A three-level npc-dab converter data-driven predictive control method and system

Abstract

The present application relates to the technical field of power electronic converter control, and particularly relates to a three-level NPC-DAB converter data-driven predictive control method and system. The method comprises synchronization, quantization processing and abnormal sample isolation based on obtained data; input-output increment data vectors are constructed based on processed sample data; a dynamic linearization input-output increment model is established based on the increment data vectors, and the pseudo partial derivative parameter vectors of the model are updated online; the candidate main phase shift angle center is determined based on the updated pseudo partial derivative parameter vectors; and a priority candidate set is generated based on the candidate main phase shift angle center. The present application uses a dynamic linearization data model to describe the local input-output relationship between the output-side DC filter inductance current and the main phase shift angle, thereby reducing the dependence of the controller on accurate physical parameters.

Description

Technical Field

[0001] This invention relates to the field of data-driven prediction technology, and in particular to a data-driven predictive control method and system for a three-level NPC-DAB converter. Background Technology

[0002] Three-level NPC-DAB converters have promising applications in AC / DC microgrids, energy storage systems, and battery charging / discharging equipment due to their lower device voltage stress and wider operating voltage range. However, these converters have multiple modulation degrees of freedom and strong system nonlinearity. Achieving accurate and robust control of the output-side DC filter inductor current under complex operating conditions such as input voltage fluctuations, load abrupt changes, and uncertain circuit parameters remains a significant challenge.

[0003] Existing mechanistic control methods for this type of converter typically exhibit strong parameter dependencies. Taking sliding mode control as an example, the paper "Non-linear Integral Sliding Mode Control Strategy for Voltage and Current Regulation of Dual Active Bridge DC-DC Converters" indicates that establishing the relevant control law usually requires Fourier series expansion of the switching function, a state-space model, and accurate circuit parameters. Under multi-degree-of-freedom modulation conditions, this type of method is prone to problems such as complex model derivation, cumbersome control parameter tuning, and high-frequency chattering caused by model mismatch. On the other hand, while model predictive control has advantages such as fast dynamic response and ease of handling constraints, its prediction process still relies on a relatively accurate mathematical model. The paper "Improved Model Predictive Control for Three-Phase Dual-Active-Bridge Converters With a Hybrid Modulation" points out that traditional model predictive control usually requires establishing a discrete predictive model that includes parameters such as leakage inductance and sampling period to achieve rolling optimization. When system parameters shift due to temperature changes, nonlinearity of magnetic components, or load disturbances, errors will occur between the predictive model and the actual object, thus affecting the control accuracy. While the problem can be improved by introducing observers or compensation loops, this increases the computational burden on the controller architecture and microprocessor, and makes it difficult to eliminate the dependence on precise parameter models.

[0004] In summary, the three-level NPC-DAB converter is subject to the combined effects of multi-degree-of-freedom modulation, strong nonlinearity, and time-varying parameter characteristics, making it difficult to establish a globally accurate analytical model. Traditional control methods based on fixed mechanism models are prone to prediction errors and performance degradation under parameter perturbations and complex operating conditions, hindering the rapid and accurate control of the output-side DC filter inductor current. Therefore, it is necessary to propose a dynamic linearized data-driven output predictive control method suitable for three-level NPC-DAB converters. This method uses the main phase shift angle as an online adjustment variable and establishes a dynamic linearized predictive model using real-time input and output data. This reduces the dependence on precise physical parameters and improves the adaptability and robustness of the output-side DC filter inductor current control. Summary of the Invention

[0005] To address the issues of existing three-level NPC-DAB converter output-side DC filter inductor current control methods being heavily reliant on precise circuit parameter models and exhibiting decreased prediction accuracy under parameter variations and operating condition disturbances, this invention provides a data-driven predictive control method and system for three-level NPC-DAB converters.

[0006] In a first aspect, the present invention provides a data-driven predictive control method for a three-level NPC-DAB converter, which adopts the following technical solution:

[0007] Acquire the output-side DC filter inductor current data and the main phase shift angle;

[0008] Synchronization, quantification, and isolation of abnormal samples are performed based on the acquired data;

[0009] Construct input and output incremental data vectors based on the processed sample data;

[0010] A dynamic linearized input-output incremental model is established based on incremental data vectors, and the pseudo-partial derivative parameter vector of the model is updated online.

[0011] Candidate principal phase shift angle centers are determined based on the updated pseudo-partial derivative parameter vector;

[0012] A priority candidate set is generated based on the candidate principal phase shift angle center;

[0013] The optimal principal phase shift angle is generated based on the preferred candidate set and the updated pseudo-partial derivative parameter vector.

[0014] The final modulation signal is generated based on the optimal main phase shift angle.

[0015] Secondly, a three-level NPC-DAB converter data-driven predictive control system includes:

[0016] The data acquisition module is configured to acquire the output-side DC filter inductor current data and the main phase shift angle;

[0017] The preprocessing module is configured to perform synchronization, quantization, and isolation of outlier samples based on the acquired data.

[0018] The incremental module is configured to construct input and output incremental data vectors based on the processed sample data.

[0019] The pseudo-partial derivative module is configured to build a dynamically linearized input-output incremental model based on the incremental data vector, and simultaneously update the pseudo-partial derivative parameter vector of the model online.

[0020] The candidate module is configured to determine the candidate principal phase shift angle center based on the updated pseudo-partial derivative parameter vector;

[0021] The priority module is configured to generate a priority candidate set based on the candidate principal phase shift angle center;

[0022] The optimal module is configured to generate the optimal principal phase shift angle based on the priority candidate set and the updated pseudo-partial derivative parameter vector.

[0023] The modulation module is configured to generate the final modulation signal based on the optimal main phase shift angle.

[0024] Thirdly, the present invention provides a computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the aforementioned three-level NPC-DAB converter data-driven predictive control method.

[0025] Fourthly, the present invention provides a terminal device, including a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, the instructions being adapted to be loaded by the processor and executed by the processor to provide a data-driven predictive control method for a three-level NPC-DAB converter.

[0026] In summary, the present invention has the following beneficial technical effects:

[0027] First, the present invention uses a dynamic linearized data model to describe the local input-output relationship between the DC filter inductor current on the output side and the main phase shift angle, without the need to establish an analytical prediction model that relies on precise leakage inductance, resistance, capacitance, device voltage drop and load parameters, thus reducing the controller's dependence on precise physical parameters.

[0028] Secondly, this invention updates the pseudo-partial derivative parameters online by sampling data in real time, enabling the prediction model to adaptively correct itself as the input voltage, load state, circuit parameters and operating conditions change, thereby reducing the impact of fixed parameter model mismatch on the output side DC filter inductor current control performance.

[0029] Third, this invention uses the deadbeat inverse solution value as the search center of a finite candidate set, predicts and calculates the cost function value one candidate at a time under the constraints of the magnitude and rate of change of the main phase shift angle, and then selects the optimal main phase shift angle. Combined with pseudo-partial derivative parameter update protection, abnormal sample isolation and safe backoff, it improves the dynamic tracking capability of the DC filter inductor current on the output side and the stability of digital implementation. Attached Figure Description

[0030] Figure 1 This is a block diagram of the main circuit topology and control system architecture of the three-level NPC-DAB converter in Embodiment 1 of the present invention;

[0031] Figure 2 This is a schematic diagram of the multi-degree-of-freedom modulation waveform and phase shift angle definition of the three-level NPC-DAB converter in Embodiment 1 of the present invention;

[0032] Figure 3 This is the overall control block diagram of the dynamically linearized data-driven output predictive control method of Embodiment 1 of the present invention;

[0033] Figure 4 This is a flowchart of the dynamic linearized input-output incremental data model and the online update of pseudo-partial derivative parameters in Embodiment 1 of the present invention;

[0034] Figure 5 This is a waveform diagram of the output response of the method of the present invention when the reference value of the DC filter inductor current on the output side changes in Embodiment 1 of the present invention.

[0035] Figure 6 This is the output response waveform of the method of the present invention when the load resistance changes in Embodiment 1 of the present invention;

[0036] Figure 7 This is a waveform diagram of the output response of the method of the present invention when the input voltage changes according to Embodiment 1 of the present invention;

[0037] Figure 8 The waveforms of the output voltage, output-side DC filter inductor current, and series inductor current of the method of the present invention at the moment of loading in Embodiment 1 of the present invention are shown.

[0038] Figure 9 The waveforms of the output voltage, output-side DC filter inductor current, and series inductor current of the method of the present invention at the moment of load reduction in Embodiment 1 of the present invention. Detailed Implementation

[0039] The present invention will be further described in detail below with reference to the accompanying drawings.

[0040] Example 1

[0041] like Figure 1As shown, this embodiment forms a closed-loop rolling control process according to S1 to S9: First, the input-output incremental history vector is constructed from the actual sampled data and the actual main phase shift angle confirmed by the modulation register, and the pseudo-partial derivative parameter vector is updated. Then, the candidate main phase shift angle center is determined by the updated pseudo-partial derivative parameter vector, all candidates are generated and predicted, the optimal main phase shift angle is selected according to the unified cost function, and the gate drive signal is output. Finally, the actual main phase shift angle confirmed by the modulation register is written back to the history buffer.

[0042] In the Within a single continuous effective control cycle, the model update uses only the historical vector of actual input-output increments and the measured DC filter inductor current increment on the output side in this cycle. Candidate predictions only replace the current candidate main phase shift angle increment component, and all candidates share the pseudo-partial derivative parameter vector updated in this cycle. The optimal main phase shift angle is only written into the cache as the actual main phase shift angle after being accepted by the modulation module and confirmed by register feedback. Candidate prediction values ​​are not used to replace the measured DC filter inductor current increment on the output side in the next control cycle.

[0043] like Figure 1 As shown, the three-level NPC-DAB converter in this embodiment is an isolated dual active bridge converter with a two-level primary side and a three-level secondary side. It includes a two-level full-bridge primary side, a three-level midpoint clamped bridge secondary side, a high-frequency transformer, a series inductor, an output-side DC bus, and an output-side DC filter inductor. The two-level full-bridge primary side is used to generate a two-level AC voltage, and the three-level midpoint clamped bridge secondary side is used to generate a multi-level AC voltage. The two bridge arms transfer energy through the high-frequency transformer and the series inductor. The output-side DC filter inductor is used to suppress the current ripple of the output-side DC filter inductor.

[0044] S1: Synchronously acquire control data and perform preliminary validity checks.

[0045] S1.1, at the preset sampling phase of each control cycle, the DC filter inductor current on the output side is acquired through a current sensor and an analog-to-digital converter. The sampling phase remains fixed relative to the carrier or switching period to reduce the impact of switching transients on the difference between adjacent samples.

[0046] S1.2, Read the actual applied main phase shift angle from the previous control cycle. Current output-side DC filter inductor current reference value and internal modulation angle , , The segmented preset values. The reference value of the DC filter inductor current on the output side is generated by the upper-level power controller, output voltage controller, or external reference unit.

[0047] S1.3 performs a preliminary check on the current sensor range, analog-to-digital conversion result, control cycle number, and modulation register feedback. If any check fails, the current cycle is marked as invalid, the pseudo-partial derivative parameter update is not performed, and the main phase shift angle is maintained at the previous valid value or switched to a preset safe value.

[0048] S2: Perform sampling synchronization, quantization processing, and isolation of abnormal samples.

[0049] S2.1, Control cycle of this embodiment 100 . No. Within each control cycle, the time relative to the carrier wave is fixed. Samples were taken at each time point, among which =3, sampling time is:

[0050] (1),

[0051] in, To control the starting point of the timing, For the first A fixed offset of each sampling point relative to the start of the control cycle. =1, 2, 3. Each It is set at a position far from the gate switching and dead zone switching; three sampled values ​​are used to form a periodic output-side DC filter inductor current sample.

[0052] S2.2, Let the number of bits in the analog-to-digital converter be... The lower and upper limits of current measurement are respectively and The current quantization step size is then:

[0053] (2),

[0054] For the original sampled values First, perform range saturation, then round down to the nearest quantization level to obtain:

[0055] (3),

[0056] S2.3, in this embodiment =3, construct the anti-outlier statistic using the median and median absolute deviation:

[0057] (4),

[0058] in, To prevent small positive numbers with a scale of zero, take = ; Let be the outlier determination coefficient, and take . =3.5; To preserve the sample index set. When all three samples are valid and no outliers are detected, It contains all three samples.

[0059] S2.4, the set of retained sample indices is denoted in ascending order as follows: = The average of the corresponding quantified samples is calculated to form the first... One control cycle is used for the periodic output samples of dynamic linearization:

[0060] (5),

[0061] in, for The Middle A reserved sample index, To retain the sample size; because >0 and >0, It must include at least the sample corresponding to the median. A control period number is appended to each period-level sample. The system includes an analog-to-digital conversion valid flag, a range flag, and a modulation register consistency flag, the logical AND of which is denoted as... Only when =1、 =1 and - When =1, the current sample and the previous sample constitute a continuous valid sample pair. If this condition is not met, the period is recorded as a cache breakpoint, the cross-breakpoint increment is not calculated, the pseudo-partial derivative parameter is not updated, and the historical cache is refilled from the next continuous valid data segment.

[0062] This embodiment takes , The number of bits and the current range of the analog-to-digital converter are determined according to the digital controller and current sensor used, and are not required inputs to the prediction model of this invention.

[0063] S3: Construct the input / output increment history vector

[0064] S3.1, using only adjacent period samples within the same continuous valid data segment, calculate the output-side DC filter inductor current increment and the actual applied main phase shift angle increment respectively:

[0065] (6),

[0066] in, It reflects the change in the DC filter inductor current on the output side during adjacent effective control cycles. This reflects the change in the actual applied principal phase shift angle between adjacent effective control cycles. (The text does not include candidate subscripts.) All indicate the first The main phase shift angle is actually written to the period and confirmed by the modulation register; candidate angles are used. express.

[0067] S3.2, In this embodiment, the output historical order is taken. =1. Principal phase shift angle and historical order =3, combine the corresponding increments in order from newest to oldest to obtain the input-output increment history vector:

[0068] (7),

[0069] In this embodiment The dimension is 4×1. Online model updates use data that has already occurred. and actual measurement During the candidate prediction phase, only the current principal phase shift increment channel is replaced with... The remaining historical components remain unchanged.

[0070] The candidate input / output increment history vector is formed by the following steps:

[0071] ,

[0072] ,

[0073] Equation (7-1) first calculates the increment of each candidate main phase shift angle relative to the actual main phase shift angle of the previous cycle, and constructs a unit selection vector for locating the current main phase shift angle increment channel; Equation (7-2) uses this selection vector to replace only the current control channel to obtain the candidate history vector, while the remaining output history and the history of the main phase shift angle that has occurred remain unchanged.

[0074] S3.3 performs integrity and activation checks on the historical vectors. If the number of consecutive valid samples is insufficient, the buffer crosses a breakpoint, or the L2 norm of the input / output incremental historical vectors is less than the activation threshold, the check will be performed. If the pseudo-partial derivative parameter update is frozen, then finite candidate predictions can still be performed near the safe backoff angle.

[0075] Continuity validity, stimulus strength, and safe rollback candidates are calculated using the following steps:

[0076] ,

[0077] ,

[0078] ,

[0079] Equation (7-3) checks the validity flags and control cycle number continuity of the data window required for constructing the history vector sample by sample; Equation (7-4) calculates the history vector energy and generates parameter update enable flags. When the update enable flag is zero, the pseudo-partial derivative parameters are frozen. Equation (7-5) uses the actual principal phase shift angle of the previous cycle as the safe backoff center and generates a small number of candidate values ​​within a preset finite symmetry index set; the safe backoff candidate step size is not greater than the maximum change in principal phase shift angle in a single cycle.

[0080] S3.4 manages historical cache and startup status in a fixed order.

[0081] The output increment and the main phase shift angle increment are stored in a first-in-first-out fixed-length buffer and concatenated according to equation (7). In this embodiment, the buffer stores at least 4 sets of consecutive valid increments. Each cycle first reads... After updating the pseudo-partial derivative parameters, construct a new parameter for candidate prediction. Finally, after the optimal main phase shift angle is accepted by the modulation module, a new one is written. .

[0082] In this embodiment, the initial pseudo-partial derivative parameter vector is taken in per-unit terms. The initial value of the current main phase shift angle increment channel is positive, corresponding to the local control direction where the increase in the main phase shift angle during forward power transmission causes an increase in the DC filter inductor current on the output side. During the startup or breakpoint recovery phase, the previous effective main phase shift angle is used to construct a safe candidate set.

[0083] S4: Establish a full-format dynamic linearized incremental model and estimate pseudo-partial derivative parameters online.

[0084] S4.1: Select the dynamic linearization model form

[0085] Based on the scope of historical input and output information usage, dynamically linearized data models include compact, partial, and full formats. The compact format retains only the current principal phase shift angle increment.

[0086] (8),

[0087] The compact format has lower computational cost, but it cannot explicitly utilize output inertia and a longer control history. This represents the equivalent perturbation residual that was not included in the historical vector, and it is bounded within the local working interval.

[0088] The partial format uses the current and multiple historical principal phase shift angle increments:

[0089] (9),

[0090] Partial formatting is suitable for situations where the output history has a weak impact but the input delay needs to be preserved; It still represents the composite residual of unmodeled dynamics, measurement noise, and external disturbances.

[0091] The full format retains both the output increment and the principal phase shift angle increment:

[0092] (10),

[0093] This embodiment preferably uses a full-format model to simultaneously preserve the output inertia and the historical influence of the principal phase shift angle. Tight and partial formats can be considered as... , Or a special case when the corresponding parameter components take simplified values.

[0094] S4.2: Establish a full-format dynamic linearized input-output incremental model

[0095] (11),

[0096] in, This is the predicted increment of the DC filter inductor current on the output side for the next control cycle. This is a 4×1 pseudo-partial derivative parameter vector. This vector is updated by real-time input / output data and is not a fixed circuit parameter.

[0097] Equation (11) yields the one-step prediction value of the DC filter inductor current on the output side for the next control cycle:

[0098] (12),

[0099] Equation (12) transforms the incremental model into an absolute prediction model that can be directly compared with the reference value of the DC filter inductor current on the output side.

[0100] S4.3: Mapping pseudo-partial derivative parameters to Adaline online estimator weights

[0101] (13),

[0102] In the formula, Indicates by the first The -1 periodic pseudo-partial derivative parameters and the history vector obtained from the first Estimated value of the incremental current of the DC filter inductor on the periodic output side; this estimate is only used to calculate the online learning error.

[0103] (14),

[0104] This represents the deviation between the measured increase in the DC filter inductor current on the output side in this cycle and the predicted value from the data model in the previous cycle. It includes the pseudo-partial derivative parameter estimation error, measurement noise, and unpredicted disturbance residuals. This error serves as a monitoring signal for Adaline's online updates.

[0105] From the Adaline structure, each node in the input layer receives... One component, the network weights, i.e., the pseudo-partial derivative parameter components, is the output of the linear activation function:

[0106] (15),

[0107] This network does not have a hidden layer, so each weight update only includes multiplication, division and limiting operations.

[0108] Using the measured increment of this period as the monitoring signal, the objective function for instantaneous squared error is defined as follows:

[0109] (16),

[0110] The objective function for the instantaneous squared error of Adaline weight update is given; using a 1 / 2 coefficient ensures that the constant 2 no longer appears after differentiating the pseudo-partial derivative parameter.

[0111] (17),

[0112] Updating along the negative gradient direction can reduce the instantaneous deviation between the measured increment and the network output; the normalization term is used to avoid directly amplifying the weight step size due to changes in the magnitude of historical vectors.

[0113] S4.4: Determine the adaptive learning factor based on the estimation error

[0114] (18),

[0115] in, Used to determine the magnitude of the learning factor, Used to determine the shape of the curve that shows the learning factor changing with error; This indicates that the calculated value is restricted to [ , Within ], and 0≤ < <2. Increase the update speed when the prediction error is large, and reduce the steady-state weight fluctuation when the error decreases.

[0116] S4.5: Calculate candidate update quantities according to the normalized least mean square rule.

[0117] (19),

[0118] in, To prevent small positive numbers with a denominator of zero, the update step size can be normalized with the magnitude of the input and output increments, since the denominator contains the energy of the historical vector.

[0119] To distinguish between insufficient single-increment amplitude and lack of effective excitation in continuous windows, in a length of Construct the data activation matrix within the sliding window:

[0120] (20),

[0121] Reflecting the length of each component of the historical vector The degree of coverage of changes within the window, among which To control the periodic summation index. The cache contains fewer than [number] valid samples. At this point, we should not judge the need for continued incentives based on this. This serves as a historical cache integrity flag; it is used when the cache already contains the construct. Set to 1 if all required consecutive valid samples are available, otherwise set to 0.

[0122] (twenty one),

[0123] To update the enable flag, This serves as a valid identifier for the sample obtained from the consistency check of sensor range, analog-to-digital conversion results, and control commands. A flag indicating the integrity of the historical cache. Sliding window matrix. It is only used to evaluate the incentive coverage and parameter identifiability of historical data, and is not a mandatory condition for performing normalization updates in each control cycle.

[0124] The pseudo-partial derivative parameter norm projection operator can be further written as:

[0125] (twenty two),

[0126] Indicates projection onto a radius of The closed ball; When =0, the first branch is satisfied, and division by zero will not occur.

[0127] For the For each pseudo-partial derivative parameter component, a single-step variation limit can also be used:

[0128] (twenty three),

[0129] Each component forms a limited vector Different channels can use different... This is to accommodate the differences in dimensions and numerical ranges between the output history channel and the main phase shift channel.

[0130] S4.6: Perform excitation protection, single-step mutation limiting, and global bounded projection on candidate values ​​of pseudo-partial derivative parameters. Maintain [the following] when excitation is insufficient. = When the excitation is effective, first restrict the single-step change of each component according to equation (23), then perform norm projection according to equation (22), and obtain the effective pseudo-partial derivative parameter of this period according to equation (24).

[0131] (twenty four),

[0132] First, the component single-step limiting of equation (23) is executed, and then the norm projection of equation (22) is executed, so that the single-step change constraint and the global bounded constraint are in the same update chain. If the protection is triggered for multiple consecutive cycles, the system is restored to the most recent effective pseudo-partial derivative parameter or initial value, and the historical cache is refilled.

[0133] S4.7: Determine the online update order and numerical consistency

[0134] In each consecutive effective control cycle, the pseudo-partial derivative parameters of the previous cycle are used first. and historical vectors calculate Then formed from measured values and Then calculate sequentially. Original update volume Component limiting results and final projection results This sequence ensures that the error in this cycle is calculated only based on the data model actually used in the previous cycle, avoiding the use of an incomplete update in this cycle to correct the same error in reverse.

[0135] Each intermediate update quantity is calculated sequentially using the following formula:

[0136] ,

[0137] ,

[0138] ,

[0139] Equation (24-1) first calculates the denominator of the historical vector energy with regularization and the increment of the original parameters for the current period; Equation (24-2) forms the original candidate values ​​based on the parameters of the previous period and performs component limiting according to the single-step change limit corresponding to each channel; Equation (24-3) performs norm projection when updating enable and retains the parameters of the previous period when updating disable. Thus, the projected candidate parameter vector before entering the sign-holding judgment is obtained.

[0140] To reduce the impact of different physical quantities on online updates, this embodiment uses a current base value. = and principal phase shift angle base value = / 2 respectively for and Standardization is performed. Pseudo-partial derivative parameters are converted according to the corresponding input-output base values, and the same set of base values ​​is used for updates, predictions, and cost calculations.

[0141] Per-unit quantity, per-unit pseudo-partial derivative parameter, and candidate output-side DC filter inductor current restoration are calculated using the following formula:

[0142] ,

[0143] ,

[0144] ,

[0145] Equation (24-4) converts the output increment and control increment into dimensionless per-unit quantities using the current base value and the main phase shift angle base value, respectively; Equation (24-5) keeps the output historical channel parameters unchanged and converts the main phase shift angle channel parameters according to the ratio of the main phase shift angle base value to the current base value; Equation (24-6) completes the candidate prediction under the same per-unit scale, and then multiplies it by the current base value to restore the actual output side DC filter inductor current prediction value.

[0146] S4.8: Sign maintenance of the control direction for the main phase shift channel.

[0147] To maintain control direction consistency during forward power transfer, a sign-holding condition is set for the pseudo-partial derivative parameter component corresponding to the current principal phase shift angle increment. If the sign of this component is opposite to the initial direction, a sign flip is only accepted after the same reverse trend is detected for three consecutive valid cycles.

[0148] The sign preservation judgment and parameter vector recombination are calculated using the following formula:

[0149] ,

[0150] ,

[0151] Equation (24-7) determines whether the projection candidate parameters show a reverse trend based on the control direction accepted in the previous cycle, and accumulates the number of consecutive reverse trend cycles; in this embodiment, the number of consecutive effective control cycles required to accept the sign flip is taken as 3. Equation (24-8) maintains the main phase shift angle channel parameters of the previous cycle when the count has not reached 3, and accepts the projection candidate value after reaching 3, and uses the unit selection vector to write the component back to the pseudo-partial derivative parameter vector.

[0152] S4.9: Boundedness Statement for Pseudo-Partial Derivative Parameter Updates

[0153] Within a local working interval, the true equivalent pseudo-partial derivative parameters are approximated as constant. Define the parameter estimation error and Lyapunov-type functions:

[0154] (25),

[0155] in, This represents the true equivalent pseudo-partial derivative parameter within the current local working interval. This indicates the parameter estimation error. This represents its square norm.

[0156] Ignoring the time-varying terms of projected saturation and pseudo-partial derivative parameters, we substitute the normalized update into... This allows us to obtain a single-step change:

[0157] (26),

[0158] Equation (26) is used to describe the single-step variation trend of error energy under ideal normalization update. Its applicability is premised on the local true equivalent pseudo-partial derivative parameter being approximately constant. =1 and projection and single-step limiting are not activated. When 0≤ <2 and When the value is greater than 0, the normalization term helps to suppress update overshoot caused by changes in the magnitude of historical vectors. In practice, when measurement noise, time-varying parameters, disturbance residuals, or projection saturation exist, this embodiment only requires that the estimator be constrained by a preset boundary, without limiting the parameters to converge to a unique constant value.

[0159] This embodiment takes =1、 =5、 =0.005、 =1.0、 = and = The single-step variation limits for the pseudo-partial derivative parameters of the output history channel and the main phase shift angle history channel are set to 0.1 and 0.01, respectively, and the norm projection radius is set to... =3 The sliding excitation matrix is ​​only used when running diagnostics, and is taken as... =5.

[0160] S5: Determine the candidate principal phase shift angle center based on the beat-free relationship.

[0161] S5.1, expand the pseudo-partial derivative parameter vector according to the output history channel and the principal phase shift angle history channel:

[0162] (27),

[0163] in, Corresponding to the current undetermined principal phase shift angle increment The remaining components correspond to known output increments or historical principal phase shift angle increments.

[0164] S5.2, using the currently measured output-side DC filter inductor current and the historical increments that have occurred, construct a baseline prediction term that does not include the current undetermined principal phase shift angle increment:

[0165] (28),

[0166] when When =1, the second summation term in equation (28) is empty and is treated as zero.

[0167] S5.3, when the reference value of the DC filter inductor current on the upper output side has a step and it is necessary to limit the change of the single-cycle control quantity, a smooth reference value for prediction is generated first:

[0168] (29),

[0169] This embodiment uses a reference smoothing coefficient. =1, meaning no additional smoothing of the upper-level reference value; when it is necessary to suppress the impact of the reference step on the candidate center. It can be selected between 0.2 and 1. Unless otherwise specified below, let 1 be the value of 0.2. = It is used as a unified reference input for candidate center solving and candidate cost calculation.

[0170] S5.4, set the predicted value of the DC filter inductor current on the output side to equal the reference value for the next control cycle. This embodiment uses a threshold with a small denominator. =0.2 =0.01; when ≥ At that time, the inverse solution yields the increment of the deadbeat reference principal phase shift angle:

[0171] (30),

[0172] right implement[- , Amplitude limiting is applied to meet the constraint of the rate of change of the principal phase shift angle.

[0173] (31),

[0174] It is the center of a finite set of candidate principal phase shift angles. When < When historical data is invalid, reference values ​​are invalid, or pseudo-partial derivative parameters are in a recovery state, the inverse division solution is not performed; instead, the following is set: = This is to avoid the control quantity from changing abruptly due to an excessively small denominator.

[0175] S5.5: Perform continuous processing and safe rollback on candidate centers.

[0176] The deadbeat inverse kinematics only determines the search center of a finite set of candidates and is not directly used as a gate instruction. The controller sequentially performs reference value preprocessing, denominator validity checking, and... Rate of change limit and Angle boundary limiting; if any check fails, rollback to the actual applied value in the previous cycle. Therefore, even when the local inverse solution is temporarily unreliable, subsequent candidate cost calculations can still operate near the known safe angle.

[0177] This embodiment only considers forward power transmission, taking... =0、 =0.45 Maximum change in principal phase shift angle in a single cycle =0.03 When the inverse solution without errors fails, the candidate center reverts to the value actually applied in the previous cycle. .

[0178] S6: Generate a finite set of candidate candidates within the feasible principal phase shift angle domain.

[0179] S6.1, with Centered on, according to a fixed step size Expanding symmetrically to both sides There are several candidate values. This embodiment takes... =5、 =0.01 And simultaneously apply angular amplitude constraints and rate of change constraints:

[0180] (32),

[0181] in, =[ , ]∩[ - , + [This represents the set of allowed principal phase shift angles for this cycle.] Only one candidate value is retained for each duplicated candidate value after clipping, resulting in a deduplicated candidate set. .

[0182] S6.2, the feasible principal phase shift angle domain for this period is written as the intersection of the allowable amplitude range and the allowable rate of change range:

[0183] (33),

[0184] If power direction, modulation pulse width, or soft-switching range constraints also need to be met, the corresponding allowable range can be continued with... I'd like to connect.

[0185] (34),

[0186] This represents the number of duplicate candidate values ​​after removing the threshold; the actual number of candidates used in cost calculation is denoted as [value]. ≤ .

[0187] S6.3: Perform candidate coverage, boundary preservation, and deduplication.

[0188] Candidate values After unfolding symmetrically around the center, first with Find the intersection, then delete the overlapping elements due to amplitude limiting, and simultaneously establish a mapping from the original candidate index to the valid candidate index. If all expanded candidates are excluded by additional constraints, then at least the actual principal phase shift angle of the previous cycle should be retained. As a fallback candidate, it ensures that the set of valid candidates is always non-empty.

[0189] S6.4 adaptively adjusts the candidate resolution based on the error band of the DC filter inductor current on the output side. When the absolute value of the error exceeds the outer layer threshold, the resolution can be increased. Or increase To expand the coverage area; when the error enters the steady-state band, it can be reduced. To improve angular resolution. Adjusting the candidate set only changes the search density, without changing the dynamic linearization model structure of equation (11).

[0190] Number of candidates Choose an odd number to include candidate centers, and satisfy the following conditions: In this embodiment, the candidate coverage radius is 0.02. It is less than the maximum change in a single period of 0.03. When the current error of the output-side DC filter inductor is large or enters the steady-state error band, it can be adjusted accordingly. Adjust the number of candidates within the odd number range of 3 to 9.

[0191] S7: Predict the output-side DC filter inductor current one candidate at a time and select the optimal main phase shift angle

[0192] S7.1, on the candidate set The first in For each principal phase shift angle, calculate its candidate increment relative to the actual principal phase shift angle of the previous cycle. = - And construct candidate history vectors:

[0193] (35),

[0194] Each candidate vector differs only in the current principal phase shift increment component, thus allowing reuse of the remaining components from the historical vectors. When =1, the input history item after the candidate increment is empty.

[0195] S7.2, using the pseudo-partial derivative parameter vector updated at the same time to predict the first... The output-side DC filter inductor current corresponding to each candidate principal phase shift angle in the next control cycle:

[0196] (36),

[0197] in, Indicates the first The predicted value of the DC filter inductor current on the output side for the next cycle corresponding to each candidate main phase shift angle.

[0198] S7.3, using current base value = and principal phase shift angle base value = / 2 Normalize the current error and angle increment of the output-side DC filter inductor:

[0199] (37),

[0200] and These are the base values ​​of current and the base values ​​of principal phase shift angle, respectively; after normalization, both types of errors are dimensionless.

[0201] S7.4, Update the model uncertainty index based on the measured incremental prediction residual:

[0202] (38),

[0203] In this embodiment, the residual smoothing coefficient is used. =0.95. The residual variance is only used for operational monitoring and model reliability assessment, and is not directly added to the candidate ranking.

[0204] The normalized output-side DC filter inductor current tracking error and main phase shift angle change shown in Equation (37) are used to construct the candidate cost function shown in Equation (39) to eliminate the differences in the dimensions and orders of magnitude of different physical quantities; the residual variance shown in Equation (38) is only used for operation monitoring and judging the credibility of the data model, and is not directly added to the online sorting of candidate main phase shift angles.

[0205] S7.5, Calculate the uniform cost function for each valid candidate value:

[0206] (39),

[0207] in, The main phase shift angle change penalty coefficient is taken in this embodiment. =0.05; To constrain penalties, candidate items that satisfy all hard constraints are selected. =0, candidates that do not meet the hard constraints are directly excluded from the set of valid candidates, and no artificial large penalty value is set.

[0208] S7.6, Traverse all valid candidate values ​​and select the one with the smallest cost function value as the optimal principal phase shift angle for this cycle:

[0209] (40),

[0210] in, Indicates the first Within each control cycle, the cost function is... Valid candidate indexes for obtaining the minimum value This represents the set of valid candidate indices that satisfy the magnitude, rate of change, and disabled state constraints. When multiple candidate values ​​have the same or approximately the same cost function value, the one with the same value is preferred. The candidate value with the smallest distance; if the distances are still the same, then the candidate with the smallest distance is selected first. The candidate value with the smallest distance is selected; if they still cannot be distinguished, the unique candidate is determined according to the fixed index order in the valid candidate set. .

[0211] S7.7: Perform candidate batch prediction, constraint determination, and deterministic decision-making.

[0212] For each candidate The output increment and the input increment that have occurred in the historical vector remain unchanged; only the current principal phase shift angle increment component is replaced. The controller can pre-calculate the contribution of all candidate shared historical terms, and then supplement it for each candidate. This reduces repeated multiplication and accumulation without changing the prediction result of equation (36). Candidate predictions use the same time step... This ensures that different candidates are comparable.

[0213] Before cost calculation, it is checked whether the candidate angles meet the constraints of amplitude, rate of change, power direction and disabled modulation state; candidates that do not meet the hard constraints are not included in the valid candidate set. Equations (37) and (38) are used for dimensional unification and model credibility monitoring, and the online candidate ranking adopts the single cost function shown in equation (39).

[0214] When multiple candidates yield the same or nearly identical cost function values, the candidate with the smallest distance from the previous actual angle is selected first, followed by the candidate with the smallest distance from the candidate center. If they still cannot be distinguished, they are selected according to a fixed index order in the valid candidate set. This fixed selection rule ensures a definite control output for identical input data and reduces the back-and-forth switching of candidate angles caused by floating-point rounding or fixed-point quantization. The final selected... Perform a boundary check again before writing to the modulation register.

[0215] S7.8: Perform two-step candidate prediction when there is a one-cycle digital control delay.

[0216] When the sampling, calculation, and modulation register refresh cause the first The principal phase shift angle calculated by the period can only be found on the 1st. When applying +1 cycles, a two-step predictive compensation is used. First, it utilizes the currently applied... predict Current at +1:

[0217] (41),

[0218] In the formula, By what has already been actually implemented and up to The measured increment of the DC filter inductor current on the output side at any given time is used to predict the values ​​before the calculation is completed. Current at time +1.

[0219] For the proposed in the +1 cycle applied the first One candidate principal phase shift angle Construct a future history vector containing the predicted increment of the DC filter inductor current on the output side:

[0220] (42),

[0221] The vector is retained respectively Each output increment component and One principal phase shift angle increment component. When When =1, only the predicted output increment is retained; when When =1, only the candidate principal phase shift angle increment is retained.

[0222] (43),

[0223] During the two-step prediction period, a frozen pseudo-partial derivative parameter assumption is used, i.e. = The error will be updated according to the measured error once the next actual sample arrives.

[0224] (44),

[0225] in, This is a constraint penalty for the two-step prediction branch; it is 0 when the hard constraints are met, and candidates that do not meet the hard constraints are excluded from the valid candidate set. The choice is to make... The smallest candidate value is taken as the expected principal phase shift angle to be applied in the following week. If the controller can complete the calculation and refresh within the same cycle, a one-step prediction branch is used instead of performing the two-step compensation.

[0226] S7.9: Managing One-Step Forecasting and Two-Step Delay Compensation Branches

[0227] The controller determines the prediction branch based on the sampling completion time, the worst-case calculation time, and the modulation register latching time. If the calculation result can take effect before the specified latching point in the current cycle, one-step candidate prediction of equations (35) to (40) is executed; if it can only be applied in the next cycle, two-step prediction of equations (41) to (44) is executed. The branch is determined during controller initialization or operating mode switching, and is not temporarily switched within the same cycle.

[0228] This embodiment uses one-step candidate prediction as the main process, and the cost function parameters are taken as follows: = , = / 2 and =0.05. When the controller calculation result can only be applied in the next control cycle, the two-step delay compensation branch shown in equations (41) to (44) is adopted, where the same normalization scale and .

[0229] S8: Generate the modulation signal and apply the optimal main phase shift angle.

[0230] S8.1, adjust the optimal principal phase shift angle. Internal modulation angle corresponding to this working range , , Input modulation module. The modulation module according to... Figure 2 The angles shown are defined to convert each angle into the turn-on and turn-off times of each switching device in the primary two-level full-bridge and the secondary three-level NPC bridge, and the gate drive signal is output after adding the dead time.

[0231] S8.2, within the same segmented operating range, the output-side DC filter inductor current closed-loop adjustment is only performed online to adjust the main phase shift angle. The internal modulation angle is used to maintain the predetermined multi-level waveform and switching sequence. When crossing a preset voltage ratio or power direction partition, the internal modulation angle group is updated first, and then S1 to S9 are executed again in the new partition.

[0232] S9: Update control status and enter the next control cycle.

[0233] S9.1, after the optimal main phase shift angle is accepted by the modulation module and confirmed by the register feedback, = It also incrementally writes to the historical cache, saving the pseudo-partial derivative parameters, learning factor, minimum cost function value, candidate index, and data validity flag for the current period.

[0234] S9.2, New data will be collected in the next control cycle. The Adaline weights are corrected by the incremental current of the DC filter inductor on the actual output side, without replacing the measured values ​​with candidate predicted values. This forms a closed-loop rolling process of "updating the model with measured data - generating and predicting candidates with the model - selecting the main phase shift angle based on the cost function - updating the model again with the actual response".

[0235] Simulation verification

[0236] To verify the output-side DC filter inductor current control effect of the method of the present invention under different operating conditions, this embodiment establishes a simulation model of a three-level NPC-DAB converter and verifies it using the parameters shown in Table 1. The rated power is... The input voltage is The output voltage range is to The transformer turns ratio is The switching frequency is The series inductance is The output DC filter inductor is The controlled variable in the simulation waveform is uniformly represented as the DC filter inductor current on the output side. The internal modulation angle is uniformly denoted as... , , And use segmented preset values.

[0237] The simulation reference input voltage in this embodiment is determined as follows: The control algorithm uses the per-unit parameters and finite candidate settings shown in Table 2.

[0238] Table 1 Main parameters of the simulation model

[0239]

[0240] Table 2 Main Parameters of Control Algorithm

[0241]

[0242] Figure 5 , Figure 6 and Figure 7 Simulation responses under three typical operating conditions are presented respectively. Figure 5The initial reference value of the DC filter inductor current on the output side is: ,exist Time becomes ,exist Time to restore to . Figure 6 The load resistor is composed of Change to It was subsequently restored to . Figure 7 Input voltage The benchmark was raised to It was subsequently restored to .Depend on Figures 5 to 7 It can be seen that under the conditions of reference current step, load change and input voltage disturbance, the DC filter inductor current on the output side can track the reference value and recover stability after the disturbance.

[0243] Figure 8 and Figure 9 The local magnified waveform during the load resistance change process is further presented. Figure 8 Output voltage at the moment of loading Output side DC filter inductor current and series inductor current , Figure 9 For the moment of load reduction , and Local waveforms are used to observe the dynamic changes in output-side variables and series inductor current at the moment of disturbance.

[0244] The simulation results above demonstrate that, while retaining the multi-degree-of-freedom modulation structure of the three-level NPC-DAB converter, the main phase shift angle can be used as an online control input to implement dynamic linearized data-driven predictive control of the output-side DC filter inductor current. Since the predictive model is constructed from real-time input and output data and updated online, the method does not require precise leakage inductance, resistance, capacitance, and load parameters as necessary modeling conditions.

[0245] Example 2

[0246] This embodiment provides a data-driven predictive control system for a three-level NPC-DAB converter.

[0247] A computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the aforementioned three-level NPC-DAB converter data-driven predictive control method.

[0248] A terminal device includes a processor and a computer-readable storage medium, the processor being configured to implement various instructions; the computer-readable storage medium being configured to store multiple instructions adapted to be loaded and executed by the processor to provide a data-driven predictive control method for a three-level NPC-DAB converter.

[0249] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A data-driven predictive control method for a three-level NPC-DAB converter, characterized in that, include: Acquire the output-side DC filter inductor current data and the main phase shift angle; Synchronization, quantification, and isolation of abnormal samples are performed based on the acquired data; Construct input and output incremental data vectors based on the processed sample data; A dynamic linearized input-output incremental model is established based on incremental data vectors, and the pseudo-partial derivative parameter vector of the model is updated online. Candidate principal phase shift angle centers are determined based on the updated pseudo-partial derivative parameter vector; A priority candidate set is generated based on the candidate principal phase shift angle center; The optimal principal phase shift angle is generated based on the preferred candidate set and the updated pseudo-partial derivative parameter vector. The final modulation signal is generated based on the optimal main phase shift angle.

2. The data-driven predictive control method for a three-level NPC-DAB converter according to claim 1, characterized in that, The process of synchronizing, quantifying, and isolating abnormal samples based on the acquired data includes setting a control period. 100 , No. Within each control cycle, the time relative to the carrier wave is fixed. Samples were taken at each time point, among which =3, sampling time is: in, To control the starting point of the timing, For the first The fixed offset of each sampling point relative to the start of the control cycle, each The location is set far from gate switching and dead-time switching; three sampled values ​​are used to form a periodic output-side DC filter inductor current sample; then, the analog-to-digital converter bit depth is set to... The lower and upper limits of current measurement are respectively and The current quantization step size is then: and the original sampled values First, perform range saturation, then round down to the nearest quantization level to obtain: Furthermore, an anti-outlier statistic is constructed using the median and median absolute deviation: in, To prevent small positive numbers with a scale of zero; The outlier determination coefficient; To preserve the sample index set, when all three samples are valid and no outliers are detected, Includes all three samples; finally, the retained sample index set is denoted in ascending order as follows: = The average of the corresponding quantified samples is calculated to form the first... One control cycle is used for the periodic output samples of dynamic linearization: in, for The Middle A reserved sample index, To retain the sample size; because >0 and >0, It must contain at least the median corresponding sample; and each periodic sample must be appended with a control period number. The system includes an analog-to-digital conversion valid flag, a range flag, and a modulation register consistency flag, the logical AND of which is denoted as... Only when =1、 =1 and - When =1, the current sample and the previous sample form a continuous valid sample pair.

3. The data-driven predictive control method for a three-level NPC-DAB converter according to claim 2, characterized in that, The construction of input / output incremental data vectors based on the processed sample data includes, to avoid discontinuous differential interference caused by buffer breakpoints or invalid samples, using only adjacent period samples within the same continuous valid data segment, calculating the output-side DC filter inductor current increment and the actual applied main phase shift angle increment, respectively, expressed as: ,in, It reflects the change in the DC filter inductor current on the output side during adjacent effective control cycles. Reflects the change in the actual applied principal phase shift angle between adjacent effective control cycles. Indicates the first The main phase shift angle is actually written to the period and confirmed by the modulation register; candidate angles are used. This indicates that, in order to simultaneously characterize the multi-cycle effects of the output-side DC filter inductor current inertia and the main phase shift angle within a finite-length history window, the output history order is taken. =1. Principal phase shift angle and historical order =3, combine the corresponding increments in order from newest to oldest to obtain the input-output increment history vector: ,in The dimension is 4×1, and the online model update uses actual occurrences. and actual measurement During the candidate prediction phase, only the current principal phase shift increment channel is replaced with... The remaining historical components remain unchanged; therefore, only the data that has already occurred is used in the model update stage, and only the current candidate control variable is changed in the candidate prediction stage, so as to maintain the consistency of the historical data source; the candidate input-output incremental historical vector is calculated according to the following steps: , in, For the first The increment of each candidate main phase shift angle relative to the actual main phase shift angle of the previous cycle. To select the unit selection vector for the current principal phase shift increment channel, For the supply of the first The input and output incremental history vectors are used for each candidate prediction; then, integrity and activation checks are performed on the history vectors. If the number of consecutive valid samples is insufficient, the buffer crosses a breakpoint, or the L2 norm of the input and output incremental history vectors is less than the activation threshold, the detection is performed. If the pseudo-partial derivative parameter update is frozen, then finite candidate predictions can still be performed near the safe backoff angle. The integrity check is used to avoid updating the model across invalid sample segments, and the activation check is used to avoid unstable parameter updates when the input-output increment is too small. The integrity and activation checks are calculated according to the following steps: ， ， ， in, The length of the continuous effective sample window required to construct the history vector. As a window integrity flag, For historical vector energy, Update the enable flag for the parameters; For a predefined finite set of symmetric indices containing zero, The candidate step size for safe backoff is no greater than the maximum change in the main phase shift angle in a single cycle.

4. The data-driven predictive control method for a three-level NPC-DAB converter according to claim 3, characterized in that, The establishment of a dynamic linearized input-output incremental model based on incremental data vectors includes dividing the dynamic linearized data model into tight format, partial format, and full format according to the usage scope of historical input and output information. The tight format only retains the current principal phase shift angle increment, and is expressed as: The partial scheme uses the current and multiple historical principal phase shift angle increments, expressed as: ; The full format retains both the output increment and the principal phase shift angle increment: Then, a full-format dynamic linearized input-output increment model is established, expressed as: ,in, This is the predicted increment of the DC filter inductor current on the output side for the next control cycle. Given a 4×1 pseudo-partial derivative parameter vector, the one-step prediction value of the DC filter inductor current on the output side in the next control cycle is obtained: The incremental model is then converted into an absolute prediction model that can be directly compared with the reference value of the DC filter inductor current on the output side. To enable the pseudo-partial derivative parameters to be identified online based on real-time input and output data, the pseudo-partial derivative parameters are mapped to Adaline online estimator weights, expressed as: In the formula, Indicates by the first The -1 periodic pseudo-partial derivative parameters and the history vector obtained from the first The estimated increment of the DC filter inductor current on the output side during the cycle; to obtain the supervision error used for weight update, the difference between the measured increment of the DC filter inductor current on the output side in the current cycle and the model estimate in the previous cycle is expressed as: in, This represents the deviation between the measured increase in the DC filter inductor current at the output side in the current cycle and the predicted value from the data model in the previous cycle, serving as a monitoring signal for the online update of Adaline. From the Adaline structure, each node in the input layer receives... One component, the network weights, i.e., the pseudo-partial derivative parameter components, is the output of the linear activation function: Where l is the input node number. Let l be the l-th pseudo-partial derivative parameter component. Let l be the l-th component of the input-output incremental history vector. To determine the linear activation output of the adaptive linear neuron; and to quantify the supervision error and obtain the gradient descent direction of the pseudo-partial derivative parameters, using the measured increment of the current period as the supervision signal, the instantaneous squared error objective function is defined as follows: ,in The objective function for the instantaneous squared error of Adaline weight updates; using a 1 / 2 coefficient so that the constant 2 no longer appears after differentiating the pseudo-partial derivative parameter, is expressed as: .

5. The data-driven predictive control method for a three-level NPC-DAB converter according to claim 4, characterized in that, The online update of the pseudo-partial derivative parameter vector of the model includes: to balance the update speed of model parameters in the dynamic stage and the suppression of fluctuations in the steady-state stage, an adaptive learning factor is determined based on the estimation error, and parameter updates are performed in the order of first calculating the adaptive learning factor and then calculating the candidate update amounts of the pseudo-partial derivative parameters. The adaptive learning factor is expressed as: ,in, >0 is used to determine the magnitude of the learning factor. Used to determine the shape of the curve that shows the learning factor changing with error; This indicates that the calculated value is restricted to [ , Within ], and 0≤ < <2; then calculate the candidate update quantity according to the normalized least mean square rule, expressed as: ,in, To prevent small positive numbers with a denominator of zero; let This serves as a historical cache integrity flag; it is used when the cache already contains the construct. The value is 1 if all consecutive valid samples are required, and 0 otherwise, as follows: ,in To update the enable flag, This serves as a valid identifier for the sample obtained from the consistency check of sensor range, analog-to-digital conversion results, and control commands. This serves as a historical cache integrity flag; the pseudo-partial derivative parameter norm projection operator is further expressed as: ,in Indicates projection onto a radius of The closed ball; finally, the first... Each pseudo-partial derivative parameter component is used, and a single-step variation limit is applied: Each component forms a limited vector. Then, excitation protection, single-step mutation limiting, and global bounded projection are performed on the candidate values ​​of the pseudo-partial derivative parameters. The online update order and numerical consistency are then determined. In each consecutive effective control cycle, the pseudo-partial derivative parameters from the previous cycle are used first. and historical vectors calculate Then formed from measured values and Then calculate sequentially. Original update volume Component limiting results and final projection results This ensures that the supervision error in this period is generated solely from the data model actually used in the previous period, and that the updated pseudo-partial derivative parameter vector is used by all candidate predictions in this period; the intermediate quantities of the pseudo-partial derivative parameter update are calculated in the following order: ， ， ， in, To normalize and update the denominator, The component limiting is applied separately to each channel as the increment of the original parameters for this period. The parameter vector is used to complete the excitation protection, component limiting and norm projection, and before entering the sign preservation judgment.

6. The data-driven predictive control method for a three-level NPC-DAB converter according to claim 5, characterized in that, The step of determining the candidate principal phase shift angle center based on the updated pseudo-partial derivative parameter vector includes expanding the pseudo-partial derivative parameter vector according to the output history channel and the principal phase shift angle history channel, as follows: ,in, Corresponding to the current undetermined principal phase shift angle increment And using the currently measured DC filter inductor current on the output side and the historical increments that have already occurred, a baseline prediction term is constructed that does not include the current increment of the principal phase shift angle to be determined: ,in, For the first Output-side DC filter inductor current increment channel parameters For the first One main phase shift angle increment channel parameter, and These are the summation indices for the output history and the principal phase shift angle history, respectively; when When =1, the second summation term in the benchmark prediction term regarding the previous main phase shift angle increment is empty and treated as zero; when the reference value of the DC filter inductor current on the upper output side has a step and it is necessary to limit the change of the single-cycle control quantity, a smooth reference value for prediction is generated first: in, This serves as the reference value for the output-side DC filter inductor current in the next control cycle after smoothing. Provide a reference value for the upper layer. The smoothing coefficient is the reference value, and ,make = It is used as a unified reference input for candidate center solving and candidate cost calculation, and a threshold with a small denominator is taken. =0.2 =0.01; when ≥ At that time, the inverse solution yields the increment of the deadbeat reference principal phase shift angle: To prevent deadbeat inversion from causing a main phase shift angle jump exceeding the converter's allowable rate of change, the following measures are taken: implement[- , The amplitude is limited to meet the constraint of the rate of change of the principal phase shift angle, expressed as: in, The center of the finite set of candidate principal phase shift angles is determined; finally, continuous processing and safe backoff are performed on the candidate centers, using the controller to sequentially perform reference value preprocessing, denominator validity check, and... Rate of change limit and Angle boundary limiting; if any check fails, rollback to the actual applied value in the previous cycle. The candidate center is obtained using the effective pseudo-partial derivative parameter vector of the current cycle obtained in claim 5, and the failure backoff value does not directly replace the actual main phase shift angle. It is only written into the historical cache after being confirmed by the modulation register.

7. The data-driven predictive control method for a three-level NPC-DAB converter according to claim 6, characterized in that, The step of generating a priority candidate set based on candidate principal phase shift angle centers includes generating a finite candidate set within the feasible principal phase shift angle domain to cover the neighborhood of the candidate centers while simultaneously satisfying the principal phase shift angle magnitude and rate of change constraints, with the priority set being... Centered on, according to a fixed step size Expanding symmetrically to both sides There are 1 candidate values, take 1 =5、 =0.01 And simultaneously apply angular amplitude constraints and rate of change constraints: in, =[ , ]∩[ - , + [This represents the set of allowed principal phase shift angles for this period; after clipping, only one candidate value is retained for each duplicate candidate value, resulting in a deduplicated candidate set.] Then, the feasible principal phase shift angle domain for this period is written as the intersection of the allowable amplitude interval and the allowable rate of change interval, expressed as: in, and These are the lower and upper limits of the principal phase shift angle amplitude, respectively. This represents the maximum change in the principal phase shift angle over a single cycle. This represents the feasible region of the main phase shift angle for this control cycle; simultaneously, if constraints on power direction, modulation pulse width, or soft-switching intervals need to be met, the corresponding allowable intervals will be extended. Find the intersection; expressed as: ,in This represents the number of duplicate candidate values ​​after removing the threshold; the actual number of candidates used in cost calculation is denoted as [value]. ≤ This involves performing candidate coverage, boundary preservation, and deduplication, with candidate values ​​being... After unfolding symmetrically around the center, first with The intersection is calculated, and overlapping elements due to amplitude limiting are deleted. Simultaneously, a mapping from the original candidate index to the valid candidate index is established to ensure that prediction and cost calculation are performed only on candidates that are non-repeating and satisfy the feasible region constraint. Finally, the candidate resolution is adaptively adjusted based on the error band of the DC filter inductor current on the output side. When the absolute value of the error is greater than the outer threshold, the resolution is increased. Or increase To expand the coverage area; when the error enters the steady-state band, reduce... To improve angular resolution, the outer layer threshold and the steady-state band threshold are respectively denoted as... and ,and The number of candidates and the step size are adjusted in stages within a preset range; the number of candidates Choose an odd number to include candidate centers, and satisfy the following conditions: .

8. The data-driven predictive control method for a three-level NPC-DAB converter according to claim 7, characterized in that, The process of generating the optimal principal phase shift angle based on the priority candidate set and the updated pseudo-partial derivative parameter vector includes, under the same model parameters, comparing the impact of each candidate principal phase shift angle on the output-side DC filter inductor current of the next control cycle, and updating the candidate set... The first in For each principal phase shift angle, calculate the candidate increment relative to the actual principal phase shift angle of the previous cycle. = - And construct candidate history vectors: in, For the first Candidate input-output increment history vectors This is a valid candidate index; the vector only replaces the current principal phase shift angle increment component with the candidate principal phase shift angle increment, and the remaining historical components are all composed of data that have actually occurred; then, the pseudo-partial derivative parameter vector updated at the same time is used to predict the first... The output-side DC filter inductor current corresponding to each candidate principal phase shift angle in the next control cycle is specifically calculated by taking the inner product of the candidate input / output increment history vector and the pseudo-partial derivative parameter vector updated in this cycle, and adding it to the currently measured output-side DC filter inductor current, expressed as: ,in, Indicates the first The predicted output-side DC filter inductor current for the next cycle corresponding to each candidate main phase shift angle; to eliminate the dimensional and order-of-magnitude differences between the output-side DC filter inductor current error and the main phase shift angle change, so that both can jointly constitute a candidate cost function, a current base value is used. = and principal phase shift angle base value = / 2 The error of the DC filter inductor current and the angle increment on the output side are normalized. The normalized error is expressed as: ,in and These are the base current value and the base main phase shift angle value, respectively. To determine a unique control quantity while considering both the tracking error of the DC filter inductor current on the output side and the amplitude of the main phase shift angle change, the online candidate selection uniformly adopts the main cost function. First, the unified cost function is calculated for each valid candidate value: ,in, The penalty coefficient for main phase shift angle change. To constrain penalties, candidate items that satisfy all hard constraints are selected. =0, candidates that do not meet the hard constraints are directly excluded from the valid candidate set; by traversing all valid candidate values, the one with the smallest cost function value is selected as the optimal principal phase shift angle for this cycle: in, Indicates the first Within each control cycle, the cost function is... Valid candidate indexes for obtaining the minimum value This represents the set of valid candidate indices that satisfy the magnitude, rate of change, and disabled state constraints.

9. A data-driven predictive control method for a three-level NPC-DAB converter according to claim 8, characterized in that, The process of generating the optimal principal phase shift angle based on the priority candidate set and the updated pseudo-partial derivative parameter vector also includes processing each candidate... The output increment and the input increment that have occurred in the historical vector remain unchanged. Only the current principal phase shift angle increment component is replaced. The contribution of all candidate shared historical terms is pre-calculated, and then supplemented for each candidate. This reduces redundant multiplication and accumulation without changing the predicted DC filter inductor current of the candidate output side, and the candidate prediction uses the same time... This ensures comparability between different candidates; when sampling, computation, and modulation register refresh cause the first... The principal phase shift angle calculated by the period can only be found on the 1st. When applying +1 cycles, a two-step predictive compensation is used. First, it utilizes the currently applied... predict Current at +1: In the formula, By what has already been actually implemented and up to The measured output-side DC filter inductor current increment at time t constitutes the calculation; for the proposed... +1 cycle applied the first One candidate principal phase shift angle Construct a future history vector containing the predicted increment of the DC filter inductor current on the output side: ， The vector is retained respectively Each output increment component and One principal phase shift angle increment component; when When =1, only the predicted output increment is retained; when When the value is 1, only the candidate principal phase shift angle increment is retained; during the two-step prediction period, the frozen pseudo-partial derivative parameter assumption is adopted, i.e. = The error is updated according to the measured error after the next actual sample arrives, as shown below: ,in, This is the constraint penalty for the two-step prediction branch; it is 0 when the hard constraint is satisfied, and candidates that do not satisfy the hard constraint are excluded from the valid candidate set. The smallest candidate value is taken as the principal phase shift angle to be applied in the following week.

10. A data-driven predictive control system for a three-level NPC-DAB converter, used to execute the dynamic linearization data-driven output predictive control method for a three-level NPC-DAB converter as described in claim 1, characterized in that, include: The data acquisition module is configured to acquire the output-side DC filter inductor current data and the main phase shift angle; The preprocessing module is configured to perform synchronization, quantization, and isolation of outlier samples based on the acquired data. The incremental module is configured to construct input and output incremental data vectors based on the processed sample data. The pseudo-partial derivative module is configured to build a dynamically linearized input-output incremental model based on the incremental data vector, and simultaneously update the pseudo-partial derivative parameter vector of the model online. The candidate module is configured to determine the candidate principal phase shift angle center based on the updated pseudo-partial derivative parameter vector; The priority module is configured to generate a priority candidate set based on the candidate principal phase shift angle center; The optimal module is configured to generate the optimal principal phase shift angle based on the priority candidate set and the updated pseudo-partial derivative parameter vector. The modulation module is configured to generate the final modulation signal based on the optimal main phase shift angle.

Images