Converter model prediction control method based on deep neural network

Abstract

The invention discloses a converter model predictive control method based on a deep neural network. The converter model predictive control method comprises the following steps that a neural network control model is adopted to control operation of a converter; analyzing the load type and the control performance of the converter by adopting a deep neural network; and when it is identified that the load type changes or the control performance is reduced, the neural network control model is switched according to rules. The deep neural network is adopted to analyze the running state and the load type of the converter, the neural network control model switching rule is designed, updating of the neural network control model is achieved, and the designed method has the advantage of rapidly and accurately achieving updating of the neural network control model.

Description

Technical Field

[0001] The present invention relates to a converter model predictive control method based on a deep neural network. Background Art

[0002] The converter model predictive control algorithm is a converter optimization control method. The converter finite control set model predictive control algorithm (FCSMPC) is a hot spot in the current converter model predictive control algorithm research and application. For the FCSMPC algorithm, multi-step prediction, error compensation and weight coefficient optimization are the main strategies to further improve the algorithm control performance. Regardless of the strategy adopted, the online calculation amount of the algorithm will be further increased, which has become a key factor restricting the improvement of the FCSMPC algorithm control performance.

[0003] Neural networks have the natural advantage of fitting nonlinear characteristics, so neural networks can be used to fit the control characteristics of the FCSMPC algorithm, and further a neural network-based converter model predictive control algorithm can be formed. In this algorithm, the FCSMPC algorithm with optimal control performance is used to obtain a data set for training the neural network, and a neural network control model that characterizes the control performance of the improved FCSMPC algorithm is obtained through offline training, and then the online control of the neural network control model is used to achieve optimal control of the converter. This algorithm does not need to consider the online calculation amount of the FCSMPC algorithm and its improved strategy, but realizes online calculation through neural networks, so it is a feasible solution to improve the control performance of the FCSMPC algorithm.

[0004] In the neural network-based converter model predictive control algorithm, the control performance and application scope of the algorithm are completely dependent on the data set used to train the neural network. In the actual operation of the converter, the load and parameters of the converter will change, so timely updating of the neural network control model is the key to achieving continuous optimization control. In the process of updating the neural network control model, based on the current operating status of the converter, it is necessary to predict the need for updating the control model in advance to avoid the adaptability and reliability problems of the control algorithm caused by the lag in the control model update. This solution uses a deep neural network to analyze the operating status and load type of the converter, designs the switching rules of the neural network control model, and realizes the update of the neural network control model. The designed method has the advantages of quickly and accurately updating the neural network control model. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a converter model predictive control method which adopts a deep neural network to analyze the converter operating status and load type, and implements the update of the neural network control model based on rules.

[0006] The technical solution of the present invention to solve the above problem is: a converter model predictive control method based on a deep neural network, comprising the following steps: S1. Using neural network control model to control converter operation; S2. Use deep neural network to analyze the load type and control performance of the converter; S3. When a change in load type or a decrease in control performance is identified, the neural network control model is switched according to the rules.

[0007] The effect of the present invention is that: for the updating problem of the neural network control model in the neural network-based converter model predictive control algorithm, compared with the traditional method, the present invention provides a converter model predictive control method that quickly and accurately updates the neural network control model. The advantages of this method are: using a deep neural network to analyze the operating status and load type of the converter, updating the neural network control model based on rules, improving the adaptability and reliability of the neural network-based converter model predictive control algorithm, solving the problem of switching the neural network control model, and reducing the risk of mismatching the neural network control model. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 Flow chart of the control method in the present invention.

[0009] Figure 2 Structure diagram of the neural network-based converter model predictive control algorithm in the present invention.

[0010] Figure 3 The structure diagram of the converter model predictive control based on deep neural network in the present invention.

[0011] Figure 4 Switching strategy diagram of the converter control model in the present invention.

[0012] Figure 5 The switching rule diagram of the neural network control model in the present invention. DETAILED DESCRIPTION

[0013] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0014] The present invention provides a converter model predictive control method based on a deep neural network, and its flow chart is as follows: Figure 1 As shown, it includes the following steps: S1. Use neural network control model to control converter operation.

[0015] The use of a neural network control model to control the operation of the converter refers to a converter control method using a neural network offline training + online control, such as Figure 2 As shown. In this method, the BP neural network is used to fit the nonlinear relationship between the system state quantity and the control quantity output by the control algorithm in the converter model predictive control algorithm, and based on the trained BP neural network, the corresponding control quantity is calculated from the sampled value of the system state to realize the online control of the converter. The design of this method includes two stages: offline training of the neural network and online control. In the offline training stage, based on the designed converter model predictive control algorithm, sample data of the converter system state quantity and the algorithm output control quantity are obtained through simulation, and the constructed BP neural network is trained with the sample data. The converter model predictive control algorithm designed in the offline training stage is a comprehensive optimization algorithm based on the converter finite control set model predictive control algorithm (Finite Control Set Model PredictiveControl, FCSMPC), and different model predictive control algorithms can be constructed according to different load types and comprehensive optimization strategies, and there is no need to consider the complexity of the algorithm and the real-time calculation problem. The external characteristics of the designed control algorithm are completely fitted by the BP neural network with the same network structure, and finally a neural network control model set with the same network structure and different parameters is obtained. In the neural network online control stage, the corresponding BP neural network control model parameters are selected according to the load type and control performance of the system, and the control quantity calculated by the neural network based on the sampling value of the system state acts on the converter system to achieve real-time control of the converter. Therefore, the complexity of the model predictive control algorithm designed in the offline training stage will not affect the real-time calculation amount of the neural network control model, that is, the online calculation amount of the neural network control model is fixed and is not affected by the complexity of the designed model predictive control algorithm, which can realize the decoupling of the converter model predictive control algorithm design and algorithm calculation amount.

[0016] S2. Use deep neural network to analyze the load type and control performance of the converter.

[0017] The use of a deep neural network to analyze the load type and control performance of the converter means that during the operation of the converter controlled by the neural network control model, there is uncertainty in the converter load type, and the control performance of the system will also decline as the system continues to operate. Therefore, it is necessary to analyze the load type, predict the downward trend of the control performance, and switch the neural network control model according to the results of the analysis and prediction, so as to improve the adaptability of the neural network control system. The deep neural network used includes a convolutional neural network and a long short-term memory neural network, wherein the convolutional neural network is used to analyze the load type of the converter, and the long short-term memory neural network is used to predict the control performance of the system. The use of a deep neural network to analyze the load type and control performance of the converter, the process is to first obtain the state quantity data of the converter system operation, and use a deep neural network on the host computer to analyze the obtained system state quantity data, and finally feed the analysis results back to the converter neural network control system. Figure 3 Shown is the structure diagram of the converter model predictive control based on deep neural network.

[0018] S3. When a change in load type or a decrease in control performance is identified, the neural network control model is switched according to the rules.

[0019] The load type change or control performance degradation refers to the load type change and control performance degradation trend analyzed by the host computer deep neural network. The load types identified by the host computer deep neural network include: no-load, resistive-inductive load and nonlinear load. The control performance degradation refers to the degradation of the waveform quality of the converter output voltage. The identification of the load type change refers to the load type analyzed by the host computer deep neural network for multiple consecutive times and the load type corresponding to the current neural network control model are inconsistent. The identification of the control performance degradation refers to the mean square value of the deviation between the converter output voltage predicted by the host computer deep neural network and the reference voltage is higher than a certain value. The neural network controller (NNC) is switched according to the rules. In the constructed neural network control model set, the neural network control models include: NNC (no load-classic FCSMPC), NNC (resistive inductive load-compensated FCSMPC), NNC (resistive inductive load-optimal FCSMPC), NNC (non-linear load-compensated FCSMPC), NNC (non-linear load-optimal FCSMPC), and the corresponding load types include: no load, resistive inductive load and non-linear load. The control strategies involved include: classic FCSMPC algorithm (classic FCSMPC), FCSMPC algorithm with prediction error compensation (compensated FCSMPC), and FCSMPC algorithm with comprehensive prediction error compensation and multi-step prediction strategy (optimal FCSMPC). The switching neural network control model according to the rule is as follows: the classic FCSMPC algorithm is used to start the converter system operation, and when the load change is identified, the compensation FCSMPC neural network control model of the corresponding load is switched to, when the system control performance is identified to be reduced, the optimal FCSMPC neural network control model of the corresponding load is switched to, if the system control performance is further identified to be reduced, the classic FCSMPC algorithm is switched to control the converter operation, and the converter control model switching strategy is as follows Figure 4 The switching rules for switching the neural network control model according to the rules also include: when the converter system is under the control of the compensation FCSMPC or the optimal FCSMPC neural network control model, when a load change is identified, switching to the neural network control model with the same control strategy for the corresponding load, and when the identified load type is no-load, uniformly switching to the no-load-classic FCSMPC neural network control model, Figure 5 Shown are the switching rules of the neural network control model.

[0020] The neural network control model is switched according to the rules, and the switching rules also include: to avoid frequent switching of the neural network control model, the load type change refers to the continuous identification of load type changes, and the control performance degradation refers to the continuous identification of system control performance degradation.

Claims

1. A converter model predictive control method based on deep neural network, characterized in that: The following steps are involved: S1. Using neural network control model to control converter operation; The use of a neural network control model to control the operation of the converter refers to a converter control method using a neural network offline training + online control; in this method, a BP neural network is used to fit the nonlinear relationship between the system state quantity and the control quantity output by the control algorithm in the converter model prediction control algorithm, and based on the trained BP neural network, the corresponding control quantity is calculated from the sampled value of the system state quantity to achieve online control of the converter; the design of this method includes two stages: offline training of the neural network and online control; In the offline training stage, based on the designed converter model predictive control algorithm, sample data of converter system state quantity and algorithm output control quantity are obtained through simulation, and the constructed BP neural network is trained with sample data; the converter model predictive control algorithm designed in the offline training stage is a kind of converter finite control set model predictive control algorithm (Finite Control Set Model Predictive Control, FCSMPC) and can build different model predictive control algorithms according to different load types and comprehensive optimization strategies, without considering the complexity of the algorithm and the real-time calculation problem. The external characteristics of the designed control algorithm are completely fitted by the BP neural network with the same network structure, and finally a set of neural network control models with the same network structure and different parameters are obtained; in the neural network online control stage, the corresponding BP neural network control model parameters are selected according to the load type and control performance of the system, and the control quantity calculated by the neural network based on the sampling value of the system state acts on the converter system to realize the real-time control of the converter; therefore, the complexity of the model predictive control algorithm designed in the offline training stage will not affect the real-time calculation amount of the neural network control model, that is, the online calculation amount of the neural network control model is fixed and is not affected by the complexity of the designed model predictive control algorithm, so the decoupling of the converter model predictive control algorithm design and the algorithm calculation amount can be realized; S2. Use deep neural network to analyze the load type and control performance of the converter; The use of a deep neural network to analyze the load type and control performance of the converter means that during the operation of the converter controlled by the neural network control model, there is uncertainty in the converter load type, and the control performance of the system will also decline with the continuous operation of the system. Therefore, it is necessary to analyze the load type, predict the downward trend of the control performance, and switch the neural network control model according to the results of the analysis and prediction to improve the adaptability of the neural network control system; the deep neural network used includes a convolutional neural network and a long short-term memory neural network, wherein the convolutional neural network is used to analyze the load type of the converter, and the long short-term memory neural network is used to predict the control performance of the system; the use of a deep neural network to analyze the load type and control performance of the converter, the process is to first obtain the state quantity data of the converter system operation, and use a deep neural network on the host computer to analyze the obtained system state quantity data, and finally feed the analysis results back to the converter neural network control system; S3. When a change in load type or a decrease in control performance is identified, the neural network control model is switched according to the rules.

2. The method for predictive control of a converter model based on a deep neural network according to claim 1, characterized in that: In the step S3, the load type change or control performance degradation refers to the load type change and control performance degradation trend analyzed by the host computer deep neural network. The load types identified by the host computer deep neural network include: no-load, resistive-inductive load and nonlinear load. The control performance degradation refers to the degradation of the waveform quality of the converter output voltage; the identification of the load type change refers to the inconsistency between the load type analyzed by the host computer deep neural network for multiple consecutive times and the load type corresponding to the current neural network control model; the identification of the control performance degradation refers to the mean square value of the deviation between the converter output voltage and the reference voltage predicted by the host computer deep neural network is higher than a certain value; the switching of the neural network control model (Neural Network Controller, NNC), in the constructed neural network control model set, the neural network control models include: NNC (no load-classic FCSMPC), NNC (resistive inductive load-compensated FCSMPC), NNC (resistive inductive load-optimal FCSMPC), NNC (non-linear load-compensated FCSMPC), NNC (non-linear load-optimal FCSMPC), the corresponding load types include: no load, resistive inductive load and non-linear load, the control strategies involved include: classic FCSMPC algorithm (classic FCSMPC), FCSMPC algorithm with prediction error compensation (compensated FCSMPC), FCSMPC algorithm with comprehensive prediction error compensation and multi-step prediction strategy (optimal FCSMPC); the switching rule of the neural network control model according to the rule is: using the classic FCSMPC algorithm to start the converter system operation, when the load change is identified, switch to the compensation FCS of the corresponding load MPC neural network control model, when it is identified that the system control performance has decreased, it switches to the optimal FCSMPC neural network control model of the corresponding load, and if it is further identified that the system control performance has decreased, it switches to the classic FCSMPC algorithm to control the operation of the converter; the switching of the neural network control model according to the rules also includes: when the converter system is under the control of the compensation FCSMPC or the optimal FCSMPC neural network control model, when it is identified that the load has changed, it switches to the neural network control model with the same control strategy of the corresponding load, and when the identified load type is no-load, it is uniformly switched to the no-load-classical FCSMPC neural network control model; the switching of the neural network control model according to the rules also includes: in order to avoid frequent switching of the neural network control model, the load type change refers to the continuous identification of the load type change for multiple times, and the control performance degradation refers to the continuous judgment of the system control performance degradation for multiple times.

Images