Abstract: This paper designs a dual closed-loop control system with current and voltage closed loops. Because this system is unstable under certain conditions, traditional PI regulators cannot optimize it. Therefore, improvements are made to this basic control system by utilizing neural network control to derive an adaptive control system based on PWM rectifier feedback, achieving better performance. Finally, based on theoretical analysis, two directly purchased diode voltage-type AC-DC-AC frequency converters were modified and tested. Relevant experimental waveforms were captured, and the experimental results were analyzed and explained, proving the accuracy and effectiveness of the design. Keywords: Voltage Source Rectifier; PWM Rectifier Feedback; Neural Network Control Abstract: This paper designs a double close-loop control system based on current and voltage. As the system is unstable in some cases, the traditional PI regulator cannot regulate the system more precisely. To achieve a more stable and better control effect, it should be improved by using Neural Network Control to obtain a feedback adaptive control system of the PWM rectifier in order to achieve better results. Finally, based on theoretic analysis, two diode voltage source type AC-DC-AC rectifiers have been reconstructed to conduct the experiment. The test waveforms and analysis results show that the design improved the accuracy and effectiveness of the control system. Keywords: Voltage Source Rectifier; PWM Rectifier Feedback; Neural Network Control Introduction In recent years, AC variable frequency speed control technology has developed rapidly in China. The superior performance of variable frequency speed control in various aspects is unmatched by other AC speed control methods. However, its drawbacks are becoming increasingly apparent, mainly manifested in: deteriorating the power supply quality of the power supply system; harmonic currents and the harmonic voltages they generate causing malfunctions in the control, protection, and measurement devices in the system; low power factor on the grid side when the rectifier is in deep phase-controlled operation, reducing the utilization rate of power generation and transmission equipment, and generating a large amount of additional losses; and due to the unidirectional conductivity of the rectifier elements, the regenerative energy from motor braking cannot be fed back to the grid. With the continuous development of power electronic devices and the continuous improvement of the performance of DSP control chips, it has become possible to implement frequency converters using dual PWM. A dual PWM frequency converter uses IGBT switching devices for both rectification and inversion, and the inverter section is already very mature in motor control, so the key to the control of the entire system lies in the rectification section. Because this technology can achieve rapid energy feedback and a high power factor, it is widely used in the control of energy feedback in small-scale hydropower and wind power generation. 1. Modeling and Analysis of the Main Circuit of a Three-Phase Bridge PWM Rectifier 1.1 Three-Phase VSR System Model The main circuit structure of the three-phase VSR is shown in Figure 2.4. T1～T6 are the rectifier power switching transistors, and D1～D6 are freewheeling diodes. When the power transistors are not conducting, the current can flow through the diodes. This figure is also the equivalent circuit diagram of the rectifier feedback section of the system design. [align=center] Figure 1 Three-phase VSR Main Circuit Structure[/align] According to the above figure and Kirchhoff's voltage law, the current balance equation yields the general mathematical model of this circuit in the three-phase stationary (a, b, c) coordinate system: Where S[sub]k[/sub] is the switching signal, defined as follows: This general mathematical model has the characteristics of clear physical meaning and intuitiveness. However, in this mathematical model, the AC side of the three-phase VSR is a time-varying AC quantity, which is not conducive to the design of the control system. 1.2 Based on the two-phase rotating coordinate system model In the three-phase stationary coordinate system, e[sub]a[/sub], e[sub]b[/sub], e[sub]c[/sub] and i[sub]a[/sub], i[sub]b[/sub], i[sub]c[/sub] are coupled. The system model can be simplified by coordinate transformation, transforming the three-phase stationary coordinate system abc to the two-phase synchronous rotating dq coordinate system. The d-axis coincides with the direction of the three-phase voltage synthesis vector and is synchronous with the angular velocity ω counterclockwise, and the q-axis leads the d-axis by 90[sup]. The transformation is divided into "equal quantity" coordinate transformation and "equal power" coordinate transformation. The transformation designed in this paper adopts the "equal quantity" coordinate transformation. The so-called "equal quantity" coordinate transformation refers to the coordinate transformation in which the general vector in a certain coordinate system is equal to the general vector in the transformed coordinate system. The relationship between the coordinate systems is shown in Figure 2.5, which shows the synthesis vector of the three-phase input voltage[22]. [align=center] Figure 2 Relationship between coordinate systems abc and dq[/align] Following the principle of equal transformation, the above transformation relationship can be described by the following transformation matrix: According to equations (1) and (3), the system model of the three-phase VSR in the rotating two-phase dq coordinate system can be obtained as follows: 2 Three-phase VSR control method From the known two-phase rotating coordinate system dq model, the differential equation of the dq two-phase current can be obtained as follows: Thus, the inner loop of the system current can be designed as follows: Based on the above analysis, the double closed-loop control structure of the converter system shown in the figure below is constructed. The outer loop is the voltage loop, which controls the output of the DC bus voltage. The system output voltage error is obtained through the DC bus voltage setpoint and feedback. The active current setpoint is calculated by the voltage regulator. Its value determines the magnitude of the active power, and its sign determines the direction of power flow. The inner loop of the system is the current loop, which controls the current response. The control block diagram is shown in Figure 3. [align=center]Figure 3 Control block diagram of three-phase VSR based on synchronous rotating converter[/align] However, the above system also has some drawbacks. To ensure the input power factor cosφ=1, the reactive current must always be kept at 0. However, this is difficult to achieve in practice, as there will always be a small amount of reactive power that cannot be eliminated, i.e., it is impossible for it to be completely zero. When the load changes abruptly, the reactive power component fed back to the current loop will be amplified, which will produce a large error, causing the system to oscillate and become unstable. At the same time, this system is a nonlinear system, and the relevant parameters are not easy to measure, making it impossible for traditional PID control to optimize. To avoid these problems, this paper introduces a neural network control method to improve the system, so as to achieve adaptive adjustment of the system and eliminate overshoot. 3 Neural Network Control Neural networks are mainly used in control system design to address the nonlinearity, uncertainty, and complexity of the system. Due to the adaptability, parallel processing capability, and robustness of neural networks, control systems using neural networks have stronger adaptability and robustness. This paper mainly introduces direct control in neural networks. 3.1 Overview of Direct Neural Network Control With the deepening of neural network research, the application of neural networks in control has gradually moved from simulation research to applied research. Among the many neural network control methods, although direct control cannot theoretically obtain a method for adjusting network weights based on the output error of an unknown nonlinear system, the Latin adage "simplicity is the mark of truth" has always influenced engineering choices. Therefore, direct neural network adaptive control is highly valued for its simple structure, concise algorithm, and ease of engineering understanding. The structure of direct neural network control is shown in Figure 4. [align=center] Figure 4 Direct Neural Network Control[/align] Based on the essence of classical PID control, a new direct neural network controller is proposed, whose network weights consist of three parts: proportional term, integral term, and derivative term. The integral term is the traditional neural network weight part, and its function is to generate memory, ensuring the continuity of control and steady-state convergence accuracy; the proportional and derivative terms ensure that the network weights have "instantaneous" responsiveness when control errors occur, thereby generating corresponding adjustment control quantities. The differential term also has a certain variable structure control capability, which can enhance the robustness of the controller. 3.2 Design of Neural Network Controller This paper constructs a neural network controller based on the discrete PID control algorithm [38] and applies it to real-time control. Usually, the PID control formula is: Replace the integral in the formula with summation and replace the differential in the formula with finite difference, then formula (7) can be approximated as: In the control process, the three weights ω1, ω2, and ω3 are corrected according to the deviation. If the deviation is less than the given error, the weights are no longer corrected. The system control composed of the network is shown in Figure 5: [align=center] Figure 5 Neural Network Controller Structure Diagram [/align] The formula for correcting the network weights is: Where Ly is the learning step size, y* is the given value, y is the output value, u is the calculated control quantity, and x1, x2, and x3 are the reference quantities to be fed in. The learning step size L[sub]y[/sub] is generally taken as a number between (0, 1). Its size determines the adjustment range of the network weights. The initial weight ω[sub]i[/sub] (0) of the network is not only related to whether the network can reach the global minimum point, but also has a significant impact on the length of the network learning time. In neural network theory, random numbers are generally taken as the initial weights. According to the characteristics of our designed neural network controller, we take the PID parameter value of the main motor to adjust the initial weights. 4 PWM Rectifier Feedback Adaptive Control System 4.1 System Design This section uses the theoretical basis of the previous chapters to design the system in detail to achieve the control requirements. The design of the rectifier section is based on the decoupling control of the PWM rectifier, which mainly includes the following parts: coordinate transformation, generation of PWM waveform, implementation of digital PI regulator, and introduction of intelligent fuzzy neural network controller, as shown in Figure 6: [align=center] Figure 6 PWM Rectifier Feedback Adaptive Control System Based on Inverter Output[/align] The power factor cosφ is calculated by the system grid side based on voltage and current sampling. The system provides a DC bus voltage command U[sup]＊[/sup][sub]dc[/sub]. The error between this command and the DC bus voltage feedback U[sub]dc[/sub] is sent to the neural network controller for calculation. The current feedback i[sub]a[/sub], i[sub]b[/sub], i[sub]c[/sub] are transformed into currents i[sub]fd[/sub], i[sub]fq[/sub] in the two-phase rotating coordinate system dq through abc/dq transformation. During this process, the current detection also obtains i[sup]＊[/sup][sub]fd[/sub] and i[sup]＊[/sup][sub]fq[/sub] in the two-phase rotating coordinate system dq via abc/dq. These two important parameters, along with the power factor cosφ, are sent to the intelligent fuzzy neural network controller for self-learning. The error between the current and the corresponding current command is then sent to the current loop PI regulator to generate voltage commands U[sup]＊[/sup][sub]pd[/sub] and U[sup]＊[/sup][sub]pq[/sub]. The calculation unit obtains voltages U[sub]pd[/sub] and U[sub]pq[/sub], which are then transformed into voltages U[sub]d[/sub] and U[sub]q[/sub] in the two-phase stationary coordinate system dq via a 2r/2s transformation. The generated PWM signals are then used to control the IGBT transistors. To determine whether energy is being rectified or inverted, and its conversion rate, the PI parameters are self-tuned to accelerate the system's response speed, enabling the system to quickly follow changes on the inverter side and achieve rapid feedback. The adjustment of i[sup]＊[/sup][sub]fd[/sub] ensures the system does not experience significant disturbances under sudden load changes. This adjustment process utilizes fuzzy control algorithms; this paper only uses a neural network control algorithm to adjust other parameters. The algorithm for adjusting i[sup]＊[/sup][sub]fd[/sub] using fuzzy control is too complex and computationally intensive, and is not discussed here, but can be further studied in the future. Using a neural network controller, nonlinear systems can be self-tuned without precisely measuring relevant parameters. Instead, the relevant quantities are fed into the controller for self-learning and then sent to the PI regulator for tuning. By introducing a neural network controller, the energy on the rectifier and inverter sides of the system remains equal, and energy can flow rapidly in both directions, preventing excess energy from being stored in the DC capacitor and wasting energy. Meanwhile, the grid-side power factor can be maintained at unity power factor, i.e., cosφ=1. 4.2. Design of System-Based Neural Network Controller According to the content of Chapter 3, the specific quantities in the system are substituted into the neural network controller to obtain the system-based controller, as shown in Figure 7. [align=center] Figure 7 Fuzzy Intelligent Neural Network Controller[/align] There are four quantities fed into the controller: the active component of the current obtained after coordinate transformation on the inverter side i[sup]＊[/sup][sub]fd[/sub], the reactive component i[sup]＊[/sup][sub]fq[/sub], the system power factor cosφ, and the error component of the output voltage. Here we give the calculation formulas of the other three parameters besides i[sup]＊[/sup][sub]fq[/sub] after the controller self-learning, where: As shown in Figure 6, the intelligent fuzzy neural network controller controls the active and reactive power loops respectively. If the active power loop passes through and performs self-learning, the network weight correction formula can be written as: 4.3 System Experiment and Result Analysis The equipment used in this experiment consists of two directly purchased diode voltage-type AC-DC-AC frequency converters connected in parallel on the DC side, with the uncontrolled rectifier section left floating. Since the inverter side uses IGBTs, the inverter side of one frequency converter serves as the rectifier side of the fast feedback system, and the inverter side of the other frequency converter serves as the inverter side of the fast feedback system, using PWM control technology. The DC motor is used as the load, and the fully digital DC speed controller is run in torque mode to provide the load for the AC motor. In the experiment, the grid-side phase voltage is 220V, and the frequency is 50Hz. The experimental waveforms were obtained using a FLUKE power quality analyzer. The experiments are all illustrated using the input voltage and current waveforms of phase a as examples, as shown in Figures 7 and 8. Figure 7 shows the rectifier input waveform of the traditional PI regulator. Figure 8 shows the rectifier input waveform based on the neural network controller. As shown in Figure 7(a), when using the traditional PI regulator, the input current value is 5A, and ripple appears on the i[sub]a[/sub] waveform. This is because the IGBT transistor cannot commutate in time. At the same time, it can be seen that the current waveform lags behind the voltage waveform, indicating that the power factor is deviated. When the input current value increases to 7A, as shown in Figure 7(b), the ripple on the i[sub]a[/sub] waveform is more obvious. At this time, i[sub]a[/sub] also lags behind u[sub]a[/sub], and the power factor deviation is more obvious. As shown in Figures 8(a) and (b), after adding a neural network controller to the control system, regardless of the current value, the ripple on the i[sub]a[/sub] waveform is eliminated, presenting a perfect sine wave, and i[sub]a[/sub] is always in phase with u[sub]a[/sub], meaning the system's power factor is 1, meeting the requirements of the system design in this project. Conclusion PWM rectifier feedback adaptive control system based on inverter output is one of the most widely studied topics in the field of power electronics due to its good sinusoidal input current, ability to achieve unity power factor, and bidirectional energy flow. It is particularly widely used in the control of energy feedback in small hydropower and wind power generation. Furthermore, with the development of power harmonic mitigation and reactive power compensation technologies, the development of three-phase high power factor reversible rectifiers has become the best choice for high-power AC-DC reversible converters and is also the core of AC-DC-AC feedback frequency converters. This paper applies an intelligent fuzzy neural network controller to a traditional dual-PWM frequency converter system. Compared with conventional PID controllers, this method can self-adjust controller parameters in real time according to system deviations, thereby improving the controller's adaptability and robustness, and significantly enhancing the system's dynamic performance. This method is also proposed to improve existing DC power supply control devices and the variable frequency speed control devices under development, demonstrating its forward-looking and pioneering role. It lays a theoretical foundation for in-depth research on the characteristics of reversible rectifiers and their control. However, further theoretical and experimental research and analysis are needed to realize the commercialization of the device, optimize the system's main circuit and control parameters, and self-tune the inverter-side current PI parameters to accelerate the system's response speed. References [1] Zhang Chongwei, Zhang Xing. PWM rectifier and its control. Beijing: Machinery Industry Press, 2003. [2] Xiong Jian, Kang Yong et al. Research on control technology of three-phase voltage PWM rectifier [J]. Power Electronics Technology, 1999 (4): 5-7. [3] Wang Zhaoan, Huang Jun. Power Electronics Technology. Beijing: Machinery Industry Press, 2000. [4] Satish Kumar. Neural Networks (English Reprint) [M] 1st Edition. Beijing: Tsinghua University Press, 2006. [5] Sato, Akira Multi-level current-source PWM rectifier based on direct power control 2007, p 1768-1773. [6] Li Sufang. Three-phase voltage type PWM rectifier [D]. Master's thesis of Huazhong University of Science and Technology, 2003.4. [7] Sun Wei, Wu Xiaojie, Dai Peng, Zhou Juan. An over view of damping methods for three-phase PWM rectifier. IEEE ICIT 2008 - Conference Proceedings, 2008. [8] Yang Xingjun, Zheng Junli. Artificial Neural Networks. Higher Education Press, 1992.9. [9] Tao Yonghua. New PID Control and Its Applications [M]. 2nd ed. Beijing: Machinery Industry Press, 1998. [10] Chen Xiaoli. Research on High Power Factor Dual PWM Variable Frequency Speed ​​Regulation System [D]. Master's Thesis, Guangxi University, 2005.6. [11] Hasan Komurcugil, Osman Kukrer. Lyapunov based control for three-phase PWM AC/DC voltage source converters. IEEE Trans. Power Electronics, 1998, 13(5): 801-813. [12] Gorecki, Krzysztof, Stepowicz, Witold J. Evaluation of macromodels of monolithic current mode PWM controllers. Proceedings of International Conference, TCSET 2006, pp. 141-144.