Adaptive weight coefficient design method and device, and multi-level converter control system

Abstract

This application relates to the field of power electronic converter control technology, specifically to an adaptive weighting coefficient design method, device, and multilevel converter control system. This method is geared towards Finite Control Set Model Predictive Control (FCS-MPC) and is applicable to multi-objective optimization control of multilevel converters. The method categorizes the four types of weighting coefficients in the cost function into two classes based on their sensitivity characteristics: sensitive parameters and inert parameters. Different strategies are then employed to determine the inert parameters. l s and G i2v By employing a four-layer nested automated simulation script... l s and G i2v Offline network format global optimization is performed, and a lookup table is built using load power factor and modulation index as indexes; sensitive parameters l fc and l np Based on real-time sampling error factor online adaptive calculation, the weights are automatically increased when the capacitor voltage deviation exceeds 5% to strengthen the balance constraint. This method significantly reduces the engineering complexity of weight coefficient tuning and provides a systematic solution for weight coefficient design in FCS-MPC multilevel converter engineering applications.

Description

Technical Field

[0001] This application relates to the field of power electronic converter control technology, and in particular to an adaptive weighting coefficient design method, device, and multilevel converter control system. Background Technology

[0002] Finite control set model predictive control (FCS-MPC) is widely used in multilevel converter control due to its fast dynamic response, inherent multi-objective optimization capability, and modulator-free structure. In high-performance applications such as new energy power generation, harmonic suppression, and electric drives, FCS-MPC needs to simultaneously optimize multiple control objectives, including output current waveform quality, flying capacitor voltage balance, midpoint potential balance, and switching frequency suppression. The weighting coefficients between these objectives directly determine the control performance, which is one of the core issues in FCS-MPC engineering applications.

[0003] Currently, the determination of FCS-MPC weight coefficients mainly relies on the following methods: (1) manual trial and error based on engineering experience; (2) offline search based on heuristic optimization algorithms (genetic algorithms, particle swarm optimization, etc.); and (3) adaptive adjustment based on fuzzy logic or neural networks. The above methods have the following common limitations: manual trial and error depends on the operator's experience and lacks systematicity and repeatability; heuristic optimization algorithms have a large computational load and the search results are at risk of getting trapped in local optima; fuzzy logic and neural network methods have complex structures, heavy real-time computational burden, and strong dependence on training data. In addition, most existing methods do not distinguish the differences in the sensitivity characteristics of different weight coefficients to control performance and adopt a uniform tuning strategy for all weight coefficients, resulting in a large tuning workload and low efficiency.

[0004] For multilevel converters with flying capacitors and DC bus capacitors (such as three-phase five-level active neutral-point clamp converters), the flying capacitor voltage balance weight in the cost function l fc Balance weights of midpoint potential l np It features real-time observability and rapid response, making it suitable for online adaptive adjustment; while the switching frequency suppression weight... l s Since the switching frequency itself is a statistical quantity over multiple control cycles, step-by-step online adjustment is prone to causing system oscillations, making offline determination more suitable. Current technology has not yet proposed a method that can systematically distinguish the characteristics of the two types of weighting coefficients mentioned above and apply optimal strategies for tuning respectively, indicating a significant technological gap. Summary of the Invention

[0005] This application provides an adaptive weighting coefficient design method, device, and multilevel converter control system, which solves the problems of existing technologies in multi-objective FCS-MPC weighting coefficient tuning, such as reliance on empirical trial and error, lack of global optimality, and insufficient real-time adaptive capability.

[0006] To address the aforementioned technical problems, in a first aspect, embodiments of this application provide an adaptive weighting coefficient design method applied to the multi-objective optimization control of a multilevel converter; the cost function of the multilevel converter includes the three-phase current tracking error weight λ. i Three-phase flying capacitor voltage balance error weighting l fc Midpoint potential balance error weight l np and switching frequency suppression weights l s The method consists of two stages: offline global optimization determination of lazy parameters. l s and G i2v Stage and sensitive parameters determined by online adaptive calculation l np and l fc Stages; among which, G i2v The coefficients are dimensional equalization coefficients; the inertial parameters are determined through offline global optimization. l s and G i2v The phase is used to apply lazy parameters using a four-level nested automated simulation script. l s and G i2v Offline network-based global optimization is performed, which includes: Step 1, classification of weight coefficient sensitivity characteristics: The weight coefficients in the cost function of the multilevel converter are classified into two categories according to their sensitivity characteristics: sensitive parameters and inert parameters. l fc and l np As a sensitive parameter, l s Dimensional Equivalence Coefficient G i2v For inert parameters; Step 2, offline global optimization of inert parameters using a network format: A four-layer nested automated simulation script is used to traverse the load power factor PF, modulation index m, etc. l s Different orders of magnitude and G i2v Different orders of magnitude, in each (PF, m, l s, G i2v The parameter combination automatically calls the converter simulation model and extracts the total harmonic distortion (THD) of the current. i Δ u f Midpoint potential balance error Δ u np Filtering that simultaneously satisfies Δ u f ≤5% and Δ u np ≤5% and comprehensively consider THD i Minimum optimal ( l s , G i2v Combinations of orders of magnitude are stored as lookup tables indexed by (PF, m); sensitive parameters are determined through online adaptive calculation. l np and l fc The phases include: online adaptive calculation to obtain sensitive parameters. l np and l fc Within each control cycle, based on the real-time sampled three-phase output current... i s,x (k) ( x = a , b , c Three-phase flying capacitor voltage U fx (k) and DC bus capacitor voltage U cup (k) U clow (k), the sensitive parameters are calculated online adaptively according to formulas (1) and (2) respectively. l np and l fc : (1) (2) in, e i , e np , e fc These are the normalized error factors for the three-phase current, the neutral point potential, and the flying capacitor voltage, respectively. G i2v Retrieved in real time from the lookup table; λ i The default value is 1; Output the weighting coefficients to the cost function: λ i =1. Obtained from online adaptive calculation l np and l fc and lookup table l s The output is fed into the FCS-MPC cost function to complete the weight coefficient configuration for the current control cycle, and participates in the optimal switching sequence search and controller output. In some exemplary embodiments, the normalized error factor e i , e np and e fc Calculate according to the following formulas: (3) (4) (5) in, This is the three-phase reference current. U dc The value represents the DC bus voltage. The denominators 0.01 and 0.05 correspond to the normalized references of 1% of the current and 5% of the capacitor voltage, respectively, which causes the error factor to increase rapidly when the capacitor voltage deviation exceeds the rated value by 5%.

[0007] In some exemplary embodiments, the offline grid-based global optimization includes a fine-grained search step of the same order of magnitude: based on the optimal order of magnitude determined in step two, a second traversal search is performed with a range of ±50% and a step size of 10%.

[0008] In some exemplary embodiments, automated simulation is implemented using MATLAB / Simulink. The four-level nested automated simulation script automatically calls the Simulink simulation model in each innermost loop to extract the THD under steady state. i Δ u f Δ u np and average switching frequency f sw Store to (PF, m , l s , G i2v ) is the result table of the index.

[0009] In some exemplary embodiments, the results table uses color coding to identify the control quality level: white indicates system out of control, gray indicates capacitor voltage balance error exceeding 5%, yellow indicates effective control with balance error not exceeding 5%, and green indicates the overall optimal combination.

[0010] In some exemplary embodiments, the online adaptive calculation of the sensitive parameters is performed to determine... l np and l fc Following this stage, the method also includes verifying the consistency between simulation and experimental results.

[0011] In some exemplary embodiments, the consistency between simulation and experimental results is verified, including: summarizing the optimal results obtained offline under each operating condition. l s , G i2v The simulation performance indicators are compared with the actual experimental performance indicators to verify the experimental applicability of the offline optimization results; if the difference between the simulation and experimental results is within a reasonable range, it is confirmed that the lookup table can be directly used for the online operation of the experimental controller.

[0012] In some exemplary embodiments, the method is applicable to various multilevel converter topologies, including a three-phase five-level active neutral point clamp converter (3P-5L-ANPC) and a three-level neutral point clamp converter (3L-NPC), as well as single-phase, three-phase, and multi-phase converter systems employing finite control set model predictive control.

[0013] Secondly, this application also provides an adaptive weight coefficient design device for finite control set model predictive control. This device is used to implement the adaptive weight coefficient design method, device, and multilevel converter control system described in the above embodiments. The device includes: a weight classification module, used to classify the weight coefficients in the cost function according to their sensitivity characteristics. l fc , l np Sensitive parameters represented by and l s , G i2v There are two types of lazy parameters, represented by [example parameter name]; the offline optimization module is used to execute a four-level nested automated simulation script, traversing (PF, [example parameter name]). m , l s , G i2v ) parameter combination space, extract simulation performance indicators and screen the globally optimal ( l s , Gi2v Order-of-magnitude combinations; lookup table construction module, used to combine the optimal results obtained from offline optimization ( l s , G i2v ) with (PF, m The index is stored as a lookup table for real-time access during online runtime; the error factor calculation module is used to calculate the normalized error factor based on the real-time sampled values ​​in each control cycle. e i , e np and e fc The online adaptive calculation module is used to calculate the normalized error factor and the result in the lookup table. G i2v The sensitive parameters are updated in real time according to formulas (1) and (2) respectively. l np and l fc The weight output module is used to convert λ i =1, calculated l np , l fc And lookup table l s The cost function is output to the finite control set model predictive control to complete the weight coefficient configuration for the current control cycle.

[0014] Thirdly, this application also provides a multilevel converter control system, the control system comprising: a sampling unit for sampling the three-phase output current, three-phase flying capacitor voltage, and DC bus capacitor voltage of the multilevel converter; and a weighting coefficient design unit, implemented using the device described in the above embodiments, for online adaptive output of the weighting coefficient λ of each control objective based on the sampled values. i , l fc , l np , l s The system includes a predictive control unit for constructing a cost function based on the weighting coefficients and searching for the optimal switching sequence using finite control set model predictive control (FCS-MPC, including finite control set model predictive control accelerated by spherical decoding algorithm: SDA-MPC); a drive output unit for converting the optimal switching sequence into drive pulse signals and outputting them to the power devices of the multilevel converter; the control system can run on a digital signal processor (DSP), field programmable gate array (FPGA) or embedded microcontroller (STM32, etc.) and supports serial or parallel computing architectures.

[0015] The technical solution provided in this application has at least the following advantages: This application provides an adaptive weighting coefficient design method, apparatus, and multilevel converter control system. The method includes two stages: offline global optimization determination of inertial parameters. l s and G i2v Stage and sensitive parameters determined by online adaptive calculation l np and l fc This application categorizes the four types of weight coefficients in the cost function into two categories based on their sensitivity characteristics: sensitive parameters and inert parameters. Different strategies are employed to determine these categories: inert parameters are determined by switching frequency suppression weights. l s Dimensional Equivalence Coefficient G i2v By employing a four-layer nested automated simulation script... l s and G i2v Offline network global optimization is performed, and a lookup table is established using load power factor and modulation index as indices; sensitive parameter: three-phase flying capacitor voltage balance weight. l fc Balance weights of midpoint potential l np Based on real-time sampling error factor online adaptive calculation, the weights are automatically increased when the capacitor voltage deviation exceeds 5% to strengthen the balance constraint. This application has clear physical basis, guaranteed global optimality, and strong robustness, significantly reducing the engineering complexity of weight coefficient tuning, and taking into account both output current waveform quality and capacitor voltage balance accuracy. It provides a systematic solution for weight coefficient design in FCS-MPC engineering applications of multilevel converters. Attached Figure Description

[0016] One or more embodiments are illustrated by way of example with reference to the accompanying drawings. These illustrations do not constitute a limitation on the embodiments, and unless otherwise stated, the figures in the drawings are not to be limited by scale.

[0017] Figure 1 This application provides an overall block diagram of an adaptive weight coefficient design method according to an embodiment, showing an offline optimization module for lazy parameters (four nested for loops) and an online adaptive calculation module for sensitive parameters (formula). l np , l fc The two-stage structure and data flow of ).

[0018] Figure 2A schematic diagram of the global order-of-magnitude optimization results for offline grid-based methods (with PF = 0.9). m = 0.3, R = 5 Ω, L = 7.7 mH, Taking the 5.4A operating condition as an example, different colors are used to distinguish them ( l s , G i2v Control quality levels of orders of magnitude combinations (white: system out of control; gray: capacitor voltage balance error > 5%; yellow: effective control and balance error ≤ 5%; green: optimal combination).

[0019] Figure 3 A schematic diagram of the global order-of-magnitude optimization results for offline grid-based methods (with PF = 0.9). m = 1.1, R = 5 Ω, L = 7.7 mH, Taking the 19.8A operating condition as an example, different colors are used to distinguish them ( l s , G i2v Control quality levels of orders of magnitude combinations (white: system out of control; gray: capacitor voltage balance error > 5%; yellow: effective control and balance error ≤ 5%; green: optimal combination).

[0020] Figure 4 A schematic diagram of the global order-of-magnitude optimization results for offline grid-based methods (with PF = 0.6). m = 0.3, R = 5 Ω, L = 21.2 mH, Taking the 3.6A operating condition as an example, different colors are used to distinguish them ( l s , G i2v Control quality levels of orders of magnitude combinations (white: system out of control; gray: capacitor voltage balance error > 5%; yellow: effective control and balance error ≤ 5%; green: optimal combination).

[0021] Figure 5 A schematic diagram of the global order-of-magnitude optimization results for offline grid-based methods (with PF = 0.6). m = 1.1, R = 5 Ω, L = 21.2 mH, Taking the 13.2A operating condition as an example, different colors are used to distinguish them ( l s ,G i2v Control quality levels of orders of magnitude combinations (white: system out of control; gray: capacitor voltage balance error > 5%; yellow: effective control and balance error ≤ 5%; green: optimal combination).

[0022] Figure 6 This is a schematic diagram illustrating the results of fine-grained optimization of the global scale in offline grid format (with PF = 0.9). m = 0.3, R =5 Ω, L = 7.7 mH, Taking the 5.4A operating condition as an example, different colors are used to distinguish them ( l s , G i2v Control quality levels of orders of magnitude combinations (white: system out of control; gray: capacitor voltage balance error > 5%; yellow: effective control and balance error ≤ 5%; green: optimal combination).

[0023] Figure 7 This is a schematic diagram illustrating the results of fine-grained optimization of the global scale in offline grid format (with PF = 0.9). m = 1.1, R =5 Ω, L = 7.7 mH, Taking the 19.8A operating condition as an example, different colors are used to distinguish them ( l s , G i2v Control quality levels of orders of magnitude combinations (white: system out of control; gray: capacitor voltage balance error > 5%; yellow: effective control and balance error ≤ 5%; green: optimal combination).

[0024] Figure 8 This is a schematic diagram illustrating the results of fine-grained optimization of the global scale in offline grid format (with PF = 0.6). m = 0.3, R =5 Ω, L = 21.2 mH, Taking the 3.6A operating condition as an example, different colors are used to distinguish them ( l s , G i2v Control quality levels of orders of magnitude combinations (white: system out of control; gray: capacitor voltage balance error > 5%; yellow: effective control and balance error ≤ 5%; green: optimal combination).

[0025] Figure 9This is a schematic diagram illustrating the results of fine-grained optimization of the global scale in offline grid format (with PF = 0.6). m = 1.1, R =5 Ω, L = 21.2 mH, Taking the 13.2A operating condition as an example, different colors are used to distinguish them ( l s , G i2v Control quality levels of orders of magnitude combinations (white: system out of control; gray: capacitor voltage balance error > 5%; yellow: effective control and balance error ≤ 5%; green: optimal combination).

[0026] Figure 10 The steady-state experimental waveform of the multilevel converter under the proposed method is shown (taking a three-phase five-level active neutral-point clamp converter as an example, PF = 0.9). m = 0.3), from top to bottom: output three-phase voltage (one phase), three-phase output current, three-phase flying capacitor voltage and DC bus capacitor voltage (corresponding to the midpoint potential).

[0027] Figure 11 The steady-state experimental waveform of the multilevel converter under the proposed method (PF = 0.9) is shown. m = 1.1), from top to bottom: output three-phase voltage (one phase), three-phase output current, three-phase flying capacitor voltage and DC bus capacitor voltage (corresponding to the midpoint potential).

[0028] Figure 12 The steady-state experimental waveform of the multilevel converter under the proposed method (PF = 0.6) is shown. m = 0.3), from top to bottom: output three-phase voltage (one phase), three-phase output current, three-phase flying capacitor voltage and DC bus capacitor voltage (corresponding to the midpoint potential).

[0029] Figure 13 The steady-state experimental waveform of the multilevel converter under the proposed method (PF = 0.6) is shown. m = 1.1), from top to bottom: output three-phase voltage (one phase), three-phase output current, three-phase flying capacitor voltage and DC bus capacitor voltage (corresponding to the midpoint potential). Detailed Implementation

[0030] As can be seen from the background technology, the existing technology has not yet proposed a method that can systematically distinguish the characteristics of the above two types of weight coefficients and use the optimal strategy to tune them respectively, which is a significant technological gap.

[0031] To address the aforementioned technical problems, this application provides an adaptive weighting coefficient design method, apparatus, and multilevel converter control system. The method comprises two stages: offline global optimization determination of inertial parameters. l s and G i2v Stage and sensitive parameters determined by online adaptive calculation l np and l fc The stage involves classifying the four types of weight coefficients in the cost function into two categories based on their sensitivity characteristics: sensitive parameters and inert parameters. Different strategies are then employed to determine the weights for each category: inert parameters are determined by suppressing the switching frequency. l s Dimensional Equivalence Coefficient G i2v By employing a four-layer nested automated simulation script... l s and G i2v Offline network global optimization is performed, and a lookup table is established using load power factor and modulation index as indices; sensitive parameter: three-phase flying capacitor voltage balance weight. l fc Balance weights of midpoint potential l np Based on real-time sampling error factor online adaptive calculation, the weights are automatically increased when the capacitor voltage deviation exceeds 5% to strengthen the balance constraint. This application provides an adaptive weight coefficient design method, device, and multilevel converter control system. This method has clear physical basis, guaranteed global optimality, and strong robustness, significantly reducing the engineering complexity of weight coefficient tuning. It balances output current waveform quality and capacitor voltage balance accuracy, providing a systematic solution for weight coefficient design in FCS-MPC engineering applications of multilevel converters.

[0032] The embodiments of this application will now be described in detail with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details have been provided in the embodiments of this application to facilitate a better understanding of the application. However, the technical solutions claimed in this application can be implemented even without these technical details and various variations and modifications based on the following embodiments.

[0033] See Figure 1 This application provides an adaptive weighting coefficient design method, apparatus, and multilevel converter control system, comprising two stages: offline global optimization determination of inertial parameters. l s and G i2v Stage and sensitive parameters determined by online adaptive calculation l np and l fc Stages; among which, G i2v is the dimensionless conversion coefficient.

[0034] The inert parameter is determined through offline global optimization. l s and G i2v The phase is used to apply lazy parameters using a four-level nested automated simulation script. l s and G i2v Perform offline grid-based global optimization, which includes steps S100 and S200.

[0035] Step S100: Classification of sensitive characteristics of weight coefficients.

[0036] The weighting coefficients in the cost function of a multilevel converter are classified into two categories based on their sensitivity characteristics: sensitive parameters and inert parameters. l fc and l np As a sensitive parameter, l s Dimensional Equivalence Coefficient G i2v This is an inert parameter.

[0037] Specifically, this step categorizes the four weight coefficients in the FCS-MPC cost function according to their sensitivity characteristics: l fc (Flying capacitor voltage balance error weighting) and l np (Midpoint potential balance error weight) is classified as a sensitive parameter; l s (Switching frequency suppression weights) and G i2v (Dimensionality equalization coefficient) is classified as an inert parameter. λ i (Current tracking weight) is set to 1 by default and does not require tuning. The classification criteria are as follows: l fc and l np The corresponding control objects (flying capacitor voltage and midpoint potential) can be directly sampled and obtained in each control cycle, which is suitable for real-time adaptive adjustment; l s The constraint is the number of switching cycles in a single control cycle, while the average switching frequency is a multi-cycle statistic. Adjusting it step-by-step online will cause oscillations, so it is more suitable to determine it offline.

[0038] Step S200: Global optimization of lazy parameters in offline network format.

[0039] A four-level nested automated simulation script is used to traverse the load power factor PF, modulation index m, and so on. l s Different orders of magnitude and G i2v Different orders of magnitude, in each (PF, m, l s , G i2v The parameter combination automatically calls the converter simulation model and extracts the total harmonic distortion (THD) of the current. i Δ u f Midpoint potential balance error Δ u np Filtering that simultaneously satisfies Δ u f ≤5% and Δ u np ≤5% and comprehensively consider THD i Minimum optimal ( l s , G i2v The order of magnitude is combined and stored as a lookup table with (PF, m) as the index.

[0040] Specifically, write a MATLAB automation script (.m file) with four nested for loops, and implement it as follows: The first for loop iterates through different load power factors PF (typical values ​​include PF = 0.9, PF = 0.6, etc.). The second for loop iterates through the entire system's scope. m ∈[0, 4 / π], with a step size of 0.1; The third for loop iterates through... l s Different orders of magnitude, such as {10 -8 , 10 -7 , 10 -6 , 10 -5 , 10 -4 , 10 - ³,10 - ², 10 - ¹, 1, 10¹, 10²}; The fourth (innermost) level is traversed using a for loop. G i2v Different orders of magnitude, such as {10 -4 , 10 -3 , 10-2 , 10 -1 , 1,10 1 , 10 2 , 10 3 , 10 4 , 10 5 , 10 6 , 10 7 , 10 8}; In the innermost loop, with the current (PF, m , l s , G i2v The simulation model parameters are initialized using parameter combinations, and the Simulink simulation model is automatically invoked and run to steady state to extract the total harmonic distortion (THD) of the current. i The maximum value of the three-phase flying capacitor voltage balance error Δ u f The maximum value of the midpoint potential balance error Δ u np and average switching frequency f sw Store to (PF, m , l s , G i2v () represents the result table of the index; Filter from the results table to find those that simultaneously satisfy Δ u f ≤5% and Δ u np ≤5% of the effective parameter combinations, select THD from the effective combinations. i Minimum and moderate average switching frequency ( l s , G i2v Combinations of orders of magnitude as (PF, m The global optimal solution under the operating conditions.

[0041] The optimal solution for each working condition is defined by (PF, m The index is stored as a lookup table for real-time retrieval during online runtime.

[0042] The sensitive parameters are determined through online adaptive calculation. l np and l fc The phase includes steps S300 and S400.

[0043] Step S300: Obtain sensitive parameters through online adaptive calculation l npand l fc .

[0044] Within each control cycle (sampling cycle), based on the real-time sampled three-phase output current... i s,x (k) ( x = a , b , c Three-phase flying capacitor voltage U fx (k) and DC bus capacitor voltage U cup (k) U clow (k), the sensitive parameters are calculated online adaptively according to formulas (1) and (2) respectively. l np and l fc : (1) (2) in, e i , e np , e fc These are the normalized error factors for the three-phase current, the neutral point potential, and the flying capacitor voltage, respectively. G i2v Retrieved in real time from the lookup table; λ i The default value is 1.

[0045] Step S400: Output the weight coefficients to the cost function.

[0046] λ i =1. Obtained from online adaptive calculation l np and l fc and lookup table l s The output is fed into the FCS-MPC cost function to complete the weight coefficient configuration for the current control cycle, and participates in the optimal switching sequence search and controller output.

[0047] In some embodiments, the normalized error factor e i , e np and e fc Calculate according to the following formulas: (3) (4) (5) in, This is the three-phase reference current. U dc The value represents the DC bus voltage. The denominators 0.01 and 0.05 correspond to the normalized references of 1% of the current and 5% of the capacitor voltage, respectively, which causes the error factor to increase rapidly when the capacitor voltage deviation exceeds the rated value by 5%.

[0048] In some embodiments, offline grid-based global optimization includes a fine-search step of the same order of magnitude: based on the optimal order of magnitude determined in step S200, a second traversal search is performed with a range of ±50% and a step size of 10%.

[0049] It should be noted that the fine-grained search step of the same order of magnitude can be omitted because simulation verification shows that different precise values ​​within the same order of magnitude affect THD. i Δ u f and Δ u np The impact difference is minimal, and omitting this step can reduce the number of offline simulation calls by 121 (11×11) times.

[0050] Switching frequency suppression weight l s The reasons for not using online adaptive computation include the following three points: Control time domain mismatch: The average switching frequency is determined by the number of switching cycles in multiple adjacent control cycles. l s The constraint is the number of times the switch is activated per cycle; adjusting the switch online step by step will cause the system to oscillate continuously. Setting reference values ​​is difficult: finite control set predictive control cannot strictly control a constant switching frequency. l s It only serves a suppression function, and the precise setting of the reference switching frequency is of little practical significance; Strong coupling between parameters: l s , G i2v It has a strong coupling relationship with other weighting coefficients, l s Incorporating online regulation will significantly increase the parameter coupling dimension and tuning workload.

[0051] The optimal result obtained by offline grid-based global optimization ( l s , G i2vThis also applies to experimental environments because: inertial parameters are not sensitive to precise values; parameter deviations between simulation and experiment caused by non-ideal factors such as switching dead time and parasitic parameters do not affect the selection of the optimal order of magnitude; and the offline simulation model already includes the same switching device dead time settings as the experimental platform.

[0052] In some embodiments, automated simulation is implemented using MATLAB / Simulink. The four-level nested automated simulation script automatically calls the Simulink simulation model in each innermost loop to extract the THD under steady state. i Δ u f Δ u np and average switching frequency f sw Store to (PF, m , l s , G i2v ) is the result table of the index.

[0053] In some embodiments, the result table uses color coding to identify the control quality level: white indicates system out of control, gray indicates capacitor voltage balance error exceeding 5%, yellow indicates effective control with balance error not exceeding 5%, and green indicates the best overall combination.

[0054] In some embodiments, the sensitivity parameter is determined through online adaptive calculation. l np and l fc Following this stage, the method also includes verifying the consistency between simulation and experimental results.

[0055] In some embodiments, verifying the consistency between simulation and experimental results includes: summarizing the optimal results obtained through offline optimization under each operating condition. l s , G i2v The simulation performance indicators are compared with the actual experimental performance indicators to verify the experimental applicability of the offline optimization results; if the difference between the simulation and experimental results is within a reasonable range, it is confirmed that the lookup table can be directly used for the online operation of the experimental controller.

[0056] In some embodiments, the method is applicable to various multilevel converter topologies, including three-phase five-level active neutral-point clamp converters (3P-5L-ANPC) and three-level neutral-point clamp converters (3L-NPC), as well as single-phase, three-phase, and multi-phase converter systems employing finite control set model predictive control. By adjusting the number of weighting coefficients and the corresponding error factor definitions in the cost function, the method can be extended to multilevel converters with voltage balance control objectives for any number of capacitors.

[0057] This application also provides an adaptive weight coefficient design device for finite control set model predictive control. This device is used to implement the adaptive weight coefficient design method, device, and multilevel converter control system described in the above embodiments. The device includes: a weight classification module, used to classify the weight coefficients in the cost function according to their sensitivity characteristics. l fc , l np Sensitive parameters represented by and l s , G i2v There are two types of lazy parameters, represented by [example parameter name]; the offline optimization module is used to execute a four-level nested automated simulation script, traversing (PF, [example parameter name]). m , l s , G i2v ) parameter combination space, extract simulation performance indicators and screen the globally optimal ( l s , G i2v Order-of-magnitude combinations; lookup table construction module, used to combine the optimal results obtained from offline optimization ( l s , G i2v ) with (PF, m The index is stored as a lookup table for real-time access during online runtime; the error factor calculation module is used to calculate the normalized error factor based on the real-time sampled values ​​in each control cycle. e i , e np and e fc The online adaptive calculation module is used to calculate the normalized error factor and the result in the lookup table. G i2v The sensitive parameters are updated in real time according to formulas (1) and (2) respectively. l np and l fc The weight output module is used to convert λ i=1, calculated l np , l fc And lookup table l s The cost function is output to the finite control set model predictive control to complete the weight coefficient configuration for the current control cycle.

[0058] This application also provides a multilevel converter control system, which includes: a sampling unit for sampling the three-phase output current, three-phase flying capacitor voltage, and DC bus capacitor voltage of the multilevel converter; and a weighting coefficient design unit, implemented using the device described in the above embodiments, for online adaptive output of the weighting coefficients λ of each control objective based on the sampled values. i , l fc , l np , l s The system includes a predictive control unit for constructing a cost function based on the weighting coefficients and searching for the optimal switching sequence using finite control set model predictive control (FCS-MPC, including finite control set model predictive control accelerated by spherical decoding algorithm: SDA-MPC); a drive output unit for converting the optimal switching sequence into drive pulse signals and outputting them to the power devices of the multilevel converter; the control system can run on a digital signal processor (DSP), field programmable gate array (FPGA) or embedded microcontroller (STM32, etc.) and supports serial or parallel computing architectures.

[0059] Using a three-phase five-level active neutral point clamp converter (3P-5L-ANPC) as a typical embodiment, and employing a TMS320F28377D digital signal processor (200 MHz) as the controller, the effectiveness of the method of the present invention was verified, and the results are shown in Tables 1 and 2.

[0060] Table 1 Simulation and Experiment Platform Parameters

[0061] Table 2

[0062] Table 1 shows the parameters of the simulation and experimental platform, and Table 2 shows four typical load conditions for verifying the effectiveness of the present invention. The control algorithm used is the Finite Control Set Model Predictive Control Algorithm (SDA-MPC) accelerated by the spherical decoding algorithm.

[0063] Complete data for offline global automated simulation optimization is listed in Figure 2 to Figure 9(The four typical operating conditions listed in Table 2 are used as verification samples). Color coding is used in the figure to distinguish control quality levels: white indicates system out of control (THD). i >20%, or >20%, or >20%, or the presence of outliers such as NaN); gray indicates that the system is effectively controlled but the capacitor voltage balance error exceeds 5%; yellow indicates that the system is effectively controlled and the capacitor voltage balance error does not exceed 5%; green indicates that THD is comprehensively balanced under the premise of meeting the capacitor voltage balance constraint. i The global optimum with the lowest minimum and lowest average switching frequency ( l s , G i2v )combination.

[0064] contrast Figure 2 -5 (order of magnitude global optimization) and Figure 6 –9 (using the same order of magnitude for fine-tuning) yields quantitative conclusions: (1) Order-of-magnitude sensitivity: Figure 2 to 5 The color distribution in the middle ranges significantly, spanning different orders of magnitude ( l s , G i2v The significant differences in control performance between the combinations indicate that the system performance is highly sensitive to the order of magnitude of the inert parameters, and offline global order of magnitude search is a necessary step to ensure the reliability of the optimization results.

[0065] (2) Inertia of the same order of magnitude: Figure 6 to 9 The same color regions are distributed in contiguous patches, and different precise values ​​within the same order of magnitude affect THD. i , and The impact difference is minimal, which quantitatively verifies the "lazy parameter" characteristic proposed in this invention—that is, offline optimization only needs to determine the optimal order of magnitude without performing a fine scan within the same order of magnitude. This characteristic makes the fine search steps within the same order of magnitude unnecessary in engineering practice, thereby reducing the number of simulation calls by about 46%.

[0066] The above-mentioned automated simulation results are highly consistent with the working principle of the converter: the higher the modulation index, the faster the capacitor charging and discharging speed per unit control cycle, the greater the capacitor voltage fluctuation amplitude, and the more difficult the balance control becomes. This is consistent with... Figure 2 to 5 The narrowing pattern of the effective region (yellow / green grids) under medium-to-high-tone regime conditions is completely consistent. The simulation results not only verify the theoretical feasibility of the proposed weight coefficient design method, but also provide sufficient evidence for the direct transfer of offline optimization results to the experimental environment.

[0067] Table 3 Simulation Experiment Results

[0068] Table 3 shows the simulation and experimental results of the Finite Control Set Model Predictive Control (SDA-MPC) algorithm accelerated by the spherical decoding algorithm under four typical operating conditions (PF=1.1 / 0.6×m=0.3 / 1.1). The simulation data comes from... Figure 6 , 7 The best among 8 and 9 ( l s , G i2v The THD corresponding to the combination i Δ u f Δ u np , f sw Simulation results and experimental data are from Figure 10 , 11 The experimental results of the Finite Control Set Model Predictive Control (SDA-MPC) algorithm accelerated by the spherical decoding algorithm shown in Figures 12 and 13 are presented under four typical operating conditions.

[0069] The simulation results in Table 3 show that, under all four operating conditions, the proposed method can stably control the three-phase flying capacitor voltage balance error and the midpoint potential balance error within 5%, while achieving a low total harmonic distortion rate of the output current and a low average switching frequency, thus verifying the feasibility and superiority of the proposed method. The comparative analysis of simulation and experimental results shows that the optimal parameter combination obtained through offline optimization is also effective in the experiment, with minimal deviation, further verifying the experimental applicability of the inertial parameters.

[0070] Compared with existing technologies, the adaptive weight coefficient design method provided in this application has the following advantages: (1) The physical basis is clear and the engineering significance is clear. The online adaptive adjustment of the sensitive parameters is directly based on the real-time ratio of the control target error, which has the intuitive physical meaning that "the larger the error, the higher the weight". It abandons the empirical trial and error that lacks physical basis in the traditional method and improves the interpretability and repeatability of the weight coefficient tuning.

[0071] (2) Global optimality is guaranteed. The lazy parameter is systematically searched globally across the entire tuning range through offline grid-based automated simulation. Compared with heuristic optimization algorithms, the method of this invention does not have the risk of getting trapped in local optima, and the search process is fully automated and does not depend on the operator's experience.

[0072] (3) Strong robustness. The characteristic of inert parameters being "sensitive to the order of magnitude but not sensitive to the precise value" ensures that the offline optimization results are robustly adaptable to parameter differences between simulation and experiment (switching dead time, parasitic parameters, etc.), small disturbances in system parameters, and changes in operating conditions. The offline results can be directly applied to the experiment without readjustment.

[0073] (4) High tuning efficiency and strong engineering applicability. The parameter search is achieved by using a four-layer nested automated script, which has low computing resource requirements (about 6 hours for a single-core CPU). It can be further accelerated by making full use of the parallel computing capabilities of modern multi-core CPUs. The whole process is standardized and reproducible, does not rely on professional experience, and can be implemented independently by any engineer.

[0074] (5) Wide range of applications. The method of this invention is not limited to a specific topology and can be extended to multilevel converters with voltage balance targets of any number of capacitors, as well as single-phase, three-phase and multi-phase converter control systems using FCS-MPC (including SDA-MPC with spherical decoding acceleration).

[0075] Based on the above technical solutions, this application provides an adaptive weighting coefficient design method, device, and multilevel converter control system. The method includes two stages: offline global optimization determination of inertial parameters. l s and G i2v Stage and sensitive parameters determined by online adaptive calculation l np and l fc This application categorizes the four types of weight coefficients in the cost function into two categories based on their sensitivity characteristics: sensitive parameters and inert parameters. Different strategies are employed to determine these categories: inert parameters are determined by switching frequency suppression weights. l s Dimensional Equivalence Coefficient G i2v By employing a four-layer nested automated simulation script... l s and G i2v Offline network global optimization is performed, and a lookup table is established using load power factor and modulation index as indices; sensitive parameter: three-phase flying capacitor voltage balance weight. l fc Balance weights of midpoint potential l npBased on real-time sampling error factor online adaptive calculation, the weights are automatically increased when the capacitor voltage deviation exceeds 5% to strengthen the balance constraint. This application has clear physical basis, guaranteed global optimality, and strong robustness, significantly reducing the engineering complexity of weight coefficient tuning, and taking into account both output current waveform quality and capacitor voltage balance accuracy. It provides a systematic solution for weight coefficient design in FCS-MPC engineering applications of multilevel converters.

[0076] Those skilled in the art will understand that the above-described embodiments are specific examples of implementing this application, and in practical applications, various changes in form and detail may be made without departing from the spirit and scope of this application. Any person skilled in the art can make their own modifications and alterations without departing from the spirit and scope of this application; therefore, the scope of protection of this application should be determined by the scope defined in the claims.

Claims

1. An adaptive weighting coefficient design method applied to multi-objective optimization control of a multilevel converter; the cost function of the multilevel converter includes the three-phase current tracking error weight λ. i Three-phase flying capacitor voltage balance error weighting λ fc Midpoint potential balance error weight λ np and switching frequency suppression weights λ s Its characteristics are, This method consists of two stages: offline global optimization determination of lazy parameters. λ s and G i2v Stage and sensitive parameters are determined through online adaptive calculation. λ np and λ fc Stages; among which, G i2v The dimensionality coefficient; The inert parameter is determined through offline global optimization. λ s and G i2v The phase is used to apply lazy parameters using a four-level nested automated simulation script. λ s and G i2v Perform offline grid-based global optimization, which includes: Step 1: Classification of Weighting Coefficient Sensitivity Characteristics: The weighting coefficients in the cost function of the multilevel converter are classified into two categories based on their sensitivity characteristics: sensitive parameters and inert parameters. λ fc and λ np As a sensitive parameter, λ s Dimensional Equivalence Coefficient G i2v Inert parameter; Step 2: Offline global optimization of lazy parameters in the network format: A four-level nested automated simulation script is used to traverse the load power factor (PF), modulation index (m), and other parameters respectively. λ s Different orders of magnitude and G i2v Different orders of magnitude, in each (PF, m, λ s , G i2v The parameter combination automatically calls the converter simulation model and extracts the total harmonic distortion (THD) of the current. i Δ u f Midpoint potential balance error Δ u np Filtering that simultaneously satisfies Δ u f ≤5% and Δ u np ≤5% and comprehensively consider THD i Minimum optimal ( λ s , G i2v The order of magnitude combinations are stored as a lookup table with (PF, m) as the index; The sensitive parameters are determined through online adaptive calculation. λ np and λ fc The phases include: Online adaptive calculation to obtain sensitive parameters λ np and λ fc Within each control cycle, based on the real-time sampled three-phase output current... i s,x (k) ( x = a , b , c Three-phase flying capacitor voltage U fx (k) and DC bus capacitor voltage U cup (k) U clow (k), the sensitive parameters are calculated online adaptively according to formulas (1) and (2) respectively. λ np and λ fc : （1） （2） in, e i , e np , e fc These are the normalized error factors for the three-phase current, the neutral point potential, and the flying capacitor voltage, respectively. G i2v Retrieved in real time from the lookup table; λ i The default value is 1; Output the weighting coefficients to the cost function: λ i =1. Obtained from online adaptive calculation λ np and λ fc and lookup table λ s The output is fed into the FCS-MPC cost function to complete the weight coefficient configuration for the current control cycle, and participates in the optimal switching sequence search and controller output.

2. The adaptive weight coefficient design method according to claim 1, characterized in that, The normalization error factor e i , e np and e fc Calculate according to the following formulas: （3） （4） （5） in, This is the three-phase reference current. U dc The value represents the DC bus voltage. The denominators 0.01 and 0.05 correspond to the normalized references of 1% of the current and 5% of the capacitor voltage, respectively, which causes the error factor to increase rapidly when the capacitor voltage deviation exceeds the rated value by 5%.

3. The adaptive weight coefficient design method according to claim 1, characterized in that, The offline grid-based global optimization includes a fine-search step of the same order of magnitude: based on the optimal order of magnitude determined in step two, a second traversal search is performed with a range of ±50% and a step size of 10%.

4. The adaptive weight coefficient design method according to claim 1, characterized in that, The automated simulation is implemented using MATLAB / Simulink. The four-level nested automated simulation script automatically calls the Simulink simulation model in each innermost loop to extract the THD under steady state. i Δ u f Δ u np and average switching frequency f sw Store to (PF, m , λ s , G i2v ) is the result table of the index.

5. The adaptive weight coefficient design method according to claim 4, characterized in that, The results table uses color coding to indicate the quality control level.

6. The adaptive weight coefficient design method according to claim 1, characterized in that, The sensitive parameters are determined through online adaptive calculation. λ np and λ fc Following this stage, the method also includes verifying the consistency between simulation and experimental results.

7. The adaptive weight coefficient design method according to claim 6, characterized in that, Verify the consistency between simulation and experimental results, including: summarizing the optimal results obtained offline under various operating conditions. λ s , G i2v The simulation performance indicators are compared with the actual experimental performance indicators to verify the experimental applicability of the offline optimization results; if the difference between the simulation and experimental results is within a reasonable range, it is confirmed that the lookup table can be directly used for the online operation of the experimental controller.

8. The adaptive weight coefficient design method according to claim 1, characterized in that, The method is applicable to various multilevel converter topologies, including the three-phase five-level active neutral point clamp converter (3P-5L-ANPC) and the three-level neutral point clamp converter (3L-NPC), as well as single-phase, three-phase and multi-phase converter systems using finite control set model predictive control.

9. An adaptive weighting coefficient design device for model predictive control of finite control sets, the device being used to implement multi-objective optimization control of a multilevel converter using the adaptive weighting coefficient design method as described in any one of claims 1 to 8, characterized in that, The device includes: The weight classification module is used to classify the weight coefficients in the cost function according to their sensitivity characteristics. λ fc , λ np Sensitive parameters represented by and λ s , G i2v Two types of inert parameters, represented by [example parameter name]; The offline optimization module is used to execute four-level nested automated simulation scripts, traversing (PF, m , λ s , G i2v ) parameter combination space, extract simulation performance indicators and screen the globally optimal ( λ s , G i2v Order of magnitude combination; The lookup table building module is used to optimize the optimal results obtained offline. λ s , G i2v ) with (PF, m The index is stored as a lookup table for real-time access during online runtime; The error factor calculation module is used to calculate the normalized error factor based on the real-time sampled values ​​in each control cycle. e i , e np and e fc ; The online adaptive calculation module is used to calculate the normalized error factor and the lookup table. G i2v The sensitive parameters are updated in real time according to formulas (1) and (2) respectively. λ np and λ fc ; The weight output module is used to convert λ i =1, calculated λ np , λ fc And lookup table λ s The cost function is output to the finite control set model predictive control to complete the weight coefficient configuration for the current control cycle.

10. A multilevel converter control system, characterized in that, The control system includes: The sampling unit is used to sample the three-phase output current, three-phase flying capacitor voltage, and DC bus capacitor voltage of the multilevel converter. The weighting coefficient design unit, implemented using the device described in claim 9, is used to adaptively output the weighting coefficients λ of each control objective online based on the sampled values. i , λ fc , λ np , λ s ; A predictive control unit is used to construct a cost function based on the weighting coefficients and to predict and control the optimal switching sequence through a finite control set model. The drive output unit is used to convert the optimal switching sequence into drive pulse signals and output them to the power devices of the multilevel converter. The control system can run on a digital signal processor (DSP), a field-programmable gate array (FPGA), or an embedded microcontroller, and supports serial or parallel computing architectures.

Images