A wide-range multi-level converter control method based on improved integral sliding mode control

Abstract

The application discloses a wide-range multi-level converter control method based on improved integral sliding mode control, which comprises the following steps: establishing a fractional-order multi-modal unified state space model of a multi-level converter control system, and constructing a distributed collaborative predictive control architecture; adopting a composite nonlinear integral sliding mode controller as an execution layer of the distributed collaborative predictive control architecture; adopting a radial basis function neural network to provide an accurate feedforward compensation signal for the multi-level converter control system; based on stability analysis, proposing an adaptive online setting rule of an effective integral coefficient of a composite nonlinear integral sliding mode surface S of the composite nonlinear integral sliding mode controller, and realizing parameter self-healing of the multi-level converter control system through a parameter identification module. The wide-range multi-level converter control method has the advantages of being capable of eliminating steady-state error, significantly improving output voltage precision and response speed, dynamically optimizing robustness and dynamic performance of the multi-level converter control system and the like.

Description

A wide-range multilevel converter control method based on improved integral sliding mode control Technical Field

[0001] This invention relates to a control technique for multilevel converters, and more particularly to a wide-range multilevel converter control method based on improved integral sliding mode control. Background Technology

[0002] As a core component at the end of the power generation industry chain, multilevel converters face increasingly higher performance requirements with rising voltage levels. Due to voltage stress limitations on switching devices, the commonly used three-level current source converter topology can no longer meet the demands of higher voltage levels. Multilevel converters, with their advantages of lower power requirements for switching devices, higher output voltage waveform quality, and lower electromagnetic interference, have become a research hotspot in academia.

[0003] Multilevel converters, as core devices for medium-to-high power density power conversion, have become key equipment in fields such as new energy power generation, smart grids, and industrial drives due to their inherent advantages in reducing switching stress and improving output waveform quality. With the continuous expansion of application scenarios, modern power electronic systems place more stringent demands on converter performance: not only must they operate stably over a wide input voltage range and under complex load conditions, but they must also achieve rapid and smooth switching between multiple operating modes, while simultaneously considering conversion efficiency, dynamic response, and electromagnetic compatibility.

[0004] Faced with the aforementioned multi-objective, strongly coupled control challenges, traditional linear control strategies are essentially based on the control theory of linear time-invariant systems. However, multilevel converters are typical complex systems with nonlinear, variable structure, and time-varying parameters. Under a wide range of operating conditions, their shortcomings—slow dynamic response and insufficient robustness to perturbations of internal system parameters and external load disturbances—become increasingly prominent, making it difficult to balance fast dynamic response and high-precision steady-state performance. Especially during the transient process of switching operating modes, traditional dual-loop PID control is prone to generating large voltage and current overshoots, and may even induce oscillations, threatening system safety.

[0005] In order to improve the control performance of nonlinear systems, sliding mode variable structure control, as a robust nonlinear control method, has been introduced into the field of power electronic converters. The core idea of ​​sliding mode variable structure control is: by designing a specific sliding surface, the system state trajectory can still be attracted and maintained near the sliding surface for a finite time when it is disturbed, thereby giving the system invariance in the sliding mode and complete robustness to matched disturbances. However, when classical sliding mode control is directly applied to multilevel converters, there are several inherent defects: (1) Due to the discontinuity of the sign function in the control law, high-frequency chattering will be triggered, which will increase the switching loss and stress of the switching devices. It may also excite unmodeled high-frequency dynamics in the system, resulting in an increase in the harmonic content of the output voltage and deterioration of electromagnetic interference performance; (2) Traditional linear sliding surfaces usually rely on the linear combination of errors. When the system state is close to the equilibrium point, the convergence speed becomes slower and the steady-state error cannot be fundamentally eliminated, which limits the control accuracy.

[0006] To address the steady-state error problem, an integral term of the error is introduced into the sliding surface to construct integral sliding mode control, theoretically achieving zero steady-state error tracking of step references or constant disturbances. However, the simple introduction of the integral term also brings new challenges: integral saturation deteriorates the dynamic response, and the performance of fixed-parameter integral sliding mode controllers remains limited when dealing with complex nonlinear and time-varying disturbances. Furthermore, most existing research is based on system models using integer-order calculus for controller design, while actual components such as inductors and capacitors exhibit fractional-order characteristics over a wide frequency range. Simple integer-order models cannot accurately describe the high-frequency dynamics of the system, limiting the theoretical upper limit of controller performance.

[0007] Furthermore, existing control schemes typically treat mode switching decisions, disturbance compensation, and the underlying tracking controller as separate or serial problems. Mode switching is often based on simple hysteresis comparisons, lacking prediction of future states, which can easily lead to frequent or inappropriate switching. While disturbance observers are effective, their accuracy and bandwidth are limited when dealing with high-frequency, discontinuous disturbances. The lack of coordination between the underlying tracking controller and the upper-level energy management or optimization objectives makes it difficult to achieve globally optimal performance. With the development of artificial intelligence and distributed computing technologies, deeply integrating these advanced concepts with traditional sliding mode control to construct a unified control framework with a multi-layered intelligent structure that considers perception, decision-making, and execution has become a key direction for overcoming existing performance bottlenecks.

[0008] Therefore, there is an urgent need to design a multilevel converter control method that can solve the key problems in the existing technology, such as inaccurate modeling, unintelligent decision-making, and under-adaptive compensation. Summary of the Invention

[0009] To avoid the shortcomings of the existing technologies, this invention provides a control method for a wide-range multilevel converter based on improved integral sliding mode control, in order to solve the problems of multi-objective collaborative optimization of wide-range multilevel converters in terms of dynamic response, steady-state accuracy, mode switching smoothness, robustness, and other aspects.

[0010] The present invention adopts the following technical solution to solve the technical problem.

[0011] This invention provides a wide-range multilevel converter control method based on improved integral sliding mode control, comprising the following steps:

[0012] Step 1: Modeling and Intelligent Prediction Steps; Establish a fractional-order multimodal unified state-space model of the multilevel converter control system and construct a distributed cooperative predictive control architecture;

[0013] Step 2: Design a composite nonlinear integral sliding mode controller and add an online learning compensation mechanism; use the composite nonlinear integral sliding mode controller as the execution layer of the distributed cooperative predictive control architecture; use a radial basis function neural network to provide accurate feedforward compensation signals for the multilevel converter control system;

[0014] Step 3: Stability analysis and design of active fault-tolerant control mechanism; based on stability analysis, an adaptive online tuning rule for the effective integral coefficient of the composite nonlinear integral sliding surface S of the composite nonlinear integral sliding mode controller is proposed, and parameter self-healing of the multilevel converter control system is realized through the parameter identification module.

[0015] The feature of the wide-range multilevel converter control method based on improved integral sliding mode control of the present invention is also that:

[0016] Further, step 1 includes the following steps:

[0017] Step 11: Establish a fractional-order multimodal unified state-space model based on fractional-order calculus;

[0018] Step 12: Mode switching predictor design steps;

[0019] Step 13: Construction steps of the distributed collaborative predictive control architecture.

[0020] Furthermore, in step 11, a fractional-order multimodal unified state-space model is established. First, a fractional-order state variable vector x is defined. α See formula (1) below;

[0021] (1);

[0022] In formula (1), It is a fractional-order state variable vector; For fractional differential operators; i L V is the current in inductor L; o V is the output voltage of the multilevel converter. cf This is the voltage across the flying capacitor C; is the fractional order, and theoretically the range of the fractional differential order is (0~1); T is the vector transpose.

[0023] Furthermore, in step 11, a fractional-order state-space equation is used to express the fractional-order multimodal unified state-space model of the multilevel converter control system.

[0024] Furthermore, the fractional-order state-space equation is shown in the following formula (2);

[0025] (2);

[0026] In formula (2), A fractional-order state variable vector of Derivative order; B is the system fractional-order state matrix of the multilevel converter control system; α The input fractional control matrix is ​​u; the system input vector of the multilevel converter control system is u. Let t be the vector of lumped nonlinear and perturbation functions, and t be time.

[0027] Furthermore, in step 12, the mode switching predictor includes a reward function r. t .

[0028] Furthermore, in step 13, the distributed collaborative predictive control architecture includes a prediction layer, an optimization layer, and an execution layer.

[0029] Furthermore, in step 2, the expression for the composite nonlinear integral sliding surface S of the composite nonlinear integral sliding mode controller is shown in the following formula (6).

[0030] (6);

[0031] In formula (6), S is the function value of the composite nonlinear integral sliding surface; k1, k2, k3, and k4 are sliding surface coefficients, all of which are positive numbers; e i For inductor current i L The tracking error and e i = i ref - i L i ref For inductor current i L Reference value; e v Output voltage V o Tracking error, ev = V ref -V o V o V is the output voltage of the multilevel converter. ref V is the output voltage of the multilevel converter. o Reference value; and It is a nonlinear saturation function used for smooth control. and These are nonlinear saturation functions. and Shape parameters, This represents the integral variable.

[0032] Furthermore, in step 3, the Lyapunov function is used for stability analysis.

[0033] Furthermore, in step 3, the calculation formula for parameter self-healing is shown in the following formula (13).

[0034] (13);

[0035] In formula (13), y(k) is the parameter estimation vector of the multilevel converter control system at time k; K(k) is the gain vector of the recursive least squares method; y(k) is the measurement output of the multilevel converter control system at time k. Let be the regression vector at time k, which consists of input and output data, and T be the matrix transpose.

[0036] Compared with existing technologies, the beneficial effects of this invention are reflected in:

[0037] This invention discloses a wide-range multilevel converter control method based on improved integral sliding mode control, comprising: establishing a fractional-order multimodal unified state-space model of the multilevel converter control system and constructing a distributed cooperative predictive control architecture; using a composite nonlinear integral sliding mode controller as the execution layer of the distributed cooperative predictive control architecture; using a radial basis function neural network to provide accurate feedforward compensation signals for the multilevel converter control system; proposing an adaptive online tuning rule for the effective integral coefficients of the composite nonlinear integral sliding surface S of the composite nonlinear integral sliding mode controller based on stability analysis, and realizing parameter self-healing of the multilevel converter control system through a parameter identification module.

[0038] The wide-range multilevel converter control method based on improved integral sliding mode control of the present invention has the advantages of eliminating steady-state error, significantly improving output voltage accuracy and response speed, and dynamically optimizing the robustness and dynamic performance of the multilevel converter control system. Attached Figure Description

[0039] Figure 1 is a control flowchart of a wide-range multilevel converter control method based on improved integral sliding mode control according to the present invention;

[0040] Figure 2 compares the output voltage waveforms of PID, traditional SMC, and the present invention under sudden load changes. PID has a slow response and significant overshoot, while traditional SMC is fast but accompanied by high-frequency chattering. The present invention has the fastest waveform response, almost no overshoot or chattering, and smooth voltage recovery, demonstrating its superior transient performance and stability under extreme dynamic conditions.

[0041] Figure 3 simulates the uncertainties in the actual system, with significant perturbations in the inductance and capacitance parameters;

[0042] Figure 4 simulates the uncertainty in the actual system, namely the periodic fluctuation of the input voltage;

[0043] Figure 5 shows that under the combined disturbance of large perturbations of inductor and capacitor parameters and periodic fluctuations of input voltage, the output voltage waveforms of PID and traditional SMC both show obvious amplitude fluctuations and phase shifts, while the output voltage of the method of the present invention remains highly stable with minimal tracking error.

[0044] Figure 6 illustrates that the adaptive gain of the present invention can be adjusted smoothly and in real time according to the system state, dynamically optimizing the controller performance;

[0045] Figure 7 illustrates that the damping coefficient of the present invention can be adjusted smoothly and in real time according to the system state, dynamically optimizing the controller performance;

[0046] Figure 8 simulates the input voltage waveform of voltage dips and harmonic interference commonly found in power grids;

[0047] Figure 9 shows that, under the input voltage conditions in Figure 8, the converter voltage output fluctuation is minimal under the control of the method of the present invention, and a high-quality sinusoidal voltage can still be output.

[0048] Figure 10 is a partial enlarged view of Figure 9, further showing that its output voltage steady-state ripple is minimal;

[0049] Figure 11 illustrates the process of online adaptive gain and damping coefficient adjustment using the method of the present invention.

[0050] Figure 12 shows the dynamic change of the sliding surface in the method of the present invention. Its sliding surface variable is constrained to near zero value, and the chattering is minimized compared with PID control and traditional SMC control.

[0051] Figure 13 shows a comprehensive quantitative comparison using a multi-index bar chart, demonstrating that the present invention is significantly superior to PID and traditional SMC in all key performance indicators.

[0052] Figure 14 compares the tracking errors of three methods (PID control, traditional SMC control, and the control method of the present invention) in wide-range voltage regulation. The method of the present invention has the smallest error amplitude and the fastest convergence speed.

[0053] Figure 15 is a comparison of the switching process of the three operating modes when adjusting voltage over a wide range.

[0054] Figure 16 shows how the adaptive gain and damping coefficient of the method of the present invention continuously change during wide-range voltage regulation to match the system operating point.

[0055] Figure 17 shows a comparison of the three methods for adjusting overshoot during wide-range voltage regulation.

[0056] Figure 18 shows a comparison of the smoothness control of the three methods during wide-range voltage regulation.

[0057] Figure 19 shows a comparison of the number of times the converter operating mode switches during the wide-range voltage regulation process;

[0058] Figure 20 shows the output voltage error of the multilevel converter using the method of the present invention when simulating the coordinated operation of photovoltaic and energy storage. The method of the present invention has the lowest amplitude and the smoothest fluctuation, which improves the grid connection and power supply quality of the multilevel converter in the new energy system.

[0059] Figure 21 is a schematic diagram of a wide-range multilevel converter circuit.

[0060] The present invention will be further described below through specific embodiments and in conjunction with the accompanying drawings. Detailed Implementation

[0061] Referring to Figure 1, this invention provides a wide-range multilevel converter control method based on improved integral sliding mode control, comprising the following steps:

[0062] Step 1: Modeling and Intelligent Prediction Steps; Establish a fractional-order multimodal unified state-space model of the multilevel converter control system and construct a distributed cooperative predictive control architecture;

[0063] The mathematical expression of the fractional-order multimodal unified state-space model described above is used to describe the multilevel converter control system (hereinafter referred to as "the system" in this invention); the above distributed cooperative predictive control architecture is used to realize the forward-looking decision-making of the control process of the multilevel converter;

[0064] Figure 21 is a circuit diagram of the wide-range multilevel converter of the present invention, and the connection relationship of each component is shown in Figure 21. The wide-range multilevel converter includes an inductor L, a power switch S, a series output voltage divider capacitor, and a flying capacitor C. The flying capacitor is used to clamp the voltage of the switch and generate an intermediate level. The inductor L is an energy storage element, and energy transfer is regulated by controlling its charging and discharging. The flying capacitor C is used to automatically balance the midpoint voltage of the bridge arm in the multilevel topology and is the key to generating high-quality multilevel output. Its voltage is one of the state variables that needs to be precisely controlled. The resistor R refers to the equivalent series resistance of the inductor, which is a non-ideal parameter that causes conduction losses and affects dynamic characteristics.

[0065] In Figure 21, V in For input voltage source; C in For input filtering capacitors. S1, S2, S3, S4, ..., Sz-1, Sz are the Z power switches that form a multi-level switching network; L1, L2, ..., L z / 2 Z / 2 energy storage inductors; inductor L is the equivalent inductance of each bridge arm, and energy transfer is regulated by controlling the charging and discharging of inductor L. Flying capacitors C include C1, C2, C3, C4, ..., Cz-1, Cz. The Z flying capacitors C are key components of the multi-level topology, used to clamp the voltage stress of the switching transistors and assist in generating multi-level output voltage waveforms. Co is the output filter capacitor, used to filter out output voltage ripple and maintain output voltage stability. R refers to the equivalent series resistance of the inductor, which is a non-ideal parameter that causes conduction losses and affects dynamic characteristics; Vo is the converter output voltage.

[0066] Step 2: Design a composite nonlinear integral sliding mode controller and add an online learning compensation mechanism; use the composite nonlinear integral sliding mode controller as the execution layer of the distributed cooperative predictive control architecture; use a radial basis function neural network to provide accurate feedforward compensation signals for the multilevel converter control system;

[0067] Step 2 involves designing a composite nonlinear integral sliding mode controller for the execution layer that combines fast dynamics, high steady-state accuracy, and strong robustness, and integrating an online learning compensation mechanism into the composite nonlinear integral sliding mode controller.

[0068] Step 3: Stability analysis and design of active fault-tolerant control mechanism; based on stability analysis, an adaptive online tuning rule for the effective integral coefficient of the composite nonlinear integral sliding surface S of the composite nonlinear integral sliding mode controller is proposed, and parameter self-healing of the multilevel converter control system is realized through the parameter identification module.

[0069] Step 3 is used to theoretically ensure the system stability of the multilevel converter control system of the present invention, design a fault response mechanism, and verify the overall performance of the multilevel converter control method of the present invention through simulation.

[0070] This invention discloses a wide-range multilevel converter control method based on improved integral sliding mode control. First, a state-space model suitable for unified multi-mode analysis is constructed, and sliding mode control laws precisely matching Buck and Boost operating modes are derived to achieve coordinated control of output voltage and inductor current. The core lies in introducing a composite nonlinear integral sliding surface, which dynamically optimizes the robustness and dynamic performance of the multilevel converter control system by online adaptive adjustment of the sliding surface gain and damping coefficient. This method not only eliminates steady-state errors and significantly improves output voltage accuracy and response speed, but also exhibits strong anti-disturbance capability and robustness under complex operating conditions such as load abrupt changes, parameter perturbations, input voltage fluctuations, and wide-range voltage regulation. Furthermore, by dynamically correcting the duty cycle and optimizing the boundary layer, the switching process is smoothed, effectively suppressing high-frequency chattering in traditional sliding mode control while reducing power device switching losses. Simulation and comparative results show that this invention achieves comprehensive performance improvements in control accuracy, dynamic quality, operating efficiency, and system reliability in extreme dynamic response, industrial harmonic environment adaptability, mode-uninterrupted switching, and photovoltaic-storage system applications.

[0071] In specific implementation, step 1 includes the following steps:

[0072] Step 11: Establish a fractional-order multimodal unified state-space model based on fractional-order calculus;

[0073] Step 12: Mode switching predictor design steps;

[0074] Step 13: Construction steps of the distributed collaborative predictive control architecture.

[0075] In specific implementation, in step 11, a fractional-order multimodal unified state-space model is established. First, a fractional-order state variable vector x is defined. α See formula (1) below;

[0076] (1);

[0077] In formula (1), It is a fractional-order state variable vector; For fractional differential operators; i L V is the current in inductor L; o V is the output voltage of the multilevel converter. cf This is the voltage across the flying capacitor C; is the fractional order, and theoretically the range of the fractional differential order is (0~1); T is the vector transpose.

[0078] Theoretical fractional derivative order The value range is (0~1).

[0079] In specific implementation, in step 11, a fractional-order state-space equation is used to express the fractional-order multimodal unified state-space model of the multilevel converter control system.

[0080] In specific implementation, the fractional-order state-space equation is shown in the following formula (2);

[0081] (2);

[0082] In formula (2), Let x be a fractional-order state variable vector. α of Derivative order; This is the system fractional-order state matrix of the multilevel converter control system; The input fractional control matrix is ​​u; the system input vector of the multilevel converter control system is u. Let t be the vector of lumped nonlinear and perturbation functions, and t be time.

[0083] Considering the non-ideal characteristics of inductance L and flying capacitor C, fractional calculus is used to establish an accurate fractional multimodal unified state space model; the fractional state space equation of formula (2) is used to express the fractional multimodal unified state space model of the multilevel converter control system.

[0084] Equation (2) is the fractional-order state-space equation of the multilevel converter control system, used to describe the fractional-order state variable vector x. α The dynamic evolution law; in formula (2), the multilevel converter control system is dynamically decomposed into a linear part and a lumped nonlinear disturbance part determined by the circuit parameters, which provides a mathematical calculation basis for subsequent separation design of equivalent control and disturbance compensation.

[0085] Based on formula (2), a fractional-order model Z of the inductor is established. L (s), see formula (3) below;

[0086] (3);

[0087] In formula (3), The fractional order is the theoretical range of the fractional differential order (0~1), but here it is taken as (0.98~1.02), which is the approximate fractional order of the actual inductor over a wide frequency range, reflecting its non-ideal characteristics. L is the inductance; s is the Laplace operator, i.e., the complex frequency variable, used to describe the fractional-order model Z of the inductor. L (s).

[0088] The fractional-order model Z of the inductor in formula (3) L (s) is the fractional order in formulas (1) and (2). The physical source of, namely, in formulas (1) and (2) It is calculated using formula (3). As can be seen from formula (3), the actual inductance L is not an ideal integer-order element in the wide frequency range; its characteristics require fractional-order calculations. This description demonstrates the necessity of using fractional calculus in this invention.

[0089] In specific implementation, in step 12, the mode switching predictor includes a reward function r. t .

[0090] To avoid transient impacts during mode switching, a deep Q-network-based intelligent mode switching predictor is designed when designing a deep reinforcement learning-based mode switching predictor. This predictor intelligently determines the timing of Buck / Boost mode switching. The state space S of the mode switching predictor is... t Includes electrical states and their rates of change, action space, and is processed through a reward function r. t Taking into account the tracking error V o -V ref Current change rate Loss P loss Wait, the reward function r t See formula (4) below;

[0091] (4);

[0092] In formula (4), r t Let w1, w2, w3, and w4 be the reward function value of the mode switching predictor at time t; w1, w2, w3, and w4 are the reward functions r, respectively. t Tracking error V o -V ref Current change rate Loss P loss and the effective value of the switching current I switch Weighting coefficients; V o V is the output voltage of the multilevel converter. ref P is the reference value for the output voltage of the multilevel converter. loss For the system power loss of the multilevel converter control system; I switch The effective value of the switching current of a multilevel converter is a parameter characterizing switching stress; the rate of change of current. Let i be the current in inductor L L The time derivative.

[0093] Formula (4) is the reward function of the mode switching predictor, which is used to quantify the complex multi-objective control problem into an optimizable scalar reward, thereby guiding the mode switching predictor to learn the optimal Buck / Boost mode switching strategy to achieve optimal global performance.

[0094] In specific implementation, step 13, the distributed collaborative predictive control architecture includes a prediction layer, an optimization layer, and an execution layer.

[0095] To generate an optimal reference trajectory for the underlying tracking controller in each control cycle of the multilevel converter control system, a rolling optimization performance index (such as output tracking error V) is designed for generalized predictive control. o -V ref Controlling incremental smoothness and current change rate (etc.) To ensure that the multi-step output can be predicted in each control cycle, and by optimizing the control sequence for a period of time in the future, the multilevel converter control system can output a smooth and accurate reference trajectory, and constrain drastic changes in control quantities such as duty cycle, reference voltage, and current command. A three-layer distributed architecture of prediction-optimization-execution is proposed. The prediction-optimization-execution three-layer distributed architecture includes a prediction layer, an optimization layer, and an execution layer.

[0096] In the prediction layer, a generalized predictive control algorithm is used to continuously optimize the performance indicators for the next N steps. The optimization layer is used to solve for the optimal control sequence (i.e., the fractional-order state variable vector) online using sequential quadratic programming. The obtained optimal control sequence is used as the optimization reference trajectory (i.e., the fractional-order state variable vector). Output. The execution layer is used to receive the reference trajectory output by the optimization layer, and the composite nonlinear integral sliding mode controller designed in step 2 of this invention tracks the reference trajectory quickly and accurately. The cost function J(k) of the generalized predictive control algorithm of the prediction layer is shown in the following formula (5);

[0097] (5);

[0098] In formula (5), J(k) is the cost function of generalized predictive control at time k; y(k+i|k) is the predicted value of the system output at time k+i in the future; and r(k+i) is the reference value of the output at time k+i in the future. Let Q be the control input increment planned at time k+j in the future; Q be the weight matrix of the error term in predictive control; R be the weight matrix of the control increment term in predictive control; N be the prediction time domain length, 1≤i≤N; M be the control time domain length, 0≤j≤M-1. J(k) is the optimization objective function used to evaluate the tracking performance.

[0099] Formula (5) receives the reward function r from the mode-switching predictor in Formula (4). t Based on the fractional-order multimodal unified state-space model of formula (2), the reference trajectory is predicted and optimized, and the optimized reference trajectory (i.e., the fractional-order state variable vector x) is output. α The target to be tracked by the subsequent composite nonlinear integral sliding mode controller can achieve decoupling and coordination between hierarchical optimization and fast tracking. The cost function J(k) is obtained by multi-step prediction according to formula (2). It controls the increment, and its relationship with u in formula (2) is as follows: That is, we get: .

[0100] In specific implementation, in step 2, the expression of the composite nonlinear integral sliding surface S of the composite nonlinear integral sliding mode controller is shown in the following formula (6).

[0101] (6);

[0102] In formula (6), S is the function value of the composite nonlinear integral sliding surface; k1, k2, k3, and k4 are sliding surface coefficients, all of which are positive numbers; e i For inductor current i L The tracking error and e i = i ref - i L i ref For inductor current i L Reference value; e v Output voltage V o Tracking error, e v = V ref -V o V o V is the output voltage of the multilevel converter. ref V is the output voltage of the multilevel converter. o Reference value; and It is a nonlinear saturation function used for smooth control. and These are nonlinear saturation functions. and Shape parameters, This represents the integral variable.

[0103] Formula (6) combines the target reference trajectory r(k+i) from Formula (5) with the adaptive parameter η from below. This is used to calculate the control inputs for fractional-order adaptive reaching laws and composite nonlinear integral sliding mode control, and also serves as the feedback signal for weight updates in the radial basis function neural network. Here, k3 and k4 are the effective integral coefficients of the composite nonlinear integral sliding surface.

[0104] Step 2 includes the following steps 21 and 22;

[0105] Step 21: Design the composite nonlinear integral sliding surface S and the composite nonlinear integral sliding control law g;

[0106] Step 211: Design a composite nonlinear integral sliding surface S containing a nonlinear saturation function;

[0107] To synthesize the current error, voltage error, and the integral term of the voltage error processed by a nonlinear function into a single scalar variable, a composite nonlinear integral sliding surface S is designed. The tanh function and... The function provides smooth control and suppresses chattering, while the integral term ensures that the system tracks step references and constant disturbances without steady-state error.

[0108] Step 212: Composite nonlinear integral sliding mode control law g;

[0109] A fractional adaptive reaching law is used to further optimize the dynamic process, as shown in the following formula (7);

[0110] (7);

[0111] In formula (7), Let ν be the fractional derivative of the composite nonlinear integral sliding surface S (i.e., the fractional adaptive reaching law); ν∈(0,1), ρ∈(0,1), where ν is the order of the fractional reaching law; For the exponential parameter of the fractional-order reaching law; and An adaptive switching gain function; It is a small positive number, used to avoid the denominator being zero; It is a symbolic function.

[0112] Formula (7) is a fractional-order adaptive reaching law that generates direct drive control quantities and serves as the basis for calculating the switching control term in the composite nonlinear integral sliding mode control law. This is achieved by designing fractional-order derivative terms. and adaptive gain and This ensures robustness in convergence within a finite time while reducing chattering caused by traditional approach laws, adapting to the dynamic requirements of different operating stages of the system. Equation (7) determines the dynamic process by which the system state trajectory approaches and remains on the composite nonlinear integral sliding surface S=0.

[0113] The adaptive gain is dynamically adjusted according to the error norm, as shown in the following formula (8).

[0114] (8);

[0115] In formula (8), and Adaptive gain η(t) and The baseline value; and Adaptive gain η(t) and The adaptive adjustment coefficient; Let the error vector be e = [e i , e v ] T The norm of; e i For inductor current i L Tracking error; e v Output voltage V o Tracking error.

[0116] Formula (8) clarifies the update rule for the adaptive reaching law gain, making the switching adaptive gain η(t) and It can adjust in real time according to the norm ||e|| of the current tracking error of the system. When the error is large, the control action is enhanced to achieve rapid convergence, and when the error is small, the control action is weakened to suppress chattering. It provides an intelligent parameter tuner for the fractional-order adaptive reaching law, thereby achieving an adaptive balance between dynamic response and steady-state accuracy.

[0117] Combining the fractional-order model, sliding surface, and reaching law, the composite nonlinear integral sliding mode control law g is derived, as shown in the following formula (9).

[0118] g=g eq +g sw +g nl +g ff (9);

[0119] In formula (9), g eq For the equivalent control quantity based on the nominal model; g sw g is the slug switching control quantity. nl g is the nonlinear damping control quantity; ff This is the feedforward compensation control variable.

[0120] Formula (9) is the synthesis expression of the final composite nonlinear integral sliding mode control law g, which is the final output and execution exit, and will be the equivalent control quantity g based on the nominal model. eq , ensuring robustness of sluice switch control quantity g sw Nonlinear damping control quantity g to enhance stability nlAnd the feedforward control quantity g of the radial basis function neural network for compensating for unknown disturbances ff The four parts are added together to form a complete, high-performance control output: the composite nonlinear integral sliding mode control law g.

[0121] Step 22: Design an online perturbation observation for a radial basis function neural network;

[0122] The unknown disturbance is approximated online using a radial basis function neural network, and its network output is shown in the following formula (10).

[0123] (10);

[0124] In formula (10), For radial basis function neural networks to handle unknown perturbations F α,unknown The estimated value; W is the output layer weight matrix of the radial basis function neural network, T is the matrix transpose; h(xα) is the response of the radial basis function neural network; ε is the approximation error of the radial basis function neural network.

[0125] g in formula (9) ff It is precisely the output of the perturbation estimator of the radial basis function neural network. The feedforward compensation term is obtained after appropriate transformation. The perturbation estimator of the radial basis function neural network learns online and approximates the unknown, slowly varying lumped perturbation in the system model. This estimate After being directly or multiplied by a feedforward gain matrix, it becomes the feedforward compensation control quantity g. ff Injected into the overall control law g.

[0126] Equation (10) defines the perturbation estimator for the radial basis function neural network. This involves online learning and approximation of the unknown, slowly varying perturbation portion in a fractional-order multimodal unified state-space model, providing accurate feedforward compensation signals for wide-range multilevel converter controllers. Within the overall architecture, The system's intelligent observer, input system state, and basis function neural network estimate the unknown disturbance, which is then injected as a feedforward term into the composite nonlinear integral sliding mode control law g to actively counteract the disturbance. Simultaneously, the output layer weight matrix W of the radial basis function neural network is updated in conjunction with the composite nonlinear integral sliding mode surface information, forming a synergistic closed loop of performance feedback and learning improvement, enhancing the system's adaptability and robustness.

[0127] To further enable the network output to quickly and accurately track changes in actual unknown disturbances, a method with... The adaptive law of the correction term updates the output layer weight matrix W in formula (10), and the update result of the output layer weight matrix W is... See formula (11) below;

[0128] (11);

[0129] In formula (11), Γ is a positive definite matrix, and hs T It is the learning rate used to update the output layer weight matrix W. The coefficients are correction terms used to ensure that the weights of the output layer weight matrix W are bounded; ||W|| F Let be the Frobenius norm of the weight matrix W. Wherein... The correction term ensures that the weights are bounded, preventing them from drifting and diverging under continuous excitation, thus ensuring the stability of the learning process. At the same time, it feeds the disturbance estimate forward to the control law, enabling real-time compensation for unknown disturbances.

[0130] In specific implementation, step 3 uses the Lyapunov function for stability analysis.

[0131] Step 3 includes the following steps 31 and 32;

[0132] Step 31: Stability analysis and parameter self-tuning based on Lyapunov functions;

[0133] The stability of the Lyapunov function is proven, and self-tuning rules for key parameters are proposed based on the analysis. The integral coefficients are adjusted online according to the moving average of the error, as shown in the following formula (12).

[0134] (12);

[0135] In formula (12), k i,eff (t) represents the effective integral coefficients at time t (i.e., k3 and k4 in Formula 6); k i0 These are the initial values ​​for the integral coefficients; For the regulating factor; |e| win k is the average of the absolute value of the error over the sliding time window. i,eff (t) is the effective integral coefficient after dynamic adjustment, corresponding to k3 and k4 in formula (6), realizing the adaptive adjustment of integral intensity.

[0136] Formula (12) gives the effective integral coefficient k i,effThe adaptive online tuning rule of (t) dynamically adjusts the strength of the integral action based on the recent average error level. When the error is large, the integral is strengthened to accelerate the elimination of steady-state error; when the error is small, the integral is weakened to suppress saturation and overshoot, thereby optimizing the dynamic process of the system. The dynamic optimization of the composite nonlinear integral sliding surface is achieved through formula (12), which transforms the effective integral coefficients k3 and k4 in the sliding surface from static to dynamic, which are adaptively adjusted according to the operating state. This completes the online fine-tuning of the control law structure and works in conjunction with the approach law gain adjustment and neural network disturbance compensation learning to form a multi-level, closed-loop controller adaptive optimization system, thereby improving the overall performance and adaptability of the system under different operating conditions.

[0137] In specific implementation, the calculation formula for parameter self-healing in step 3 is shown in the following formula (13).

[0138] (13);

[0139] In formula (13), y(k) is the parameter estimation vector of the multilevel converter control system at time k; K(k) is the gain vector of the recursive least squares method; y(k) is the measurement output of the multilevel converter control system at time k. Let be the regression vector at time k, which consists of input and output data, and T be the matrix transpose.

[0140] Step 32: Design of parameter self-healing and active fault-tolerant control mechanism;

[0141] This design employs a recursive least squares method with a forgetting factor to identify key circuit parameters (including inductor L and capacitor C) online. It identifies the inductor L and capacitor C in the critical circuits of a multilevel converter in real time and monitors their drift. If the changes in inductor L and capacitor C exceed threshold values, an alarm can be triggered or the parameters of the multilevel converter control system (including inductor L and capacitor C) can be automatically adjusted, thus achieving a self-healing function for the multilevel converter control system.

[0142] Formula (13) serves as the parameter identification module for the multilevel converter control system. Its identification results are used to verify and update the linear matrices Aα and Bα and the lumped nonlinear disturbance matrix F of the fractional state-space equations in Formula (1). α (x α The parameters of the composite nonlinear integral sliding mode surface and control law (i.e., the composite nonlinear integral sliding mode controller) are corrected to ensure that the multilevel converter control system always operates based on an accurate fractional-order multimodal unified state-space model, thus improving long-term reliability. The parameter identification module is a software functional module integrating algorithms such as recursive least squares (RLS), responsible for online identification of key system parameters, and is the core of realizing the system's self-sensing and self-maintenance functions.

[0143] Linear matrices in fractional-order multimodal unified state-space models and It is directly composed of circuit parameters. When the parameter identification module outputs new estimated values ​​for inductance L, flying capacitor C, resistance R, etc., these estimated values ​​are substituted into the linear matrix. and The parsing expressions of each element are recalculated.

[0144] Parameter updates enable the fractional-order multimodal unified state-space model to more accurately describe the dynamic behavior of the multilevel converter under the current actual parameters, providing an accurate mathematical model for controllers and predictors based on the fractional-order multimodal unified state-space model, and ensuring that the long-term performance of the composite nonlinear integral sliding mode controller is not degraded due to parameter drift. It typically includes unmodeled dynamics, external disturbances, and nonlinear terms. Some nonlinear terms are parameter-dependent, such as the nonlinearity caused by saturated inductance, which can be updated through parameter identification. The parameters are related to improve the accuracy of disturbance compensation.

[0145] The ultimate goal of parameter identification and fractional-order multimodal unified state-space model updating is to ensure that the design basis model and internal parameters of the composite nonlinear integral sliding mode controller remain accurate, thereby ensuring that its long-term control performance does not degrade.

[0146] The controller based on the fractional-order multimodal unified state-space model refers to the control component calculated based on the fractional-order state-space equation of formula (2) in the composite nonlinear integral sliding mode controller, i.e., the equivalent control u. eq ;u eq It is the theoretical control quantity required to accurately maintain the system state on the sliding surface (s=0) under the assumption that the system is free from any disturbances and uncertainties in the sliding mode control law. eq The calculation directly depends on the system's fractional-order state matrix in the fractional-order state-space equations. and input fractional control matrix B α ;u eq This determines the basic control actions of the system under ideal conditions.

[0147] In this invention, the equivalent control u eq As a predictor of the fractional-order multimodal unified state-space model, that is, as the system prediction model on which the generalized predictive control algorithm in the prediction layer depends, that is, the fractional-order multimodal unified state-space model expressed by the fractional-order state-space equation of formula (2).

[0148] The process of modifying the control parameters in the composite nonlinear integral sliding surface and control law includes the following steps:

[0149] Step 321: First, the sliding surface coefficients k1~k4 and the reaching law reference gain η0, It can be readjusted or fine-tuned using an adaptive law based on the feature values ​​or bandwidth requirements of the new model;

[0150] Step 321: The adaptive tuning rule for the integral term (Equation 12) will dynamically adjust the integral strength k based on the error characteristics under the new model. i,eff ;

[0151] Step 321: The neural network perturbation observer (Equation 10) compensates for the nonlinear and perturbation components in the model in a feedforward form. Its weights W are updated online using new data. The closed-loop parameter correction mechanism ensures that the controller is always matched to the physical system, thus maintaining excellent dynamic and static performance and robustness under parameter changes.

[0152] Active fault tolerance is achieved by designing a fault diagnosis unit based on a sliding mode observer. When a fault is detected, the reconfiguration control law is triggered, as shown in the following formula (14).

[0153] (14);

[0154] In formula (14), g ft For fault-tolerant control laws; g normal This is the control law under normal conditions; K f Fault-tolerant control gain matrix, r f This is the residual signal for fault diagnosis.

[0155] Formula (14) defines the active fault-tolerant control law. After the fault diagnosis unit detects a fault in the system, it immediately adds a fault residual r to the original normal control law. f Driven compensation terms It reconfigures control functions to maintain the basic operational capabilities of the system and achieves performance degradation under fault conditions.

[0156] Formula (14) is a safety net to ensure the reliability of system operation. When the system is normal, it is not activated. Once a fault occurs, it immediately corrects the control signal output by the composite nonlinear integral sliding mode control law and combines it with the parameter self-healing function to form a full spectrum reliability solution for parameters to slowly drift to sudden component failure.

[0157] This invention presents a control method for a wide-range multilevel converter based on improved integral sliding mode control, which can solve the challenges of multi-objective coordinated optimization of wide-range multilevel converters in terms of dynamic response, steady-state accuracy, mode switching smoothness, and robustness. Its core lies in constructing a three-layer intelligent control architecture integrating perception, decision-making, and execution, and its effectiveness is verified through rigorous theoretical analysis and complete simulation.

[0158] As shown in Figure 1, in the wide-range multilevel converter control method based on improved integral sliding mode control of the present invention, a deeply collaborative intelligent composite control architecture is designed, which combines fractional-order accurate modeling, look-ahead smooth switching of working modes, robust sliding mode control and online disturbance compensation, to achieve unified optimization of multi-level control objectives from reference tracking, switching transients to disturbance rejection performance, and comprehensively improve the dynamic and static quality of the system.

[0159] This invention utilizes a sliding mode controller design based on a fractional-order adaptive reaching law to effectively suppress chattering while maintaining strong robustness. Furthermore, by combining neural network feedforward compensation, a feedback-feedforward collaborative mechanism is constructed to enhance the adaptability to parameter changes and unknown disturbances.

[0160] This invention integrates intelligent prediction and parameter self-tuning strategies based on deep reinforcement learning to achieve optimal smooth switching of working modes and online optimization of the controller, ensuring stable, efficient and low-loss operation of the system across the entire operating range.

[0161] This invention is specifically designed for three-level Buck-Boost converters. Based on traditional integral sliding mode control, it integrates deep reinforcement learning and adaptive neural networks to improve control performance under multi-mode switching and wide-range parameter variation environments.

[0162] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0163] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A wide-range multilevel converter control method based on improved integral sliding mode control, characterized in that, The process includes the following steps: Step 1: Modeling and intelligent prediction steps; establishing a fractional-order multimodal unified state-space model of the multilevel converter control system and constructing a distributed cooperative predictive control architecture; Step 2: Designing a composite nonlinear integral sliding mode controller and adding an online learning compensation mechanism; using a composite nonlinear integral sliding mode controller as the execution layer of the distributed cooperative predictive control architecture; using a radial basis function neural network to provide accurate feedforward compensation signals for the multilevel converter control system; the expression of the composite nonlinear integral sliding mode surface S of the composite nonlinear integral sliding mode controller is shown in the following formula (6). (6); In formula (6), S is the function value of the composite nonlinear integral sliding surface; k1, k2, k3 and k4 are sliding surface coefficients, all of which are positive numbers; e i For inductor current i L The tracking error and e i = i ref - i L i ref For inductor current i L Reference value; e v Output voltage V o Tracking error, e v = V ref -V o V o V is the output voltage of the multilevel converter. ref V is the output voltage of the multilevel converter. o Reference values; tanh(·) and is a nonlinear saturation function used for smooth control; β and γ are the nonlinear saturation functions tanh(·) and γ, respectively. The shape parameter, τ, represents the integral variable; the perturbation estimator of the radial basis function neural network learns online and approximates the unknown, slowly varying lumped perturbation in the system model. unknown; this estimate The unknown value, after being directly or multiplied by a feedforward gain matrix, is used as the feedforward compensation control quantity g. ff Injected into the overall control law g; Step 3: Stability Analysis and Design of Active Fault-Tolerant Control Mechanism; Based on stability analysis, an adaptive online tuning rule for the effective integral coefficients of the composite nonlinear integral sliding surface S of the composite nonlinear integral sliding mode controller is proposed, and parameter self-healing of the multilevel converter control system is achieved through a parameter identification module; Lyapunov function is used for stability analysis; whereby... The integral coefficient is adjusted online based on the moving average of the error, as shown in the following formula (12); (12); In formula (12), k i,eff (t) represents the effective integral coefficient at time t; k i0 These are the initial values ​​for the integral coefficients; For the regulating factor; |e| win k is the average of the absolute value of the error over the sliding time window. i,eff (t) is the effective integral coefficient after dynamic adjustment; the calculation formula for parameter self-healing is shown in the following formula (13); (13); in formula (13), y(k) is the parameter estimation vector of the multilevel converter control system at time k; K(k) is the gain vector of the recursive least squares method; y(k) is the measurement output of the multilevel converter control system at time k. Let be the regression vector at time k, which consists of input and output data, and T be the matrix transpose.

2. The wide-range multilevel converter control method based on improved integral sliding mode control according to claim 1, characterized in that, Step 1 includes the following steps: Step 11: Establish a fractional-order multimodal unified state space model based on fractional-order calculus; Step 12: Design steps for the mode switching predictor; Step 13: Construction steps for the distributed collaborative predictive control architecture.

3. The wide-range multilevel converter control method based on improved integral sliding mode control according to claim 2, characterized in that, In step 11, a fractional-order multimodal unified state-space model is established. First, a fractional-order state variable vector x is defined. α See formula (1) below; (1); In formula (1), x α D is a fractional-order state variable vector; α For fractional differential operators; i L V is the current in inductor L; o V is the output voltage of the multilevel converter. cf α is the voltage across the flying capacitor C; α is the fractional order, and theoretically, α can take values ​​from 0 to 1; T is the vector transpose.

4. A wide-range multilevel converter control method based on improved integral sliding mode control according to claim 2, characterized in that, In step 11, a fractional-order state-space equation is used to express the fractional-order multimodal unified state-space model of the multilevel converter control system.

5. A wide-range multilevel converter control method based on improved integral sliding mode control according to claim 4, characterized in that, The fractional state-space equation is shown in the following formula (2); (2); In formula (2), D α x α Let x be a fractional-order state variable vector. α The α-derivative; A α B is the system fractional-order state matrix of the multilevel converter control system; α The input fractional-order control matrix is ​​denoted by ; u is the system input vector of the multilevel converter control system; F α (x α ,t) is the lumped nonlinearity and disturbance function vector, where t is time.

6. A wide-range multilevel converter control method based on improved integral sliding mode control according to claim 2, characterized in that, In step 12, the mode switching predictor includes a reward function r. t .

7. A wide-range multilevel converter control method based on improved integral sliding mode control according to claim 2, characterized in that, In step 13, the distributed collaborative predictive control architecture includes a prediction layer, an optimization layer, and an execution layer.

Images