Two-level bus voltage balancing device and balancing method thereof

Abstract

The invention discloses a two-level bus voltage balancing device and a balancing method thereof, and relates to the technical field of circuit voltage balancing, and the method comprises a positive bus capacitor, a negative bus capacitor, a flyback transformer, a first switching device and a second switching device. The flyback transformer comprises a first winding and a second winding; the first winding is connected between the positive bus capacitor and the first switching device, and the second winding is connected between the negative bus capacitor and the second switching device; when the voltage of the positive bus capacitor is higher than a preset threshold value, the first switching device is triggered to be conducted to form an energy transfer loop; when the voltage of the negative bus capacitor is higher than a preset threshold value, the second switching device is triggered to be conducted to form an energy transfer loop; the first switching device and the second switching device are electrically connected with a controller, and the controller outputs PWM control signals according to real-time voltage values of the positive bus capacitor and the negative bus capacitor. The method has the effects of improving the energy utilization efficiency and reducing the cost.

Description

Technical Field

[0001] This application relates to the field of circuit voltage equalization technology, and in particular to a two-level bus voltage equalization device and its equalization method. Background Technology

[0002] Currently, two-level inverter circuits hold a crucial position in the field of power electronics, widely used in scenarios such as two-level inverters and UPS circuits. With the continuous development of power electronics technology, two-level inverter circuits play a key role in various industrial and civil equipment, providing efficient solutions for power conversion and transmission, and driving the development of related industries. They possess certain advantages in achieving power conversion and control, can meet the needs of different loads, and are of great significance in improving energy utilization efficiency and equipment performance.

[0003] In two-level inverter circuits, existing technologies often employ a bleeder circuit to address the voltage imbalance between the positive and negative buses. When the positive bus voltage is high, the positive-side bleeder circuit is activated to discharge. Specifically, this is done by opening the corresponding switch, allowing the positive bus capacitor, switch, and resistor to form a discharge loop, thereby reducing the positive bus voltage. Conversely, when the negative bus voltage is high, the negative-side bleeder circuit is activated to discharge, allowing the negative bus capacitor, resistor, and switch to form a discharge loop, thus reducing the negative bus voltage. This method can regulate the bus voltage to a certain extent, maintaining the basic operation of the circuit.

[0004] However, the existing technique of regulating bus voltage through a discharge circuit has significant drawbacks. During the regulation process, excess energy is wasted in the form of heat, etc., and this energy cannot be fully utilized, resulting in a huge waste of energy and hindering the improvement of energy efficiency and cost reduction. Summary of the Invention

[0005] In order to overcome the above-mentioned technical problems, improve energy utilization efficiency and reduce costs, this application provides a two-level bus voltage equalization device and its equalization method.

[0006] Firstly, the objective of this invention is achieved through the following technical solution: A two-level bus voltage balancing device includes: a positive bus capacitor, a negative bus capacitor, a flyback transformer, a first switching device, and a second switching device. The flyback transformer includes a first winding and a second winding; the first winding is connected between the positive bus capacitor and the first switching device, and the second winding is connected between the negative bus capacitor and the second switching device. When the voltage of the positive bus capacitor is higher than a preset threshold, the first switching device is triggered to conduct to form an energy transfer loop; when the voltage of the negative bus capacitor is higher than a preset threshold, the second switching device is triggered to conduct to form an energy transfer loop. The first switching device and the second switching device are electrically connected to a controller, which outputs a PWM control signal based on the real-time voltage values ​​of the positive bus capacitor and the negative bus capacitor.

[0007] By adopting the above technical solution, when the voltage of the positive or negative bus capacitor is higher than a preset threshold, the first or second switching device is triggered to form an energy transfer loop, thereby realizing energy transfer between the positive and negative bus capacitors. This avoids faults caused by bus voltage imbalance and does not waste energy, making full use of energy. This invention utilizes a positive bus capacitor, a negative bus capacitor, a flyback transformer, a first switching device, and a second switching device. When the voltage of the positive or negative bus capacitor is higher than a preset threshold, an energy transfer loop is formed, achieving voltage balance between the positive and negative bus capacitors. Based on this, a controller is set up to output a PWM control signal according to the real-time voltage values ​​of the positive and negative bus capacitors, enabling more precise control of the conduction of the first and second switching devices, thus more effectively achieving energy transfer and bus voltage balance.

[0008] In a preferred embodiment of this application: the turns ratio of the first winding to the second winding is 1:1; both the first switching device and the second switching device are MOSFET devices.

[0009] By adopting the above technical solution, the first winding and the second winding of the flyback transformer are designed with a 1:1 turns ratio, and MOSFETs are selected as the first and second switching devices. On the one hand, this ensures that the voltage gain is 1 when energy is transferred bidirectionally between the positive and negative buses, simplifies the control logic and avoids voltage mismatch; on the other hand, MOSFETs have the advantages of low on-resistance, high switching speed and easy driving.

[0010] Secondly, the objective of this invention is achieved through the following technical solution: A balancing method applied to the two-level bus voltage balancing device described above, the method comprising: The voltage sampling module collects the voltage values ​​of the positive bus capacitor and the negative bus capacitor in real time, and calculates the bus voltage deviation. Based on the bus voltage deviation and historical operating data, a self-learning neural network controller is used to generate initial equalization control parameters. The initial equalization control parameters are input to the model predictive equalization controller. Based on the current load current, system temperature, and flyback transformer core saturation state, a multi-objective optimization function is constructed, and the optimal PWM control sequence is obtained by solving the function. In the low-frequency operating range, a space vector modulation strategy is used to reduce the switching frequency, and in the high-frequency operating range, a pulse sequence optimization strategy is switched to suppress transient oscillations, achieving millisecond-level bidirectional energy transfer. When a load change event or bus voltage deviation exceeds a preset dynamic threshold is detected, a feedforward compensation command is triggered. A nonlinear extended state observer is used to estimate the total amount of external disturbance, and the total amount of external disturbance is superimposed on the output of the model predictive equalization controller to reconstruct the equalization control path. After each equalization control cycle, the network weight parameters of the self-learning neural network controller are updated based on the actual bus voltage convergence speed and the theoretical prediction error.

[0011] By adopting the above technical solutions, a closed-loop intelligent equalization system is constructed by integrating a self-learning neural network controller, a model predictive equalization controller, a hybrid modulation strategy, and a feedforward compensation mechanism. The self-learning neural network generates initial control parameters based on historical data, shortening response latency; the model predictive controller combines multiple physical quantities to construct a multi-objective optimization function, achieving synergistic optimization of accuracy and efficiency; space vector modulation and pulse sequence optimization are switched according to frequency bands, balancing low loss and disturbance rejection; resonant injection and timing adjustment accelerate energy transfer; feedforward compensation handles sudden load changes; and online weight updates enable continuous evolution. Overall, millisecond-level dynamic response and high energy efficiency are achieved.

[0012] In a preferred embodiment of this application, the initial equalization control parameters include a first duty cycle initial value of the first switching device, a second duty cycle initial value of the second switching device, and an enable flag for the resonant assisted energy transfer mode; The self-learning neural network controller adopts a convolutional-long short-term memory hybrid network structure. The input variables of the self-learning neural network controller include: the current bus voltage deviation, the bus voltage deviation change rate, the load power change rate, the ambient temperature, and the actual duty cycle sequence of the first and second switching devices in the past N control cycles. The output variables of the self-learning neural network controller include normalized initial values ​​of the first duty cycle, the second duty cycle, and an enable flag for the resonant assisted energy transfer mode. The self-learning neural network controller loads pre-trained network weights during the system power-on phase and calculates the loss function based on the actual equilibrium performance index at the end of each equilibrium control cycle, and updates the network weight parameters using an online fine-tuning algorithm.

[0013] By adopting the above technical solution, the specific input-output structure of the self-learning neural network controller is defined—taking multi-dimensional states such as bus voltage deviation, rate of change, and load power change rate as inputs, and outputting the initial duty cycle value and resonance activation flag. A CNN-LSTM hybrid architecture is used to capture spatiotemporal correlation features, enabling the initial control strategy to have operational adaptability. Pre-trained weights ensure cold-start performance, while the online fine-tuning mechanism allows the controller to continuously approach the optimal mapping as it accumulates operational experience.

[0014] In a preferred embodiment of this application, the multi-objective optimization function constructed by the model predictive equilibrium controller is defined as: in, The weighting coefficients are dynamic and adjusted in real time based on the system stability margin; k+1 represents the next control step. This is the predicted value of the bus voltage deviation for the next control step; These are the equivalent on-state currents of the first switching device Q1 and the second switching device Q2, respectively. These are the on-resistances of the first switching device Q1 and the second switching device Q2, respectively; Let these be the first and second duty cycles of the current period, respectively. This is the value corresponding to the previous period; The model predictive equalization controller continuously optimizes the control sequence for the next M steps in each control cycle, and executes only the first step of the control input.

[0015] By adopting the above technical solution, a multi-objective optimization function is constructed, incorporating voltage deviation, switching losses, and control rate of change. Weighting coefficients based on dynamically adjusted system stability margins are introduced, enabling the model predictive equalization controller to automatically balance the three objectives of "fast equalization," "low loss," and "smooth control" under different operating conditions. The rolling optimization mechanism ensures control foresight, while executing only the first output step also considers real-time performance.

[0016] In a preferred embodiment of this application, the conditions for enabling the resonant-assisted energy transfer mode include: The absolute value of the bus voltage deviation is greater than the first voltage threshold, and the rate of change of the bus voltage deviation is greater than zero.

[0017] In a preferred example, after triggering a feedforward compensation instruction, the method includes: A nonlinear extended state observer is used to jointly estimate the unmodeled dynamic characteristics and external load disturbances, and outputs the total disturbance estimate. Based on the total disturbance estimate and the preset feedforward gain coefficient, the feedforward compensation amount is calculated; The feedforward compensation is superimposed on the output control quantity of the model predictive equalization controller to generate the final control quantity; the preset feedforward gain coefficient is adaptively adjusted according to the system closed-loop bandwidth.

[0018] By adopting the above technical solution, when disturbances such as sudden load changes occur, the total disturbance is estimated in real time by a nonlinear extended state observer, and a feedforward compensation is generated and superimposed on the main controller output, thereby achieving active cancellation of the disturbance rather than passive adjustment. The feedforward gain coefficient is adaptively adjusted according to the system closed-loop bandwidth to avoid over-compensation that could cause oscillations.

[0019] In a preferred embodiment, this application automatically switches to a security balancing mode during system startup or fault recovery, the security balancing mode including: The maximum duty cycle of the first and second switching devices is limited to no more than 30%. Disable resonant-assisted energy transfer mode; Energy transfer is performed using a fixed-frequency, low-duty-cycle exploratory PWM signal until the absolute value of the bus voltage deviation is less than the preset voltage deviation threshold. Then, the workflow of the self-learning neural network controller and the model predictive equalization controller is activated.

[0020] By adopting the above technical solution, during highly uncertain phases such as system startup or fault recovery, the system automatically switches to a safe balancing mode: limiting the switch duty cycle, disabling resonance, and employing tentative low-power PWM, effectively preventing overcurrent, overvoltage, or magnetic saturation risks caused by unknown initial states. The advanced intelligent control module is activated only after the bus voltage deviation converges to within the safe threshold; this significantly improves the reliability and user safety of the two-level bus voltage balancing device throughout its entire lifecycle.

[0021] In summary, this application includes at least one of the following beneficial technical effects: 1. Avoid energy waste and make full use of excess energy. By triggering the corresponding switching devices to form an energy transfer circuit, the excess energy of the capacitor on the high-voltage side of the bus is transferred to the capacitor on the other side for charging, changing the way that traditional discharge circuits waste energy. 2. Balance the positive and negative bus voltages. Trigger the corresponding switching devices to conduct according to the positive and negative bus capacitor voltages to keep the positive and negative bus voltages balanced and prevent faults caused by bus voltage imbalance. Attached Figure Description

[0022] Figure 1 This is a circuit diagram of a two-level bus voltage equalization device in one embodiment of this application; Figure 2 This is a flowchart of an equalization method applied to a two-level bus voltage equalization device in one embodiment of this application. Detailed Implementation

[0023] The present application will be further described in detail below with reference to the accompanying drawings.

[0024] In one embodiment, such as Figure 1 As shown, this application discloses a two-level bus voltage equalization device, which... Figure 1 Taking the circuit diagram shown as an example, the two-level bus voltage balancing device includes a positive bus capacitor C1, a negative bus capacitor C2, a flyback transformer TX1, a first switching device Q1, and a second switching device Q2. The flyback transformer TX1 includes a first winding N1 and a second winding N2. The turns ratio of the first winding N1 to the second winding N2 is 1:1. Both the first switching device Q1 and the second switching device Q2 are MOSFET devices; in this embodiment, an NMOS transistor is used. The first winding N1 is connected between the positive bus capacitor C1 and the first switching device Q1, and the second winding N2 is connected between the negative bus capacitor C2 and the second switching device Q2. When the voltage of the positive bus capacitor C1 is higher than a preset threshold, the first switching device Q1 is triggered to conduct to form an energy transfer loop; when the voltage of the negative bus capacitor C2 is higher than a preset threshold, the second switching device Q2 is triggered to conduct to form an energy transfer loop.

[0025] The first switching device Q1 and the second switching device Q2 are electrically connected to a controller (not shown in the figure). The controller outputs a PWM control signal based on the real-time voltage values ​​of the positive bus capacitor C1 and the negative bus capacitor C2. For example, the controller uses a TITMS320F28379D DSP chip with a sampling frequency of 10kHz and a PWM resolution of 12 bits.

[0026] like Figure 1 As shown, the positive terminal of the positive bus capacitor C1 is connected to the high-voltage terminal VDC+ of the DC bus, and the negative terminal is grounded; the positive terminal of the negative bus capacitor C2 is grounded, and the negative terminal is connected to the low-voltage terminal VDC− of the DC bus. The source of the first switching device Q1 is grounded, and its gate receives the PWM signal from the controller. The source of Q2 is grounded, and its gate is independently controlled by the controller. Under normal operating conditions, the controller acquires the voltage VC1 across the positive bus capacitor C1 and the voltage VC2 across the negative bus capacitor C2 in real time through voltage sensors (such as a resistor divider network or a Hall voltage sensor), and calculates the bus voltage deviation ΔV = VC1 − VC2. When |ΔV| exceeds a preset threshold of ±3V and VC1>VC2, the controller outputs a high-frequency PWM signal to drive Q1 to conduct. Current flows from the positive bus capacitor C1 through the first winding N1 to the first switching device Q1 and then to ground, causing the TX1 core to store energy. When the first switching device Q1 is turned off, the first winding N1 induces a reverse electromotive force, which generates a positive voltage on the second winding N2 side through parasitic capacitance or auxiliary rectification path, causing energy to transfer to the negative bus capacitor C2.

[0027] Conversely, when VC2 > VC1, the controller drives the second switching device Q2 to operate, and energy is transferred in the reverse direction from the negative bus capacitor C2 through N2 → N1 to the positive bus capacitor C1. The entire process requires no external power supply, utilizing only the excess energy of the bus itself to complete the rebalancing, significantly improving system energy efficiency. The flyback transformer TX1 uses an EE or PQ type ferrite core, with an operating frequency set to 50–200kHz. The controller can use a TI TMS320F280049C DSP chip, with a built-in 12-bit ADC module, a sampling period of 10μs, and a PWM resolution of 150ns.

[0028] The output terminals of the first winding N1 and the second winding N2 are respectively connected to diodes D1 and D2, forming a full-bridge flyback topology with the first switching device Q1 and the second switching device Q2, thereby realizing unidirectional energy flow control.

[0029] The operating principle of a two-level bus voltage equalization device according to an embodiment of this application is as follows: When the voltage of the positive bus capacitor C1 is detected to be higher than a preset threshold (e.g., ±5V), the controller outputs a high-level PWM signal to turn on the first switching device Q1. At this time, current flows from C1 through the first switching device Q1 into the first winding N1, establishing a magnetic field to store energy. When the first switching device Q1 is turned off, the magnetic field collapses, and the induced electromotive force charges the negative bus capacitor C2 through the diode D1, realizing the transfer of energy from the positive bus to the negative bus. Conversely, when the voltage of the negative bus capacitor C2 is too high, the controller triggers the second switching device Q2 to turn on, and energy is transferred in the reverse direction to the positive bus capacitor C1 through the second winding N2 and the diode D2.

[0030] In another embodiment, such as Figure 2 As shown, this application also discloses a balancing method for a two-level bus voltage balancing device, which is applied to a two-level bus voltage balancing device as described above. The balancing method for a two-level bus voltage balancing device specifically includes the following steps: S1: The voltage values ​​of the positive bus capacitor and the negative bus capacitor are collected in real time based on the voltage sampling module, and the bus voltage deviation is calculated.

[0031] In this embodiment, the two-level bus voltage balancing device of this application is applicable to multi-level power electronic systems such as three-phase inverters and cascaded H-bridges. The controller is a digital signal processor (DSP), microcontroller (MCU), or field-programmable gate array (FPGA), and has analog signal sampling, logic judgment, and PWM signal generation functions.

[0032] S2: Based on the bus voltage deviation and historical operating data, a self-learning neural network controller is used to generate initial equalization control parameters.

[0033] In this embodiment, the initial equalization control parameters include the initial value of the first duty cycle of the first switching device, the initial value of the second duty cycle of the second switching device, and the enable flag of the resonant assisted energy transfer mode; the self-learning neural network controller is a lightweight deep learning model deployed on an embedded processor, which has online inference and incremental learning capabilities. Its input is a multi-dimensional system state vector, and its output is the preliminary control command for the next control cycle; the initial equalization control parameters refer to the initial duty cycle value and the resonant mode enable flag quickly generated by the data-driven model before fine optimization is performed, which are used to shorten the convergence time of the model predictive controller and improve the accuracy of the first response; the historical operating data includes the voltage, current, temperature and control action sequence in the past several control cycles, which constitute time series features.

[0034] Specifically, the self-learning neural network controller adopts a convolutional-long short-term memory hybrid network structure. The input variables of the self-learning neural network controller include: the current bus voltage deviation, the rate of change of bus voltage deviation, the rate of change of load power, the ambient temperature, and the actual duty cycle sequence of the first and second switching devices in the past N control cycles. The output variables of the self-learning neural network controller include the normalized initial values ​​of the first and second duty cycles, and the activation flag of the resonant assisted energy transfer mode. The self-learning neural network controller loads pre-trained network weights during the system power-on phase and calculates the loss function based on the actual equalization performance index at the end of each equalization control cycle, and updates the network weight parameters using an online fine-tuning algorithm.

[0035] Furthermore, the convolutional-long short-term memory hybrid network structure is a CNN-LSTM hybrid neural network, comprising: one one-dimensional convolutional layer (kernel size = 3, output channels = 16), one LSTM layer (hidden units = 32), and two fully connected layers (output dimension = 3). The input vector has a dimension of 7, including: the current bus voltage deviation ΔV, the rate of change of bus voltage deviation d(ΔV) / dt (calculated through third-order difference), the rate of change of load power dP / dt, the ambient temperature T, and the duty cycles of Q1 and Q2 over the past three cycles. The output enable flag is {0, 1}.

[0036] In this embodiment, the activation conditions for the resonant assisted energy transfer mode include: the absolute value of the bus voltage deviation is greater than the first voltage threshold, and the rate of change of the bus voltage deviation is greater than 1.

[0037] Specifically, the first voltage threshold is set to 5.0V. The controller acquires the voltage values ​​of the positive and negative bus capacitors using a 16-bit ADC at a sampling rate of 20kHz, and calculates the bus voltage deviation rate using a third-order Savitzky-Golay filter. This embodiment employs a second-order DPLL structure, including a phase detector (PD), a loop filter (LF), and a numerically controlled oscillator (NCO). The DPLL uses... For the input signal, the frequency estimate is updated every 100μs. When resonant mode is enabled, DPLL starts tracking: initial settings. =100kHz, PD output phase error The frequency correction is generated by a PI-type LF (proportional gain Kp=0.8, integral gain Ki=200), and the NCO adjusts the output frequency accordingly. After approximately 2ms of convergence, the DPLL locks the actual resonant frequency at 102.3kHz. The controller then configures the carrier generator of the PWM module, switching its frequency from the original 60kHz to 102.3kHz while maintaining the duty cycle command unchanged. At this time, Q1 turns on when the resonant current naturally crosses zero, and the measured turn-on loss is reduced by 63%, and the bidirectional energy transfer time is shortened from 5.2ms to 1.8ms.

[0038] In this embodiment, the self-learning neural network controller can be implemented by those skilled in the art, and its complete training and deployment process is as follows: First, in the offline stage, a high-fidelity model containing the nonlinear magnetization curve of the flyback transformer, parasitic parameters of the switching devices, and typical load disturbances is constructed based on a system simulation platform (such as MATLAB / Simulink+SimscapeElectrical). This model generates 10,000 sets of operating condition data, each set containing an input vector and corresponding expert labels—that is, the duty cycle sequence used by a traditional PI controller to achieve optimal balanced performance under the same operating conditions. The Adam optimizer (initial learning rate = 0.001, batch size = 64, 200 iterations) is used to train the CNN-LSTM network end-to-end, with the loss function being the mean squared error (MSE). After training, the network weights are solidified and deployed to the controller's Flash memory. During the online phase, at the end of each control cycle, the controller performs a single-step fine-tuning: it calculates the deviation between the current equilibrium performance metric (such as the absolute integral value of the bus voltage deviation over the next 5ms) and the predicted value as the Huber loss, and updates the weights of the fully connected layer using stochastic gradient descent (SGD, learning rate = 0.0005) with the most recent 100 samples stored in RAM.

[0039] S3: Input the initial equalization control parameters into the model predictive equalization controller. Based on the current load current, system temperature and flyback transformer core saturation state, construct a multi-objective optimization function and solve for the optimal PWM control sequence.

[0040] In this embodiment, the Model Predictive Balancing Controller (MPBC) is a finite-time rolling optimization controller based on the system dynamic model. Its core is to predict the system behavior in the next M steps within each control cycle and solve for the control sequence that minimizes the multi-objective cost function. The multi-objective optimization function comprehensively considers the conflicting objectives of voltage balance speed, switching loss, and control smoothness. The saturation state of the flyback transformer core is estimated by the flux linkage observer or current slope to constrain the maximum conduction time.

[0041] Specifically, MPBC uses a discrete state-space model: x(k+1)=Ax(k)+Bu(k)+Ed(k) Where the state vector x= A is the system matrix with a dimension of 2×2, which is obtained by discretizing the continuous system model (such as circuit differential equations); Let B be the transformer flux linkage, and B be the input matrix describing the strength and direction of the effect of the control input u(k) on the state change, with a dimension of 2×2; the control vector u= The disturbance d is the load current; E is the disturbance matrix with a dimension of 2×1, which describes how the disturbance is coupled into the state equation.

[0042] Furthermore, the system matrix A, input matrix B, and disturbance matrix E in the discrete state-space model are not derived theoretically, but are updated in real time through online system identification. Specifically, during system idle periods (such as when the bus voltage deviation |ΔV| < 1V and the load is stable), the controller injects a small-amplitude pseudo-random binary sequence (PRBS) signal as a test stimulus, and simultaneously acquires the estimated values ​​of ΔV and flux linkage. (Obtained by voltage integration method), control quantity u and load current The response data; using recursive least squares (RLS) with a forgetting factor λ=0.98, the parameters of the following linear regression model were identified: The identification results are directly filled into matrices A, B, and E, and used for MPC prediction in the next control cycle.

[0043] The identification results are directly filled into matrices A, B, and E, and used for MPC prediction in the next control cycle.

[0044] The multi-objective optimization function (cost function) for constructing the model predictive equilibrium controller is defined as follows: in, The weighting coefficients are dynamic and adjusted in real time based on the system stability margin; k+1 represents the next control step. This is the predicted value of the bus voltage deviation for the next control step; These are the equivalent on-state currents of the first switching device Q1 and the second switching device Q2, respectively. These are the on-resistances of the first switching device Q1 and the second switching device Q2, respectively; Let these be the first and second duty cycles of the current period, respectively. The value corresponds to the previous cycle; the model predictive equalizer continuously optimizes the control sequence for the next M steps in each control cycle, and only executes the first step of the control input.

[0045] For example, a C code solver was generated using the ACADO toolbox on a TI TMS320F28388D DSP, with prediction time domain M=5 and control time domain N=2. Using the normalized duty cycle of Q1 (S2 output) of 0.38 as the initial value, the optimal sequence was obtained within 45μs for QP, and the first output was taken. ( To ultimately output the optimal control quantity in the first step, which serves as the control command for Q1 and Q2 in the next cycle, core saturation is achieved through the limit. Implementation, in which The instantaneous voltage of the first winding (N1) The maximum permissible volt-second product is the hard upper limit for the safe operation of the magnetic core. If an overshoot is predicted, the duty cycle is forcibly truncated.

[0046] S4: In the low-frequency operating range, a space vector modulation strategy is used to reduce the switching frequency, and in the high-frequency operating range, a pulse sequence optimization strategy is switched to suppress transient oscillations, achieving millisecond-level bidirectional energy transfer.

[0047] In this embodiment, the low-frequency operating range refers to the operating state with a switching frequency below 50kHz. At this frequency, the switching loss is small, and the frequency is preferentially reduced to improve efficiency. The high-frequency operating range refers to the state with a switching frequency above 100kHz, where it is necessary to suppress voltage overshoot and EMI caused by di / dt. Space vector modulation (SVM) achieves efficient low-frequency control by synthesizing a virtual voltage vector. Pulse sequence optimization (PTO) is a non-uniform PWM technique that suppresses harmonics in specific frequency bands by adjusting the pulse position and width. The LC resonant circuit consists of the leakage inductance of the flyback transformer. With external resonant capacitor Its impedance phase angle θ = arctan(ω) / (1 / ω This determines whether it is near the resonance point.

[0048] For example, the frequency threshold is set to 75kHz. This is when the equivalent switching frequency of the MPBC output... =60kHz, less than 75kHz, the controller enables a low-frequency optimization strategy based on the space vector modulation concept: the low-frequency optimization strategy will use a single duty cycle command =0.41 is converted to a drive sequence containing one main pulse and one narrow auxiliary pulse within one PWM cycle, for example, the main pulse width is 0.35. The auxiliary pulse width is 0.06. Interval dead time. When =120kHz, switch to Pulse Sequence Optimization (PTO) strategy mode: split a single wide pulse into 3 narrow pulses (e.g., 0.15 + 0.11 + 0.15), the total duty cycle remains unchanged, but high-frequency harmonic energy is dispersed, reducing the stress on the LC filter. Simultaneously, the resonant capacitor voltage is sampled using a high-speed ADC (1MSPS). With bus voltage Calculate the phase difference Δϕ=∠( )−∠( If |Δϕ| < π / 6 and |ΔV > 5V, then resonant injection is initiated: the PWM carrier frequency is adjusted to be locked by the digital phase-locked loop (DPLL). This allows Q1 / Q2 to turn on under zero-voltage conditions, reducing switching losses by more than 60%. For example, enabling resonance at ΔV=16.6V shortens the energy transfer time from 8ms to 2.3ms.

[0049] Furthermore, the LC resonant parameters are designed as follows: =1.1 μ H, external resonant capacitor =220nFQ value>50; DPLL uses a second-order loop filter with a bandwidth of 5kHz.

[0050] S5: When a load change event or bus voltage deviation exceeds the preset dynamic threshold is detected, a feedforward compensation command is triggered. The nonlinear extended state observer is used to estimate the total amount of external disturbance, and the total amount of external disturbance is superimposed on the output of the model predictive equalization controller to reconstruct the equalization control path.

[0051] In this embodiment, a load mutation event refers to a significant step change in load power within a short period of time, such as motor start-up and shutdown or photovoltaic cloud shading; the preset dynamic threshold is an upper limit of voltage deviation that is adaptively adjusted according to the operating conditions, such as ±3V in steady state and ±15V in transient state; the nonlinear extended state observer (NESO) is a high-order sliding mode observer that treats unmodeled dynamics, parameter perturbations and external disturbances as a unified "total disturbance" for real-time estimation; feedforward compensation refers to directly applying the disturbance estimate to the control quantity.

[0052] Specifically, the load mutation event is the load mutation rate | If / dt>5A / ms, and the preset dynamic threshold is a bus voltage deviation of 12V lasting for 2 cycles, then feedforward compensation is triggered; the three state change vectors of the NESO structure include: Where the error signal e = ΔV− fal(⋅) is a nonlinear function. For the total disturbance estimate The observer gain parameter is set to The bandwidth is approximately 300kHz. (Estimated) After reaching 4.2V, calculate the feedforward compensation. Among them, the feedforward compensation increment =0.02 is determined by the system closed-loop bandwidth (5kHz). The final control quantity is the sum of the feedforward compensation increment and the basic control quantity output by the model predictive controller, such as... =0.41.

[0053] S6: After each equalization control cycle, update the network weight parameters of the self-learning neural network controller based on the actual bus voltage convergence speed and the theoretical prediction error.

[0054] In this embodiment, the actual bus voltage convergence speed refers to the time taken from triggering equalization to |ΔV| decreasing to the target threshold (e.g., 2V); the theoretical prediction error refers to the deviation between the ΔV(k+1) predicted by the model-predicted equalization controller and the measured value; updating network weight parameters refers to fine-tuning some or all of the parameters of the self-learning neural network using an online learning algorithm such as gradient descent, so that its output is closer to the optimal initial value under the current operating conditions.

[0055] Specifically, in this embodiment, the following operations are performed at the end of each control cycle (Ts=100μs): record the start time of this equalization process. and the time when |ΔV(t) < 2V Calculate the measured convergence time Simultaneously read MPBC in The convergence time of the time-prospective calculation (based on the integral estimation of the optimized trajectory). Define the loss function: Where λ=0.5. The Adam optimizer (learning rate=0.0005, first moment decay rate=0.9, second moment decay rate=0.999) is used to update the weights of the last fully connected layer of the neural network step by step. For example, in a certain equilibrium... =3.2ms, =2.8ms, the prediction error is relatively large, and the initial value under the same working conditions after the update is closer to the optimum. To save computing power, the update is only triggered when L>0.1.

[0056] In one embodiment, after triggering the feedforward compensation command, the method includes: S501: Employs a nonlinear extended state observer to jointly estimate unmodeled dynamic characteristics and external load disturbances, outputting a total disturbance estimate.

[0057] In this embodiment, the Nonlinear Extended State Observer (NESO) is a high-gain state observer; unmodeled dynamic characteristics refer to dynamic behaviors not included in the MPBC prediction model due to model simplification (such as ignoring high-frequency parasitic oscillations and hysteresis effects); external load disturbances mainly refer to the power step changes of the inverter or DC / DC converter connected to the DC bus, manifested as instantaneous charging and discharging current impacts on the bus capacitor; the total disturbance estimate is the third state variable output by NESO. The unit is volts (V), representing the equivalent effect of the total disturbance estimate on the bus voltage deviation ΔV.

[0058] Specifically, in this embodiment, a third-order NESO is deployed on a TI TMS320F28388D DSP, and its discretization form is as follows: The sampling period =10μs, observation error e(k)=ΔV(k)− (k), control gain b=1.0 (normalized), and nonlinear function fal(⋅) parameters as described above. When the system detects a sudden load change (such as a motor starting causing the bus current to jump from 15A to 60A), NESO converges within 20μs and outputs the total disturbance estimate. =4.7V, indicating that the current disturbance is equivalent to increasing the bus voltage deviation by an additional 4.7V.

[0059] Furthermore, the gain parameters of the nonlinear extended state observer It is not a fixed empirical value, but rather dynamically tuned based on the system's current operating point. The tuning method is based on the bandwidth configuration principle: first, the natural frequency corresponding to the system's dominant pole is obtained through the aforementioned online system identification. Then, set the desired bandwidth for the observer. This ensures the speed is much faster than the control bandwidth; finally, calculate the gain using the following formula: .

[0060] S502: Calculate the feedforward compensation based on the total disturbance estimate and the preset feedforward gain coefficient.

[0061] In this embodiment, a preset feedforward gain coefficient is used. It is a dimensionless scaling factor used to convert disturbance voltage quantities into equivalent duty cycle compensation quantities; feedforward compensation quantity. It is the increment directly superimposed on the control command. Preset feedforward gain coefficient. The adaptive adjustment can be achieved through a pre-set closed-loop bandwidth-gain mapping table. The contents of the closed-loop bandwidth-gain mapping table are as follows: when the closed-loop bandwidth is 1.0kHz, the preset feedforward gain coefficient is... The preset feedforward gain coefficient is 0.010; when the closed-loop bandwidth is 3.0kHz. The preset feedforward gain coefficient is 0.016; when the closed-loop bandwidth is 5.0kHz, the preset feedforward gain coefficient is... The preset feedforward gain coefficient is 0.020; when the closed-loop bandwidth is 8.0kHz. It is 0.025.

[0062] Specifically, the closed-loop bandwidth of the system is obtained through online identification. The controller injects a small-amplitude sinusoidal test signal (amplitude 0.5%, frequency 1–10kHz), and measures the amplitude-frequency response characteristic of ΔV. The frequency corresponding to the −3dB point is taken as the frequency. If currently =5.2kHz, then according to the empirical formula =0.018+0.0004× Calculated =0.0201. Subsequently, the feedforward compensation amount is calculated as follows: =0.0201×4.7≈0.0945. This value indicates that approximately 9.45% of the on-time needs to be added to the original duty cycle to completely offset the disturbance. To prevent actuator saturation, set | |≤0.2.

[0063] S503: The feedforward compensation is superimposed on the output control quantity of the model predictive equalizer to generate the final control quantity; the preset feedforward gain coefficient is adaptively adjusted according to the system closed-loop bandwidth.

[0064] In this embodiment, the output control quantity of the model predictive equalizer controller refers to the optimal duty cycle command calculated by the MPBC in step S3. The final control quantity is the PWM command actually applied to the switching device drive circuit; the superposition operation only applies to the switching devices that dominate the energy transfer direction. The preset feedforward gain coefficient is adaptively adjusted according to the system closed-loop bandwidth, which has been implemented in S502. Here, we emphasize its synergistic relationship with the final control synthesis.

[0065] Specifically, assuming the MPBC output And since the current ΔV = +12.5V > 0, it indicates that energy needs to be transferred from the positive bus to the negative bus. Therefore, the feedforward compensation only applies to the duty cycle of Q1. The final control quantity is calculated as follows: This instruction is sent to the ePWM module to generate the gate signals driving Q1 and Q2. Actual measurements show that without feedforward enabled, the peak value of |ΔV| reaches 18.3V after a sudden load change, and it takes 6.1ms to recover to ±2V; with feedforward enabled, the peak value is suppressed to 9.8V, and the recovery time is shortened to 2.4ms. Furthermore, the system automatically performs closed-loop bandwidth identification every 500ms and updates... This ensures the accuracy of compensation during long-term operation.

[0066] Specifically, after the system exits the safe equilibrium mode, the workflow of the self-learning neural network controller and the model predictive equilibrium controller can be reactivated only if the following two conditions are met simultaneously: The absolute value of the bus voltage deviation remains below the second voltage threshold (e.g., 2.0V) for more than 100ms; The load power change rate |dP / dt| is less than 5% / ms of the rated power. This dual criterion ensures that the system only switches back to the advanced intelligent control mode after fully recovering to steady state, avoiding instability during mode switching.

[0067] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0068] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0069] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A two-level bus voltage balancing device, characterized in that, include: Positive bus capacitor, negative bus capacitor, flyback transformer, first switching device and second switching device; The flyback transformer includes a first winding and a second winding; the first winding is connected between the positive bus capacitor and the first switching device, and the second winding is connected between the negative bus capacitor and the second switching device. When the voltage of the positive bus capacitor is higher than a preset threshold, the first switching device is triggered to conduct to form an energy transfer loop; when the voltage of the negative bus capacitor is higher than a preset threshold, the second switching device is triggered to conduct to form an energy transfer loop. The first switching device and the second switching device are electrically connected to a controller, which outputs a PWM control signal based on the real-time voltage values ​​of the positive bus capacitor and the negative bus capacitor.

2. The two-level bus voltage equalization device according to claim 1, characterized in that, The turns ratio of the first winding to the second winding is 1:1; both the first switching device and the second switching device are MOSFET devices.

3. A balancing method applied to a two-level bus voltage balancing device as described in any one of claims 1 to 2, characterized in that, The methods include: The voltage sampling module collects the voltage values ​​of the positive bus capacitor and the negative bus capacitor in real time, and calculates the bus voltage deviation. Based on the bus voltage deviation and historical operating data, a self-learning neural network controller is used to generate initial equalization control parameters. The initial equalization control parameters are input to the model predictive equalization controller. Based on the current load current, system temperature, and flyback transformer core saturation state, a multi-objective optimization function is constructed, and the optimal PWM control sequence is obtained by solving the function. In the low-frequency operating range, a space vector modulation strategy is used to reduce the switching frequency, and in the high-frequency operating range, a pulse sequence optimization strategy is switched to suppress transient oscillations, achieving millisecond-level bidirectional energy transfer. When a load change event or bus voltage deviation exceeds a preset dynamic threshold is detected, a feedforward compensation command is triggered. A nonlinear extended state observer is used to estimate the total amount of external disturbance, and the total amount of external disturbance is superimposed on the output of the model predictive equalization controller to reconstruct the equalization control path. After each equalization control cycle, the network weight parameters of the self-learning neural network controller are updated based on the actual bus voltage convergence speed and the theoretical prediction error.

4. The equalization method applied to a two-level bus voltage equalization device according to claim 3, characterized in that, The initial equalization control parameters include the first duty cycle initial value of the first switching device, the second duty cycle initial value of the second switching device, and the activation flag of the resonant assisted energy transfer mode; The self-learning neural network controller adopts a convolutional-long short-term memory hybrid network structure. The input variables of the self-learning neural network controller include: the current bus voltage deviation, the bus voltage deviation change rate, the load power change rate, the ambient temperature, and the actual duty cycle sequence of the first and second switching devices in the past N control cycles. The output variables of the self-learning neural network controller include normalized initial values ​​of the first duty cycle, the second duty cycle, and an enable flag for the resonant assisted energy transfer mode. The self-learning neural network controller loads pre-trained network weights during the system power-on phase and calculates the loss function based on the actual equilibrium performance index at the end of each equilibrium control cycle, and updates the network weight parameters using an online fine-tuning algorithm.

5. The equalization method applied to a two-level bus voltage equalization device according to claim 3, characterized in that, The multi-objective optimization function constructed by the model predictive equilibrium controller is defined as follows: in, The weighting coefficients are dynamic and adjusted in real time based on the system stability margin; k+1 represents the next control step. This is the predicted value of the bus voltage deviation for the next control step; These are the equivalent on-state currents of the first switching device Q1 and the second switching device Q2, respectively. These are the on-resistances of the first switching device Q1 and the second switching device Q2, respectively; Let these be the first and second duty cycles of the current period, respectively. This is the value corresponding to the previous period; The model predictive equalization controller continuously optimizes the control sequence for the next M steps in each control cycle, and executes only the first step of the control input.

6. The equalization method applied to a two-level bus voltage equalization device according to claim 3, characterized in that, The activation conditions for the resonant-assisted energy transfer mode include: The absolute value of the bus voltage deviation is greater than the first voltage threshold, and the rate of change of the bus voltage deviation is greater than zero.

7. The equalization method applied to a two-level bus voltage equalization device according to claim 3, characterized in that, After triggering the feedforward compensation command, the methods include: A nonlinear extended state observer is used to jointly estimate the unmodeled dynamic characteristics and external load disturbances, and outputs the total disturbance estimate. Based on the total disturbance estimate and the preset feedforward gain coefficient, the feedforward compensation amount is calculated; The feedforward compensation is superimposed on the output control quantity of the model predictive equalization controller to generate the final control quantity; the preset feedforward gain coefficient is adaptively adjusted according to the system closed-loop bandwidth.

8. The equalization method applied to a two-level bus voltage equalization device according to claim 3, characterized in that, During system startup or fault recovery, the system automatically switches to a security balancing mode, which includes: The maximum duty cycle of the first and second switching devices is limited to no more than 30%. Disable resonant-assisted energy transfer mode; Energy transfer is performed using a fixed-frequency, low-duty-cycle exploratory PWM signal until the absolute value of the bus voltage deviation is less than the preset voltage deviation threshold. Then, the workflow of the self-learning neural network controller and the model predictive equalization controller is activated.

Images