A three-phase two-level inverter control optimization system based on ensemble learning

Abstract

This invention relates to the technical field of inverter control and discloses a three-phase two-level inverter control optimization system based on ensemble learning. The system includes multiple inverters that communicate with a remote monitoring and optimization workstation via a remote data communication unit. The remote monitoring and optimization workstation is used to acquire the operating data sequence of each inverter online. First, it uses an ensemble learning method to calculate the pseudo-partial derivative matrix of the dynamic linear model (DL), thereby optimizing the finite set model predictive control algorithm (FCS-DLMPC). Then, it uses the optimized FCS-DLMPC algorithm to control the inverters.

Description

Technical Field

[0001] This invention belongs to the technical field of inverter control, specifically relating to a three-phase two-level inverter control optimization system based on ensemble learning. Background Technology

[0002] Finite Control Set Model Predictive Control (FCS-MPC) has gained increasing attention due to its multi-objective constraint optimization capabilities, excellent dynamic performance, and thorough consideration of the discrete characteristics and finite switching states of power electronic converters. In practical applications, the actual values ​​of parameters such as inductance inevitably deviate from their nominal values ​​due to factors such as temperature drift and aging. FCS-MPC is a control method based on predictive models. When there is a mismatch between the parameter values ​​in the model and the actual values, it can lead to the accumulation of current prediction errors, increased steady-state errors, and deterioration of total electrical discharge temperature (THD). If the parameter mismatch exceeds a certain level, it can even cause system instability. Summary of the Invention

[0003] To address the aforementioned technical challenges, this invention proposes an inverter control optimization system based on ensemble learning (EL), which combines dynamic linearization (DL) technology with finite control set dynamic linearization model predictive control (FCS-DLMPC). This system can collect inverter operating data online, optimize and update inverter control parameters using ensemble learning, thereby resolving the model dependency problem of traditional finite control set model predictive control (FCS-MPC) and ensuring long-term inverter performance.

[0004] To achieve the above objectives, the present invention provides the following technical solution: A three-phase two-level inverter control optimization system based on ensemble learning includes multiple inverters that communicate with a remote monitoring and optimization workstation via a remote data communication unit. The remote monitoring and optimization workstation is used to acquire the operating data sequence of each inverter online. First, it uses the ensemble learning method to calculate the pseudo-partial derivative matrix of the dynamic linear model DL, and then optimizes the finite set model predictive control algorithm FCS-DLMPC. Finally, it uses the optimized finite set model predictive control algorithm FCS-DLMPC to control the inverter.

[0005] Furthermore, the current output voltage increment vector of the inverter is used as the input and the output current change is used as the output to train multiple base models. Then, the training results of all base models are combined using simple averaging and weighted averaging respectively, and the combination result with the minimum global loss is selected as the final training result output.

[0006] Furthermore, when using weighted averages for combination, the global loss of each base model on the entire dataset is first calculated using the following formula. , Then, the weighted average combination result is calculated using the following formula. , Where M represents the total number of base models, This represents the training result of the m-th base model, and there are a total of M base models; Calculate the combined result after simple averaging using the following formula. , .

[0007] Furthermore, the global loss is calculated for the combined results obtained by the two ensemble strategies, simple averaging and weighted averaging, respectively. The combined result with the smaller global loss is selected as the final training result for subsequent parameter optimization.

[0008] Furthermore, five base models are set, all of which are set as linear mapping models.

[0009] Furthermore, the remote data transmission unit includes a data transmission unit and a cloud server. The data transmission unit is responsible for realizing the bidirectional conversion between serial communication and wireless communication between each inverter and the cloud server. The cloud server serves as a relay station for remote data communication, responsible for bidirectional data interaction between the data transmission unit and the remote monitoring and optimization workstation.

[0010] Compared with the prior art, the beneficial effects of the present invention are: 1. Constructing an intelligent control closed loop: By providing a three-phase two-level inverter control system solution that integrates data acquisition, parameter optimization, and online updates, the inverter has the ability to continuously evolve through "operation-monitoring-optimization," thereby ensuring its output performance and reliability in the long term while significantly reducing maintenance costs.

[0011] 2. Overcoming the bottleneck of model dependence: The designed FCS-DLMPC method uses dynamic linearization technology to transform the control problem into online optimization based on input and output data, which fundamentally gets rid of the dependence of traditional FCS-MPC on precise mathematical models and system parameters, and greatly enhances the adaptability and robustness of the control system to changes in operating conditions and non-ideal factors.

[0012] 3. Precisely addressing the challenge of performance degradation: The proposed parameter optimization method based on ensemble learning can effectively mine complex coupling relationships from actual operating data, intelligently calibrate parameter mismatch problems caused by device aging, temperature drift, etc., accurately suppress the resulting decrease in control accuracy and harmonic degradation, and significantly extend the life cycle of the inverter for efficient operation. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the overall framework of the present invention. Figure 1 ; Figure 2 This is a schematic diagram of the topology of the three-phase two-level inverter of the present invention; Figure 3 This is a schematic diagram of the control flow of the existing FCS-DLMPC technology; Figure 4 This is a schematic diagram of the communication process between the inverter and the workstation of the present invention; Figure 5 This is a schematic diagram of the inverter control parameter optimization framework based on ensemble learning of the present invention. Detailed Implementation

[0014] To make the technical means, creative features, objectives and effects of this invention easier to understand, the following embodiments, in conjunction with the accompanying drawings, specifically illustrate the integrated learning-based three-phase two-level inverter control optimization system of this invention. It should be noted that the description of these embodiments is for the purpose of helping to understand this invention, but does not constitute a limitation of this invention.

[0015] When traditional Finite Control Set Model Predictive Control (FCS-MPC) is applied to a three-phase two-level inverter system, performance inevitably suffers from degradation due to system parameter mismatch and drift issues. To address this problem, this invention provides a three-phase two-level inverter control optimization system based on ensemble learning, the system structure of which is shown in the attached diagram. Figure 1 As shown.

[0016] The three-phase two-level inverter control optimization system based on ensemble learning mainly consists of three parts: a three-phase two-level inverter, a remote data transmission unit, and a remote monitoring and optimization workstation. The specific functions of each part are described below: 1. Three-phase two-level inverter Three-phase two-level inverters are among the most common power electronic converters in industry. They are simple and reliable in structure, converting direct current (DC) into three-phase alternating current (AC) with adjustable voltage and frequency. They are the core components of equipment such as frequency converters, wind power grid-connected inverters, photovoltaic grid-connected inverters, and ocean energy grid-connected inverters. The topology of a three-phase two-level inverter is shown in the attached figure. Figure 2 As shown, where, It is the DC bus voltage. It is a DC bus capacitor; each bridge arm has two switching devices with complementary control signals. It is the output inductor, It is the load resistor.

[0017] The inverter is the optimization target of this system, specifically the optimizer control parameters. The control method is the core of the inverter, directly determining the quality of the output power (harmonic level) and the dynamic response speed of the system. FCS-MPC, with its intuitive algorithm structure, excellent dynamic response speed, and superior multi-objective control capabilities, has attracted much attention in the field of power electronic converter control method research. However, model dependence is an inherent problem of FCS-MPC, and parameter drift or time drift is unavoidable in practical applications, significantly limiting the performance of this method. To address this issue, Dynamic Linearization (DL) technology is combined with FCS-MPC to implement Finite Control Set Dynamic Linearization Model Predictive Control (FCS-DLMPC). The DL model is constructed only based on the system's input and output data, effectively eliminating the model dependence problem of FCS-MPC. The control flowchart of FCS-DLMPC is attached. Figure 3 As shown.

[0018] Dynamic linearization is a model reduction method within the theoretical framework of Model-Free Adaptive Control (MFAC). It includes three equivalent models: Compact Form Dynamic Linearization (CFDL), Partial Form Dynamic Linearization (PFDL), and Full Form Dynamic Linearization (FFDL). Among these, PFDL has strong model expressive power and moderate computational complexity. Therefore, this invention adopts the following PFDL model as the prediction model: (1) In equation (1), This represents the system state, specifically the change in output current. It is the system output, indicating the difference between the sampled values ​​at the current and next time points. It is sampled once per sampling period. , ; L represents the depth of the historical data involved in the calculation, where L>0; The output voltage increment vector represents the system state and is the system input. , Each output voltage increment represents the difference between the sampled value at the current moment and the next moment, and is sampled once per sampling period; This is a pseudo-bias matrix, also known as the weight matrix, used to describe the relationship between the system output and the system control input, as detailed below: (2) in, This is the most critical parameter in the FCS-DLMPC method, designed to describe the relationship between system output and system control input, and determines the final effect of FCS-DLMPC. Optimization using appropriate methods is crucial. This can directly help improve the performance of three-phase two-level inverters based on the FCS-DLMPC strategy. A detailed explanation of how to optimize the weight matrix using ensemble learning methods will follow. .

[0019] 2. Remote data communication unit The remote data communication unit consists of a data transfer unit and a cloud server. It is responsible for establishing the remote data interaction link in the three-phase two-level inverter control and optimization system. The data transfer unit (DTU) is responsible for enabling bidirectional serial and wireless communication between the three-phase two-level inverter and the cloud server. The cloud server (CS) acts as a relay station for remote data communication, responsible for bidirectional data interaction between the data transfer unit and the remote monitoring and optimization workstation. A flowchart of the communication process between the inverter and the workstation is attached. Figure 4 As shown.

[0020] In practical applications, digital signal processors (DSPs) manufactured by Texas Instruments (TI) are often used as inverter controllers. Considering the communication interface resources of the DSP and the need for remote wireless communication, the DTU typically implements conversion between serial communication (RS485, etc.) and wireless communication (4G, etc.). The core advantage of the Message Queuing Telemetry Transport (MQTT) protocol lies in its lightweight, efficient, and reliable communication architecture specifically designed for IoT scenarios. Therefore, MQTT cloud servers are a perfect fit for three-phase two-level inverter control optimization systems, easily enabling one-to-many three-phase two-level inverter control optimization systems.

[0021] 3. Remote monitoring and workstation optimization The remote monitoring and optimization workstation is a crucial component of the three-phase two-level inverter control optimization system, representing the point where artificial intelligence technology (ensemble learning) is applied to inverter control optimization. Formally, the remote monitoring and optimization workstation is system software running on the user's computer. It can acquire inverter operating data sequences online, utilize ensemble learning methods to optimize the inverter's control (predictive model) parameters based on these sequences, and update them online. It mainly includes the following interfaces and functions: ① Main Interface: The software will analyze and display the real-time status data transmitted back by the inverter on the main interface, allowing users to observe the real-time values ​​of key statuses such as phase current and line voltage.

[0022] ② Curve Viewing Interface: All returned numerical statuses can be plotted as curves in the curve viewing interface, allowing users to perceive the changing trends of each state of the inverter through the numerical curves.

[0023] ③ Control Optimization Interface: The functions of the control optimization interface include reading the current inverter control (predictive model) parameters, obtaining inverter operating data, optimizing control (predictive model) parameters, and updating them. Specific descriptions of each function are as follows: Read current control (predictive model) parameters: Users can read the current control (predictive model) parameters by pressing a button; these parameters will be used in the optimized calculation of the controller to reflect the performance difference before and after the optimization of the control parameters, and provide guidance for users to update the control parameters of the inverter.

[0024] Obtaining inverter operating data: Users can click the button at any time to obtain a batch of inverter operating data; the inverter operating data obtained here is used to optimize the control (predictive model) parameters, and its format and content are closely related to the DL predictive model shown in equation (1); Optimize controller (predictive model) parameters: After obtaining the inverter operating data and the current control (predictive model) parameters, the user can click this button to optimize the inverter control through the ensemble learning method, i.e., in equation (1). The merits of the control parameters before and after optimization can be judged by the calculation results in the interface.

[0025] Update control (predictive model) parameters: After optimizing the control parameters, if the optimized control (predictive model) parameters perform better, the user can click this button to update the control parameters.

[0026] The core advantage of artificial intelligence technology in inverter control parameter optimization lies in its strong model approximation capability. Methods such as neural networks can directly approximate the inherent dynamic characteristics of an inverter using only input and output data, thereby locking in the optimal control parameters without relying on a precise physical model. This significantly improves the system's dynamic response and steady-state accuracy. This invention proposes an inverter control parameter optimization method based on ensemble learning. The framework of the designed ensemble learning-based inverter control parameter optimization method is attached. Figure 5 As shown, its core idea is: based on the input and output data collected during the actual operation of the inverter, an ensemble learning strategy is used to study the pseudo-partial derivative matrix. Tuning and optimization are performed to accurately describe the precise relationship between the system output and input, enhancing the FCS-DLMPC controller's ability to resist parameter mismatch and ensuring its control performance during application. After obtaining the operating data of the three-phase two-level inverter system, the process for optimizing the inverter control parameters based on this parameter optimization method is as follows: (1) Data preprocessing Raw data collected at each sampling time ,in, and Let be the output current at the current time (sampling period) and the next time (sampling period), respectively. Then, the change in output current is: (3) Standard training sample pairs are constructed as follows: (4) The input and output are Z-score standardized separately, and then divided into training sets in an 8:2 ratio. and verification set .

[0027] (2) Base model training Five base models were set for ensemble learning, all of which are linear mapping models. For a single base model, the optimization objective is to minimize the mean squared error on the training set. (5) The initial weight matrix is ​​generated using the He initialization strategy. Then, the Adam optimizer is used for mini-batch gradient descent iterations until convergence. For batch... The gradient is calculated as follows: (6) Adam adaptively updates the learning rate of each parameter based on the gradient, balancing convergence speed and stability.

[0028] An early stopping mechanism based on the validation set is introduced during training: the validation set loss is calculated after each iteration. If continuous Wheel (default) If the validation set loss does not improve (e.g., =500), then terminate training and return the weight matrix with the minimum validation set loss. This is to suppress overfitting.

[0029] (3) Integrated learning strategy To enhance the robustness and generalization ability of the optimization results, an ensemble learning strategy is adopted: Step 1: Multi-base model training use (default Five different random seeds, i.e., different initial weight matrices, are used, each corresponding to one of the five base models. The base model training process is repeated to obtain the base model set. The different base models differ due to variations in initialization and data shuffling order.

[0030] Step 2: Base Model Evaluation Calculate the training results of each base model. Different weight matrices ) in the entire dataset Global loss : (7) Step 3: Integration and optimization of dual strategies.

[0031] Two integration strategies are adopted: Simple average: (8) Loss-weighted average: (9) in, .

[0032] Then, the combined results of the two integration strategies mentioned above are calculated using formula (7). , The global loss is used to select the result with the smaller loss as the final training output. : (10) It is important to note that the schemes and arrangements of this application shown in the exemplary embodiments are merely exemplary. Although only a few embodiments are described in detail in this disclosure, those who consult this disclosure will readily understand that many modifications are possible (e.g., variations in various parameter values ​​(temperature, power, humidity, etc.), installation arrangements, names, colors, logical orders, etc.) without substantially departing from the novel teachings and advantages of the subject matter described in this application. Therefore, all such modifications are also included within the scope of the invention, and the order or sequence of any process or method steps may be changed or rearranged according to alternative embodiments. In the claims, any "apparatus plus function" clause is intended to cover the structure described herein for performing the function, and not only structurally equivalent but also equivalent in structure. Other substitutions, modifications, alterations, and omissions may be made in the design, operation, and arrangement of the exemplary embodiments without departing from the scope of the invention. Therefore, the invention is not limited to the particular embodiments but extends to a variety of modifications that still fall within the scope of the appended claims.

[0033] Furthermore, in order to provide a concise description of exemplary embodiments, not all features of actual embodiments (i.e., those features that are not relevant to the best mode of carrying out the invention as currently considered, or those features that are not relevant to implementing the invention) may be omitted.

[0034] It should be understood that numerous specific implementation decisions can be made during the development of any practical implementation, such as in any engineering or design project. Such development efforts may be complex and time-consuming, but for those skilled in the art who benefit from this disclosure, the development effort will be a routine work of design, manufacturing, and production without requiring much experimentation.

[0035] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An ensemble learning based three-phase two-level inverter control optimization system, characterized in that: It includes multiple inverters that communicate with a remote monitoring and optimization workstation via a remote data communication unit. The remote monitoring and optimization workstation is used to acquire the operating data sequence of each inverter online. First, it uses the ensemble learning method to calculate the pseudo-partial derivative matrix of the dynamic linear model DL, and then optimizes the finite set model predictive control algorithm FCS-DLMPC. Finally, it uses the optimized finite set model predictive control algorithm FCS-DLMPC to control the inverter.

2. The ensemble learning based three-phase two-level inverter control optimization system of claim 1, wherein: First, the current output voltage increment vector of the inverter is used as the input, and the output current change is used as the output. Multiple base models are trained respectively. Then, the training results of all base models are combined using simple averaging and weighted averaging respectively. The combined result with the minimum global loss is selected as the final training result output.

3. The three-phase two-level inverter control optimization system based on ensemble learning according to claim 2, characterized in that: In the case of combining using weighted average, first calculate the global loss of each base model on the entire data set using the following formula , Then, the weighted average combination result is calculated using the following formula. , Where M represents the total number of base models, This represents the training result of the m-th base model, and there are a total of M base models; Calculate the combined result after simple averaging using the following formula. , 。 4. The three-phase two-level inverter control optimization system based on ensemble learning according to claim 2, characterized in that: The global loss is calculated for the combined results obtained by the two ensemble strategies, simple averaging and weighted averaging. The combined result with the smaller global loss is selected as the final training result for subsequent parameter optimization.

5. The three-phase two-level inverter control optimization system based on ensemble learning according to claim 2, characterized in that: Five base models are set, all of which are set as linear mapping models.

6. The three-phase two-level inverter control optimization system based on ensemble learning according to claim 1, characterized in that: The remote data transmission unit includes a data transmission unit and a cloud server. The data transmission unit is responsible for realizing the bidirectional conversion between serial communication and wireless communication between each inverter and the cloud server. The cloud server serves as a relay station for remote data communication, responsible for bidirectional data interaction between the data transmission unit and the remote monitoring and optimization workstation.

Images