A model predictive control method based on residual neural network optimization
By constructing a residual parallel neural network to optimize the weight factor of MPC, the problem of insufficient control performance of MPC under complex operating conditions is solved, and efficient and accurate optimal weight factor prediction is achieved. It is suitable for multi-level converter systems and simplifies control system.
Patent Information
- Application Number
- CN202211211043.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-09-30
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2042-09-30
AI Technical Summary
The existing model predictive control (MPC) methods lack mature theoretical methods in weight factor optimization, resulting in limited control performance. The traditional optimization methods increase the computational volume and complexity, making it difficult to adapt to multi-level converter systems. Complex operating conditions.
Using the optimization method based on residual neural network, by constructing residual parallel neural network, using the Adam algorithm to train the network, optimize the cost function and weight factor of the MPC algorithm to achieve optimal control under different working conditions without increasing the computational burden. .
It realizes efficient prediction of optimal weight factors without increasing the computational burden, improves the control performance of MPC, is suitable for the prediction control of traction motor models under complex operating conditions, simplifies the complexity of the control system, and improves the Prediction accuracy.
Smart Images

Figure CN115685747B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of model predictive control, and in particular relates to a model predictive control method based on residual neural network optimization. Background Art
[0002] Model Predictive Control (MPC) is a motor drive control strategy that has emerged in recent years and has been gradually applied in industrial fields such as wind power generation, ship propulsion, and traction transmission. Compared with traditional vector control, MPC has multi-objective, multi-variable, and multi-constraint control characteristics, and is more suitable for multi-level converters. At the same time, MPC can greatly improve the torque response speed of the traction motor, balance the DC side midpoint voltage, and reduce the current harmonics and switching losses of the converter. In addition, MPC can easily handle multi-objective optimization problems, which is better than traditional field-oriented vector control and direct torque control. Therefore, MPC is considered to be one of the motor control strategies with the most development potential. However, the control performance of MPC is limited by the design of the cost function and its weight factor. At present, there are few methods for the design and optimization of weight factors, and there is still a lack of mature theoretical methods.
[0003] Conventional weight factor optimization methods mainly start from avoiding the use of cost functions and state observers. The former uses voltage signal injection, reference vector merging, and redundant vector reconstruction to achieve the purpose of canceling weight factors, avoiding the selection of weight factors, but sacrificing the dynamic performance of the system. The latter monitors the state information under different working conditions through the observer, and adjusts the weight factor of the cost function online according to the model prediction control performance, thereby realizing online adjustment and optimization of the weight factor. However, this method increases the complexity and calculation amount of the control system, reduces the response speed of MPC, and when used in multi-level converter systems, the amount of calculation 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, improve the MPC control performance, and meet the traction requirements of traction trains under complex operating conditions. In recent years, with the vigorous development of artificial neural networks (ANN), the nonlinear fitting characteristics of ANN have been widely used in pattern recognition, natural language processing, biotechnology, and transportation. Summary of the invention
[0004] 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 and weight factor of the MPC algorithm to achieve optimal control under various operating conditions without increasing the computational burden of the MPC algorithm.
[0005] A model predictive control method based on residual neural network optimization of the present invention comprises the following steps:
[0006] Step 1: Establish a prediction model for system state variables, adjust the cost function of the model predictive control algorithm, and record the current harmonic distortion rate and switching frequency composition data samples under different weight factors.
[0007] Step 2: Based on the collected data, an optimization objective function is established, and the data set is divided into a training set, a validation set, and a test set for network evaluation and training.
[0008] Step 3: Construct a residual parallel neural network, use the weight factor as the network input layer and the objective function as the network output layer, train the network based on the data set and the Adam algorithm, adjust the residual module, network depth and other hyperparameters of the neural network according to the evaluation indicators, and select the network with the best indicators.
[0009] Step 4: Use the weight factor with a 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 achieve optimal control of the state variables.
[0010] Furthermore, the method for establishing the cost function of the above model predictive control algorithm is specifically as follows:
[0011] Taking the permanent magnet motor as the control object, the discrete sub-current prediction model of the motor drive system in the rotating coordinate system is established, which is expressed as:
[0012]
[0013] Among them, i d (k+1), i q (k+1) represents the predicted value of the dq-axis stator current at the k+1 sampling time, i d (k), i q (k) represents the dq axis stator current sampling value at the current k sampling time, u d (k) and u q (k) represents the dq-axis stator voltage value at the current k sampling time, T s represents the sampling period, R s Indicates the stator winding resistance, L d , L q represents the stator inductance in the dq axis, ω e Expressed as the electrical angular velocity of the permanent magnet motor, ψ f Represents the flux linkage of a permanent magnet.
[0014] According to the current value flowing through the upper and lower capacitors of the DC bus, a discrete prediction model of the DC bus side capacitor is established to predict the voltage values of the upper and lower capacitors on the DC bus side; the calculation formula is:
[0015]
[0016] Among them, i c1 (k), i c2 (k) represents the current value of the upper and lower capacitors on the DC bus, v c1 (k) and v c2 (k) represents the voltage value of the upper and lower capacitors on the DC bus side at the current k sampling moment, 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 the k+1 sampling time, and C is the capacitance of the DC side capacitor.
[0017]
[0018] Among them, J i represents the cost function of tracking current error, J dc The cost function representing the midpoint voltage deviation value, S x (k-1) represents the normalized phase voltage at the previous sampling time, S x (k) represents the normalized phase voltage at the current sampling moment.
[0019] According to the established cost functions of tracking current error, midpoint voltage deviation and switching frequency tracking error, the total cost function is established as:
[0020] J=J i +λ dc J dc +λ sw J sw
[0021] Among them, J is the total cost function, λ dc , sw They represent the weight factors of neutral point potential balance and switching frequency regulation respectively; according to Simulink simulation, the corresponding current harmonic distortion THD and switching frequency f under different weight factor combinations are obtained sw , forming a data set.
[0022] Furthermore, the above step 2 is specifically as follows:
[0023] The data set is processed for outliers, and the outliers are represented as:
[0024] I=Q3-Q1
[0025] O l =Q3-1.5×I
[0026] O u =Q3+1.5×I
[0027] Among them, Q3, I and Q1 are the upper quartile, median and lower quartile respectively. u or less than O l When the outliers are cleaned, the data set will be divided into training set, cross-validation set and test set according to the proportion of 70%, 15% and 15%.
[0028] The training set and cross-validation set need to be normalized to improve the training speed of the residual network; the range change method is used to normalize the data and map the data to [0-1]. The process is expressed as:
[0029]
[0030] 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, By normalizing individual values.
[0031] The training set is used to train 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:
[0032]
[0033] Among them, RMSE is a quantitative indicator. is the predicted value of the neural network, y i is the true value in the data set, and m is the number of samples in the data set.
[0034] Furthermore, 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.
[0035] The input layer consists of model predictive control weight factors, namely midpoint voltage balance λ dc and 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:
[0036]
[0037] g [l] =max(0,Z j [l] ) j=1...N l
[0038] a [l] =g[l] (Z j [l] )
[0039] Among them, g [l] is the expression of the activation function ReLU, when Z j [l] The value reaches g [l] When the threshold is reached, the output of the neuron is [l] Will be passed to the next layer, N l is the number of neurons in layer l, j is the number of neurons in the previous layer, and w j [l] and b j [l] are the weights and biases of the fully connected layer.
[0040] The parallel network unit and the residual connection together constitute a residual parallel network. The residual connection enables the parallel network unit to skip the next layer and connect the interlayer to the accumulation layer.
[0041] According to the collected data set, the objective function of the residual network is defined as:
[0042] f ANN =α t ×THD+α f ×f sw
[0043] Where, THD is the total harmonic distortion of current, f sw is the switching frequency of the model predictive control, α t With α f are the weight coefficients of current harmonic distortion and switching frequency, α t The larger the value of is, the greater the weight of optimizing the total harmonic distortion of current in the objective function is. f The same is true with reducing the switching frequency.
[0044] According to the collected data, the weight factor is used as the input layer of the residual parallel network, and the current harmonic distortion is used as the network output layer. The Adam algorithm is combined for back-propagation training. The trained network is evaluated according to the RMSE. The hyperparameters of the parallel structure are adjusted according to the training results to select the optimal network.
[0045] Furthermore, the weight factor with a shorter step size is used as the input layer of the optimal network. The optimal network will give the corresponding objective function value, and the weight factor that minimizes the objective function is the optimal weight factor under the objective function.
[0046] The beneficial technical effects of the present invention are:
[0047] (1) The model predictive control method proposed in the present invention constructs a residual parallel neural network and performs offline training on the neural network so that it has the ability to predict the optimal weight factor. This method does not increase the computational burden of the model predictive control algorithm and 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, and there is no need to change the model predictive control algorithm itself.
[0048] (2) Compared with the traditional BP neural network, the present invention has higher accuracy. The residual connection avoids the problems of gradient vanishing and network degradation in the deep network. The parallel structure enables the residual network to have better model expression ability and higher prediction accuracy under the same parameter scale. At the same time, the parallel structure simplifies the difficulty of neural network adjustment.
[0049] (3) The invention is based on a data-driven optimization method and has good robustness to data differences. Therefore, the method is more suitable for model predictive control strategies under complex operating conditions, such as the traction motor model predictive control system. The complex operating conditions can be equivalent to several objective functions, thereby simplifying the complexity of the control system. BRIEF DESCRIPTION OF THE DRAWINGS
[0050] Figure 1 The model predictive control weight factor optimization flow chart of the present invention;
[0051] Figure 2 The block diagram of the permanent magnet motor drive system based on the three-level neutral point clamped inverter;
[0052] Figure 3 It is a simulation diagram of the model predictive control of the present invention;
[0053] Figure 4 It is a structural diagram of the residual parallel neural network of the present invention;
[0054] Figure 5 This is a diagram showing the effect of residual parallel neural network training of the present invention;
[0055] Figure 6 This is a diagram of the residual parallel neural network optimization results of the present invention. DETAILED DESCRIPTION
[0056] The present invention is further described in detail below with reference to the accompanying drawings and specific implementation methods.
[0057] The present invention takes a permanent magnet motor drive system fed by a three-level neutral point clamped inverter as an example and provides a model predictive control method based on residual neural network optimization, such as Figure 1 As shown, specifically:
[0058] S1. Establish a prediction model for system state variables, adjust the cost function of the model predictive control algorithm, and record the current harmonic distortion rate and switching frequency composition data samples under different weight factors.
[0059] The present invention first establishes a mathematical model of the permanent magnet motor drive system, such as Figure 2 As shown in FIG. 1 , a three-level diode neutral point clamped inverter permanent magnet motor drive system according to this embodiment is shown. In the inverter topology, each phase bridge arm includes four switching devices with the same switching frequency, wherein the switch tube T x1 and T x3 For complementary devices, T x2 and T x4 As complementary devices, the gate switching signals are mutually inverted. The output voltage corresponding to different switching states of the inverter topology is given in Table 1 below:
[0060] Table 1 Switching status of three-level inverter
[0061]
[0062] In the table, V dc is the voltage value provided by the DC bus side; the normalized vector S in the table x It can be further expressed as:
[0063] S x =[S a S b S c ]
[0064] In the formula, S a , S b , S c The value range of is (1, 0). Then, the mathematical model of permanent magnet motor is established:
[0065]
[0066] Among them, u d 、u q represents the stator voltage under dq axis, R s represents the stator winding resistance, i d 、i q Indicates the stator current in the dq axis, L d , L q represents the stator inductance in the dq axis, ω e Expressed as the permanent magnet motor electrical angle, ψ f Represents the permanent magnet flux. 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, which is expressed as:
[0067]
[0068] where i d (k+1), i q (k+1) represents the predicted value of the dq-axis stator current at the k+1 sampling time, i d (k), i q (k) represents the dq axis stator current sampling value at the current k sampling time, u d (k) and u q (k) represents the dq-axis stator voltage value at the current k sampling time, T s Indicates the sampling period.
[0069] In practical applications, due to the existence of calculation sampling delay, the delay needs to be compensated. By compensating the sampling delay and control delay in one step, the predicted value of the dq-axis stator current at the k+2 sampling time is obtained, which is expressed as:
[0070]
[0071] Among them, i d (k+2), i q (k+2) represents the predicted value of the dq-axis stator current at the k+1 sampling time, u d (k+1),u q (k+1) represents the dq-axis stator voltage value to be adopted at the k+1 sampling time.
[0072] According to the switching state and phase current of the inverter, the DC bus capacitor current is calculated and expressed as:
[0073]
[0074] Among them, i c1 (k), i c2 (k) represents the current value of the upper and lower capacitors on the DC bus, i dc (k) represents the current value flowing through the DC bus capacitor, S a1 , S b1 , S c1 , S a4 , S b4 , S c4 They 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, i a (k), i b (k), i c (k) respectively represent the output current value of the corresponding bridge arm.
[0075] According to the current value flowing through the upper and lower capacitors on the DC bus, a discretized prediction model of the capacitor on the DC bus side of the inverter is established to predict the voltage value of the upper and lower capacitors on the DC bus side; the calculation method is:
[0076]
[0077] Among them, v c1 (k) and v c2 (k) represents the voltage value of the upper and lower capacitors on the DC bus side at the current k sampling moment, 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 the k+1 sampling time, and C is the capacitance of the DC side capacitor.
[0078]
[0079] Among them, J i represents the cost function of tracking current error, J dc The cost function representing the midpoint voltage deviation value, J sw represents the cost function of switching frequency tracking error, S x (k-1) represents the normalized phase voltage at the previous sampling time, S x (k) represents the normalized phase voltage at the current sampling moment.
[0080] 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:
[0081] J=J i +λ dc J dc +λ sw J sw
[0082] Among them, J is the total cost function, λ dc , sw They represent the weight factors for neutral point potential balance and switching frequency regulation, λ dc and λ sw The value ranges of are 0-5 and 0-10, respectively, and the step length is 0.5. Figure 3 As shown, according to Simulink simulation, different weight factors (λ dc and λ sw ) and the corresponding total harmonic distortion (THD) of the current and the switching frequency (f sw ), and compose them into a simulation data set.
[0083] After the data set required by the residual parallel network is established, the present invention cleans and normalizes the outliers of the data set according to subsequent steps.
[0084] S2. Based on the collected data, an optimization objective function is established and the data set is divided into a training set, a validation set, and a test set for grid evaluation and training.
[0085] According to the collected data set, the objective function of the residual network is defined, which can be expressed as:
[0086] f ANN =α t ×THD+α f ×f sw
[0087] Where THD is the total harmonic distortion of current, f sw is the switching frequency of the model predictive control. t With α f are the weight coefficients of current harmonic distortion and switching frequency, α t The larger the value of , the greater the weight of optimizing the total harmonic distortion of current in the objective function. f The same is true for reducing the switching frequency. In this case, α t is 5, α f is 1. In addition, the data set is cleaned of outliers, and the outliers are represented as:
[0088] I=Q3-Q1
[0089] O l =Q3-1.5×I
[0090] O u =Q3+1.5×I
[0091] Where Q3, I and Q1 are the upper quartile, median and lower quartile respectively. u or less than O l 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 proportion of 70%, 15% and 15%. The training set and cross validation set need to be normalized to speed up the training of the residual network. The present invention adopts the extreme difference change method to normalize the data and maps the data to [0-1]. The process can be expressed as:
[0092]
[0093] 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, Through the normalized individual values. 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:
[0094]
[0095] Among them, RMSE is a quantitative indicator. is the predicted value of the neural network, y i is the true value in the data set, and m is the number of samples in the data set. After processing and normalizing the outliers in the data set, the present invention uses MATLAB to construct a residual parallel network according to the data distribution.
[0096] S3. Construct a residual parallel network, use the weight factor as the network input layer and the objective function as the output layer, train the network based on the data set and the Adam algorithm, adjust the residual module, network depth and other hyperparameters of the neural network according to the evaluation indicators, and select the residual network with the best indicators.
[0097] Residual network Figure 4 As shown, it mainly includes 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:
[0098] The input layer 1 consists of the model predictive control weight factors, which are midpoint voltage balance λ dc and 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
[0099]
[0100] g [l] =max(0,Z j [l] ) j=1...N l
[0101] a [l] =g [l] (Z j [l] )
[0102] Among them, g [l] is the expression of the activation function ReLU, when Z j [l] The value reaches g [l] When the threshold is reached, the output of the neuron is [l] Will be passed to the next layer, N l is the number of neurons in layer l, j is the number of neurons in the previous layer, and w j[l] and b j [l] are the weights and biases of the fully connected layer; in addition, the parallel structure enables the fully connected network to have better model expression capabilities under the same parameter scale.
[0103] The parallel network unit 6 and the residual connection 7 together form 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. This avoids the problem of gradient disappearance and network degradation during neural network training, and enables the deep network to have better training performance.
[0104] According to the network training set, back propagation training is performed in combination with the Adam algorithm. The pseudo code of the Adam algorithm is shown in Table 2. j [l] and b j [l] , according to the RMSE indicator, 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 network, without adjusting the number of neurons in the hidden layer, thereby simplifying the adjustment process of the neural network. After adjustment, the optimal structure of the residual parallel network consists of two parallel network units residually connected, and the parallel network unit is connected by four fully connected networks and an accumulation layer 5. In addition, the input layer of the residual parallel network is a weight factor (λ dc and λ sw ), the output layer 8 is the optimization objective function f ANN The training performance of the residual parallel network is similar to the RMSE of the traditional BP network training. Figure 5 shown.
[0105] Table 2 Adam back propagation algorithm
[0106]
[0107] 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. The predicted value of the optimal network is inversely normalized to obtain the true value of the objective function, and then find the weight factor that minimizes the objective function to achieve optimal control of the state variables.
[0108] Among them, the weight factor range remains unchanged, the step size is 0.01, and the predicted f ANNFind the minimum value in , and its corresponding weight factor is the optimal weight factor under the objective function. Substitute the predicted weight factor into the cost function of model predictive control, and then Figure 6 It can be seen that compared with the traditional BP network, the residual parallel network can accurately find the optimal weight factor corresponding to the objective function, thereby improving the control performance of the model predictive control and is suitable for complex working conditions of train traction.
Claims
1. A model predictive control method based on residual neural network optimization, characterized in that: The following steps are involved: Step 1: Establish a prediction model for system state variables, adjust the cost function of the model predictive control algorithm, and record data samples of current harmonic distortion rate and switching frequency composition under different weight factors; 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 of the motor drive system in the rotating coordinate system is established, which is expressed as: Among them, i d (k+1), i q (k+1) represents the predicted value of the dq-axis stator current at the k+1 sampling time, i d (k), i q (k) represents the dq axis stator current sampling value at the current k sampling time, u d (k),u q (k) represents the dq-axis stator voltage value at the current k sampling time, T s represents the sampling period, R s Indicates the stator winding resistance, L d , L q represents the stator inductance in the dq axis, ω e Expressed as the electrical angular velocity of the permanent magnet motor, ψ f represents the flux linkage of permanent magnet; According to the current value flowing through the upper and lower capacitors of the DC bus, a discrete prediction model of the DC bus side capacitor is established to predict the voltage values of the upper and lower capacitors on the DC bus side; the calculation formula is: Among them, i c1 (k), i c2 (k) represents the current value of the upper and lower capacitors on the DC bus, v c1 (k) and v c2 (k) represents the voltage value of the upper and lower capacitors on the DC bus side at the current k sampling moment, v c1 (k+1) and v c2 (k+1) represents the predicted values of the upper and lower capacitor voltages on the DC bus side at the k+1 sampling time, respectively, and C is the capacitance of the DC side capacitor; in, and They 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 tracking current error, J dc The cost function representing the midpoint voltage deviation value, J sw represents the cost function of switching frequency tracking error, S x (k-1) represents the normalized phase voltage at the previous sampling time, 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 established as: J=J i +λ dc I dc +λ sw I sw Among them, J is the total cost function, λ dc , sw They represent the weight factors of neutral point potential balance and switching frequency regulation respectively; according to Simulink simulation, the corresponding current harmonic distortion THD and switching frequency f under different weight factor combinations are obtained sw , forming them into a data set; Step 2: Based on the collected data, an optimization objective function is established, and the data set is divided into a training set, a validation set, and a test set for network evaluation and training; Step 3: Construct a residual parallel neural network, use the weight factor as the network input layer, and the objective function as the network output layer. Train the network based on the data set and the Adam algorithm. Adjust the residual module, network depth and other hyperparameters of the neural network according to the evaluation index, and select the network with the best index. The constructed residual parallel neural network includes an input layer (1), a fully connected layer (2), an activation function layer (3), a fully connected network (4), an accumulation layer (5), a parallel network unit (6), a residual connection (7), and an output layer (8); The input layer (1) consists of the model predictive control weight factors, which are midpoint voltage balance λ dc and switching frequency limit λ sw ; The fully connected layer (2) and the activation function ReLU (3) 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 Z j [l] The value reaches g [l] When the threshold is reached, the output of the neuron is [l] Will be passed to the next layer, N l is the number of neurons in layer l, j is the number of neurons in the previous layer, and 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) enables the parallel network unit (6) to skip the next layer, and the interlayer is connected to the accumulation layer (5). According to the collected data set, the objective function of the residual network is defined as: f ANN =a t ×THD+α f ×f sw Where, THD is the total harmonic distortion of current, f sw is the switching frequency of the model predictive control, α t With α f are the weight coefficients of current harmonic distortion and switching frequency, α t The larger the value of is, the greater the weight of optimizing the total harmonic distortion of current in the objective function is. f The same is true with reducing the switching frequency; According to the collected data, the weight factor is used as the input layer of the residual parallel network, and the current harmonic distortion is used as the network output layer. Back propagation training is performed in combination with the Adam algorithm. The trained network is evaluated based on RMSE, and the hyperparameters of the parallel structure are adjusted according to the training results to select the optimal network. Step 4: Use the weight factor with a 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 achieve optimal control of the state variables.
2. A model predictive control method based on residual neural network optimization according to claim 1, characterized in that: The step 2 is specifically as follows: The data set is processed for outliers, and the outliers are represented as: I=Q3-Q1 O l =Q3-1.5×I O u =Q3+1.5×I Among them, Q3, I and Q1 are the upper quartile, median and lower quartile respectively. u or less than O l When the outliers are cleaned, the data set will be divided into training set, cross-validation set and test set according to the proportion of 70%, 15% and 15% respectively. The training set and cross-validation set need to be normalized to improve the training speed of the residual network; the range change method is used to normalize the data and map the data 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, By normalizing individual values; The training set is used to train 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: Among them, RMSE is a quantitative indicator. is the predicted value of the neural network, y i is the true value in the data set, and m is the number of samples in the data set.
3. The model predictive control method based on residual neural network optimization according to claim 1 is characterized in that: The weight factor with a shorter step size is used as the input layer of the optimal network. The optimal network will give the corresponding objective function value. The weight factor that minimizes the objective function is the optimal weight factor under the objective function, thus achieving optimal control of the state variables.
