CN115685747A - A Model Predictive Control Method Based on Residual Neural Network Optimization - Google Patents

Abstract

The invention discloses a model prediction control method based on residual error neural network optimization, which specifically comprises the following steps: establishing a prediction model of system state variables, adjusting a cost function of a model predictive control algorithm, and recording current harmonic distortion rate and switching frequency composition data samples under different weight factors; establishing an optimization objective function according to the collected data, and dividing a data set into a training set, a verification set and a test set; constructing a residual parallel neural network, taking a weight factor as a network input layer and a target function as a network output layer, training the network by relying on a data set and an Adam algorithm, adjusting hyper-parameters such as a residual module, network depth and the like of the neural network according to evaluation indexes, and selecting a residual network with optimal indexes; and taking the weight factor with shorter step length as an input layer of the optimal network, predicting the target function, searching the weight factor which enables the target function to be minimum, and realizing the optimal control of the state variable. The invention realizes the optimization of the weight factors under different working conditions and different performance requirements, and effectively improves the control performance and robustness of the model predictive control algorithm.

Description

A Model Predictive Control Method Based on Residual Neural Network Optimization

technical field

The invention belongs to the field of model predictive control, in particular to a model predictive control method based on residual neural network optimization.

Background technique

Model predictive control (Model Predictive Control, referred to as MPC) is a motor drive control strategy that has emerged in recent years. It has been gradually applied in industrial fields such as wind power generation, ship propulsion, and traction drive. Compared with traditional vector control, MPC has multi-objective , multi-variable and multi-constraint control characteristics, more suitable for multi-level converters. At the same time, MPC can greatly increase the torque response speed of the traction motor, balance the midpoint voltage of the DC side, and reduce the current harmonics and switching losses of the converter. In addition, MPC can handle multi-objective optimization problems conveniently, outperforming traditional field-oriented vector control and direct torque control. Therefore, MPC is considered to be one of the most promising motor control strategies. However, the control performance of MPC is limited by the design of the cost function and its weight factors. At present, there are few design and optimization methods of weight factors, and there is still a lack of mature theoretical methods.

Conventional weight factor optimization methods mainly start from avoiding the use of cost functions and state observers. The former uses methods such as voltage signal injection, reference vector merging, and redundant vector reconstruction to achieve the purpose of canceling the weight factor, avoiding the selection of the weight factor, but sacrificing the system dynamic performance. The latter monitors the state information under different working conditions through the observer, adjusts the weight factor of the cost function online according to the model prediction control performance, and then realizes the online adjustment and optimization of the weight factor, but this method increases the complexity and calculation of the control system. The amount reduces the response speed of the MPC, and when it is used in a multilevel converter system, the calculation amount will increase sharply. Therefore, it is urgent to study a weight factor optimization method that does not increase the amount of calculation and has universality, so as to improve the control performance of MPC and meet the traction requirements of traction trains under complex operating conditions. In recent years, with the vigorous development of artificial neural network (ANN), the nonlinear fitting characteristics of ANN have been widely used in pattern recognition, natural language processing, biotechnology and transportation.

Contents of the invention

In view of the above-mentioned deficiencies, the present invention provides a model predictive control method based on residual neural network optimization, which utilizes the residual neural network to optimize the cost function of the MPC algorithm and its weighting factors to achieve optimal performance under various operating conditions. control without increasing the computational burden of the MPC algorithm.

A kind of model predictive control method based on residual neural network optimization of the present invention comprises the following steps:

Step 1: Establish a prediction model of system state variables, adjust the cost function of the model predictive control algorithm, and record the current harmonic distortion rate and switching frequency under different weight factors to form data samples.

Step 2: Based on the collected data, establish an optimization objective function, and at the same time divide the data set into training set, verification set and test set for network evaluation and training.

Step 3: Build a residual parallel neural network, use the weight factor as the network input layer, and the objective function as the network output layer, rely on the data set and the Adam algorithm to train the network, and adjust the hyperparameters such as the residual module and network depth of the neural network according to the evaluation index , select the index optimal network.

Step 4: Use the weight factor with shorter step size as the input layer of the optimal network, predict the objective function, find the weight factor that minimizes the objective function, and realize the optimal control of the state variable.

Further, the establishment method of the above model predictive control algorithm cost function is as follows:

Taking the permanent magnet motor as the control object, the fixed and discrete sub-current prediction model of the motor drive system in the rotating coordinate system is established, expressed as:

Among them, i _d (k+1), i _q (k+1) represent the dq axis stator current prediction value at k+1 sampling time, and i _d (k), i _q (k) represent the current k sampling time dq axis stator current sampling value, u _d (k) and u _q (k) represent the dq axis stator voltage value at the current k sampling moment, T _s represents the sampling period, R _s represents the stator winding resistance, L _d and L _q represent The stator inductance under the dq axis, ω _e represents the electrical angular velocity of the permanent magnet motor, and ψ _f represents the flux linkage of the permanent magnet.

According to the current value flowing through the DC bus upper and lower capacitors, a discrete prediction model for the DC bus side capacitors is established to predict the voltage values of the DC bus upper and lower capacitors; the calculation formula is:

Among them, i _c1 (k) and i _c2 (k) represent the current values of the upper and lower capacitors on the DC bus, respectively, and v _c1 (k) and v _c2 (k) represent the upper and lower capacitors on the DC bus side at the current k sampling time The voltage values, v _c1 (k+1) and v _c2 (k+1) represent the predicted voltage values of the upper and lower capacitors on the DC bus side at sampling time k+1 respectively, and C is the capacitance of the capacitor on the DC side.

Among them, J _i represents the cost function of the tracking current error, J _dc represents the cost function of the midpoint voltage deviation value, S _x (k-1) represents the normalized phase voltage at the last sampling moment, and S _x (k) represents The normalized phase voltage at the current sampling moment.

According to the established cost functions of tracking current error, midpoint voltage deviation and switching frequency tracking error, the total cost function is expressed as:

J＝J _i +λ _dc J _dc +λ _sw J _sw

Among them, J is the total cost function, λ _dc and λ _sw respectively represent the weight factors of neutral point potential balance and switching frequency adjustment; according to Simulink simulation, the corresponding current harmonic distortion THD and switching frequency f _sw under different weight factor combinations are obtained , to form a dataset.

Further, the above step 2 is specifically:

The data set is subjected to outlier processing, and the outlier is expressed as:

I=Q ₃ -Q ₁

O _l =Q ₃ -1.5×I

O _u =Q ₃ +1.5×I

Among them, Q ₃ , I and Q ₁ are the upper quartile, the median and the lower quartile respectively, when the value of the data set is greater than O _u or less than O _l , it will be regarded as an outlier; cleaning outliers After that, the data set will be divided into training set, cross-validation set and test set according to the ratio of 70%, 15%, and 15%.

The training set and the cross-validation set need to be normalized to improve the training speed of the residual network; the data is normalized by using the range change method, and the data is mapped to [0-1]. The process is expressed as :

通过归一化后的个体值。Among them, x _i is the individual value of the data sample, x _min is the minimum value in the data set, x _max is the maximum value,

Individual values after normalization.

The training set is used to train the neural network, and the cross-validation set and verification set are used to evaluate the accuracy and generalization ability of the network respectively. The quantitative indicators of accuracy and generalization ability are expressed as:

为神经网络预测值，y_i为数据集中的真实值，m为数据集的样本数。Among them, RMSE is a quantitative index,

is the predicted value of the neural network, y _i is the real value in the data set, and m is the number of samples in the data set.

Further, the residual parallel neural network constructed in the above step 3 includes an input layer, a fully connected layer, an activation function layer, a fully connected network, an accumulation layer, a parallel network unit, a residual connection, and an output layer.

The input layer is composed of model predictive control weight factors, which are the midpoint voltage balance λ _dc and the switching frequency limit λ _sw ; the fully connected layer and the activation function ReLU form a fully connected network, where the information transfer formula of the fully connected layer is expressed as:

g ^[l] = max(0,Z _j ^[l] ) j = 1...N _l

a ^[l] = g ^[l] (Z _j ^[l] )

Among them, g ^[l] is the expression of the activation function ReLU. When the value of Z _j ^[l] reaches the threshold of g ^[l] , the output a ^[l] of the neuron will be passed to the next layer, and N _l is the l layer The number of neurons, j is the number of neurons in the previous layer, w _j ^[l] and b _j ^[l] are the weights and biases of the fully connected layer.

The parallel network unit and the residual connection constitute the residual parallel network, and the residual connection makes the parallel network unit skip the next layer, and the interlayer is connected to the accumulation layer.

According to the collected data set, define the objective function of the residual network, expressed as:

f _ANN ＝α _t ×THD+α _f ×f _sw

Among them, THD is the total current harmonic distortion, f _sw is the switching frequency of model predictive control, α _t and α _f are the weight coefficients of current harmonic distortion and switching frequency, respectively, the larger the value of α _t is, the optimized current harmonic total The larger the weight of distortion in the objective function, the more the relation of α _f with reducing the switching frequency.

According to the collected data, the weight factor is used as the residual parallel network input layer, and the current harmonic distortion is used as the network output layer, combined with the Adam algorithm for backpropagation training, the trained network is evaluated according to the RMSE, and the parallel structure is adjusted according to the training results. parameters to select the optimal network.

Furthermore, the weight factor with a shorter step size is used as the input layer of the optimal network, and the optimal network will give the corresponding objective function value, so that the weight factor with the smallest objective function is the optimal weight factor under the objective function.

The beneficial technical effect of the present invention is:

(1) The model predictive control method proposed by the present invention carries out off-line training to the neural network by constructing a residual parallel neural network, so that it has the ability to predict the optimal weight factor. This method will not increase the computational burden of the model predictive control algorithm, It can objectively and efficiently predict the optimal weight factor under different working conditions. Since this method is a data-driven algorithm, it only needs to collect and define the data of the network input layer and output layer, without changing the model predictive control algorithm itself.

(2) Compared with the traditional BP neural network, the present invention has higher accuracy, and the residual connection avoids the problems of deep network gradient disappearance and network degradation, and the parallel structure enables the residual network to have better Model expression ability and higher prediction accuracy, while the parallel structure simplifies the difficulty of neural network adjustment.

(3) The invention is based on a data-driven optimization method, which has good robustness to data differences, so this method is more suitable for model predictive control strategies under complex operating conditions, such as traction motor model predictive control systems, the Complex working conditions can be equivalent by several objective functions, thus simplifying the complexity of the control system.

Description of drawings

Fig. 1 is the optimization flowchart of model predictive control weight factor of the present invention;

Figure 2 is a block diagram of a permanent magnet motor drive system based on a three-level neutral-point clamped inverter;

Fig. 3 is a model predictive control simulation diagram of the present invention;

Fig. 4 is residual parallel neural network structural diagram of the present invention;

Fig. 5 is the residual parallel neural network training rendering of the present invention;

Fig. 6 is a diagram of the optimization result of the residual parallel neural network of the present invention.

Detailed ways

The present invention will be described in further detail below in conjunction with the accompanying drawings and specific implementation methods.

The present invention takes a permanent magnet motor drive system fed by a three-level neutral point clamping inverter as an example, and provides a model predictive control method based on residual neural network optimization, as shown in Figure 1, specifically:

S1. Establish a prediction model of system state variables, adjust the cost function of the model prediction control algorithm, and record the current harmonic distortion rate and switching frequency under different weight factors to form data samples.

The present invention first establishes the mathematical model of the permanent magnet motor drive system, as shown in FIG. 2 , which is the three-level diode midpoint clamp type inverter permanent magnet motor drive system involved in this embodiment. In the inverter topology, each phase bridge arm contains four switching devices with the same switching frequency. Among them, the switching tubes _Tx1 and _Tx3 are complementary devices, _Tx2 and _Tx4 are complementary devices, and the gate switching signals are reversed. Table 1 gives the output voltage corresponding to different switching states of the inverter topology:

Table 1 Switching status of three-level inverter

In the table, V _dc is the voltage value provided by the DC bus side; the normalized vector S _x in the table can be further expressed as:

S _x = [S _a S _b S _c ]

In the formula, the value ranges of S _a , S _b , and S _c are all (1, 0). Subsequently, the mathematical model of the permanent magnet motor is established:

Among them, u _d and u _q represent the stator voltage under the dq axis, R _s represents the stator winding resistance, id _and i _q represent the stator current under the dq axis, L _d and L _q represent the stator inductance under the dq axis, ω _e Expressed as the electrical angle of the permanent magnet motor, ψ _f represents the flux linkage of the permanent magnet. At the same time, the first-order forward Euler method is used to discretize the stator current model of the permanent magnet motor in the rotating coordinate system, and a discretized stator current prediction model in the rotating coordinate system is established, expressed as:

Among them, i _d (k+1), i _q (k+1) represent the dq-axis stator current prediction value at k+1 sampling time, and i _d (k), i _q (k) represent dq at the current k sampling time axis stator current sampling value, u _d (k) and u _q (k) represent the dq axis stator voltage value at the current k sampling moment, and T _s represents the sampling period.

In practical applications, due to the calculation and sampling delay, the delay needs to be compensated. By one-step compensation for the sampling delay and the control delay, the predicted value of the dq axis stator current at the k+2 sampling time is obtained, expressed as:

Among them, i _d (k+2), i _q (k+2) represent the dq axis stator current prediction value at k+1 sampling time, u _d (k+1), u _q (k+1) represent k+ 1 The dq axis stator voltage value to be adopted at the sampling moment.

According to the switching state and phase current of the inverter, the DC bus capacitor current is calculated, expressed as:

Among them, i _c1 (k) and i _c2 (k) represent the current value of the DC bus capacitor and the lower capacitor respectively, _idc (k) represents the current value flowing through the DC bus capacitor, S _a1 , S _b1 , S _c1 , S _a4 , S _b4 , and S _c4 represent the switch status values of the corresponding bridge arm switch tubes, 1 means the switch tube is turned on, 0 means the switch tube is turned off, and _ia (k), _ib (k), and _ic (k) respectively Indicates the output current value of the corresponding bridge arm.

According to the current value flowing through the DC bus upper and lower capacitors, a discrete prediction model for the inverter DC bus side capacitors is established to predict the voltage values of the upper and lower capacitors on the DC bus side; the calculation method is:

Among them, v _c1 (k) and v _c2 (k) represent the voltage values of the upper and lower capacitors on the DC bus side at the current k sampling time respectively, and v _c1 (k+1) and v _c2 (k+1) respectively represent k+1 The predicted values of the upper and lower capacitor voltages on the DC bus side at the sampling moment, and C is the capacitance of the DC side capacitor.

Among them, J _i represents the cost function of the tracking current error, J _dc represents the cost function of the midpoint voltage deviation value, J _sw represents the cost function of the switching frequency tracking error, S _x (k-1) represents the normalized value at the previous sampling time Normalized phase voltage, S _x (k) represents the normalized phase voltage at the current sampling moment.

According to the established cost function of establishing tracking current error, midpoint voltage deviation and switching frequency tracking error, the total cost function is expressed as:

J＝J _i +λ _dc J _dc +λ _sw J _sw

Among them, J is the total cost function, λ _dc and λ _sw represent the weight factors of neutral point potential balance and switching frequency adjustment respectively, and the value ranges of λ _dc and λ _sw are 0-5 and 0-10 respectively, and the step size Both are 0.5. As shown in Figure 3, according to the Simulink simulation, the corresponding current total harmonic distortion (THD) and switching frequency (f _sw ) under different combinations of weight factors (λ _dc and λ _sw ) are obtained, and they are composed into a simulation data set.

After the data sets required by the residual parallel network are established, the present invention cleans and normalizes the abnormal values of the data sets according to the subsequent steps.

S2. Based on the collected data, an optimization objective function is established and the data set is divided into a training set, a verification set and a test set for grid evaluation and training.

According to the collected data set, define the objective function of the residual network, which can be expressed as:

f _ANN ＝α _t ×THD+α _f ×f _sw

Among them, THD is the total harmonic distortion of the current, and f _sw is the switching frequency of the model predictive control. α _t and α _f are the weight coefficients of current harmonic distortion and switching frequency respectively. The larger the value of α _t is, the greater the weight of the optimized current harmonic total distortion in the objective function will be. The same is true for α _f with decreasing switching frequency. In this example, α _t is 5 and α _f is 1. In addition, the data set is cleaned of outliers, and the outliers are expressed as:

I=Q ₃ -Q ₁

O _l =Q ₃ -1.5×I

O _u =Q ₃ +1.5×I

Among them, Q ₃ , I and Q ₁ are the upper quartile, the median and the lower quartile respectively, and when the value of the data set is greater than _Ou or less than _Ol , it will be regarded as an outlier. After cleaning the outliers, the data set will be divided into training set, cross-validation set and test set according to the ratio of 70%, 15%, and 15%. The training set and the cross-validation set need to be normalized to speed up the training of the residual network. The present invention adopts the range change method to normalize the data, and maps the data to [0-1]. The process can be expressed as:

通过归一化后的个体值。此外，训练集训练神经网络，交叉验证集和验证集分别用于评估网络准确率和泛化能力。准确率与泛化能力的量化指标表示为：Where x _i is the individual value of the data sample, x _min is the minimum value in the data set, x _max is the maximum value,

Individual values after normalization. In addition, the training set trains the neural network, and the cross-validation set and validation set are used to evaluate the network accuracy and generalization ability, respectively. The quantitative indicators of accuracy and generalization ability are expressed as:

为神经网络预测值，y_i为数据集中的真实值，m为数据集的样本数。将数据集异常值处理和归一化之后，本发明将根据数据分布，利用MATLAB构建残差并联网络。Among them, RMSE is a quantitative index,

is the predicted value of the neural network, y _i is the real value in the data set, and m is the number of samples in the data set. After processing and normalizing the outliers of the data set, the present invention uses MATLAB to construct a residual parallel network according to the data distribution.

S3. Build a residual parallel network, use the weight factor as the network input layer, and the objective function as the output layer, rely on the data set and the Adam algorithm to train the network, adjust the hyperparameters such as the residual module and network depth of the neural network according to the evaluation index, and select the index Optimal Residual Networks.

The residual network is shown in Figure 4, mainly including input layer 1, fully connected layer 2, activation function layer 3, fully connected network 4, accumulation layer 5, parallel network unit 6, residual connection 7, and output layer 8:

The input layer 1 is composed of model predictive control weight factors, which are respectively the midpoint voltage balance λ _dc and the switching frequency limit λ _sw ; the fully connected layer 2 and the activation function ReLU3 form a fully connected network 4, where the information transfer formula of the fully connected layer is expressed as

g ^[l] = max(0,Z _j ^[l] ) j = 1...N _l

a ^[l] = g ^[l] (Z _j ^[l] )

Among them, g ^[l] is the expression of the activation function ReLU. When the value of Z _j ^[l] reaches the threshold of g ^[l] , the output a ^[l] of the neuron will be passed to the next layer, and N _l is the l layer The number of neurons in j is the number of neurons in the previous layer, w _j ^[l] and b _j ^[l] are the weights and biases of the fully connected layer; in addition, the parallel structure makes the fully connected network network with the same parameters Under the scale, it has better expressive power of the model.

The parallel network unit 6 and the residual connection 7 together constitute a residual parallel network. The residual connection 7 enables the parallel network unit 6 to skip the next layer, and the interlayer is connected to the accumulation layer 5, weakening the strong connection between each layer. It avoids the problems of gradient disappearance and network degradation during the training process of the neural network, so that the deep network has better training performance.

According to the network training set, combined with the Adam algorithm for backpropagation training. The pseudo code of the Adam algorithm is shown in Table 2. Update w _j ^[l] and b _j ^[l] through backpropagation, according to the RMSE index, combined with the Experiment Manager in the MATLAB deep learning toolbox, adjust the residual parallel network structure, and select the network with the smallest RMSE. Compared with the number of neurons in the hidden layer, the fully connected network in the residual parallel network can improve the prediction accuracy of the residual parallel network. Therefore, it is only necessary to adjust the depth width (fully connected network) and depth (parallel network unit) of the residual disease to your network, without adjusting the number of neurons in the hidden layer, thus simplifying the adjustment process of the neural network. After adjustment, the optimal structure of the residual parallel network is residually connected by two parallel network units, and the parallel network unit is connected by four fully connected networks and accumulation layer 5. In addition, the input layer of the residual parallel network is the weight factors (λ _dc and λ _sw ), and the output layer 8 is the optimization objective function f _ANN . The training performance of the residual parallel network and the traditional BP network training RMSE are shown in Fig. 5.

Table 2 Adam backpropagation algorithm

S4. Use the weight factor with a shorter step size as the input layer of the optimal network. The optimal network will give the corresponding objective function value, and inverse normalize the predicted value of the optimal network to obtain the true value of the objective function. Then find the weight factor that minimizes the objective function to achieve the optimal control of the state variables.

Among them, the range of the weight factor remains unchanged, the step size is 0.01, and the minimum value is found in the predicted f _ANN , and the corresponding weight factor is the optimal weight factor under the objective function. Bring the predicted weight factor into the cost function of model predictive control. It can be seen from Figure 6 that compared with the traditional BP network, the residual parallel network can accurately find the optimal weight factor corresponding to the objective function. Therefore, the control performance of model predictive control is improved, and it is suitable for complex working conditions of train traction.

Claims (5)

1. A model predictive control method based on residual neural network optimization, is characterized in that, comprises the following steps: Step 1: Establish a prediction model of system state variables, adjust the cost function of the model predictive control algorithm, and record the current harmonic distortion rate and switching frequency under different weight factors to form data samples; Step 2: Based on the collected data, establish an optimization objective function, and at the same time divide the data set into training set, verification set and test set for network evaluation and training; Step 3: Build a residual parallel neural network, use the weight factor as the network input layer, and the objective function as the network output layer, rely on the data set and the Adam algorithm to train the network, and adjust the hyperparameters such as the residual module and network depth of the neural network according to the evaluation index , select the optimal index network; Step 4: Use the weight factor with shorter step size as the input layer of the optimal network, predict the objective function, find the weight factor that minimizes the objective function, and realize the optimal control of the state variable. 2. a kind of model predictive control method based on residual neural network optimization according to claim 1, is characterized in that, the establishment method of described model predictive control algorithm cost function is specifically: Taking the permanent magnet motor as the control object, a discretized stator current prediction model of the motor drive system in the rotating coordinate system is established, expressed as:

Among them, i _d (k+1), i _q (k+1) represent the dq axis stator current prediction value at k+1 sampling time, and i _d (k), i _q (k) represent the current k sampling time dq axis stator current sampling value, u _d (k) and u _q (k) represent the dq axis stator voltage value at the current k sampling moment, T _s represents the sampling period, R _s represents the stator winding resistance, L _d and L _q represent The stator inductance under the dq axis, ω _e represents the electrical angular velocity of the permanent magnet motor, and ψ _f represents the flux linkage of the permanent magnet; According to the current value flowing through the DC bus upper and lower capacitors, a discrete prediction model for the DC bus side capacitors is established to predict the voltage values of the DC bus upper and lower capacitors; the calculation formula is:

Among them, i _c1 (k) and i _c2 (k) represent the current values of the upper and lower capacitors on the DC bus, respectively, and v _c1 (k) and v _c2 (k) represent the upper and lower capacitors on the DC bus side at the current k sampling time Voltage values, v _c1 (k+1) and v _c2 (k+1) respectively represent the predicted values of the upper and lower capacitor voltages on the DC bus side at sampling time k+1, and C is the capacitance of the DC side capacitor;

in,

and

Respectively represent the stator current reference values of the d-axis and q-axis in the rotating coordinate system, J _i represents the cost function of the tracking current error, J _dc represents the cost function of the midpoint voltage deviation value, S _x (k-1) represents the last sampling The normalized phase voltage at the moment, S _x (k) represents the normalized phase voltage at the current sampling moment; According to the established cost function of tracking current error, midpoint voltage deviation and switching frequency tracking error, the total cost function is expressed as: J＝J _i +λ _dc J _dc +λ _sw J _sw Among them, J is the total cost function, λ _dc and λ _sw respectively represent the weight factors of neutral point potential balance and switching frequency adjustment; according to Simulink simulation, the corresponding current harmonic distortion THD and switching frequency f _sw under different weight factor combinations are obtained , to form a dataset. 3. a kind of model predictive control method based on residual neural network optimization according to claim 1, is characterized in that, described step 2 is specifically: The data set is subjected to outlier processing, and the outlier is expressed as: I=Q ₃ -Q ₁ O _l =Q ₃ -1.5×I O _u =Q ₃ +1.5×I Among them, Q ₃ , I and Q ₁ are the upper quartile, the median and the lower quartile respectively, when the value of the data set is greater than O _u or less than O _l , it will be regarded as an outlier; cleaning outliers After that, the data set will be divided into training set, cross-validation set and test set according to the ratio of 70%, 15%, and 15%; The training set and the cross-validation set need to be normalized to improve the training speed of the residual network; the data is normalized by using the range change method, and the data is mapped to [0-1]. The process is expressed as :

Among them, x _i is the individual value of the data sample, x _min is the minimum value in the data set, x _max is the maximum value,

Individual values after normalization; The training set is used to train the neural network, and the cross-validation set and verification set are used to evaluate the accuracy and generalization ability of the network respectively. The quantitative indicators of accuracy and generalization ability are expressed as:

Among them, RMSE is a quantitative index,

is the predicted value of the neural network, y _i is the real value in the data set, and m is the number of samples in the data set. 4. a kind of model predictive control method based on residual neural network optimization according to claim 1, is characterized in that, the residual parallel neural network that described step 3 builds comprises input layer (1), fully connected layer (2 ), activation function layer (3), fully connected network (4), accumulation layer (5), parallel network unit (6), residual connection (7), output layer (8); The input layer (1) is composed of model predictive control weight factors, which are respectively the midpoint voltage balance λ _dc and the switching frequency limit λ _sw ; the fully connected layer (2) and the activation function ReLU (3) form a fully connected network (4), in which The information transfer formula of the connection layer is expressed as

g ^[l] = max(0,Z _j ^[l] ) j = 1...N _l a ^[l] = g ^[l] (Z _j ^[l] ) Among them, g ^[l] is the expression of the activation function ReLU. When the value of Z _j ^[l] reaches the threshold of g ^[l] , the output a ^[l] of the neuron will be passed to the next layer, and N _l is the l layer The number of neurons, j is the number of neurons in the previous layer, w _j ^[l] and b _j ^[l] are the weights and biases of the fully connected layer; The parallel network unit (6) and the residual connection (7) together constitute a residual parallel network, the residual connection (7) makes the parallel network unit (6) skip the next layer, and the interlayer is connected to the accumulation layer (5); According to the collected data set, define the objective function of the residual network, expressed as: f _ANN ＝α _t ×THD+α _f ×f _sw Among them, THD is the total current harmonic distortion, f _sw is the switching frequency of model predictive control, α _t and α _f are the weight coefficients of current harmonic distortion and switching frequency, respectively, the larger the value of α _t is, the optimized current harmonic total The greater the weight of distortion in the objective function, the relationship between α _f and reducing the switching frequency is also the same; According to the collected data, the weight factor is used as the residual parallel network input layer, and the current harmonic distortion is used as the network output layer, combined with the Adam algorithm for backpropagation training, the trained network is evaluated according to the RMSE, and the parallel structure is adjusted according to the training results. parameters to select the optimal network. 5. a kind of model predictive control method based on residual neural network optimization according to claim 4, is characterized in that, the weight factor with the shorter step size is used as the input layer of optimal network, and optimal network will give corresponding The value of the objective function, so that the minimum weight factor of the objective function is the optimal weight factor under the objective function, and the optimal control of the state variable is realized.

Images