A control method of a shunt active power filter, an electronic device and a medium

Abstract

The application discloses a control method of a parallel active power filter, electronic equipment and a medium, and is applied to the fields of power electronics technology and power quality control. The control method of the parallel active power filter provided by the application collects system operation data in real time, obtains a reference compensation current according to a mathematical model of the parallel active power filter in a d-q coordinate system, constructs an adaptive optimization model in combination with a Markov decision mechanism, trains an intelligent agent offline through a two-stage curriculum learning mechanism, and obtains an optimal strategy model. When the system is online, target PI parameters that adapt to real-time working conditions are dynamically output, fixed parameters are replaced, and precise closed-loop regulation and control are completed in cooperation with decoupling control and space vector pulse width modulation. The application effectively solves the problems of poor working condition adaptability, slow dynamic response and low harmonic compensation accuracy of traditional fixed PI control, avoids control instability defects caused by working condition fluctuations, and greatly improves the dynamic compensation performance and operation stability of the filtering system.

Description

Technical Field

[0001] This application relates to the fields of power electronics technology and power quality control, and in particular to a control method, electronic equipment and medium for a parallel active power filter. Background Technology

[0002] With the rapid development of power electronics technology, various nonlinear loads such as rectifiers and frequency converters have been widely used in power systems. These loads are prone to injecting harmonic currents into the power grid during operation, causing severe fluctuations in reactive power and resulting in serious power quality pollution. This problem not only reduces power grid efficiency, exacerbates equipment losses, and shortens equipment lifespan, but also interferes with the normal operation of the power system, seriously threatening the safe, stable, and efficient operation of the power grid. The shunt active power filter (SAPF), as a core power quality management device for harmonic suppression and dynamic reactive power compensation, directly determines the effectiveness of power quality management through its control performance, making it a key device for ensuring the reliable and efficient operation of the power grid.

[0003] In current industrial applications, SAPF (Proportional-Integral) dual-loop control strategies are often adopted, which are simple in structure and easy to implement in engineering. However, this approach can only achieve basic compensation functions under stable steady-state conditions. Since the parameters of the PI controller are fixed after tuning and lack dynamic adaptive adjustment capabilities, it suffers from poor dynamic response, low compensation accuracy, and insufficient robustness when facing complex scenarios such as load changes, grid voltage fluctuations, and operating disturbances. The effects of harmonic mitigation and reactive power compensation are significantly limited, making it difficult to adapt to complex and ever-changing grid operating conditions. To overcome the shortcomings of traditional PI control, existing improved control strategies still have many inherent weaknesses: 1. Fuzzy adaptive control relies on manual experience for modeling, resulting in poor generalization ability under operating conditions; 2. Model predictive control and deadbeat control are highly dependent on accurate mathematical models, and are prone to model mismatch problems due to equipment parameter drift and line impedance changes; 3. Sliding mode control has inherent chattering defects, and suffers from high computational load and poor real-time performance under complex topologies; 4. Traditional supervised learning methods such as neural networks rely on offline sample training, lack online adaptive optimization capabilities, and are difficult to adapt to dynamic disturbance scenarios in the power grid.

[0004] Deep Reinforcement Learning (DRL) provides a novel technical approach for intelligent adaptive control of SAPF (Power Grid Protection). Among them, the DDPG (Deep Deterministic Policy Gradient) algorithm is adapted to the characteristics of continuous action space regulation, which meets the control requirements of dynamic compensation in SAPF. However, the traditional DDPG algorithm has inherent problems such as Q-value overestimation, unstable training convergence, and weak disturbance rejection, which cannot meet the requirements of high-precision and high-stability power quality management and control of SAPF under complex power grid conditions.

[0005] Given the above-mentioned technologies, finding an SAPF optimization control method with low model dependence, strong adaptability, high control accuracy, and excellent robustness is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0006] The purpose of this application is to provide a control method, electronic device and medium for parallel active power filters, which can solve the problem that the traditional DDPG algorithm cannot meet the requirements of high precision and high stability power quality management and control of SAPF under complex power grid conditions in the prior art.

[0007] To address the aforementioned technical problems, this application provides a control method for a parallel active power filter, applied to a power filtering system including a three-phase three-wire circuit topology, a sampling and detection unit, and a pulse width modulation drive unit. The power filtering system is configured with a dual-closed-loop control unit and a dual-delay reinforcement learning adaptive PI optimization unit. The control hierarchy of the power filtering system adopts a four-level hierarchical architecture: voltage outer loop, current inner loop, dual-delay deep deterministic strategy parameter optimization layer, and pulse width modulation execution layer. The method includes: The system operation data of the power filtering system is collected through the sampling and detection unit; Based on system operation data, a mathematical model of a parallel active power filter in a synchronously rotating dq coordinate system is constructed to obtain the reference compensation current. Based on the mathematical model of the parallel active power filter and the reference compensation current, a corresponding Markov decision-making dual-delay deep strategy adaptive optimization model is constructed. The dual-delay deep policy adaptive optimization model based on Markov decision-making adopts a two-stage course learning mechanism to train the dual-delay deep policy agent offline. The parameters of the dual-delay deep policy agent are iteratively optimized through the pre-training stage and the fine-tuning stage to obtain the optimal dual-delay deep policy model. The optimal dual-delay deep strategy model is integrated into the power filtering system for online invocation, and the target PI parameters are output to replace the fixed PI parameters of the dual closed-loop control unit. Based on the updated target PI parameters, the corresponding power transistor drive pulse is generated by combining decoupling control and space vector pulse width modulation.

[0008] Preferably, a mathematical model of a parallel active power filter in a synchronously rotating dq coordinate system is constructed to obtain the reference compensation current, including: Based on the three-phase grid voltage, three-phase load current, filter compensation current and DC bus voltage in the system operation data, a mathematical model circuit equation for a parallel active power filter in a three-phase static coordinate system is established. The electrical quantities in the circuit equations of the mathematical model of the parallel active power filter are converted into electrical quantities in a two-phase stationary coordinate system by Clark coordinate transformation, and then converted into DC electrical quantities in a synchronous rotating dq coordinate system by Park coordinate transformation, so as to construct the mathematical model of the parallel active power filter. Based on the DC electrical quantities in the synchronous rotating dq coordinate system, a current decoupling model is constructed to correspond to the mathematical model of the parallel active power filter. The d-axis current coupling component and the q-axis current coupling component are decoupled to obtain the independent d-axis current component and the independent q-axis current component without coupling. Based on independent d-axis current components and independent q-axis current components, combined with the voltage regulation of the outer voltage loop, the corresponding reference compensation current is obtained through the command current generation model; wherein, the reference compensation current includes: d-axis reference compensation current and q-axis reference compensation current.

[0009] Preferably, a dual-delay deep policy adaptive optimization model for Markov decision-making is constructed, including: Based on the mathematical model of the parallel active power filter and the reference compensation current, and combined with the operating deviation characteristics of the power filter system, a four-dimensional state vector with the deviation signal of the power filter system as the core and the state space corresponding to the four-dimensional state vector are constructed. Based on the state space and combined with the adjustable parameter characteristics of the dual-loop control unit, all the core PI parameters to be optimized in the dual-loop control unit are constructed into a six-dimensional continuous action vector, and an action space corresponding to the six-dimensional continuous action vector is constructed. Based on the state space and action space, and combining the harmonic compensation of the parallel active power filter, the voltage regulation control target, and the system safety operation constraints of the power filter system, a composite reward function containing the dominant performance term, action smoothing term, and boundary constraint term is configured to construct the corresponding reward mechanism. Based on the reward mechanism, a dual anti-jitter mechanism combining reward function constraints and hardware filtering is constructed to build a corresponding Markov decision-making dual-delay deep policy adaptive optimization model.

[0010] Preferably, the expression for the four-dimensional state vector is: ; The expression for a six-dimensional continuous action vector is: ; The expression for the composite reward function is: ; in, It is a four-dimensional state vector; This refers to the d-axis harmonic current error. This refers to the q-axis harmonic current error. This refers to the DC-side voltage error. This represents the rate of change of DC-side voltage. For transpose; It is a six-dimensional continuous action vector; and These are two parameters of the outer voltage loop; , , , These are the four parameters of the inner current loop; For a composite reward function; The primary performance parameter; For motion smoothing parameters; These are the boundary constraint parameters.

[0011] Preferably, the optimal dual-latency deep policy agent model is obtained by iteratively optimizing the parameters through a pre-training phase and a fine-tuning phase, including: In the pre-training phase, a piecewise error penalty reward function is used to constrain the state of the power filtering system to within the safe operating range; During the fine-tuning phase, the pre-trained weights from the pre-training phase are loaded, and the system switches to the full reward function, which focuses on minimizing the total harmonic distortion rate. Based on two-stage hierarchical training using a piecewise error penalty reward function and a complete reward function, the optimal dual-delay deep policy model is obtained through iterative convergence.

[0012] Preferably, the optimal dual-delay deep strategy model is integrated into the online application of the power filtering system, including: Load the optimal dual-delay deep policy model that has converged during offline training, retain the execution network built into the optimal dual-delay deep policy model for real-time inference, and remove the evaluation network and target network used for offline iterative training. The processed optimal dual-delay deep strategy model is then integrated into the control program of the power filtering system.

[0013] Preferably, the target PI parameter is output to replace the fixed PI parameter of the dual closed-loop control unit, including: The operating status data of the power filtering system is acquired and input into the optimal dual-delay deep strategy model; The execution network in the optimal dual-delay deep strategy model outputs target PI parameters that are adapted to the current operating conditions in real time. The target PI parameter is used to replace the preset fixed PI parameter inside the dual closed-loop control unit.

[0014] Preferably, based on the updated target PI parameters, a corresponding power transistor drive pulse is generated by combining decoupling control and space vector pulse width modulation, including: Based on the updated target PI parameters, determine the corresponding modulation voltage command; The modulation voltage command is decoupled to obtain the decoupled voltage modulation signal; The decoupled voltage modulation signal is input to the pulse width modulation drive unit so that the corresponding power transistor drive pulse can be generated by space vector pulse width modulation.

[0015] On the other hand, this application also provides an electronic device, including a memory for storing computer programs; The processor is used to execute computer programs to implement the steps of the control method for the parallel active power filter described above.

[0016] On the other hand, this application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the control method for the parallel active power filter described above.

[0017] This application provides a control method for a parallel active power filter, specifically applied to a power filtering system including a three-phase three-wire circuit topology, a sampling and detection unit, and a pulse width modulation (PWM) drive unit. This power filtering system is configured with a dual-closed-loop control unit and a dual-delay reinforcement learning adaptive PI optimization unit. The control hierarchy of the power filtering system adopts a four-level hierarchical architecture: voltage outer loop, current inner loop, dual-delay deep deterministic strategy parameter optimization layer, and PWM execution layer. This application optimizes the system control logic through a four-level hierarchical control architecture. First, it collects real-time system operation data of the power filtering system. Then, it obtains the reference compensation current based on the mathematical model of the parallel active power filter in the synchronously rotating dq coordinate system. Finally, it constructs a dual-delay deep strategy adaptive optimization model adapted to the operating conditions using a Markov decision mechanism. Simultaneously, in the Markov decision-based dual-delay deep strategy adaptive optimization model, a two-stage learning mechanism is used to perform offline iterative training on the dual-delay deep strategy agent, selecting the optimal dual-delay deep strategy model. When the power filtering system is running online, the trained optimal dual-delay deep strategy model is invoked to output target PI parameters adapted to the current operating conditions in real time. These parameters dynamically replace the fixed PI parameters of the dual closed-loop control unit. Combined with decoupling control and space vector pulse width modulation, precise power transistor drive pulses are generated to complete the system's closed-loop regulation. This application utilizes a Markov decision-based dual-delay deep strategy adaptive optimization model and a two-stage offline training optimization method, combined with real-time dynamic adjustment of PI parameters. This overcomes the shortcomings of traditional dual closed-loop control, which has fixed parameters and cannot adapt to dynamic grid conditions. It can match the optimal control parameters in real time according to grid load fluctuations and complex operating conditions, effectively compensating for the slow dynamic response and insufficient harmonic compensation accuracy of traditional control strategies. This significantly improves the dynamic response performance and harmonic compensation accuracy of parallel active power filters under varying operating conditions, effectively avoiding control instability problems that easily occur with fixed parameters under extreme and fluctuating conditions, and greatly enhancing the environmental adaptability and operational reliability of the power filtering system. Attached Figure Description

[0018] To more clearly illustrate the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating a control method for a parallel active power filter provided in this application embodiment; Figure 2 The main circuit topology diagram of the three-phase three-wire parallel active power filter provided in this application; Figure 3The SAPF voltage and current dual closed-loop control structure diagram provided in this application; Figure 4 The overall block diagram of the TD3-based SAPF adaptive PI control provided in this application; Figure 5(a) is a first schematic diagram of the TD3-based training convergence process provided in this application; Figure 5(b) is a second schematic diagram of the TD3-based training convergence process provided in this application; Figure 6 A structural diagram of an electronic device provided in another embodiment of this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of this application.

[0021] The core of this application is to provide a control method, electronic device, and medium for a parallel active power filter.

[0022] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0023] Figure 1 A flowchart illustrating a control method for a parallel active power filter provided in this application embodiment is shown below. Figure 1 As shown, it includes the following steps: S10: Collect system operation data of the power filtering system through the sampling and detection unit.

[0024] S11: Based on the system operation data, construct a mathematical model of the parallel active power filter in the synchronous rotating dq coordinate system to obtain the reference compensation current.

[0025] S12: Based on the mathematical model of the parallel active power filter and the reference compensation current, construct the corresponding Markov decision-making dual-delay deep strategy adaptive optimization model.

[0026] S13: A dual-delay deep policy adaptive optimization model based on Markov decision-making. A two-stage course learning mechanism is used to train the dual-delay deep policy agent offline. The parameters of the dual-delay deep policy agent are iteratively optimized through the pre-training stage and the fine-tuning stage to obtain the optimal dual-delay deep policy model.

[0027] S14: Integrate the optimal dual-delay deep strategy model into the power filtering system for online invocation, and output the target PI parameters to replace the fixed PI parameters of the dual closed-loop control unit.

[0028] S15: Based on the updated target PI parameters, generate the corresponding power transistor drive pulse by combining decoupling control and space vector pulse width modulation.

[0029] In a specific embodiment, the optimized control method for the parallel active power filter provided in this application is specifically applied to a power filtering system including: a three-phase three-wire circuit topology, a sampling and detection unit, and a pulse width modulation (PWM) drive unit. The power filtering system is configured with a dual-closed-loop control unit and a twin-delayed deep deterministic policy gradient (TD3) adaptive PI optimization unit. Furthermore, the control hierarchy of the power filtering system adopts a four-level hierarchical architecture: voltage outer loop, current inner loop, twin-delayed deep deterministic policy gradient parameter optimization layer, and pulse width modulation execution layer. Figure 2 , Figure 3 and Figure 4 As shown. Figure 2 The main circuit topology diagram of the three-phase three-wire parallel active power filter provided in this application is used to represent the controlled object (physical system). Figure 3 The SAPF voltage and current dual closed-loop control structure diagram provided in this application is used to represent the controller (algorithm implementation). Figure 4 The overall block diagram of the TD3-based SAPF adaptive PI control provided in this application is used to illustrate the complete adaptive PI control architecture of the parallel active power filter, which consists of sampling acquisition, coordinate transformation, dual closed-loop basic control, dual-delay deep strategy parameter adaptive optimization, space vector pulse modulation, and power device driving. Based on... Figure 4 Real-time correction of the added adaptive parameter optimization branch Figure 3 The PI control parameters within, Figure 3 Combined with the updated control parameters, according to Figure 2 The optimal pulse width modulation control signal is calculated from the feedback voltage and current, and then sent back. Figure 2 Execution forms a complete closed-loop control system equipped with adaptive parameter tuning logic.

[0030] according to Figure 2It can be seen that the power filtering system mainly consists of a three-phase AC power supply, a nonlinear load that generates distorted current, and a parallel-connected SAPF main circuit. The working principle of SAPF is to first calculate the reference compensation current based on the voltage, current, and DC voltage; then, the current tracking controller compares the reference current with the actual current to generate a control signal; finally, the drive circuit outputs PWM pulses to control the converter, causing it to inject anti-phase harmonic current, thereby achieving the purification of the grid current.

[0031] according to Figure 3 As can be seen, this power filtering system adopts a four-level architecture: voltage outer loop, current inner loop, TD3 parameter optimization layer, and PWM execution layer. The voltage outer loop is responsible for stabilizing the DC bus voltage and outputting the d-axis active current reference component. The current inner loop realizes rapid tracking of the reference compensation current and completes the harmonic and reactive power compensation tasks. The TD3 parameter optimization layer senses the system operating status in real time and outputs the optimal PI parameter combination. The PWM execution layer converts the control signal into power transistor drive pulses to complete the direct control of the inverter. The six parameters output by the TD3 agent—the voltage outer loop proportional-integral parameters and the current inner loop d-axis and q-axis proportional-integral parameters—directly replace the traditional fixed PI parameters, realizing online dynamic adjustment of the control gain, so that the controller always matches the current system operating conditions.

[0032] What can be determined is that the three-phase three-wire circuit topology corresponds to Figure 2 The components include a three-phase power supply, a nonlinear load, an L / R filter inductor, a three-phase inverter bridge composed of Insulated Gate Bipolar Transistors (IGBTs), and a DC bus capacitor; the sampling and detection unit corresponds to... Figure 2 The voltage sensor and current sensor in the system are used to detect the effective phase voltage of the three-phase power grid. , and and DC bus voltage Current sensors are used to detect three-phase nonlinear load current ( , , ) and the compensation current injected into the grid by the three-phase SAPF ( , , Dual closed-loop control unit corresponding to Figure 3 The voltage outer loop PI, current inner loop PI (d-axis and q-axis), and feedforward decoupling ( Cross-decoupling gain), coordinate transformation ( ); PWM drive unit corresponding Figure 2 The drive circuit and inverter switching transistors in the TD3. The adaptive PI optimization unit in... Figure 4 The six PI parameters are drawn in the middle. , , , , , ).

[0033] besides, Figure 2 In These are the three-phase current values; These are the three-phase voltage values; This is the actual value of the DC bus voltage (feedback). This is the actual compensation current along the d-axis; This is the actual compensation current along the q-axis; This is the d-axis current reference value (reactive / harmonic components). This is the reference value for the q-axis current (reactive / harmonic components). Figure 3 In This is the reference value (target setting value) for the DC bus voltage. This is a reference value for the DC bus current; This is the actual compensation voltage along the d-axis; This is the actual compensation voltage along the q-axis; Angular frequency; For time; for Shaft reference voltage; for Shaft reference voltage; Figure 4 In It is a four-dimensional state vector; It is a six-dimensional continuous action vector.

[0034] according to Figure 2-4As shown in the attached diagram, the principle of this method is as follows: The power filtering system first collects core system operating data such as voltage and current on the grid side and load side in real time through the sampling and detection unit, accurately determining real-time operating condition information such as grid harmonic disturbances and load fluctuations, providing a real data foundation for subsequent modeling and control calculations; then, based on the collected system operating data, a high-precision parallel active power filter mathematical model is built in the synchronously rotating dq coordinate system; based on the parallel active power filter mathematical model, a reference compensation current adapted to the current grid operating conditions is obtained, and the required harmonic compensation control target of the system is determined. Furthermore, to address the issues of traditional fixed PI parameters failing to adapt to dynamic operating conditions and insufficient compensation accuracy, this application constructs a dual-delay deep strategy adaptive optimization model integrating Markov decision-making mechanisms based on the established mathematical model of the parallel active power filter and the obtained reference compensation current. This model transforms changes in grid operating conditions, adjustments to control parameters, and the quality of compensation into an iteratively optimized decision-making process, achieving dynamic matching between the control strategy and actual operating conditions. On this basis, a two-stage learning mechanism is employed to conduct offline hierarchical training of the dual-delay deep strategy agent. The pre-training stage completes the initial parameter screening and model convergence for basic operating conditions, while the fine-tuning stage iteratively optimizes the core parameters of the agent for complex fluctuating operating conditions, effectively avoiding single-training... The training method avoids problems such as underfitting and poor generalization ability, and ultimately converges to obtain an optimal dual-delay deep strategy model that adapts to multiple operating conditions. After completing offline training, the optimal dual-delay deep strategy model is connected to the power filter control system for online real-time invocation. This can be understood as deploying the optimal dual-delay deep strategy model online, relying on the model's adaptive decision-making capability to output target PI parameters that match the current power grid operating state in real time, dynamically replacing the fixed PI parameters in the traditional dual closed-loop control unit, breaking the control limitations of fixed parameters. Finally, the system completes closed-loop decoupling control of the voltage outer loop and current inner loop based on the real-time updated target PI parameters, and combines Space Vector Pulse Width Modulation (SVPWM) to generate high-precision, low-harmonic power transistor drive pulses to drive the orderly switching of insulated gate bipolar transistors, ultimately achieving accurate dynamic compensation for power grid harmonics. This effectively improves the dynamic response speed and harmonic compensation accuracy of the parallel active power filter under complex operating conditions such as load fluctuations and power grid disturbances, ensuring the long-term stable and efficient operation of the power filter system.

[0035] Therefore, the control method for a parallel active power filter provided in this application optimizes the system control logic through a four-level hierarchical control architecture. First, it collects real-time system operation data of the power filter system. Then, it obtains the reference compensation current based on the mathematical model of the parallel active power filter in the synchronously rotating dq coordinate system. Finally, it constructs an adaptive optimization model for the dual-delay deep strategy adapted to the operating conditions using a Markov decision mechanism. Simultaneously, in the Markov decision-based dual-delay deep strategy adaptive optimization model, a two-stage learning mechanism is used to perform offline iterative training on the dual-delay deep strategy agent, selecting the optimal dual-delay deep strategy model. During online operation of the power filter system, the trained optimal dual-delay deep strategy model is invoked to output the target PI parameters adapted to the current operating conditions in real time, dynamically replacing the fixed PI parameters of the dual closed-loop control unit. Finally, it combines decoupling control and space vector pulse width modulation to generate precise power transistor drive pulses, completing the system's closed-loop regulation. This application utilizes a Markov decision-making-based dual-delay deep strategy adaptive optimization model and a two-stage offline training optimization method, combined with a real-time dynamic adjustment of PI parameters. This overcomes the shortcomings of traditional dual-closed-loop control, which has fixed parameters and cannot adapt to dynamic power grid conditions. It can match the optimal control parameters in real time according to power grid load fluctuations and complex operating conditions, effectively compensating for the shortcomings of traditional control strategies, such as slow dynamic response and insufficient harmonic compensation accuracy. This significantly improves the dynamic response performance and harmonic compensation accuracy of parallel active power filters under varying operating conditions, effectively avoiding control instability problems that easily occur with fixed parameters under extreme and fluctuating conditions, and greatly enhancing the environmental adaptability and operational reliability of the power filtering system.

[0036] Based on the above embodiments, as a preferred embodiment, a mathematical model of a parallel active power filter in a synchronously rotating dq coordinate system is constructed to obtain the reference compensation current, including: Based on the three-phase grid voltage, three-phase load current, filter compensation current and DC bus voltage in the system operation data, a mathematical model circuit equation for a parallel active power filter in a three-phase static coordinate system is established. The electrical quantities in the circuit equations of the mathematical model of the parallel active power filter are converted into electrical quantities in a two-phase stationary coordinate system by Clark coordinate transformation, and then converted into DC electrical quantities in a synchronous rotating dq coordinate system by Park coordinate transformation, so as to construct the mathematical model of the parallel active power filter. Based on the DC electrical quantities in the synchronous rotating dq coordinate system, a current decoupling model is constructed to correspond to the mathematical model of the parallel active power filter. The d-axis current coupling component and the q-axis current coupling component are decoupled to obtain the independent d-axis current component and the independent q-axis current component without coupling. Based on independent d-axis current components and independent q-axis current components, combined with the voltage regulation of the outer voltage loop, the corresponding reference compensation current is obtained through the command current generation model; wherein, the reference compensation current includes: d-axis reference compensation current and q-axis reference compensation current.

[0037] In a specific embodiment, the mathematical model of the parallel active power filter includes a three-phase stationary coordinate system model, Clarke-Park coordinate transformation, dq-axis decoupling model (i.e., current decoupling model), and harmonic detection and command current generation model, providing a theoretical basis for the control strategy.

[0038] 1. Three-phase stationary coordinate system model: Based on Kirchhoff's voltage law and instantaneous power conservation, the equations for the AC and DC sides are obtained: ; in, , and This refers to the instantaneous phase voltage values ​​of a three-phase power grid. , and This is the three-phase compensation voltage; , and This is for three-phase compensation current; This is the DC bus voltage; For AC side filter inductance; The equivalent resistance is the sum of the effects of high-frequency switching and line losses. This is a DC-side energy storage capacitor.

[0039] 2. Coordinate Transformation: To achieve zero steady-state error control of time-varying AC signals, the mathematical model in the stationary coordinate system described above needs to be transformed into a synchronously rotating coordinate system. Let the general variables be... Represents the phase voltage in the system or phase current First, the variables are transformed from the equal-power Clarke transform. Coordinate system mapped to two stationary phases Coordinate system: ; Subsequently, using the fundamental angular frequency of the power grid Let the rotational speed be the synchronous phase-locked angle. Transform it to using the Park transform. Synchronous rotating coordinate system: ; 3. Decoupling mathematical model of dq axis: By applying the coordinate transformation rules to the above formula, the SAPF model can be derived. External dynamic equations of the AC side in a rotating coordinate system: ; in, and for Inductive coupling of the shaft; and for Compensation current for the shaft; and for Instantaneous values ​​of phase voltages in the grid; and for Compensation voltage for the shaft.

[0040] Based on this formula, it can be seen that... The mathematical model in the coordinate system has coupling between the inductance of the d-axis and q-axis, and the larger the current, the greater the coupling effect, which is detrimental to the design of the controller.

[0041] Similarly, applying coordinate transformation to the AC / DC power balance equations, under the principle of equal power transformation, the dynamic analysis equation for the DC-side capacitor can be rigorously derived as follows: ; As can be seen, the above formulas together constitute the complete SAPF. The mathematical model in the coordinate system provides a physical foundation for subsequent controller design and state space reconstruction.

[0042] 4. Harmonic Detection and Command Current Generation: The primary task of SAPF is to extract harmonics and reactive currents generated by nonlinear loads in real time. Given that the control strategy of this application... The expansion is performed in a rotating coordinate system. This application adopts... For harmonic detection algorithms, the frequency domain decomposition of nonlinear load current can be expressed as: ; in, and for Nonlinear load current of the shaft; and for The corresponding fundamental DC component (fundamental active current) of the axis. and for The axis contains the AC components of each harmonic (compensating harmonic active components).

[0043] Then, a low-pass filter (LPF) is used to extract the fundamental active current. To achieve unity power factor and maintain DC bus voltage. For stability, SAPF needs to compensate for harmonic active power on the d-axis. And superimposed the active power regulation output from the outer voltage loop. .

[0044] Eliminate all reactive power and harmonics along the q-axis. The final result can be obtained by combining the results. Shaft reference command current (i.e., reference compensation current): ; in, This is the d-axis reference compensation current; This serves as the q-axis reference compensation current, which is also the precise tracking target for the subsequent TD3 current inner loop control algorithm.

[0045] It should be noted that the embodiments provided in this application are only one possible implementation method, but are not limited to this only implementation method. Users can set their own implementation methods according to their needs.

[0046] This embodiment constructs a circuit mathematical model by collecting operating data such as grid voltage, load current, compensation current, and DC bus voltage. Using Clarke coordinate transformation and Parker coordinate transformation, the three-phase AC electrical quantities are converted into... The use of direct current in a coordinate system simplifies the calculation of power grid harmonics and current regulation. Simultaneously, decoupling the d-axis and q-axis currents eliminates current coupling interference between coordinate axes, avoiding current regulation deviations and response lags caused by coupled components. Furthermore, by combining this with voltage outer-loop regulation commands to generate precise d-axis and q-axis reference compensation currents, the acquired compensation commands better reflect actual power grid operating conditions, improving the accuracy and reliability of the reference current commands. This lays the foundation for subsequent adaptive PI parameter optimization and high-precision harmonic compensation control.

[0047] Based on the above embodiments, as a preferred embodiment, a corresponding Markov decision-making dual-delay deep policy adaptive optimization model is constructed, including: Based on the mathematical model of the parallel active power filter and the reference compensation current, and combined with the operating deviation characteristics of the power filter system, a four-dimensional state vector with the deviation signal of the power filter system as the core and the state space corresponding to the four-dimensional state vector are constructed. Based on the state space and combined with the adjustable parameter characteristics of the dual-loop control unit, all the core PI parameters to be optimized in the dual-loop control unit are constructed into a six-dimensional continuous action vector, and an action space corresponding to the six-dimensional continuous action vector is constructed. Based on the state space and action space, and combining the harmonic compensation of the parallel active power filter, the voltage regulation control target, and the system safety operation constraints of the power filter system, a composite reward function containing the dominant performance term, action smoothing term, and boundary constraint term is configured to construct the corresponding reward mechanism. Based on the reward mechanism, a dual anti-jitter mechanism combining reward function constraints and hardware filtering is constructed to build a corresponding Markov decision-making dual-delay deep policy adaptive optimization model.

[0048] In a specific embodiment, the construction of the corresponding Markov decision-making dual-delay deep policy adaptive optimization model includes: state space design, action space design, and reward function design.

[0049] 1. State Space Design: To enable the agent to accurately capture the real-time operating characteristics of the system, a four-dimensional state vector with the deviation signal as its core is constructed: ; in, It is a four-dimensional state vector; This refers to the d-axis harmonic current error. This refers to the q-axis harmonic current error. This refers to the DC-side voltage error. This represents the rate of change of DC-side voltage. This is a transpose. In addition, Shaft harmonic current error and DC side voltage error Provides a real-time performance reference for parameter tuning; DC-side voltage change rate Used to capture system inertial dynamics, it endows the dual-delay deep strategy (TD3) agent with the ability to predict voltage fluctuation trends, enabling it to adjust the gain in advance according to the error change rate, thereby enhancing system damping and effectively suppressing transient overshoot caused by load changes.

[0050] 2. Motion Space Design: The motion space is defined as the six-dimensional continuous motion vector of the six core PI parameters in the dual closed-loop control system. ; in, For transpose; It is a six-dimensional continuous action vector; and These are two parameters of the outer voltage loop; , , , These are the four parameters of the inner current loop.

[0051] 3. Reward Function Design: The reward function is the core component that transforms physical constraints into policy optimization gradients. To enable the TD3 agent to collaboratively optimize steady-state accuracy and dynamic performance while satisfying the safety constraints of the power electronic system, a composite reward function consisting of a dominant performance term, an action smoothing term, and a boundary constraint term is adopted during the fine-tuning phase. ; in, For a composite reward function; The dominant performance parameter (dominant performance item); For motion smoothing parameters (motion smoothing terms); These are boundary constraint parameters (boundary constraint terms).

[0052] Its dominant performance items Includes quadratic penalty for DC-side voltage error, THD (Total Harmonic Distortion) index for grid current, and over-limit step penalty. The definition is as follows: ; ; ; ; in, To encourage the reduction of harmonic distortion; To penalize DC voltage deviation; and for Quantity. In addition, When the error reaches the boundary, a heavy penalty is imposed, forcing the agent to quickly correct its actions and ensuring that it can quickly recover to a steady state under external disturbances.

[0053] Motion smoothing item Apply a quadratic penalty to the changes in motion between adjacent time steps to suppress high-frequency chattering of the output parameters: ; in, This is the proportionality coefficient; This represents the output value of the i-th action channel at the current time t. Let be the output value of the i-th action channel at the previous time t-1.

[0054] Boundary constraint terms Apply a step penalty when the action output approaches the saturation region to prevent the neural network output from saturating: ; To achieve a reasonable balance between motion smoothness and transient response capability, this application introduces a first-order low-pass filter in the motion output channel during the later stages of training after the policy has initially converged. ; in, The transfer function represents the relationship between the input and output of the filter in the complex frequency domain (s-domain); It is a time constant; is the Laplace operator, representing the complex frequency.

[0055] Among them, the time constant is set to The corresponding cutoff frequency is approximately 500 rad / s. This cutoff frequency is much higher than the voltage outer loop bandwidth (typically tens of rad / s) and the current inner loop bandwidth (typically hundreds of rad / s). The phase lag introduced within the target control bandwidth is negligible, while effectively attenuating high-frequency parameter chattering outside the bandwidth. After the introduction of the first-order low-pass filter, the TD3 agent continues training for hundreds of rounds under new action conditions to ensure that the policy fully adapts to the filtered closed-loop response characteristics, ensuring consistency between the training and deployment environments.

[0056] It should be noted that the embodiments provided in this application are only one possible implementation method, but are not limited to this only implementation method. Users can set their own implementation methods according to their needs.

[0057] The core consideration presented in this application is: if we simply rely on increasing the motion smoothness term... The penalty coefficient used to suppress high-frequency jitter will also penalize the rapid parameter adjustment actions necessary for the agent during load changes, resulting in sluggish dynamic response. A first-order low-pass filter is introduced as a physical-level smoothing mechanism, complementing the action regularization term in the reward function: the regularization term guides the strategy towards smoothness, while the first-order low-pass filter eliminates residual high-frequency components at the execution level. Their roles are clearly defined, jointly ensuring the smoothness and transient response capability of the output PI parameters.

[0058] Based on the above embodiments, as a preferred embodiment, it iteratively optimizes the parameters of the dual-delay deep policy agent through a pre-training phase and a fine-tuning phase to obtain the optimal dual-delay deep policy model, including: In the pre-training phase, a piecewise error penalty reward function is used to constrain the state of the power filtering system to within the safe operating range; During the fine-tuning phase, the pre-trained weights from the pre-training phase are loaded, and the system switches to the full reward function, which focuses on minimizing the total harmonic distortion rate. Based on two-stage hierarchical training using a piecewise error penalty reward function and a complete reward function, the optimal dual-delay deep policy model is obtained through iterative convergence.

[0059] In a specific embodiment, to initialize the TD3 agent strategy within a stable operating region, this application employs a two-stage training strategy based on course learning. First, in the pre-training stage, a piecewise penalty-type reward function is constructed, centered on DC-side voltage and harmonic current errors. A small penalty is applied when the error is within the normal range, while the penalty term weight is significantly increased when the error exceeds a preset threshold, thereby strengthening the suppression of out-of-limit states and driving the TD3 agent to quickly pull the system state back to a safe operating range. Based on this, the fine-tuning stage loads the network weights obtained from the pre-training (i.e., the pre-trained weights) and switches to a complete reward function primarily guided by minimizing total harmonic distortion (THD). This stage further refines the PI parameters based on the established stable control, thereby improving steady-state accuracy and dynamic performance while ensuring stable system operation, accelerating convergence, and effectively avoiding getting trapped in local optima.

[0060] After constructing the model as described above, it is necessary to verify the effectiveness of the TD3 adaptive control strategy (optimal dual-delay deep strategy model). This can be achieved by building a SAPF simulation model and controller training environment in appropriate software (e.g., MATLAB / Simulink). The Actor network (execution network) employs a two-layer fully connected hidden layer structure with 128 and 200 neurons respectively. The activation function for both hidden layers is ReLU, and the output layer uses the Tanh function to constrain actions to a bounded continuous interval. The Critic network (evaluation network) adopts a two-stream structure: the state input and action input undergo feature extraction through independent paths, are then fused through an additive layer, and output Q-value estimation through a fully connected layer. The hidden layers of both Critic networks also use the ReLU activation function. To enhance training stability, gradient clipping (threshold of 1) is introduced into the Critic network. Regarding the selection of exploration noise, although standard TD3 typically uses independent and identically distributed Gaussian noise, this application selects Ornstein-Uhlenbeck (OU) noise as the exploration mechanism. Compared to Gaussian noise, OU noise exhibits temporal autocorrelation, enabling the generation of continuous and smooth exploration trajectories in the parameter space. This avoids the interference of random jumps in PI parameters during adjacent training steps on the stability of the control loop, thereby improving the convergence stability of the training process. Then, the RL (Reinforcement Learning) control system is trained using preset parameters. When both indicators synergistically show that the policy network and the value function have fully converged, the TD3 agent obtains a stable and reliable control policy. Figures 5(a) and 5(b) are schematic diagrams of the TD3-based training convergence process provided in this application. In Figure 5(a), the average reward gradually increases from initial large fluctuations and converges, stabilizing after approximately 450 rounds. In Figure 5(b), the Q-value estimation of the dual-Critic network continuously improves and converges to a stable level. Furthermore, in Figure 5(a), the horizontal axis represents the number of iterations, and the vertical axis represents the average reward value; in Figure 5(b), the horizontal axis represents the number of iterations, and the vertical axis represents the Q-value.

[0061] Based on the above embodiments, as a preferred embodiment, the optimal dual-delay depth strategy model is integrated into the online invocation of the power filtering system, including: Load the optimal dual-delay deep policy model that has converged during offline training, retain the execution network built into the optimal dual-delay deep policy model for real-time inference, and remove the evaluation network and target network used for offline iterative training. The processed optimal dual-delay deep strategy model is then integrated into the control program of the power filtering system.

[0062] The output target PI parameter, used to replace the fixed PI parameter of the dual closed-loop control unit, includes: The operating status data of the power filtering system is acquired and input into the optimal dual-delay deep strategy model; The execution network in the optimal dual-delay deep strategy model outputs target PI parameters that are adapted to the current operating conditions in real time. The target PI parameter is used to replace the preset fixed PI parameter inside the dual closed-loop control unit.

[0063] Specifically, based on the updated target PI parameters, the corresponding power transistor drive pulse is generated by combining decoupling control and space vector pulse width modulation, including: Based on the updated target PI parameters, determine the corresponding modulation voltage command; The modulation voltage command is decoupled to obtain the decoupled voltage modulation signal; The decoupled voltage modulation signal is input to the pulse width modulation drive unit so that the corresponding power transistor drive pulse can be generated by space vector pulse width modulation.

[0064] In a specific embodiment, after offline training is completed, the system enters the online deployment and real-time operation phase of the TD3 adaptive PI control (optimal dual-delay deep strategy model). The entire system achieves lightweight and high-efficiency adaptive control under actual operating conditions. During the online deployment phase, only the trained Actor network (execution network) is loaded, discarding the Critic network (evaluation network) and the target network, significantly reducing online computational resource consumption. The TD3 agent collects the system state information in real time, including d-axis current tracking error, q-axis current tracking error, DC bus voltage error, and DC voltage change rate. The resulting state vector is input into the Actor network, which directly outputs six optimal PI parameters (target PI parameters) through forward propagation. These target PI parameters are substituted into the SAPF dual closed-loop control system, replacing the traditional fixed PI parameters. Combined with voltage feedforward and decoupling compensation terms, a precise voltage modulation signal is generated. After space vector pulse width modulation processing, the voltage modulation signal generates corresponding power transistor drive pulses, directly controlling the switching of the three-phase bridge inverter. This ultimately achieves rapid and precise adjustment of the compensation current, completing adaptive compensation for grid harmonics and stable control of the DC bus voltage.

[0065] Therefore, this application, by constructing a dual-delay deep strategy adaptive optimization model based on Markov decision-making and an optimization method of two-stage offline training, combined with a control method of real-time dynamic adjustment of PI parameters, overcomes the shortcomings of traditional dual closed-loop control parameters being fixed and unable to adapt to dynamic power grid conditions. It can match the optimal control parameters in real time according to power grid load fluctuations and complex operating conditions, effectively making up for the defects of slow dynamic response and insufficient harmonic compensation accuracy of traditional control strategies. It significantly improves the dynamic response performance and harmonic compensation accuracy of parallel active power filters under variable operating conditions, effectively avoids the control instability problem that is prone to occur under extreme and fluctuating conditions with fixed parameters, and greatly enhances the environmental adaptability and operational reliability of the power filtering system.

[0066] Figure 6 A structural diagram of an electronic device provided in another embodiment of this application, such as... Figure 6 As shown, the electronic device includes: a memory 20 for storing computer programs; The processor 21 is used to execute a computer program to implement the steps of the control method for the parallel active power filter as described in the above embodiments.

[0067] The electronic devices provided in this embodiment may include, but are not limited to, smartphones, tablets, laptops, or desktop computers.

[0068] The processor 21 may include one or more processing cores, such as a quad-core processor or an octa-core processor. The processor 21 may be implemented using at least one of the following hardware forms: Digital Signal Processor (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 21 may also include a main processor and a coprocessor. The main processor, also known as the Central Processing Unit (CPU), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, the processor 21 may integrate a Graphics Processing Unit (GPU), which is responsible for rendering and drawing the content to be displayed on the screen. In some embodiments, the processor 21 may also include an Artificial Intelligence (AI) processor, which is used to handle computational operations related to machine learning.

[0069] The memory 20 may include one or more computer-readable storage media, which may be non-transitory. The memory 20 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory devices. In this embodiment, the memory 20 is used to store at least the following computer program 201, which, after being loaded and executed by the processor 21, is capable of implementing the relevant steps of the control method for the parallel active power filter disclosed in any of the foregoing embodiments. In addition, the resources stored in the memory 20 may also include an operating system 202 and data 203, and the storage method may be temporary storage or permanent storage. The operating system 202 may include Windows, Unix, Linux, etc.

[0070] In some embodiments, the electronic device may further include a display screen 22, an input / output interface 23, a communication interface 24, a power supply 25, and a communication bus 26.

[0071] Those skilled in the art will understand that Figure 6 The structures shown do not constitute a limitation on electronic devices and may include more or fewer components than those shown.

[0072] The electronic device provided in this application includes a memory and a processor. When the processor executes the program stored in the memory, it can implement the above-described control method for the parallel active power filter.

[0073] Finally, this application also provides an embodiment corresponding to a computer-readable storage medium. The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps described in the above method embodiments.

[0074] It is understood that if the methods in the above embodiments are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and executes all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0075] The control method, electronic device, and medium for a parallel active power filter provided in this application have been described in detail above. The various embodiments in the specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and relevant parts can be referred to in the method section. It should be noted that those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of the claims of this application.

[0076] It should also be noted that, in this specification, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

Claims

1. A control method for a parallel active power filter, characterized in that, A method is applied to a power filtering system comprising a three-phase three-wire circuit topology, a sampling and detection unit, and a pulse width modulation (PWM) driving unit, wherein the power filtering system is configured with a dual-closed-loop control unit and a dual-delay reinforcement learning adaptive PI optimization unit, and the control hierarchy of the power filtering system adopts a four-level hierarchical architecture consisting of a voltage outer loop, a current inner loop, a dual-delay deep deterministic strategy parameter optimization layer, and a PWM execution layer. The method includes: The sampling and detection unit collects system operation data of the power filtering system. Based on the system operation data, a mathematical model of a parallel active power filter in a synchronously rotating dq coordinate system is constructed to obtain the reference compensation current. Based on the mathematical model of the parallel active power filter and the reference compensation current, a corresponding Markov decision-making dual-delay deep strategy adaptive optimization model is constructed. Based on the Markov decision-making dual-delay deep policy adaptive optimization model, a two-stage course learning mechanism is used to train the dual-delay deep policy agent offline. The parameters of the dual-delay deep policy agent are iteratively optimized through the pre-training stage and the fine-tuning stage to obtain the optimal dual-delay deep policy model. The optimal dual-delay depth strategy model is connected to the power filtering system for online invocation, and the target PI parameter is output to replace the fixed PI parameter of the dual closed-loop control unit. Based on the updated target PI parameters, the corresponding power transistor drive pulse is generated by combining decoupling control and space vector pulse width modulation.

2. The control method for a parallel active power filter according to claim 1, characterized in that, The construction of a mathematical model for a parallel active power filter in a synchronously rotating dq coordinate system to obtain the reference compensation current includes: Based on the three-phase grid voltage, three-phase load current, filter compensation current and DC bus voltage in the system operation data, establish the mathematical model circuit equation of the parallel active power filter in the three-phase stationary coordinate system. The electrical quantities in the circuit equations of the mathematical model of the parallel active power filter are converted into electrical quantities in a two-phase stationary coordinate system by Clark coordinate transformation, and then converted into DC electrical quantities in the synchronous rotating dq coordinate system by Park coordinate transformation, so as to construct the mathematical model of the parallel active power filter. Based on the DC electrical quantities in the synchronous rotating dq coordinate system, a current decoupling model corresponding to the mathematical model of the parallel active power filter is constructed. The d-axis current coupling component and the q-axis current coupling component are decoupled to obtain an independent d-axis current component and an independent q-axis current component without coupling. Based on the independent d-axis current component and the independent q-axis current component, and combined with the voltage regulation of the outer voltage loop, the corresponding reference compensation current is obtained through the command current generation model; wherein, the reference compensation current includes: d-axis reference compensation current and q-axis reference compensation current.

3. The control method for a parallel active power filter according to claim 1, characterized in that, The construction of the corresponding Markov decision-making dual-delay deep policy adaptive optimization model includes: Based on the mathematical model of the parallel active power filter and the reference compensation current, and combined with the operating deviation characteristics of the power filter system, a four-dimensional state vector with the deviation signal of the power filter system as the core and a state space corresponding to the four-dimensional state vector are constructed. Based on the state space and combined with the adjustable parameter characteristics of the dual closed-loop control unit, all the core PI parameters to be optimized in the dual closed-loop control unit are constructed into a six-dimensional continuous action vector, and an action space corresponding to the six-dimensional continuous action vector is constructed. Based on the state space and the action space, and combining the harmonic compensation of the parallel active power filter, the voltage regulation control target, and the system safety operation constraints of the power filter system, a composite reward function containing dominant performance terms, action smoothing terms, and boundary constraint terms is configured to construct a corresponding reward mechanism. Based on the aforementioned reward mechanism, a dual anti-jitter mechanism combining reward function constraints and hardware filtering is constructed to build the corresponding dual-delay deep policy adaptive optimization model for the Markov decision.

4. The control method for a parallel active power filter according to claim 3, characterized in that, The expression for the four-dimensional state vector is: ； The expression for the six-dimensional continuous action vector is: ； The expression for the composite reward function is: ； in, The four-dimensional state vector; This refers to the d-axis harmonic current error. This refers to the q-axis harmonic current error. This refers to the DC-side voltage error. This represents the rate of change of DC-side voltage. For transpose; The six-dimensional continuous action vector; and These are the two parameters of the outer voltage loop; , , , These are the four parameters of the inner current loop; The composite reward function; The primary performance parameter; For motion smoothing parameters; These are the boundary constraint parameters.

5. The control method for a parallel active power filter according to claim 1, characterized in that, The process of iteratively optimizing the parameters of the dual-latency deep policy agent through pre-training and fine-tuning phases to obtain the optimal dual-latency deep policy model includes: In the pre-training phase, a piecewise error penalty reward function is used to constrain the state of the power filtering system to a safe operating range; In the fine-tuning phase, the pre-trained weights from the pre-training phase are loaded, and the system switches to the full reward function with the minimization of total harmonic distortion as its core function. Based on the two-stage hierarchical training of the piecewise error penalty reward function and the complete reward function, the optimal dual-delay deep policy model is obtained through iterative convergence.

6. The control method for a parallel active power filter according to claim 1, characterized in that, The step of integrating the optimal dual-delay depth strategy model into the power filtering system for online invocation includes: Load the optimal dual-delay deep policy model that has converged during offline training, retain the execution network built into the optimal dual-delay deep policy model for real-time inference, and remove the evaluation network and target network used for offline iterative training; The processed optimal dual-delay depth strategy model is then integrated into the control program of the power filtering system.

7. The control method for a parallel active power filter according to claim 6, characterized in that, The output target PI parameter, used to replace the fixed PI parameter of the dual closed-loop control unit, includes: The operating status data of the power filtering system is obtained and input into the optimal dual-delay depth strategy model; The execution network in the optimal dual-delay depth strategy model outputs the target PI parameter adapted to the current operating condition in real time. The target PI parameter is used to successively replace the fixed PI parameter preset inside the dual closed-loop control unit.

8. The control method for a parallel active power filter according to any one of claims 1 to 7, characterized in that, The step of generating the corresponding power transistor drive pulse based on the updated target PI parameters, combined with decoupling control and space vector pulse width modulation, includes: Based on the updated target PI parameters, the corresponding modulation voltage command is determined; The modulation voltage command is decoupled to obtain a decoupled voltage modulation signal; The decoupled voltage modulation signal is input to the pulse width modulation drive unit so that the corresponding power transistor drive pulse is generated by space vector pulse width modulation.

9. An electronic device, characterized in that, Includes memory used to store computer programs; A processor, configured to execute the computer program to implement the steps of the control method for a parallel active power filter as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the control method for the parallel active power filter as described in any one of claims 1 to 8.

Images