'and b' are updated weights and deviations, respectively; alpha is learning rate, h (x) is output of the FC layer;
for ease of implementation and automatic adjustment of the learning rate, adam back propagation algorithm may be selected. The residual network realizes identity mapping by using residual links, and the method does not bring additional parameters to the residual network and does not increase calculation burden. The backward propagation gradient drop is corrected as:
wherein H (x) and H' (x) represent the outputs of the previous FC layer and the next FC layer, respectively; x is the output of the residual connection; d (D) 1 and D2 Is the gradient of a conventional neural network; as the number of hidden layers increases, the gradient will decrease, resulting in the disappearance of the gradient of the neural network. But by means of residual connection, when D 1 and D2 When it becomes zero, the problem of gradient extinction can be avoided.
The hyper-parameters of the residual network are adjusted with the minimized root mean square error RMSE as a quantization index, RMSE expressed as:
in order to avoid the network from falling into a locally optimal solution, the residual network is trained multiple times by using different initial values. The deep learning tool box Experiment Manager in the MATLAB2020b is selected as an automatic machine learning technology (autopl), so that the optimal super parameters of the residual network can be automatically found, and the problem that the super parameters of the residual network are complex and time-consuming to adjust is solved.
Based on the established data set, the weight factor lambda is calculated dc 、λ sw As a residual network input layer, the total distortion THD of current harmonic waves c Midpoint voltage deviation V ag And an average switching frequency f sw As a network output layer, adam algorithm is selected for back propagation training to minimize root mean square error RMSE to evaluate the trained network and adjust the super parameters.
Further, the implementation of the dynamic optimization of the weight factor is specifically: a lookup table is constructed, which consists of predicted optimal weight factors and defined working states, and the working states are simplified into rotor rotating speed N ref And a load torque T e Is a combination of (a); the method comprises the following specific steps:
1) And (3) data generation: the range of weight factors and the data set for training and evaluating the depth residual network are generated by simulation.
2) Training and prediction: and adjusting the super-parameters according to the training performance of the designed residual error network, wherein the optimal super-parameters are automatically found by an experiment manager in a deep learning tool box, and the lookup table consists of predicted weight factors.
3) Dynamic optimization: according to the current working state, the dynamic optimization of the MPC is realized by selecting the optimal weight factors from the lookup table.
The beneficial technical effects of the invention are as follows:
(1) According to the dynamic optimization method for the predictive control weight factors, disclosed by the invention, the network is trained by constructing the depth residual error network and carrying out forward propagation and Adam reverse propagation, compared with the traditional BP neural network, the problem of gradient disappearance of the depth neural network in the training process is avoided by residual error connection, and the accuracy and generalization capability of the predictive weight factors are improved.
(2) The invention belongs to a data-driven optimization algorithm, which utilizes a design and training network to analyze and predict data, does not change a model prediction control algorithm, has good robustness on data difference, and simplifies the complexity of a control system, so that the method is more suitable for application scenes of complex operation conditions, such as model prediction control of a multi-level inverter.
(3) The invention can realize the dynamic optimization of the model predictive control weight factor. The lookup table consists of weight factors predicted by the residual network and different working states, is easy to realize, requires less logic resources, and realizes dynamic optimization of the predicted control weight factors under different operation conditions by on-line adjustment of the predicted weight factors in the lookup table.
Drawings
FIG. 1 is a flow chart of dynamic optimization of model predictive control weighting factors in accordance with the present invention.
Fig. 2 is a block diagram of a permanent magnet motor drive system based on a three level active neutral point clamped inverter.
FIG. 3 is a block diagram of model predictive control in accordance with the present invention.
Fig. 4 is a diagram of a depth residual network structure according to the present invention.
Fig. 5 is a diagram of the depth residual network training effect of the present invention.
Fig. 6 is a graph of the depth residual network optimization result of the present invention.
Detailed Description
The invention will be described in further detail with reference to the accompanying drawings and the detailed description.
The invention aims at a permanent magnet motor driving system fed by a three-level Active Neutral Point Clamped (ANPC) inverter, provides a dynamic optimization method for predicting control weight factors based on a depth residual error network, utilizes the residual error network to predict the optimal weight factors of an MPC cost function, realizes optimal control under different working conditions, realizes the training of the residual error network by adopting an off-line mode, and avoids bringing extra calculation burden to MPC. According to typical operation conditions, a database is built with the predicted weight factors, and dynamic adjustment of the weight factors is realized through a table look-up method. The flow is shown in fig. 1, and specifically comprises the following steps:
step 1: establishing a three-level inverter feed permanent magnet motor driving system prediction mathematical model, predicting a cost function of a control algorithm according to a control target design model, and obtaining corresponding motor stator current harmonic total distortion THD under different weight factor combinations through MATLAB/Simulink simulation c Neutral point voltage deviation V of DC side bus ag And average switching frequency f sw Data samples are composed.
As shown in fig. 2, the permanent magnet motor driving system fed by the three-level ANPC inverter has six power switching tubes (T x1 —T x6 ) Composition, wherein T x1 and Tx2 Is a complementary switching tube T x3 and Tx4 Is a complementary switching tube T x5 and Tx6 As complementary switching transistors, table 1 shows the switching states and corresponding output voltages of the three-level ANPC inverter.
Table 1 switching states and output voltages of three-level ANPC inverter
In the table, "1" means on, "0" means off, u a 、u b 、u c E (-1, 0, 1), the normalized output voltage vector u is further expressed as:
u＝[u a u b u c ] T
next, a mathematical model of the permanent magnet motor stator voltage is built, expressed as:
wherein ,ud 、u q Representing stator voltage under dq coordinate system axis, R s Representing stator winding resistance, i d 、i q Representing stator current in dq axis, L d 、L q Representing stator inductance, ω, in the dq axis r Expressed as the mechanical angular velocity of the permanent magnet motor, ψ f Representing the permanent magnet flux linkage.
Performing a first-order forward Euler method on a voltage equation under a rotating coordinate system, and establishing a discretization stator current prediction model of the system under the rotating coordinate system, wherein the discretization stator current prediction model is expressed as follows:
wherein k represents a time counter, T s For sampling period, i d (k+1)、i q (k+1) represents the dq-axis stator current prediction value at the sampling time k+1, i d (k)、i q (k) Representing the dq-axis stator current sampling value at the current sampling time k, u d (k)、u q (k) Representing the dq-axis stator voltage value at the current sampling instant k.
According to the voltage values of the upper capacitor and the lower capacitor at the side of the direct current bus, a discretization prediction model of the voltage difference value of the capacitor at the side of the direct current bus is established, wherein the prediction model of the point potential is expressed as:
wherein ,Cdc Representing the capacitance of the DC side capacitor, u x (k) Normalized phase voltage, i, representing current sampling instant k x (k) Phase current, deltav, representing the current sampling instant k n(k) and Δvn (k+1) represents the upper and lower capacitance voltage deviations of the current sampling time k and the sampling time k+1, respectively;
in order to achieve comprehensive optimization of the predictive control algorithm, a cost function is defined by considering tracking current, midpoint potential balance and switching penalty as:
wherein ,Ji Cost function representing tracking current error, J dc Cost function, J, representing midpoint voltage deviation sw Cost function, i, representing switching frequency tracking error sref Representing the reference phase current, i s (k+1)＝[i d (k+1)，i q (k+1)] T Represents the stator current in the rotating coordinate system at the next sampling time dq, u x (k+1) represents the normalized phase voltage at the next sampling time k+1.
The method comprises the steps of establishing a total cost function according to the established cost functions of tracking current errors, midpoint voltage deviations and switching frequency tracking errors, wherein the total cost function is expressed as:
J＝J i +λ dc J dc +λ sw J sw
wherein J is the total cost function, lambda dc 、λ sw The weight factors respectively representing neutral point potential balance and switching frequency adjustment can be adjusted empirically through simulation and experimental tests. Fig. 3 is a block diagram of model predictive control, which is based on an established predictive model and cost function, and exhaustively enumerates 27 voltage vectors, calculates a cost function J, and selects a vector that minimizes the cost function J as an optimal output of the current control period. Lambda is defined by simulation dc 、λ sw The ranges of (2) are 0-5, and the step sizes are 0.25. Collecting current harmonic distortion THD under different weight factor combinations c Voltage deviation V of midpoint ag And a switching frequency f sw A dataset is composed.
Step 2: based on the control target of the permanent magnet motor driving system, an objective function is defined to evaluate the control performance, then a data set is established and normalized, and meanwhile, the data set is divided into a training set and a cross verification set in proportion and is respectively used for network design, training and network performance evaluation.
To facilitate evaluation of control performance, the depth residual network objective function is defined as:
wherein , and />The normalized current harmonic total distortion, the midpoint voltage deviation and the average switching frequency are respectively; alpha 1 、α 2 And alpha is 3 For weights predefined in the objective function, alpha 1 The larger the value of (2), the larger the weight of the optimized current harmonic total distortion in the objective function; alpha 2 The larger the value of (2), the larger the weight of the control midpoint voltage deviation in the objective function; alpha 3 The larger the value of (2), the larger the weight of the average switching frequency at the objective function.
And carrying out normalization processing on the data set to eliminate interference of different units and amplitudes, thereby improving the training speed of the residual error network. The collected data sets are divided into a training set and a cross-validation set according to the proportion of 70% and 30%, wherein the training set is used for training a designed network, and the cross-validation set is used for evaluating the network performance.
After normalization processing and division are carried out on the data, the invention constructs a depth residual error network according to the data set by utilizing MATLAB/Simulink simulation software.
Step 3: constructing a depth residual error network, taking the weight factors as an input layer of the network, and THD c 、V ag and fsw As an output layer of the network, training the network according to the data set and Adam algorithm, determining the hyper-parameters of the residual network by minimizing the root mean square error RMSE, and selecting the prediction index of the minimized objective function as the optimal weight factor combination of the model predictive control algorithm.
The constructed depth residual network mainly comprises an input layer 1, a residual block 2, a residual connection 3, an accumulation layer 4 and an output layer 5 as shown in fig. 4.
Input layer 1 is the weight factor for neutral voltage balance and switching frequency adjustmentSub lambda dc and λsw The method comprises the steps of carrying out a first treatment on the surface of the The full connection layer (1), the activation function (2), the accumulation layer (3) and the residual connection (4) form a residual block 2, wherein the information transfer formula of the residual block 2 is as follows:
f [l] ＝max(0,Z j [l] ) j＝1...N l
a [l] ＝f [l] (Z j [l] )
wherein ,f[l] Is the functional expression of the activation function ReLU, Z j [l] To activate the input of the function, when Z j [l] When the value of (a) reaches the threshold of ReLU, the neuron output is a [l] And will pass to the accumulation layer and then to the next residual block on the same principle. N (N) l The number of neurons in layer I, j is the number of neurons in the upper layer, and w j [l] and bj [l] Weights and deviations for the full connection layer.
Compared with a common network, the residual connection is added between every two layers/three layers of the residual network, the input and the output are superimposed point by point, the addition does not add extra parameters and calculation amount to the network, meanwhile, the training speed of a model can be greatly improved, the training effect is improved, when the number of layers of the network is deepened, the structure can prevent gradient from disappearing, and the problem of model degradation in a deep network is solved.
The training process of the residual error network is to update network parameters by using an Adam algorithm of a back propagation algorithm, the residual error network realizes identity mapping by using residual error links, and gradient descent correction of back propagation is as follows:
wherein H (x) and H' (x) represent outputs of the previous FC layer and the next FC layer; x is the output of the residual connection. D (D) 1 and D2 Is the gradient of a conventional neural network. As the number of hidden layers increases, the gradient will decrease, resulting in the disappearance of the gradient of the neural network. But by means of residual connection, when D 1 and D2 When it becomes zero, the gradient can be prevented from disappearing.
The hyper-parameters of the residual network are determined with a minimized Root Mean Square Error (RMSE) as a quantization index, the RMSE expressed as:
wherein ,for neural network predictors, y i Being the true value of the dataset, m is the number of samples of the dataset.
In order to avoid the network from falling into a locally optimal solution, the residual network is trained multiple times by using different initial values. The deep learning tool box Experiment Manager in the MATLAB2020b is selected as an automatic machine learning technology (AutoML), so that the optimal super-parameters for minimizing RMSE can be automatically found, and the problem of complex super-parameter adjustment of a residual network is solved.
Two residual blocks and one accumulation layer are connected by one residual, and a plurality of the residual blocks and one accumulation layer are connected in series to form a depth residual network. The input layer 1 of the depth residual network is a weight factor (lambda) dc and λsw ) The output layer 5 is an optimization objective function (THD c 、V ag and fsw ). Fig. 5 is a graph of training effects of a depth residual network, which has better evaluation index and smaller minimized Root Mean Square Error (RMSE) than the conventional BP neural network.
Will weight factor lambda dc 、λ sw The range settings are 0-5, the step length is 0.25, and the current harmonic total distortion THD is used as the input layer of the residual parallel network c Midpoint voltage deviation V ag And an average switching frequency f sw As a network output layer, the Adam algorithm is selected for back propagation training to minimize root mean square error(RMSE) to evaluate the trained network and adjust the super-parameters. The trained network can predict the weight factor minimizing the objective function, and the prediction index is the optimal weight factor of the model predictive control algorithm. i.e b
FIG. 6 shows the performance of the target function predicted by the present invention corresponding to the optimal weighting factor under different working conditions, and the result shows that the trained residual error network can accurately predict the optimal weighting factor under different working conditions. The superiority of the depth residual error network prediction weight factor is further explained, the performance of model prediction control is improved, and the method is suitable for complex application working conditions of a multi-level converter system.
Step 4: and constructing a data set by using the optimal weight factors under different working conditions and the corresponding working conditions, and simultaneously realizing dynamic optimization of the model predictive control weight factors under different working conditions by using a table look-up method.
A look-up table is designed consisting of predicted optimal weight factors and defined operating states. The operating state is reduced to rotor speed (N) ref ) And load torque (T) e ) Is a combination of (a) and (b). For example, there are 25 different operating states, and then there are 25 combinations of weighting factors. Through 10 trained networks, the optimal weight factors under 25 defined working states can be predicted. And then automatically finding out 10 trained networks corresponding to 25 working states through an AutoML technology, and automatically collecting the 10 trained networks by a lookup table. According to the current working state, the prediction weight factors in the lookup table are adjusted on line, so that the dynamic optimization of the model prediction control performance can be realized.
The method comprises the following specific steps:
1) And (3) data generation: generating a range of weight factors and a dataset for training and evaluating a depth residual network by simulation;
2) Training and prediction: and adjusting the super parameters according to the training performance of the designed residual error network. The optimal hyper-parameters can be automatically found by an experiment manager in the deep learning tool box, and the lookup table consists of predicted weight factors.
3) Dynamic optimization: according to the current working state, the dynamic optimization of the MPC is realized by selecting the optimal weight factors from the lookup table.
Claims (5)
1. The dynamic optimization method of the predictive control weight factor based on the depth residual error network is characterized by comprising the following steps:
step 1: establishing a three-level inverter feed permanent magnet motor driving system prediction mathematical model, predicting a cost function of a control algorithm according to a control target design model, and obtaining corresponding motor stator current harmonic total distortion THD under different weight factor combinations through MATLAB/Simulink simulation c Neutral point voltage deviation V of DC side bus ag And average switching frequency f sw Forming a data sample;
step 2: defining an objective function based on a control target of a permanent magnet motor driving system to evaluate control performance, then establishing a data set, carrying out normalization processing on the data set, and dividing the data set into a training set and a cross verification set according to a proportion for respectively designing, training and evaluating network performance;
step 3: constructing a depth residual error network, taking the weight factors as an input layer of the network, and THD c 、V ag and fsw As the output layer of the network, training the network according to the data set and Adam algorithm, determining the super parameter of the residual network by minimizing the Root Mean Square Error (RMSE), and selecting the prediction index of the minimized objective function as the optimal weight factor combination of the model prediction control algorithm;
step 4: and constructing a data set by using the optimal weight factors under different working conditions and the corresponding working conditions, and simultaneously realizing dynamic optimization of the model predictive control weight factors under different working conditions by using a table look-up method.
2. The method for dynamically optimizing the predictive control weight factor based on the depth residual network according to claim 1, wherein the establishing of the predictive mathematical model and the cost function in the step 1 is specifically as follows:
the method comprises the steps of taking a permanent magnet motor fed by a three-level active neutral point clamped inverter as a control object, performing discretization processing on a motor stator voltage equation under a dq rotating coordinate system by adopting a first-order forward Euler method, and establishing a discretization stator current prediction model of a permanent magnet motor driving system under the rotating coordinate system, wherein the discretization stator current prediction model is expressed as follows:
wherein k represents the sampling time, T s For sampling period, i d (k+1)、i q (k+1) represents the dq-axis stator current prediction value at the sampling time k+1, i d (k)、i q (k) Representing the dq-axis stator current sampling value at the current sampling time k, u d (k)、u q (k) Representing the dq-axis stator voltage value at the current sampling time k, L d 、L q Representing dq-axis stator inductance in rotating coordinate system, R s Representing the resistance value, ω, of the stator winding r Representing the electrical angular velocity of a permanent magnet motor, ψ f Representing permanent magnet flux linkage;
according to the voltage values of the upper capacitor and the lower capacitor at the side of the direct current bus, a discretization prediction model of the voltage difference value of the capacitor at the side of the direct current bus is established, wherein the prediction model of the potential of the neutral point is expressed as follows:
wherein ,vc1 and vc2 Representing the upper and lower capacitance voltages, v n Represents neutral point potential, C dc Representing the capacitance of the DC side capacitor, u x Representing normalized output phase voltage, i x Representing phase current, u x (k) Representing the normalized phase voltage, i, at the current sampling instant k x (k) Represents the phase current, deltav, at the current sampling instant k n(k) and Δvn (k+1) represents the upper and lower capacitance voltage deviations of the current sampling time k and the sampling time k+1, respectively;
the number of on and off switching times of the inverter switching tube is expressed as:
Δu x (k+1)＝||u x (k+1)-u x (k)||
wherein ,ux (k+1) represents the normalized phase voltage at the next sampling time k+1;
to achieve comprehensive optimization of multiple objectives, a cost function is defined by considering stator current tracking, midpoint potential balancing, and switching adjustment as:
wherein ,Ji Cost function representing tracking current error, J dc Cost function, J, representing midpoint voltage deviation sw Cost function, i, representing switching frequency tracking error sref Representing the reference phase current, i s (k+1)＝[i d (k+1)，i q (k+1)] T Representing a stator current predicted value under a next sampling time dq rotation coordinate system;
the method comprises the steps of establishing a total cost function according to the established cost functions of tracking current errors, midpoint voltage deviations and switching frequency tracking errors, wherein the total cost function is expressed as:
J＝J i +λ dc J dc +λ sw J sw
wherein J is the total cost function, lambda dc 、λ sw Weight factors respectively representing neutral point potential balance and switching frequency adjustment; and enumerating 27 space voltage vectors, calculating a cost function J, and selecting the voltage vector with the smallest J as the optimal output of the current control period.
3. The method for dynamically optimizing predictive control weight factors based on depth residual network according to claim 1, wherein the step 2 defines an objective function, establishes a data set, and divides the data set for network training and evaluation specifically comprises:
defining a depth residual network objective function as:
wherein , and />Respectively predicting, controlling and normalizing the current harmonic total distortion, midpoint voltage deviation and average switching frequency of the model; alpha 1 、α 2 And alpha is 3 For weights predefined in the objective function, alpha 1 The larger the value of (2), the larger the weight of the optimized current harmonic total distortion in the objective function; alpha 2 The larger the value of (2), the larger the weight of the control midpoint voltage deviation in the objective function; alpha 3 The larger the value of (2), the larger the weight of the average switching frequency at the objective function;
the weight ranges are 0-5, and the step sizes are 0.25; collecting corresponding current harmonic distortion THD c Voltage deviation V of midpoint ag And a switching frequency f sw Forming a data set, and carrying out normalization processing on the data set;
the data sets were divided into training sets for training the designed network and cross-validation sets for evaluating network performance at 70% and 30% duty cycles.
4. The method for dynamically optimizing predictive control weight factors based on depth residual error network according to claim 3, wherein the depth residual error network constructed in the step 3 is specifically:
the depth residual network consists of a fully connected FC layer and an activation function ReLU, expressed as:
f(x)＝max(0,x)
the training process of the residual network is to update parameters by using gradient descent of a back propagation algorithm, and is expressed as follows:
wherein ω 'and b' are updated weights and deviations, respectively; alpha is learning rate, h (x) is output of the FC layer;
the backward propagation gradient drop is corrected as:
wherein H (x) and H' (x) represent the outputs of the previous FC layer and the next FC layer, respectively; x is the output of the residual connection; d (D) 1 and D2 Is the gradient of a conventional neural network;
the hyper-parameters of the residual network are adjusted with the minimized root mean square error RMSE as a quantization index, RMSE expressed as:
the deep learning tool box Experiment Manager in MATLAB2020b is selected as an automatic machine learning technique AutoML;
based on the established data set, the weight factor lambda is calculated dc 、λ sw As a residual network input layer, the total distortion THD of current harmonic waves c Midpoint voltage deviation V ag And an average switching frequency f sw As a network output layer, selecting an Adam algorithm for back propagationTraining to minimize root mean square error RMSE, evaluating the trained network, and adjusting the super-parameters.
5. The method for dynamically optimizing predictive control weight factors based on depth residual network according to claim 4, wherein the implementing the dynamic optimization of weight factors is specifically: a lookup table is constructed, which consists of predicted optimal weight factors and defined working states, and the working states are simplified into rotor rotating speed N ref And a load torque T e Is a combination of (a); the method comprises the following specific steps:
1) And (3) data generation: generating a range of weight factors and a dataset for training and evaluating a depth residual network by simulation;
2) Training and prediction: according to the designed training performance of the residual error network, the super-parameters are adjusted, the optimal super-parameters are automatically found by an experiment manager in a deep learning tool box, and a lookup table consists of predicted weight factors;
3) Dynamic optimization: according to the current working state, the dynamic optimization of the MPC is realized by selecting the optimal weight factors from the lookup table.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310830632.XA CN116954072A (en) | 2023-07-07 | 2023-07-07 | Prediction control weight factor dynamic optimization method based on depth residual error network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310830632.XA CN116954072A (en) | 2023-07-07 | 2023-07-07 | Prediction control weight factor dynamic optimization method based on depth residual error network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116954072A true CN116954072A (en) | 2023-10-27 |
Family
ID=88457631
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310830632.XA Pending CN116954072A (en) | 2023-07-07 | 2023-07-07 | Prediction control weight factor dynamic optimization method based on depth residual error network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116954072A (en) |
-
2023
- 2023-07-07 CN CN202310830632.XA patent/CN116954072A/en active Pending
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111541411B (en) | Double-three-level inverter open winding motor model control method | |
| CN112003491B (en) | Control method and system of modular parallel three-phase three-level inverter | |
| WO2022151609A1 (en) | Dual three-phase permanent magnet synchronous motor control method for alternately executing sampling and control program | |
| CN112910297B (en) | Three-level SNPC converter system and two-stage model prediction control method | |
| Yao et al. | ANN optimization of weighting factors using genetic algorithm for model predictive control of PMSM drives | |
| Kamalapur et al. | A novel neuro fuzzy method DTC of induction motor drive for efficient torque response | |
| CN112701951B (en) | Inverter voltage state prediction control method based on tolerant hierarchical sequence method | |
| CN116954072A (en) | Prediction control weight factor dynamic optimization method based on depth residual error network | |
| CN115685747A (en) | Model prediction control method based on residual error neural network optimization | |
| CN109617438B (en) | Control method of pure electric vehicle modular multilevel converter | |
| CN117097189A (en) | Control method of photovoltaic inverter | |
| CN110518625B (en) | Grid-connected inverter direct-current component suppression method with variable learning rate BP-PID control | |
| Sivaraju et al. | Improving the efficiency of induction motor drive by flux and torque control: A hybrid LSE-RERNN approach | |
| Muneer et al. | Optimal machine-learning-based controller for shunt active power filter by auto machine learning | |
| CN116232132A (en) | Rotating speed control method of brushless direct current motor of electric automobile | |
| CN112531782B (en) | Method for suppressing voltage fluctuation of bus of wind power converter | |
| CN116208017A (en) | Inverter predictive control method based on optimal switching sequence model | |
| CN111769770B (en) | Linear induction motor multi-step finite set model prediction control method and system | |
| CN114710055A (en) | Two-parallel power converter model prediction control method based on finite set single vector | |
| CN114785093A (en) | MMC model prediction control method of direct-drive permanent magnet wind power system | |
| CN114301062A (en) | Distributed energy optimization treatment device | |
| CN112182945A (en) | Test microgrid construction and control method capable of being used for verification of various microgrid structures | |
| Tian et al. | Weighting factors design in model predictive direct torque control based on cascaded neural network | |
| Yang et al. | Analytic Hierarchy Process-based Control System Optimization for Buck Circuit using Genetic Algorithm | |
| CN115622465B (en) | Multi-level converter weightless factor model prediction control method |
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 |