Layered cooperative control method for cascaded H-bridge frequency converters

Abstract

The invention discloses a hierarchical cooperative control method for a cascaded H-bridge frequency converter, and the method comprises the steps: distributing a disturbance signal following a spiral phase coding rule for each H-bridge unit, and enabling the vector sum of the phase of the group of signals to generate a plurality of energy quasi-zero moments in a preset disturbance period through design; at the quasi-zero moment when the energy is greatly suppressed, synchronously triggering all units to perform high signal-to-noise ratio data acquisition, and accurately identifying transient parameters representing dynamic characteristics of each unit based on the acquired data; and finally, performing closed-loop adjustment on an upper-layer power distribution strategy, a bottom-layer current loop and the like according to the identified transient parameters so as to compensate the dynamic response difference of each unit. According to the method, high-precision online parameter identification can be realized on the premise of not influencing stable operation of the system, and the power balance capability and transient stability of the system are remarkably improved.

Description

Technical Field

[0001] This invention relates to technologies related to frequency converters, and in particular to a hierarchical collaborative control method for cascaded H-bridge frequency converters. Background Technology

[0002] Cascaded H-bridge (CHB) frequency converters have become key power electronic equipment in medium- and high-voltage, high-power applications due to their outstanding advantages in modular design, high output voltage quality, and ease of implementation in high-voltage and high-capacity applications. They are widely used in core scenarios such as flexible DC transmission (HVDC-Light / VSC-HVDC), large-scale energy storage systems (BESS), static synchronous compensators (STATCOM), and renewable energy grid connection interfaces. In these applications, the frequency converter is not only the core of energy conversion but also the cornerstone for maintaining grid stability and ensuring power quality. Therefore, researching and improving cascaded H-bridge frequency converters, especially their control performance, operational stability, and reliability under complex dynamic conditions, has significant theoretical research value and engineering application value for promoting the construction of the energy internet, facilitating the consumption of renewable energy, and ensuring national energy security.

[0003] Currently, significant progress has been made in control technology for cascaded H-bridge frequency converters. In terms of system control architecture, a hierarchical control approach is commonly adopted, where a central controller or upper-level controller is responsible for overall power command calculation and distribution, while the local controllers of each H-bridge unit are responsible for executing specific voltage and current control tasks. Regarding unit voltage equalization control, strategies such as sorting-based equalization and modulation wave compensation strategies based on carrier phase shifting or phase sequence rotation have been developed. In the field of parameter identification, to obtain the mathematical model of the system for optimized control, researchers have applied various online identification methods, such as applying recursive least squares (RLS) to identify system parameters under steady-state or specific operating conditions, and analyzing the system's impedance characteristics by injecting disturbance signals of specific frequencies. These technologies, to a certain extent, ensure the basic operation of cascaded H-bridge frequency converters.

[0004] However, with the increase in system capacity and the increasingly stringent requirements for dynamic performance, existing technologies have revealed some deep-seated technical problems when dealing with large-scale, highly dynamic application scenarios. These problems are not isolated but intertwined, jointly restricting further improvements in system performance. The core challenge lies in the difficulty of existing technologies to accurately obtain the true dynamic parameters of each module during power transients without sacrificing system stability. Summary of the Invention

[0005] The purpose of this invention is to resolve the contradiction between parameter identification accuracy and system disturbance under dynamic operating conditions by providing a hierarchical collaborative control method for cascaded H-bridge frequency converters.

[0006] Technical solution: A hierarchical collaborative control method for cascaded H-bridge frequency converters, wherein the cascaded H-bridge frequency converters include at least three H-bridge units, including: Each H-bridge unit is assigned a perturbation signal with a predetermined phase relationship, so that the vector sum of the perturbation signals of each unit satisfies the minimum condition at a predetermined time, and multiple energy quasi-zero time points are generated within a preset perturbation period. At the near-zero energy point, multiple H-bridge units are simultaneously triggered to acquire data with high signal-to-noise ratio, and transient parameters characterizing the dynamic characteristics of each unit are identified based on the acquired data. Based on transient parameters, the multi-level control law of the cascaded H-bridge frequency converter is adjusted in a closed loop to generate and execute optimized control commands.

[0007] The beneficial effects include the ability to achieve high-precision online parameter identification without affecting the stable operation of the system, and to significantly improve the system's power balancing capability and transient stability. Attached Figure Description

[0008] Figure 1 This is a flowchart of the present invention.

[0009] Figure 2 This is a flowchart of the spatiotemporal collaborative planning of disturbances according to the present invention.

[0010] Figure 3 This is a flowchart of the present invention for establishing the phase evolution equation for each H-bridge unit.

[0011] Figure 4 This is a flowchart of the present invention for solving the energy quasi-zero constraint equation. Detailed Implementation

[0012] In order to solve the problems existing in the current technology, the applicant conducted in-depth research and found that: The challenges of the background technology are specifically manifested in two interconnected and profound technical problems. The first problem is the inherent contradiction between signal-to-noise ratio and system disturbance in online parameter identification. To accurately identify parameters such as the equivalent series resistance (ESR) of a unit, which have a significant but small impact on dynamic response, a disturbance signal with sufficient energy and a wide bandwidth needs to be injected into the system so that the response signal can be effectively extracted from the background of strong switching noise and power fluctuations. However, in transient processes with rapidly changing high power, injecting such a strong disturbance signal itself poses a serious threat to the output power quality and stability of the system, which is unacceptable in engineering applications. Conversely, if a weak disturbance with a small amplitude is used, its response signal will be completely submerged in noise, resulting in huge errors in the identification results and rendering it useless. The second problem stems from the inherent heterogeneity and asynchronicity of large-scale modular systems. Due to manufacturing tolerances, uneven aging of components (especially DC-side support capacitors), and differences in heat dissipation conditions of each module, the actual dynamic characteristics (such as the equivalent RC time constant) of each H-bridge unit are unique and change over time. Meanwhile, in distributed control networks, communication delays and timing jitter between units introduce asynchronicity in control. When the system encounters sudden power command changes, traditional control strategies that treat all units as homogeneous, or strategies that rely on delayed feedback information for adjustment, are unable to effectively manage such complex heterogeneous and asynchronous responses. This often leads to instantaneous power distribution imbalances and voltage or current overshoots in individual units, posing a potential threat to the safe and stable operation of the system.

[0013] Example 1: This example describes the overall process of a hierarchical collaborative control method for cascaded H-bridge frequency converters. The cascaded H-bridge frequency converter includes at least three H-bridge units.

[0014] Specifically, it includes the following steps: Step 101: Assign a disturbance signal with a predetermined phase relationship to each H-bridge unit, so that the vector sum of the disturbance signals of each unit satisfies the minimum condition at a predetermined time, so as to generate multiple energy quasi-zero time points within a preset disturbance period.

[0015] Alternatively, a spiral phase encoding trajectory is designed for the disturbance signal of each H-bridge unit, so that the disturbance phase of each unit changes dynamically in the time dimension according to the spiral evolution law including the initial phase, rotation frequency and periodic modulation term. By solving the constraint equation of the complex vector sum of the disturbance signal of each unit, the moment when the vector sum satisfies the preset minimum value condition is determined, and multiple energy quasi-zero point moments are generated within the preset disturbance period. In this embodiment, the quasi-zero energy time refers to a specific sampling time t. kAt this moment, the disturbance signals injected by all H-bridge units cancel each other out due to their specific phase relationship, causing the total energy fluctuation caused by the disturbance at the system's common coupling point (such as the DC bus) to reach a minimum. Specifically, this minimum condition can be quantified as the magnitude of the vector sum of all disturbance signals in the complex domain being much smaller than their algebraic sum, for example, less than 1% of the algebraic sum, i.e., satisfying |Σ i A i •exp(jθ i (t k ))|<0.01•Σ i |A i | Where i is the sequence number of the H-bridge unit; A i Let θ be the amplitude of the disturbance signal in the i-th unit; i (t k Let ) be the disturbance signal of the i-th unit at time t. k The phase. To achieve this goal, the perturbation phase θ for each unit can be... i (t) The evolution equations that follow the spiral phase coding law, such as θ i (t)=φ i +ω i t+β i •sin(2πt / Tsweep). By solving the above constraint equations, multiple energy quasi-zero point times can be calculated and determined in advance, forming a timetable.

[0016] The purpose of this step is to proactively and deterministically create several quiet moments during normal system operation. Under normal operating conditions, the internal electrical quantities of a cascaded H-bridge frequency converter (especially the DC-side voltage) are affected by various noise sources such as switching ripple and load fluctuations, resulting in a very low signal-to-noise ratio. Direct parameter identification would introduce significant errors. By causing the disturbance energy to converge to a minimum point at a specific moment, it is equivalent to suppressing the main noise source (i.e., the ripple generated by the injected disturbance itself) by more than 40dB in the time domain, thus creating ideal conditions for subsequent high signal-to-noise ratio data acquisition and high-precision parameter identification.

[0017] Step 102: At the quasi-zero energy point, the multiple H-bridge units are simultaneously triggered to perform high signal-to-noise ratio data acquisition, and transient parameters characterizing the dynamic characteristics of each unit are identified based on the acquired data. Alternatively, at the quasi-zero energy point, the multiple H-bridge units are simultaneously triggered to perform data acquisition, wherein the noise caused by the disturbance at the quasi-zero energy point is suppressed to the background noise level, and transient parameters characterizing the dynamic characteristics of each unit are identified based on the data acquired under high signal-to-noise ratio conditions. In specific implementation, the system, based on the energy quasi-zero timetable generated in step 101, uses a high-precision synchronous triggering mechanism (e.g., a hardware trigger signal generated by the FPGA) to trigger the energy quasi-zero timetable at a predetermined time t. k Simultaneously start the analog-to-digital converters (ADCs) of all H-bridge units at all times, and acquire the DC-side capacitor voltage Vdc_i(t) of each unit. k ) and DC current idc_i(t k Electrical quantities such as [list of quantities]. Because the broadband noise caused by the disturbance is greatly suppressed at this time, the collected data has a very high signal-to-noise ratio.

[0018] Based on this, transient parameters are identified. These transient parameters mainly refer to the equivalent electrical parameters of the H-bridge unit under conditions of rapid high-power changes, such as the equivalent capacitance value Ceq_i and the equivalent series resistance Resr_i. One identification method involves using voltage and current data collected from multiple (e.g., three) consecutive energy quasi-zero points, solving difference equations to obtain Ceq_i, and then using the principle of energy conservation to calculate Resr_i.

[0019] This step utilizes the time window created in step 101 to address the signal-to-noise ratio challenge in parameter identification. Compared to identification at random moments or under steady-state conditions, parameters identified at the energy quasi-zero point are more accurate and better reflect the true characteristics of the system during dynamic processes, providing reliable data input for subsequent control optimization. For example, the parameter identification accuracy can be better than 5%.

[0020] Step 103: Based on the identified transient parameters, the multi-level control law of the cascaded H-bridge frequency converter is adjusted in a closed loop to generate and execute optimized control commands.

[0021] Alternatively, based on the identified transient parameters, the multi-level control law of the cascaded H-bridge frequency converter is adjusted in a closed loop, including updating the upper-level power distribution strategy, the middle-level coordination control parameters, and the lower-level controller parameters, and generating and executing optimized control commands.

[0022] The multi-level control law here typically includes an upper-level power controller, a middle-level coordination controller, and a lower-level PWM controller. The specific method of closed-loop adjustment is to update the mathematical model of each level controller with the transient parameter set P_transient=[Ceq_i,Resr_i,...] identified in step 102.

[0023] For example, the upper-level power controller can optimize the power allocation coefficient of each unit based on the latest Resr_i and Ceq_i values, so that units with slower dynamic response (time constant τ) can achieve better power allocation. iThe system allocates smaller power steps (=Resr_i×Ceq_i is relatively large) to compensate for the differences in dynamic response between units. The underlying PWM controller can then synchronously tune the PI parameters (such as Kp_i, Ki_i) of its internal current loop based on the updated unit equivalent impedance characteristics to maintain optimal dynamic response and stability.

[0024] This step is used to achieve online self-adaptation of control system parameters. Device parameters in power electronic systems change due to temperature, aging, and operating point variations; fixed controller parameters cannot always guarantee optimal performance. By periodically executing a closed-loop process of disturbance injection-zero sampling-parameter identification-model update, the control system can perceive its own changes in real time and make corresponding adjustments, thereby maintaining a high level of power balance and operational stability throughout its entire lifecycle and under various operating conditions.

[0025] Example 2: System Initialization and Transient Event Detection. This example describes the preparatory technical steps performed before executing the method described in Example 1, providing the necessary data foundation and trigger signals for subsequent disturbance planning and parameter identification.

[0026] Step 201: Perform high-speed synchronous data acquisition.

[0027] In this embodiment, a 16-bit high-precision analog-to-digital converter (ADC) is configured for each H-bridge unit, and the instantaneous values ​​of the three-phase output current are synchronously acquired at a sampling rate of, for example, 100 kHz. The parameters are ia_i(t), ib_i(t), ic_i(t), and the DC-side capacitor voltage Vdc_i(t). To ensure synchronization, a 100MHz global master clock can be generated using a phase-locked loop (PLL) of a central field-programmable gate array (FPGA). This master clock is distributed to each H-bridge unit through a dedicated clock distribution network to ensure that the deviation of the ADC sampling time of each unit is less than 100 nanoseconds (ns). The acquired multi-channel raw data stream can be transmitted to the FPGA through a high-speed serial interface (such as JESD204B), and the deterministic delay characteristics of the interface and the data alignment FIFO (First-In-First-Out queue) inside the FPGA are used to achieve precise alignment of the data of each channel in terms of timestamps.

[0028] This step obtains high-quality raw data that accurately reflects the instantaneous state of the system. In power electronic control, especially in scenarios involving phase coordination and dynamic identification, even small temporal deviations in the data from each unit can lead to significant errors in the calculation results. Precise hardware-level clock synchronization and data alignment provide a fundamental guarantee for the accuracy of all subsequent calculations.

[0029] Step 202: Perform transient event detection.

[0030] Specifically, the main controller in the system receives active power commands from the upper-level system via the Controller Area Network (CAN) bus. P_ref(t) and reactive power command Q_ref(t). To detect abrupt changes in the power command, the controller internally runs a digital filtering algorithm to calculate the rate of change of the power command. For example, a 5-point differential algorithm can be used to calculate the rate of change of active power dP / dt. The calculation formula is: dP / dt = (P_ref(t) - P_ref(t - 4*Ts)) / (4*Ts); where Ts is the sampling period of the power command. The controller monitors the absolute value of this rate of change, |dP / dt|, in real time and compares it with a preset threshold. When |dP / dt| exceeds the threshold, for example, exceeding 20% ​​of the rated power per millisecond (0.2 pu / ms), the system determines that a transient event has occurred. At this time, the controller immediately sets a transient trigger flag, for example, Flag_transient = 1, and records the current time as the transient start time t0.

[0031] This step ensures that the high-precision identification method proposed in this invention is activated only when most needed, rather than running continuously. Continuous disturbance injection would introduce additional losses and minor disturbances to the system. By accurately detecting transient events such as power surges, the application of this method can be focused on the moments when system parameters are most likely to change and have the greatest impact on control performance, thus achieving effective savings in computational resources and system losses.

[0032] Step 203: Extract initial state features.

[0033] After the transient trigger flag is set in step 202, the system uses steady-state data collected within a certain period before triggering (e.g., 10 milliseconds) to perform feature extraction. Specifically, for each H-bridge unit i, the mean Vdc_avg_i and standard deviation σ_Vdc_i of its DC voltage are calculated. Based on this, a key feature parameter—voltage imbalance εi—can be calculated, with the formula εi=(Vdc_i-Vdc_mean) / Vdc_mean×100%; where Vdc_mean is the average value of the DC voltage of all N units. Simultaneously, by performing a synchronous rotating coordinate system (dq) transformation on the collected three-phase currents, the direct-axis current component Id_i and quadrature-axis current component Iq_i of each unit at the steady-state operating point are obtained. Finally, these extracted feature parameters, together with information such as the current carrier phase θcarrier_i read from the PWM generator, can be assembled into an N×6 initial state matrix X0, which has the form X0=[Vdc_avg,σ_Vdc,ε,Id,Iq,θcarrier].

[0034] This step provides precise initial conditions for subsequent adaptive perturbation design. The initial state matrix X0 acts like a snapshot of the system before the transient occurs, containing key information about the imbalance between units (such as εi). The subsequent spiral phase encoding process utilizes this information to tailor the initial phase and contraction velocity of the perturbation signal for each unit, making the injection of perturbation energy more targeted and enabling faster compensation of the system's initial imbalance—a core prerequisite for the adaptive nature of this invention. Optionally, richer multidimensional imbalance features, such as voltage gradients and clustering features, can be extracted to construct a more comprehensive feature vector F_imb, providing input for higher-order compensation algorithms.

[0035] Example 3: Generation of quasi-zero energy moments based on spiral phase coding. This example illustrates how to specifically design and plan the phase of the disturbance signal in step 101 of Example 1, and accurately solve for the quasi-zero energy moments.

[0036] Step 301: Based on the initial system state, establish a phase evolution equation for each H-bridge unit that follows the spiral phase coding law. This set of phase evolution equations together defines the dynamic trajectory of the disturbance phase of each unit in the time dimension.

[0037] In this embodiment, the spiral phase encoding rule refers to a specific mathematical form, which is the phase θ of the disturbance signal of the i-th H-bridge unit. i (t) imparts dynamic characteristics that evolve over time. This evolution is composed of a unique initial phase, a unique rotation frequency, and a common periodic modulation term, which causes the vector tips representing the perturbation signals of each unit in the complex plane to exhibit a spiral convergent or divergent motion trajectory.

[0038] Specifically, the phase evolution equation can be expressed as: θ i (t)=φ i +ω i t+β i •f(t); where i is the index of the H-bridge element, and t is time; φ i ω is the initial phase of the i-th unit; i β is the perturbation angular frequency of the i-th element; i Let f(t) be the adaptive shrinkage factor for the i-th unit; f(t) is a periodic function, such as sin(2πt / Tsweep), where Tsweep is the preset perturbation scan period. The process for establishing these parameters is as follows: Establish the initial phase φ i φ iThe design aims to compensate for the initial imbalance of the system. First, the voltage imbalance ε of each unit is extracted from the initial state matrix X0 obtained in Example 2. i Then, an initial phase compensation term Δφ is calculated using a nonlinear compensation function. i For example, Δφ i =-arctan(ε i / 10), this function ensures that cells with higher voltages receive a phase lead to accelerate energy release. Finally, combined with a global phase offset φ_offset_opt obtained through convex optimization, the final initial phase φ is calculated. i =2πi / N+Δφ i +φ_offset_opt.

[0039] Where N is the total number of H-bridge units.

[0040] Establish angular frequency ω i To avoid coupling with the system's inherent resonant frequency, the fundamental rotational angular frequency ω0 is set within a safe region, for example, 2π•100Hz. The frequency ω of each unit... i Based on this, interval allocation is performed, ω i =ω0+(i-1)•Δω, where the frequency interval Δω is, for example, 2π•50Hz / N, to ensure that the perturbation frequencies of each unit have sufficient separation.

[0041] Establish adaptive contraction factor β i It is used to regulate the convergence pattern of disturbance energy in the spatiotemporal dimension. Its calculation comprehensively considers the voltage imbalance and the differences in the dynamic characteristics of each unit. Specifically, β i =β0•exp(-i / τ_base)•(1+0.1•εi)•(τeq_mean / τeq_i) 0.5 Where β0 is the basic contraction factor; τ_base is the basic contraction time, which is related to the system's average equivalent time constant; τeq_i is the equivalent time constant of the i-th unit, which can be identified online; and τeq_mean is the average of the equivalent time constants of all units.

[0042] This step abandons the simple, fixed perturbation signal of traditional methods, and instead adopts an adaptive, carefully designed perturbation strategy that is deeply coupled with the system state. This is achieved by tailoring the phase evolution trajectory of each unit, particularly utilizing the initial phase φ. i To actively combat initial imbalance, an adaptive contraction factor β is used. iTo compensate for the differences in dynamic response among the units, the energy balance of the system can be accelerated in the initial stage of disturbance injection, and the disturbance vectors of the units can be precisely and cooperatively canceled at a predetermined time, even if the response rates of the units are different. This lays a solid foundation for generating high-quality quasi-zero energy points.

[0043] Optionally, the periodic function Besides a sine function, f(t) can also be a triangular wave function, a sawtooth wave function, or any other suitable periodic function. The initial phase compensation function can also be a piecewise linear function or a higher-order nonlinear function to achieve different compensation intensities across different imbalance ranges.

[0044] Step 302: Based on the set of phase evolution equations, construct an energy quasi-zero constraint equation and solve the energy quasi-zero constraint equation to determine the energy quasi-zero time.

[0045] In this step, the quasi-zero energy constraint equation refers to a equation that describes the disturbance signal vectors of all H-bridge units and the energy at a specific time t. k A mathematical expression that satisfies the minimum condition. Specifically, the equation can be expressed as: |Σ i A i •exp(jθ i (t k ))|<ε•Σ i |A i |;Among them, A i Let be the disturbance amplitude of the i-th unit, the magnitude of which can be related to the voltage imbalance ε. i Positive correlation, for example, A i =0.02•Irated•(1+ε i / 20), where Irated is the rated current of the unit; j is the imaginary unit; ε is a preset energy suppression ratio threshold, for example, ε=0.01, which corresponds to a relative suppression ratio of 40dB.

[0046] The specific implementation method for solving this constraint equation is as follows: Numerical Iteration Solution: Since this equation is a transcendental equation, it is usually solved using numerical iteration methods. For example, the Newton-Raphson iteration method can be used.

[0047] First, the complex equation F(t) = Σ i A i •exp(jθ i(t) is linearized using Taylor expansion near the candidate zeros. Then, within each scan cycle [0, Tsweep], multiple initial guess points are selected, and the iterative algorithm tk^(n+1)=tk^(n)-Re[F] / Re[J]-j•Im[F] / Im[J] is executed, where J is the Jacobian matrix of F(t), and Re[] and Im[] represent taking the real and imaginary parts, respectively. When |F(tk)| is less than the convergence criterion, the iteration stops, and a candidate zero set T_candidate is obtained.

[0048] Validity verification: For each candidate zero t in the candidate zero set... k Effectiveness verification is performed. This verification process verifies at least two conditions: one is the energy suppression ratio threshold, i.e., SR(t) k )=|Σ i A i •exp(jθ i (t k ))| / (Σ i |A i |)<ε; Second, the phase spatial dispersion threshold, i.e., the phase θ of each unit. i (t k The standard deviation std(θi(tk)) in the interval [0,2π] needs to be greater than a preset value, such as π / 3, to ensure that the formation of zeros is due to the uniform dispersion and mutual cancellation of the vectors, rather than accidental aggregation in a few directions.

[0049] Constructing a timetable: From the candidate zeros that have passed the validity screening, further optimization can be performed based on application requirements, such as selecting zeros that are more evenly distributed within the perturbation period, or based on the zero quality (i.e., SR(t)). k The zero points are sorted by size, and finally M (e.g. 20-30) of the best zero points are selected to construct the final energy quasi-zero time table T_zero.

[0050] This step transforms a physical problem (finding the minimum point of energy fluctuations in a system) into a mathematical problem with well-defined constraints. Unlike simple signal superposition analysis, this step, by establishing and solving constraint equations, can accurately and deterministically predict the timing of quasi-zero energy points. In particular, the validity screening process, by introducing physical constraints (energy suppression ratio) and statistical constraints (phase dispersion), can eliminate spurious and unstable zeros, greatly improving the reliability and practicality of the generated timetable and ensuring that subsequent sampling operations can truly be performed during high-quality quiet moments.

[0051] Alternatively, in addition to the Newton-Raphson method, other multidimensional root-finding algorithms, such as the Broyden method or particle swarm optimization, can be used for the solution. In the validity assessment, a zero-point stability criterion can be further introduced, for example, requiring the rate of change of the energy suppression ratio near the zero point, |dSR / dt|, to be less than a certain threshold, to ensure that the sampling time point is located at a flat energy valley.

[0052] Preferably, the method further includes step 303, which involves pre-compensating the theoretical phase trajectory before constructing the energy quasi-zero constraint equation.

[0053] This step is used to correct discrepancies between theoretical models and physical reality. Specifically, before constructing the energy quasi-zero constraint equations, it also includes: Establish a phase distortion prediction model: A predictive model, M_distort, is established to quantify the phase distortion caused by the nonlinear characteristics of power devices (such as IGBTs). This model can comprehensively consider multiple factors, for example, Δθ_total = Δθ_switch + Δθ_dead + Δθ_parasitic. Here, Δθ_switch represents the distortion caused by switching delay (inconsistent turn-on and turn-off times); Δθ_dead represents the distortion caused by dead-time effects; and Δθ_parasitic represents the distortion caused by parasitic parameters such as stray inductance. The internal coefficients of this model can be updated online using recursive least squares to adapt to device aging and changes in operating conditions.

[0054] Calculate and superimpose the phase pre-compensation amount: Based on the model M_distort and current operating conditions such as the current change rate di / dt, the predicted value of the phase distortion, i.e., the phase pre-compensation amount Δθ_total, can be calculated in real time. Then, this pre-compensation amount is subtracted from the theoretical phase trajectory to form a compensated phase trajectory θ. i _comp(t)=θ i (t)-Δθ_total(t).

[0055] The constraint equations are constructed using the compensated phase trajectory: Finally, this compensated phase trajectory θ, which is closer to the physical reality, is used. i The energy quasi-zero constraint equations mentioned above are constructed and solved using _comp(t).

[0056] Since the theoretically calculated phase evolution equations are based on ideal device models, while actual power semiconductor devices exhibit nonlinear effects such as switching delays and dead zones, a small but crucial deviation occurs between the actual disturbance phase injected into the circuit and the phase commanded by the controller. Without compensation, even if a perfect zero point is theoretically calculated, the energy cancellation effect in the actual system will be significantly reduced. By establishing a distortion model and performing pre-compensation, this physical deviation is predicted and canceled in advance, enabling the actual disturbance phase of each unit to reach the preset value more accurately at the target time. This results in a deeper and more perfect quasi-zero energy point, further improving the signal-to-noise ratio and accuracy of subsequent parameter identification.

[0057] Alternatively, the phase distortion model can also take a simpler form, such as a model that only considers the dead zone effect, or a lookup table (LUT) form based on measured data, interpolating the compensation amount according to the current operating conditions (such as current and junction temperature).

[0058] Example 4: Distributed Phase Coordination and Synchronization Mechanism. This example illustrates how, during the injection of a disturbance signal, the system uses a distributed coordination mechanism to ensure that the actual phase of all H-bridge units can accurately track the theoretical trajectory, overcoming interference from real-world factors such as communication delays.

[0059] Step 401: Before synchronously triggering the multiple H-bridge units to perform high signal-to-noise ratio data acquisition, each H-bridge unit performs a distributed phase coordination process. This process generates a phase correction amount based on the deviation between the actual phase obtained from the adjacent unit and the theoretical phase it follows, and combines the theoretical phase with the phase correction amount to form an injected phase.

[0060] In this embodiment, the distributed phase coordination process is a decentralized control strategy. The local controller of each H-bridge unit only needs to communicate with its neighboring units in the topology at low bandwidth (e.g., via RS485 bus at a frequency of 1kHz) to exchange their respective actual phase information θ_actual. It does not require a central master node to coordinate all units, thus exhibiting strong robustness and scalability.

[0061] In practice, the controller of each unit i performs the following operations: First, it receives data from neighboring unit j (where j belongs to the neighbor set N of i). iThe actual phase θj_actual of the unit is calculated. Then, the synchronization error between its actual phase and the phase of its neighboring units is calculated. This error is usually defined as sin(θj_actual - θi_actual - φji), where φji is the theoretically expected phase difference between units j and i. The controller performs a weighted summation of all synchronization errors from neighbors to generate a phase correction Δθ_sync_i. Finally, this real-time calculated correction is superimposed on the theoretical phase to form the final phase to be injected: θi_inject(t) = θi_theory(t) + Δθ_sync_i.

[0062] This step establishes a dynamic negative feedback adjustment loop to maintain phase synchronization of the entire system. The theoretical phase trajectory θi_theory(t) is an ideal open-loop command. However, in actual systems, due to factors such as the slight drift of the local clock of each unit, the difference in the execution time of the control loop, and communication delays, the actual phase of each unit will gradually deviate from the theoretical trajectory. The distributed coordination process enables each unit to sense the phase deviation between itself and its neighbors and actively correct it. This local correction behavior, transmitted through the network, eventually achieves global synchronization of the entire system. This effectively overcomes the problem of error accumulation in open-loop control, ensuring that even during disturbance periods lasting several seconds, the actual phase of all units can be locked on the predetermined theoretical trajectory, with the error controlled to, for example, within 5 degrees.

[0063] Alternatively, the communication topology does not necessarily have to be a fixed physical connection; it can also be adaptive. For example, the system can periodically measure the communication latency and packet loss rate between units through ping tests, and dynamically construct high-quality communication links using algorithms such as minimum spanning tree to optimize synchronization performance.

[0064] Step 402: The distributed phase coordination process dynamically updates a phase elasticity coefficient for each unit through a preset fuzzy adaptive rule, and applies an improved Kuramoto synchronization model that takes into account communication delay to finally calculate the phase correction amount.

[0065] The improved Kuramoto synchronization model here is the core algorithm of the aforementioned cooperative process. It originates from a mathematical model of nonlinear dynamic systems and has been engineered to be applicable to power electronic systems. Its discrete form can be expressed as: δθi=Kp•Σ j∈N ia ij •sin(θ j_actual -θ i_actual -φ ji ); where δθ i It is the intermediate value of the phase correction; Kp is the proportional gain; a ijIt is the coupling strength related to the topological distance. The phase elasticity coefficient ξi is an adaptive gain used to adjust the corrected response characteristics, and the final phase correction Δθ_sync_i is derived by combining ξi and δθi, etc.

[0066] The specific implementation method is as follows: An improved Kuramoto model is applied: this model explicitly accounts for communication delay. When calculating the synchronization error, the neighbor phase θj used is the value at time t-τij, where τij is the communication delay between units i and j. To compensate for this delay, the theoretical phase difference φji includes a compensation term, such as -ωi•τij.

[0067] The phase elasticity coefficient ξi is dynamically updated using a fuzzy adaptive rule. First, the instantaneous synchronization error ei(t) and its rate of change dei / dt of each unit are evaluated in real time. Then, inference is performed according to a preset fuzzy rule. For example, if ei is large and dei / dt is positive (indicating a large and continuously increasing error), ξi decreases (indicating a need to strengthen coupling and apply a stronger correction force); if ei is small and dei / dt is negative (indicating a small and converging error), ξi increases (indicating that coupling can be relaxed, reducing control actions to avoid overshoot). Through membership functions and defuzzification, the state of (ei, dei / dt) is mapped in real time to a specific value of ξi, for example, dynamically changing within the range of [0.6, 1.4].

[0068] Traditional fixed-gain synchronization controllers struggle to adapt to complex and ever-changing system conditions. Fluctuations in communication delay and variations in system noise levels can all affect synchronization performance. Introducing an improved, delay-considering Kuramoto model makes the algorithm more closely resemble physical reality. Furthermore, the introduction of a fuzzy adaptive phase elasticity coefficient ξi on this basis endows the controller with intelligence. It can gently fine-tune when the system is stably synchronized, reducing unnecessary control jitter and energy loss; and decisively and forcefully correct large phase deviations, ensuring system robustness. This adaptive adjustment between strong and weak coupling allows the synchronization mechanism to simultaneously achieve stability and speed, significantly improving the accuracy of phase locking and the system's dynamic adaptability.

[0069] Preferably, the method further includes step 403, which introduces a momentum term that increases with the number of iterations when calculating the phase correction, and uses an event-triggered mechanism for updating and broadcasting. This step is a further optimization of the core algorithm in step 402, aiming to improve efficiency and reduce resource consumption.

[0070] Introducing a momentum term: A momentum term is introduced into the iterative algorithm for calculating the phase correction. Its update rule can be expressed as: δθ i k+1 )=δθik +α•Σj...+β•(δθ i k -δθ i k-1 ); where k is the number of iterations or time steps; α is the learning rate; and β is the momentum coefficient. This momentum coefficient β can be designed to increase with the number of iterations, for example, β=0.3•(1-exp(-k / 10)), where the momentum is small in the early stage of iteration and gradually increases in the later stage.

[0071] An event-triggered mechanism is adopted: the communication method is changed from fixed-period broadcasting to event-triggered broadcasting. Specifically, after calculating the new phase correction amount δθi, the controller of each unit does not broadcast it immediately, but first determines whether the change exceeds a preset trigger threshold θ_threshold. Only when |δθi|>θ_threshold will the update be performed and broadcast externally.

[0072] The introduction of a momentum term borrows from optimization algorithms. It helps the update of the correction quantity overcome small local oscillation points and accelerates convergence in the correct direction, much like a ball with inertia rolling down a hill. This speeds up the convergence of the synchronization process. The motivation for using an event-triggered mechanism is to save valuable communication bandwidth. When the system is close to or in a synchronized state, the phase changes of each unit are very small. At this time, continuously broadcasting almost the same data at a fixed period (e.g., 1 kHz) is a huge waste. The event-triggered mechanism ensures that communication only occurs when necessary (i.e., when a significant phase change occurs), reducing communication overhead by up to 60% without sacrificing synchronization performance. This is especially important for cascaded systems with a large number of units.

[0073] Alternatively, the momentum coefficient β can also be a variable that adaptively adjusts based on the system's synchronization state. The conditions for event triggering can also be more complex, for example, by combining a weighted average of local and neighbor errors to achieve smarter triggering decisions.

[0074] Example 5: Phase-Locked Perturbation Injection and High Signal-to-Noise Ratio Data Acquisition. This example provides a specific and preferred implementation method for the data acquisition process. It elaborates on how to accurately inject the theoretical phase trajectory designed in Example 3 into the physical system under the guarantee of the phase coordination mechanism in Example 4, and complete the crucial high signal-to-noise ratio data acquisition.

[0075] Step 501: Perform precise timing control of disturbance injection.

[0076] In this embodiment, precise timing control refers to ensuring that the injected phase value θi_inject(t) at each moment, which is obtained through complex calculations, can be applied to the pulse width modulator (PWM) of the corresponding H-bridge unit without delay or jitter, and is strictly synchronized with the carrier period of the PWM.

[0077] In practice, each H-bridge unit runs a phase update state machine in its local digital signal processor (DSP) or FPGA. This state machine is strictly synchronized with the interrupt service routine (ISR) of the PWM module. At the start of a PWM cycle, the state machine reads the current theoretical phase θi_theory(t) from the theoretical phase trajectory matrix Φ_traj generated in Embodiment 3 from shared memory, and obtains the latest phase correction Δθ_sync_i and phase elasticity coefficient ξi from the coordination mechanism in Embodiment 4. Then, it calculates the final injected phase θi_inject = θi_theory + ξi•Δθ_sync_i. This calculation process is required to be completed within a very short time window (e.g., 10 microseconds) before the next PWM compare register update. Subsequently, the new phase values ​​calculated by all N units are batch-processed and synchronously written to the phase registers of their respective PWM generators within the same clock cycle via the direct memory access (DMA) controller.

[0078] The timeliness of phase information is crucial for achieving near-zero energy. Any delay in phase application or asynchrony between units will directly disrupt the target time t. k The phase cancellation condition is achieved. By employing a state machine that is strictly synchronized with the PWM interrupt and DMA batch update technology, the timing error caused by the uncertainty of software instruction execution time can be eliminated to the greatest extent, ensuring the consistency of phase update actions of all units in time (up to the nanosecond level), thus physically faithfully reproducing the phase evolution process of the theoretical design.

[0079] Optionally, in some systems requiring extremely high real-time performance, phase information can be generated and fed to the PWM module in real time directly by dedicated hardware logic (HDL) in the FPGA, based on the phase trajectory matrix and correction amount, without DMA. This can achieve lower latency and higher timing accuracy.

[0080] Step 502: Perform disturbance superposition and amplitude limiting protection on the modulated wave.

[0081] Here, the modulation wave mi_original refers to the original control signal calculated by the upper-level controller (such as the power loop) to determine the PWM duty cycle when no disturbance is applied, and its amplitude is usually between [-1,1].

[0082] In practice, the disturbance component is generated based on the disturbance amplitude vector A and the current injection phase θi_inject(t), i.e., di(t) = Ai•sin(ωit + θi_inject(t)). This disturbance component is then superimposed on the original modulated wave to form a temporary modulated wave mi_temp = mi_original + di(t). To ensure safe system operation, mi_temp requires dual limiting protection. The first layer is hard limiting, which enforces the constraint |mi_new| ≤ M_max, where M_max is a safe upper limit to prevent overmodulation, for example, 0.95. The second layer is preferred soft limiting, which is handled by a smooth saturation function (such as the hyperbolic tangent function tanh), mi_new = M_max•tanh(mi_temp / M_max). This method smoothly limits the waveform when the limiting value is reached, rather than producing a sharp clipping. Simultaneously, the system records the frequency and amplitude of limiting events as a basis for adjusting the disturbance amplitude A in the next cycle.

[0083] To address the safety and fidelity issues associated with disturbance injection, direct linear superposition can lead to the synthesized modulated wave exceeding the safe operating range of the PWM module (i.e., overmodulation). This can cause significant current surges and voltage stresses, potentially damaging power devices. Limiting protection is an essential safety measure. Compared to traditional hard limiting, soft limiting maintains the continuity of the modulated wave, avoids introducing high-frequency harmonic components generated by waveform clipping, thereby reducing unnecessary electromagnetic interference to the system and ensuring the purity of the injected disturbance.

[0084] Optionally, the upper limit M_max can also be a dynamic value. For example, it can be adaptively adjusted based on the real-time monitored DC bus voltage. When the bus voltage is low, M_max is appropriately relaxed; when the bus voltage is too high, M_max is tightened to achieve more refined safety management.

[0085] Step 503: Perform synchronous triggering and high-precision sampling at the zero point.

[0086] In this embodiment, the synchronous triggering mechanism specifically refers to a global digital signal generated by the central FPGA and broadcast through dedicated hardware. This signal is used to command the ADC modules of all H-bridge units to start data conversion at exactly the same time.

[0087] In practice, the FPGA internally maintains a timer synchronized with the energy quasi-zero timetable T_zero. When the time t is equal to the next zero-point time... k After a fixed preparation time (e.g., 50 microseconds), the FPGA sends a pre-synchronization signal to all ADC modules, putting them into a ready state. At the precise zero-point time t...k The FPGA sends a rising-edge trigger signal, Trig_zero, to all ADC modules simultaneously via a dedicated, equally-length hardware trigger line. Upon receiving this signal, all ADC channels (such as Vdc_i and idc_i) immediately initiate one or more sampling operations. The FPGA appends a high-precision timestamp (e.g., 1 nanosecond resolution) to each acquired data point. After acquisition, the system can immediately verify the validity of the data. For example, it calculates the instantaneous noise level Noise_rms of the DC bus voltage at the zero-point and verifies whether it is less than a preset threshold (e.g., Noise_rms < 0.001, i.e., 0.1%) to confirm the actual achievement of the quasi-zero energy level. This ultimately forms a high-precision sampling dataset D_zero containing information such as (Vdc(tk), idc(tk), timestamp, and Noise_rms).

[0088] This step ensures that the data acquisition measurement action occurs precisely at the moment when the measured object is cleanest and has the highest signal-to-noise ratio through precise hardware-level synchronization. Software triggering inevitably suffers from command delays and timing jitter, failing to meet nanosecond-level synchronization requirements. Using hardware trigger lines is a classic and reliable method for achieving ultra-high-precision synchronization among multiple distributed modules. All the previous complex phase encoding design and collaborative control efforts are transformed into a high-value, low-noise dataset D_zero containing the true dynamic characteristics of the system, providing unprecedentedly high-quality raw materials for subsequent parameter identification.

[0089] Example 6: High-precision identification and adaptive compensation of transient parameters. This example illustrates how to identify high-precision transient parameters from the high-precision sampling dataset D_zero generated in Example 5 and then adaptively compensate for them.

[0090] Step 601: Using voltage and current data collected from multiple energy quasi-zero moments, the initial equivalent capacitance is obtained by solving the difference equation, and the initial equivalent series resistance is obtained by the principle of energy conservation.

[0091] The initial equivalent capacitance Ceq and initial equivalent series resistance Resr here refer to the equivalent parameter values ​​that reflect the current operating conditions of the system, calculated directly from the current measurement data without considering the effects of slowly changing factors such as temperature and aging.

[0092] In practice, the identification process is as follows: Ciq identification: Extract voltage-current data pairs {(V1,i1,t1),(V2,i2,t2),(V3,i3,t3)} at three consecutive zero-point times from the high-precision sampling dataset D_zero. To improve the stability and accuracy of numerical calculations, a second-order central difference scheme is used to solve the definition equation of capacitance i=C•dV / dt. The calculation formula is: Ceq_i=(idc_i(t3)-idc_i(t1))*(t3-t1) / (Vdc_i(t3)-Vdc_i(t1)). Before calculation, the extracted voltage sequence (V1,V2,V3) can be subjected to Savitzky-Golay filtering to smooth out residual quantization noise.

[0093] Resr identification: Resr identification utilizes the principle of energy conservation. Over one or more disturbance cycles, the total system loss Ploss equals the input power Pin minus the output power Pout. This total loss mainly consists of switching loss Psw and conduction loss, with the portion of conduction loss related to ESR being idc. 2 •Resr. By establishing the loss equation Ploss_measured=i 2 • Resr can be calculated by combining Resr and Psw_model, and then fitting the data from multiple sampling points using the least squares method. The switching loss Psw_model can be obtained from a pre-calibrated model or lookup table based on the current switching frequency, voltage, current, and other operating conditions.

[0094] This step employs the most suitable physical principle for identification based on different parameter characteristics. For capacitor Ceq, its iv relationship is differential, so solving it using difference equations is a natural and direct method. For resistor Resr, its characteristics are reflected in energy consumption; therefore, macroscopic identification methods based on energy conservation are much more robust than attempting to extract Resr from instantaneous voltage drops (a method easily overwhelmed by noise). By selecting appropriate identification principles and numerically stable calculation methods, highly confident initial parameter values ​​can be extracted from high signal-to-noise ratio data.

[0095] Optionally, to avoid numerical instability caused by the difference method when the voltage change is extremely small (i.e., the denominator is close to zero), a judgment condition can be set to automatically expand to a difference format of 5 points or more when |V3-V1| is less than a certain threshold, so as to enhance the robustness of the algorithm.

[0096] Step 602: Obtain the real-time temperature and device running time of each unit, and apply a temperature and device aging compensation model to correct the initial equivalent capacitance and initial equivalent series resistance.

[0097] The temperature and device aging compensation model is a mathematical model used to correct identified parameters by transforming parameter values ​​from current, specific operating conditions (such as high temperature, aging) to a standard or more general state.

[0098] In practice, firstly, the real-time temperature Ths_i is acquired using a temperature sensor installed on the heatsink of each H-bridge unit, and the equivalent junction temperature Tj_i of the capacitor core is estimated based on the power loss and thermal resistance model. Simultaneously, the cumulative operating time t_operation_i of that unit is read from the system's non-volatile memory. Then, a compensation model is applied for correction. Temperature compensation: For electrolytic capacitors, the characteristic of their capacitance changing with temperature can be approximated by a quadratic polynomial: Ceq_T=Ceq•(1+αT•(Tj_i-25)+βT•(Tj_i-25)) 2 Where Ceq is the initial value identified in step 601; αT and βT are temperature coefficients, whose initial values ​​can be obtained from the device datasheet (e.g., αT = -0.002 / °C, βT = 5e-6 / °C). 2 ).

[0099] Aging compensation: The aging of capacitors (decrease in capacitance and increase in ESR) is usually related to operating time. A simplified model can be: Ceq_final=Ceq_T•(1-γ•(t_operation_i)^0.5); where γ is an aging factor calibrated through long-term experimental data.

[0100] Because the parameters of power electronic components are not static, the capacitance and ESR of capacitors are highly sensitive to temperature and lifespan. Without compensation, the parameters identified under different ambient temperatures or after a year of equipment operation will vary significantly, causing the control model based on these parameters to lose consistency and accuracy. By using a compensation model, the identified values ​​can be corrected to the current actual operating conditions, or normalized to a standard reference condition (such as 25°C, brand new condition), improving the performance consistency and robustness of the control system throughout its entire lifespan and over a wide temperature range.

[0101] Preferably, the method further includes step 603, which involves continuously recording the historically identified transient parameters and corresponding real-time temperature data, and using the recursive least squares method to update the internal coefficients of the compensation model online.

[0102] This step represents a further intelligent upgrade to the compensation model in step 602. Specifically, the system maintains a historical database to continuously record parameter and operating condition data pairs after each successful identification, for example, (Ceq_identified, Tj). When the database reaches a certain size, a background self-calibration task is activated. This task uses the Recursive Least Squares (RLS) algorithm, employing historical data as the training set, to iteratively update the temperature compensation model C(T) = C25•(1+αT•(T-25)+βT•(T-25)) in step 602 online. 2 The internal coefficients αT and βT of the model are used. The goal of the RLS algorithm is to find a new set of (αT, βT) that minimizes the mean square error between the model's predictions and all historical measurements. The updated coefficients replace the original initial values ​​from the datasheet and are used in subsequent temperature compensation calculations.

[0103] Because the temperature coefficient and other parameters provided in the component datasheet are only typical or average values ​​for that specific component model, and the characteristics of components from different batches and individuals vary. Furthermore, the aging process itself can alter the temperature coefficient. Relying on fixed, non-personalized compensation coefficients introduces systematic errors. The RLS online update mechanism allows the compensation model to learn and adapt to the actual physical characteristics of the specific component it serves. The system can break free from dependence on general component data, establishing a fully personalized, time-evolving, ultra-high-precision parameter model, thereby achieving optimal control performance.

[0104] Alternatively, in addition to RLS, other online system identification methods such as the Extended Kalman Filter (EKF) can be used to update the coefficients of the compensation model. The EKF may exhibit better performance when dealing with nonlinear and noisy systems.

[0105] Example 7: Closed-Loop Optimization and Adjustment of Multi-Level Control Laws. This example illustrates how, after identifying high-precision transient parameters through Example 6, the system utilizes these parameters to optimize and adjust its multi-level control laws, thereby ultimately transforming the results of parameter identification into improved control performance and forming a complete identification-control closed loop.

[0106] Step 701: Based on the equivalent capacitance and equivalent series resistance of each unit contained in the transient parameters, optimize the upper-layer power allocation strategy to compensate for the dynamic response differences between units.

[0107] In this embodiment, the upper-layer power allocation strategy refers to the algorithm adopted by the upper-layer controller (or the middle-layer coordination controller) to determine how to allocate the power increment among the N H-bridge units when the total system power command changes.

[0108] In practice, this optimization process further includes: Calculate the dynamic response time constant: First, using the equivalent capacitance Ceq_i and equivalent series resistance Resr_i in the compensated transient parameter set P_transient output in Example 6, calculate the dynamic response time constant τ for each H-bridge unit i. i =Resr_i×Ceq_i. This time constant τ i It intuitively reflects how quickly the unit responds to power commands, τ. i The larger the value, the slower the response.

[0109] Determine the power allocation weights: Then, determine a set of power allocation weights w that are inversely proportional to the dynamic response time constant. i For example, the weight can be set to w. i =1 / (τ i )^γ, where the exponent γ is an adjustable parameter used to balance response speed and stability, typically ranging from 0.5 to 1, for example, 0.7.

[0110] Update the upper-layer power allocation strategy: Finally, based on this set of power allocation weights, update the upper-layer power allocation strategy. Specifically, the weights can be normalized to form a new power allocation matrix K_power, where the diagonal element K_power_ii = w i / (Σkw k When performing power allocation, the power increment allocated to the i-th unit is the total power increment multiplied by its corresponding weight K_power_ii. To prevent system oscillations caused by sudden changes in control parameters, a smoothing factor can be introduced when updating the allocation matrix, for example, K_power_new = 0.8 × K_power_old + 0.2 × K_power_calculated, to achieve a smooth transition between the old and new parameters.

[0111] This step transforms the control strategy from a static, uniform allocation to a dynamic, tailored allocation. In traditional control, power commands are typically distributed evenly, but this ignores the differences in dynamic response capabilities among units due to variations in manufacturing tolerances, operating environments, and aging processes. This step, by identifying the true time constant online, allows the power allocation strategy to predict which units are faster and which are slower. Therefore, during transients, faster-responding units are assigned slightly more tasks, and slower-responding units are assigned slightly fewer tasks. This enables the entire H-bridge cluster to coordinate and uniformly handle power surges, preventing individual units from experiencing overcurrent or overvoltage due to excessively fast response, or from dragging down overall performance due to excessively slow response. This significantly improves the system's power balance and transient stability.

[0112] Optionally, the power allocation weight w i The calculation can also incorporate other factors, such as the real-time heat sink temperature of each unit or the estimated operating efficiency, thereby achieving a more comprehensive multi-objective optimization allocation strategy that takes into account dynamic response and thermal balance, as well as loss balancing.

[0113] Step 702: Synchronously tune the underlying current loop parameters of each unit to match the updated equivalent impedance characteristics of each unit.

[0114] The underlying current loop parameters specifically refer to the parameters of the PI (proportional-integral) controller inside the local controller of each H-bridge unit, which is used to accurately track current commands, namely the proportional gain Kp_i and the integral time constant Ki_i.

[0115] In practice, after identifying the new element parameters Ceq_i and Resr_i, the system will automatically perform the following tuning process: Update the controlled object model: First, update the control object transfer function model of the current loop with the new Ceq_i and Resr_i. The equivalent impedance characteristic of this model becomes Zeq_i(s) = Resr_i + 1 / (s•Ceq_i).

[0116] Retuning the PI parameters: Based on the updated Zeq_i(s) model, an automatic tuning algorithm is used to recalculate the optimal Kp_i and Ki_i. For example, a frequency domain design method can be used to set the desired control performance indicators, such as phase margin PM (e.g., 45°) and cutoff frequency fc (e.g., one-tenth of the switching frequency fsw). Based on these indicators and the new transfer function, the new Kp_i and Ki_i can be solved using standard control theory formulas.

[0117] Implement smooth parameter switching: To avoid sudden changes and oscillations in the current loop caused by parameter updates, the newly calculated PI parameters are not applied immediately. Instead, they are gradually changed from the old value to the new value over several to dozens of control cycles through a smooth switching logic.

[0118] Any high-performance feedback controller has parameters optimally designed for the specific characteristics of the controlled object. When the controlled object (i.e., the equivalent impedance of the H-bridge unit) changes due to temperature or aging, the original optimal PI parameters are no longer optimal, potentially leading to slower current tracking, overshoot, or even instability. This step, through the linkage of identification and tuning, ensures that the current loop controller of each unit always matches its latest and true physical characteristics. This guarantees that the system maintains fast, accurate, and stable current control capabilities throughout its entire lifecycle, regardless of changes in individual units, which is the foundation for achieving high-performance power conversion.

[0119] Alternatively, in addition to the PI controller, if a more advanced control strategy such as model predictive control (MPC) is used at the underlying level, the implementation of this step becomes directly updating the identified Ceq_i and Resr_i into the internal state space model of the MPC controller (such as correcting the relevant terms in the A and B matrices), which can also achieve the effect of matching the model and optimizing the control.

[0120] Step 703: After executing the optimized control command, evaluate the power balance index of the system, and correct the initial parameters used for the spatiotemporal dimension collaborative planning based on the feedback of the index, thereby realizing closed-loop optimization across cycles.

[0121] Cross-cycle closed-loop optimization is a higher-level adaptive mechanism. Its optimization target is no longer the current control command, but the perturbation design process itself used to generate the next round of energy quasi-zero point.

[0122] In practice, this process includes: Real-time performance evaluation: After applying the optimized control commands in steps 701 and 702, the system continuously monitors the actual output power Pi_actual of each unit and calculates the power distribution deviation δPi. Based on these deviations, one or more indicators characterizing the overall power balance performance of the system are statistically derived, such as the standard deviation σ_P of the power deviations of all units.

[0123] Feedback correction of spiral coding parameters: When the evaluated performance index σ_P exceeds a preset threshold (e.g., 5%), the system determines that the current control optimization effect has not met expectations and triggers a feedback correction mechanism for the spiral coding parameters in Example 3. For example, the gradient descent method is used to update the basic contraction factor β0, and its update law is β0_new=β0-α•∂σ_P / ∂β0, where α is the learning rate, and the gradient ∂σ_P / ∂β0 can be numerically estimated through small perturbations. In addition, if a continuous oscillation at a specific frequency is detected by Fast Fourier Transform (FFT) analysis of the power waveform, and this frequency is related to the perturbation scan frequency 1 / Tsweep, the system will automatically fine-tune Tsweep to avoid potential resonance points.

[0124] If the system's power balance remains poor after one round of identification-control optimization, it may mean that the parameters identified in this round were inaccurate. The root cause might be that the disturbance design in this round (determined by parameters such as β0 and Tsweep) is not optimal for the current system state. This step, by evaluating the final control effect, corrects the meta-parameters used to generate the identification conditions, achieving a leap from optimized control to optimized identification. The system can not only adapt to its own parameter changes but also learn how to better understand itself, forming a higher-dimensional, continuously self-improving intelligent closed loop.

[0125] Example 8: Iterative Execution and Convergence Framework.

[0126] Step 801, the method is executed iteratively, wherein: the transient parameters identified at the end of each iteration are used as the basis for updating spatiotemporal dimension co-planning in the next iteration.

[0127] In this embodiment, iterative execution refers to treating the complete process described in Embodiments 2 to 7 (from transient event detection to control law optimization) as a complete identification-control cycle or an iteration, and this cycle can be executed repeatedly.

[0128] Specifically, the execution flow of the entire method constitutes an iterative framework: Startup (nth iteration): After detecting a transient event, the system initiates the nth iteration. It first executes Example 3 based on the current system model (or the result of the previous iteration n-1) to design a spiral phase coding scheme.

[0129] Execution and Identification (nth Iteration): Subsequently, the system executes Examples 4, 5, and 6, namely, performing distributed coordination, injecting perturbations, sampling at the energy quasi-zero point, and finally identifying the transient parameter set P_transient[n] in the current state.

[0130] Control optimization (nth iteration): Next, execute Example 7 to optimize the multi-level control law using P_transient[n].

[0131] Iterative connection: The key is that the parameter set P_transient[n] identified in this iteration is not only used for the current control, but also used as the basis for updating to start the next (n+1) iteration. For example, the unit time constants τi[n] contained therein will be used to calculate the more targeted adaptive shrinkage factor βi[n+1] in the n+1 iteration.

[0132] Through iteration, the system achieves a self-improvement and model refinement capability. The first identification may be based on a relatively coarse initial model; while the identification result is better than no identification, it may still contain errors. The second iteration, however, is designed based on a more accurate model after the first identification. Therefore, its perturbation injection scheme is superior, producing higher-quality energy quasi-zero points, thus identifying more accurate parameters. This iterative process of using the output of the previous iteration to optimize the input of the next iteration allows the system's understanding of its own model to gradually approach the real physical situation.

[0133] Step 802: The iterative process continues until the transient parameters meet the preset convergence conditions.

[0134] The convergence condition is a logical criterion used to determine whether the iterative process has reached sufficient precision to stop.

[0135] In practice, after each iteration ends and a new parameter set P_transient[n] is calculated, the controller compares it with the result P_transient[n-1] of the previous iteration. The convergence condition can be set as follows: the norm difference between the parameter vectors identified in two consecutive iterations is less than a preset minimum value δ_p, i.e., ||P_transient[n]-P_transient[n-1]||<δ_p. Alternatively, the convergence condition can be that the final system performance index (such as the power imbalance σ_P calculated in Example 7) changes only slightly in two consecutive iterations and is below the target threshold. Once the convergence condition is met, the system can pause the iteration process and enter a normal operating state that only uses the current optimal control law until the next major transient event or a preset timer triggers the iteration process again.

[0136] This step provides a clear termination mechanism for the iterative process, avoiding endless computation and resource consumption. After ensuring that the system model accuracy meets the requirements, the identification operations such as perturbation injection are stopped in a timely manner, allowing the system to return to the operating state with the lowest loss and highest efficiency, thus achieving a balance between high-precision identification and efficient operation.

[0137] Step 803: During the iteration process, specific mathematical evolutionary laws are applied.

[0138] Specifically, in step 1 (i.e., the design phase of each iteration): The perturbation phase of each unit follows θ i (t)=φ i +ω i t+β i • The evolution law of f(t), where i is the unit number, φ i For the initial phase, ω i Let β be the rotation frequency. i Let f(t) be the contraction factor, and f(t) be a periodic function.

[0139] By solving the equation |Σ i A i •exp(jθ i (t))|<ε•Σ i |A i |Determine the quasi-zero energy point, where A i Let ε be the perturbation amplitude of the i-th unit, and ε be the preset suppression ratio threshold.

[0140] The core object of the iterative solution of this invention is clearly identified, namely, finding a solution that satisfies specific constraints and follows specific evolutionary laws.

[0141] Example 9: This example illustrates the core technical concept of the present invention in a concrete way, particularly the process of generating the quasi-zero energy moment based on spiral phase encoding as described in Example 3.

[0142] Step 901: Set up the scene and obtain the initial state.

[0143] This case study sets up a cascaded H-bridge frequency converter system, which contains N=5 H-bridge units. Some of the system's rated parameters are set as follows: DC side rated voltage Vdc_rated = 800V, unit rated output current Irated = 50A. Before a certain transient event is triggered, using the method described in Example 2, an imbalance in the DC side voltage values ​​of each unit is detected, specifically as follows: Vdc = [805V, 795V, 810V, 790V, 800V].

[0144] Based on the above settings, the initial state characteristics are quantified and calculated: Calculate the average voltage Vdc_mean: Vdc_mean=(805+795+810+790+800) / 5=800V.

[0145] Calculate the voltage imbalance εi of each unit: According to the formula εi=(Vdc_i-Vdc_mean) / Vdc_mean×100%, the voltage imbalance vector ε is obtained. ε1=(805-800) / 800=+0.625%; ε2=(795-800) / 800=-0.625%; ε3=(810-800 ) / 800=+1.250%; ε4=(790-800) / 800=-1.250%; ε5=(800-800) / 800=0.0%.

[0146] The voltage imbalance vector ε will serve as the core input for subsequent adaptive parameter calculations.

[0147] Establish a realistic, quantifiable starting point for computation that includes imbalances. All subsequent parameter calculations will be based on this specific initial state.

[0148] Step 902: Calculate the key parameters of the spiral phase encoding. This step, based on the method described in Embodiment 3, transforms the initial state features obtained in Step 901 into specific parameters required for the phase evolution equation.

[0149] Calculate the disturbance amplitude A i To correlate the disturbance with the degree of imbalance, formula A is used. i =0.02*Irated*(1+ε i / 20). In this case, the base amplitude is 0.02 * 50A = 1A. The disturbance amplitude vector A of each element is calculated as follows: A1=1*(1+0.00625 / 20)≈1.0003A; A2=1*(1-0.00625 / 20)≈0.9997A; A3=1*(1+0.0125 / 20)≈1.0006A; A4=1*(1-0.0125 / 20)≈0.9994A; A5 = 1 * (1 + 0.0 / 20) = 1.0000A; Calculate the initial phase φ i : First, calculate the initial phase compensation term Δφ. i Using the formula Δφ i =-arctan(ε i / 10). The unit is radians (rad).

[0150] Δφ1=-arctan(0.00625 / 10)≈-0.000625rad; Δφ2=-arctan(-0.00625 / 10)≈+0.000625rad; Δφ3=-arctan(0.0125 / 10)≈-0.001250rad; Δφ4=-arctan(-0.0125 / 10)≈+0.001250rad; Δφ5=-arctan(0.0 / 10)=0.0rad.

[0151] Secondly, it is assumed that the global phase offset φ_offset_opt calculated by the convex optimization method described in Example 3 is a small value, for example, φ_offset_opt=0.01rad.

[0152] Finally, calculate the complete initial phase φ. i =2πi / N+Δφ i +φ_offset_opt.

[0153] φ1=2π*1 / 5+(-0.000625)+0.01≈1.266rad.

[0154] φ2=2π*2 / 5+(0.000625)+0.01≈2.524rad.

[0155] φ3=2π*3 / 5+(-0.001250)+0.01≈3.779rad.

[0156] φ4=2π*4 / 5+(0.001250)+0.01≈5.039rad.

[0157] φ5=2π*5 / 5+(0.0)+0.01≈6.293rad.

[0158] Set other helical parameters: Based on the technical solution, set other parameters as follows: basic contractility factor β0 = 0.5 (and assume that all β... i (All values ​​are used to simplify calculations), scan period Tsweep = 0.002s (2ms), fundamental angular frequency ω0 = 2π•100rad / s, frequency interval Δω = 2π•50 / 5 = 2π•10rad / s. Each calculated group (A) i ,φ i ,ω i ,β i Both are related to the initial imbalance ε i The correlation directly reflects the adaptability of the disturbance design.

[0159] Step 903: Solve and verify the quasi-zero energy time. This step will use the parameters calculated in step 902 to construct and solve the quasi-zero energy constraint equations to find a specific quasi-zero energy time t. k And its effectiveness is verified.

[0160] Constructing the complete equation: Substitute all the above parameters into the phase evolution equation θ i (t)=φ i +ω i t+β i • sin(2πt / Tsweep) and the quasi-zero energy constraint equation |Σ i A i •exp(jθ i (t))|<ε•Σ i |A i |in the middle.

[0161] Solving for time t k This transcendental equation system requires a solution. By applying a numerical solver such as the Newton-Raphson algorithm within the time interval t∈[0,2ms] under the parameter settings of this case, a candidate energy quasi-zero time can be obtained, for example, t k ≈0.000785s (i.e. 0.785ms).

[0162] Verification time t k Validity: Calculated at t k Phase of each unit at time t: kSubstituting 0.000785s into the phase evolution equation θ of each unit i (t)

[0163] θ1(t k )≈1.266+(2π•100)•(0.000785)+0.5•sin(2π•0.785 / 2)≈2.25rad (approximately 129°); θ2(t k )≈2.524+(2π•110)•(0.000785)+0.5•sin(2π•0.785 / 2)≈3.56rad (approximately 204°); θ3(t k )≈3.779+(2π•120)•(0.000785)+0.5•sin(2π•0.785 / 2)≈4.86rad (approximately 278°); θ4(t k )≈5.039+(2π•130)•(0.000785)+0.5•sin(2π•0.785 / 2)≈6.17rad (approximately 354°); θ5(t k )≈6.293+(2π•140)•(0.000785)+0.5•sin(2π•0.785 / 2)≈1.18rad (approximately 68°) (Note: The final result is modulo 2π).

[0164] Calculate the energy suppression ratio SR(t) k ): Calculate the complex vector sum F(t) k )=Σ i A i •exp(jθ i (t k The calculated A above i and θ i (t k Substituting the value, we can obtain F(t). k )≈0.003+j0.004.

[0165] Calculate its modulus |F(t) k )|=sqrt(0.003 2 +0.004 2 =0.005.

[0166] Calculate the algebraic sum of the disturbance amplitudes Σ|A i |≈1.0003+0.9997+1.0006+0.9994+1.0000=5.0.

[0167] Calculate the energy suppression ratio SR(t)k )=|F(t k )| / Σ|A i |=0.005 / 5.0=0.001.

[0168] Verification result: 0.001 is much smaller than the preset suppression ratio threshold ε=0.01. Therefore, the energy suppression ratio condition is met.

[0169] Verifying phase spatial dispersion: Observing the five phase values ​​calculated above (129°, 204°, 278°, 354°, 68°), they are quite dispersed on the unit circle [0, 360°], without any clustering. Their standard deviation can be further calculated, and the result will be greater than the preset phase spatial dispersion threshold (e.g., π / 3 rad, i.e., 60°). Therefore, the phase dispersion condition is met.

[0170] Supplementary Example 10 describes the implementation mechanism of this core architecture with dual time scales, specifically explaining how the fast local control at 100kHz and the slow collaborative correction at 1kHz work together organically.

[0171] Step 1001: Perform fast local control (100kHz timescale).

[0172] Fast local control refers to a control loop that operates at a high speed on the microsecond (μs) level within the local controller of each H-bridge unit and does not rely on real-time communication.

[0173] The core of this timescale is the PWM interrupt service routine (ISR), which executes at the same frequency as the PWM switching frequency, for example, 100kHz (with a period of 10μs). Within this ISR, the controller performs the following tasks: Obtain the latest, synchronously sampled voltage and current values ​​from the ADC.

[0174] The underlying current loop (e.g., the PI controller described in Example 7) is run to calculate the original modulation wave mi_original for the next cycle.

[0175] From the pre-loaded phase trajectory matrix Φ_traj, the theoretically injected phase θi_theory(t) is read according to the current microsecond-level timestamp.

[0176] Read the phase correction Δθ_sync_i updated by the slow cooperative loop from a specific shared memory address.

[0177] The final injection phase θi_inject is synthesized and superimposed on mi_original to generate the final PWM comparison value, and the hardware registers are updated.

[0178] Power electronic converters exhibit extremely fast electrical dynamics, requiring closed-loop control on a microsecond-scale to ensure stability. A fast local control loop handles all high-bandwidth control tasks directly related to the unit's stability, such as current tracking and PWM generation, ensuring the system's basic stable operation.

[0179] Step 1002: Perform slow co-correction (1kHz timescale).

[0180] Slow Coordination Correction refers to a coordination and synchronization loop that operates at the millisecond (ms) level between H-bridge units and relies on low-bandwidth communication.

[0181] The controller runs a low-priority periodic task with an execution frequency of 1 kHz (period of 1 ms). Within this task, the controller performs the following: It broadcasts its average actual phase information to its topological neighbors via CAN or RS485 bus.

[0182] Receive phase information from neighbors.

[0183] Running the distributed phase coordination algorithm described in Example 4 (such as the improved Kuramoto model), an updated phase correction Δθ_sync_i is calculated based on the collected neighbor information and its own deviation.

[0184] Write the newly calculated Δθ_sync_i to the specific shared memory address mentioned in step 1001.

[0185] Phase synchronization errors between units are typically a slow drift process, requiring no microsecond-level ultra-high frequency correction. Using a slow 1kHz scale for coordination enables: 1) significantly reduced communication bandwidth requirements, making the use of inexpensive and reliable industrial fieldbuses possible; 2) reduced computational burden on each unit's DSP, as the complex coordination algorithm executes only once every 1ms; and 3) higher tolerance for communication latency, with millisecond-level delays being entirely acceptable and compensable for a 1kHz control cycle.

[0186] Step 1003: Implement the interface and data exchange between the two time scales.

[0187] This step describes how the fast and slow loops can be decoupled and exchange data safely.

[0188] The core of data exchange is the shared memory address that is used as a mailbox.

[0189] Slow write, fast read: The slow co-current loop (1kHz) is the only writer, writing a new Δθ_sync_i to the address every 1ms. The fast local loop (100kHz) is the only reader, reading the value at the address once every 10μs of its own cycle.

[0190] Data validity: This means that within any 1ms slow cycle, the fast loop will read the same Δθ_sync_i value 100 times consecutively and use it to correct the theoretical phase of its high-frequency evolution. When the next 1ms arrives, the slow loop will provide an updated correction value, which the fast loop will use in the next 100 cycles.

[0191] This slow-write, fast-read asynchronous data interface mechanism successfully decouples two control loops with different rates. The fast loop does not need to wait for the slow loop's computation or communication, ensuring its hard real-time performance. The slow loop, in turn, provides a quasi-static, low-frequency bias or correction signal. This hierarchical, decoupled architecture is a classic method for designing complex control systems. It ensures both the high dynamic response and stability of the lower-level system and the efficient, robust, and scalable collaboration of the upper-level system, thus realizing the core concept of this invention in a highly efficient and feasible engineering manner.

[0192] According to one aspect of this application, in Embodiment 3, the value range of β0 is exemplarily limited to [0.2, 0.8]. This range is chosen based on a trade-off between the convergence rate of the perturbation energy and system stability. β0 directly determines the convergence (or divergence) rate of the spiral phase trajectory in the time dimension. A small β0 (e.g., <0.2) will make the convergence process of the perturbation energy too slow, potentially failing to form a sufficiently deep energy zero point within one scan period Tsweep. An excessively large β0 (e.g., >0.8) may cause the perturbation energy to change too drastically, equivalent to applying an excessively strong dynamic shock to the system, potentially exciting high-frequency oscillations and affecting system stability.

[0193] The optimal value of β0 is related to the system's equivalent time constant τ_eq and switching frequency fsw. In engineering practice, it can be determined through simulation or by conducting small-range frequency sweep tests on the actual system. The general principle is to choose the largest possible value of β0 to obtain the optimal energy zero depth in the shortest time, while ensuring system stability (i.e., without causing oscillations). [0.2, 0.8] is an empirical range given for a typical megawatt-level cascaded H-bridge frequency converter system, which has been verified by numerous experiments.

[0194] According to one aspect of this application, in step 703 of Embodiment Seven, the adaptive update of the spiral coding parameters employs gradient descent, with the learning rate α calculated as α = 0.1•exp(-σ_P / 10). This formula is designed to achieve adaptive adjustment of the learning rate. When the power imbalance σ_P is large, the exp term is close to 1, and the learning rate α is large (maximum 0.1), enabling the system to converge quickly towards better parameters. As σ_P gradually decreases and tends to stabilize, the exp term tends to 0, and the learning rate α also decreases accordingly. This helps the system perform a fine search near the optimal point, avoiding oscillations caused by an excessively large learning rate, thereby ensuring the stability of convergence.

[0195] The coefficient 0.1 defines the maximum value of the base learning rate, which determines the initial convergence speed. This value is a typical, relatively conservative setting, achieving faster convergence while ensuring stability. The coefficient 10, located in the denominator of the exponent, defines the sensitivity of the learning rate to decay as performance improves. This value means that as the power imbalance σ_P decreases from 10% to near zero, the learning rate undergoes a complete and smooth decay process. Both coefficients can be further fine-tuned through simulation or experimentation to match the dynamic characteristics of a specific system.

[0196] According to one aspect of this application, the noise mentioned in step 101 of Embodiment 1 and other related descriptions is suppressed by more than 40 dB, the calculation method being based on the energy suppression ratio (SR).

[0197] The noise suppression effect (unit: decibels dB) is defined as: Suppression_dB = 20•log 10 (1 / SR). Wherein, SR is the energy suppression ratio defined in step 302 of Example 3, SR=|Σ i A i •exp(jθ i (t k ))| / (Σ i |A i |). When SR is successfully suppressed to 0.01, Suppression_dB = 20•log 10 (1 / 0.01) = 20•log 10 (100) = 20 * 2 = 40 dB. Therefore, noise suppression of 40 dB is another equivalent quantitative expression of the energy quasi-zero depth reaching 1%. It should be clarified that what is mainly suppressed here is the synthetic ripple noise at the common coupling point caused by the applied disturbance signal itself with known frequency and phase, rather than all background white noise in the system.

[0198] According to one aspect of this application, in step 302 of embodiment three, the method is used to solve the nonlinear equation system F(t) = 0. Its convergence is mathematically conditional, mainly depending on the selection of the initial guess point and the characteristics of the function itself. In this application, its convergence can be guaranteed in the following ways: Initial point selection: Instead of being randomly selected within the interval [0, Tsweep], it is intelligently selected based on physical laws. For example, the approximate location of the zero point can be estimated based on a simplified linear model, or the zero point location of the current period can be predicted using the zero point distribution of the previous period, thus ensuring that the initial guessed point falls within the convergence region.

[0199] Convergence criteria: In addition to |F(tk)| being less than the threshold, a maximum iteration limit (e.g., 20 iterations) should be added, as well as a check for the non-singularity of the Jacobian matrix J(t) (i.e., the determinant is not zero) to avoid iterations at points where the derivative is zero.

[0200] According to one aspect of this application, in Embodiment 4, the improved Kuramoto model describes a discrete-time dynamic system. Its stability conditions have been extensively studied in the literature. For the discretized model in this embodiment, a sufficient condition for its stability is |1+Δt•ΣⱼK i ⱼ•cos(θⱼ-θ i -φ i ⱼ)|<1, where Δt is the update step size. By appropriately selecting the coupling strength K i The sufficiently small value of ⱼ and the update cycle ensure that the system converges stably to a synchronized state. Furthermore, the adaptive mechanism of the introduced phase elasticity coefficient ξi is itself a measure to enhance stability; it can suppress oscillations by relaxing coupling when the system exhibits an oscillating tendency.

[0201] According to one aspect of this application, in step 703 of embodiment seven, the adaptive update of the spiral coding parameters can indeed lead to system oscillations. To avoid this problem, several stabilization measures are employed: Low learning rate: The base value of the learning rate α is set at a small level (0.1) and automatically decreases as performance improves. Smoothing / filtering: All updated parameters (such as K_power and PI parameters) are smoothed or low-pass filtered to avoid step changes in parameters. Dead-band / hysteresis: A dead-band can be added to the logic that triggers parameter updates, meaning that updates are only initiated after the performance deviation σ_P exceeds a certain threshold for a considerable period of time, avoiding overreaction to small, temporary performance fluctuations.

[0202] According to one aspect of this application, an exemplary and reasonable scope of key hardware requirements is now provided: ADC Requirements: To accurately capture voltage and current dynamics, it is recommended that the ADC have a vertical resolution of at least 14 bits and a sampling rate of at least 50kHz. The 16-bit, 100kHz configuration given in the example is the preferred configuration for achieving high-precision identification.

[0203] Communication Requirements: For the distributed collaboration in Example 4, the communication link needs to have deterministic and relatively low latency. For example, using RS485 or CAN-FD bus, it is feasible to achieve 1kHz neighbor information exchange with a scale of N=20, and the communication latency should be on the order of 1 millisecond.

[0204] Computational requirements: The local controller (DSP / FPGA) needs sufficient computing power. The FPGA is mainly responsible for parallel tasks such as high-speed synchronous sampling, data alignment, PWM generation, and hardware triggering. The DSP is responsible for running more complex algorithms, including the calculation of spiral phase encoding parameters, distributed consensus algorithms, least-squares fitting for parameter identification, and online tuning of the control law. Taking TI's C2000 series DSP as an example, its floating-point operation capability needs to reach the level of hundreds of MFLOPS to ensure that all control and identification tasks can be completed within the specified time scale.

[0205] According to one aspect of this application, by designing each H-bridge unit in accordance with θ i (t)=φ i +ω i t+β i • The spiral phase trajectory of the evolution law of sin(2πt / Tsweep) and the exact solution of the constraint equation |Σ i A i •exp(jθ i (t))|<ε•Σ i |A i |, so that the disturbance signal injected into all units at a specific time t k Due to the precise phase relationship, they cancel each other out, achieving a near-zero energy level with broadband noise suppressed by more than 40dB. This time-domain noise separation mechanism solves the technical challenge of severe DC bus voltage ripple and extremely low signal-to-noise ratio (typically below 20dB) in cascaded H-bridge frequency converters under high-power transient conditions, leading to parameter identification errors exceeding 20%. Compared to traditional filtering or averaging methods that sacrifice response speed or increase hardware costs, this method improves parameter identification accuracy to within 5% without affecting system dynamic performance, and requires no additional hardware filters, reducing system costs by approximately 15%.

[0206] According to one aspect of this application, the instantaneous synchronization error e is evaluated in real time. i (t) and its rate of change de i / dt, using fuzzy rules to dynamically adjust the phase elastic coefficient ξ i ∈[0.6,1.4], which makes the improved Kuramoto synchronization model δθ i =Kp•Σⱼ∈N i a i ⱼ•sin(θⱼ-θ i -φⱼ i It can adaptively switch between strong and weak coupling. When the synchronization error is large and continues to increase, ξ i Automatic reduction to strengthen coupling ensures that phase synchronization accuracy remains within 5° even in harsh industrial environments with communication delays of up to 10ms and packet loss rates of up to 5%; when the system approaches synchronization, ξ i This increases the number of unnecessary control actions and reduces communication bandwidth usage by 60%. This solves the technical bottleneck of traditional fixed-gain synchronization algorithms, which suffer from synchronization failure and system oscillation due to accumulated communication delays when the number of cascaded units reaches more than 20. This allows the method to be extended to ultra-large-scale systems with more than 50 H-bridge units.

[0207] According to one aspect of this application, by extracting the voltage imbalance ε i Gradient of adjacent units ∇V i A multidimensional feature vector F_imb is constructed using the local clustering feature cluster_id and the temporal correlation corr_coef, and a piecewise nonlinear compensation function (|ε) is designed. i Mild compensation when |<2%, 2%≤|ε i When |<5%, secondary enhancement occurs; |ε i (Saturation limiting at ≥5%), calculate accurate initial phase compensation Δφ for each cell. i This multi-dimensional compensation strategy enables the system to reduce the imbalance to below 0.5% within the first disturbance cycle (2ms) under severe imbalance conditions, such as system startup or sudden load changes causing voltage deviations of ±10V (1.25%) between units. Compared to traditional voltage equalization control, which requires an adjustment time of over 100ms, this represents a 50-fold improvement in response speed. Simultaneously, it avoids potential overcurrent surges during rapid voltage equalization, improving system reliability under frequent start-stop conditions.

[0208] According to one aspect of this application, by constructing a vector sum equation F(t)=Σ in the complex field... i A i •exp(jθ i (t) is then linearized using Taylor expansion, and the Newton-Raphson iterative algorithm, which separates the real and imaginary parts, is employed. k ^(n+1)=t k^(n)-Re[F] / Re[J]-j•Im[F] / Im[J], precisely solving for 20-30 quasi-zero energy points within each 2ms scan cycle. Combined with a triple verification mechanism (energy suppression ratio SR<0.01, phase dispersion std(θ)... i With a zero-point stability of |dSR / dt| < 0.1 / ms (> π / 3), the effectiveness of each zero point exceeds 99%. This high-density, high-quality zero-point distribution allows for sufficient high signal-to-noise ratio sampling points within a single disturbance cycle, supporting the use of statistical methods such as least squares to improve the robustness of parameter identification. Even if individual zero points fail due to sudden disturbances, the reliability of the identification results can still be guaranteed, increasing the confidence level of parameter identification from 80% in traditional methods to over 95%.

[0209] According to one aspect of this application, a temperature compensation model Ceq_T = Ceq•(1+αT•(T-25)+βT•(T-25)) is established, which includes a linear term αT and a quadratic term βT. 2 The system employs a recursive least squares method to update model coefficients online based on historical data, achieving an accurate mapping of parameter identification results from specific operating conditions to standard conditions. When the junction temperature of the H-bridge unit varies within the range of 25°C to 85°C, the uncompensated capacitance value deviation can reach ±15%, which is reduced to within ±2% after temperature compensation. Simultaneously considering the device aging effect C_aged = C_new•(1-γ•t_operation^0.5), the system can maintain a parameter identification accuracy better than 5% even after 5 years of operation. This solves the problem that traditional offline calibration methods cannot adapt to parameter drift caused by temperature and aging in actual operation, ensuring the performance consistency of the control system throughout its entire life cycle and extending the equipment overhaul cycle from 2 years to more than 5 years.

[0210] According to one aspect of this application, the dynamic response time constant τ of each H-bridge element is calculated in real time. i =Resr_i×Ceq_i, and design a power allocation weight w that is inversely proportional to the time constant. i =1 / τ iThe exponent of 0.7 allows faster-responding units to handle more transient power changes, while slower-responding units receive relatively smoother power commands. The choice of an exponent of 0.7 balances dynamic performance and stability margin, and, combined with a smoothing factor K_power_new = 0.8 × K_power_old + 0.2 × K_power_calculated, prevents abrupt parameter changes. When the system power command steps at a rate of 0.2 pu / ms, this optimization strategy reduces the current overshoot of each unit from ±20% in the traditional equal-sharing strategy to within ±8%, and the maximum transient imbalance from 15% to 5%. This effectively avoids overload tripping caused by slow response in some severely aged units, improving the availability and reliability of the entire cascaded system under complex operating conditions.

[0211] According to one aspect of this application, by evaluating the power balance index σ_P after optimized control, the basic shrinkage factor of the spiral encoding is updated in real time using the gradient descent method β0_new=β0-α•∂σ_P / ∂β0, where the learning rate α=0.1•exp(-σ_P / 10) automatically decreases with performance improvement to enhance convergence stability. When FFT analysis detects a continuous oscillation related to the scan frequency 1 / Tsweep, Tsweep is automatically adjusted to avoid the resonance point, preventing disturbance injection from triggering system resonance. This closed-loop mechanism of "identification-control-re-optimization identification" enables the system to autonomously learn and adapt to constantly changing operating conditions. After 1000 hours of continuous operation, the parameter identification accuracy not only did not decrease but increased from the initial 5% to 3%, truly achieving the self-evolutionary characteristic of "becoming more accurate with use," which is unattainable by traditional open-loop identification methods.

[0212] This invention fundamentally solves the technical challenge of low parameter identification accuracy in cascaded H-bridge frequency converters under high-power transient conditions by creating an innovative method for quasi-zero energy points through spiral phase encoding. By carefully designing spiral phase trajectories for each H-bridge unit that follow specific evolutionary laws and solving the constraint equations for complex domain vector sums, precise cancellation of disturbance signals is achieved at predetermined times, creating a high signal-to-noise ratio sampling window with significantly suppressed noise. Combined with a distributed Kuramoto synchronization mechanism using adaptive phase elasticity coefficients, high-precision phase synchronization can be maintained even in complex systems with communication delays and a large number of units. Based on the highly accurate identified transient parameters, the system dynamically optimizes the power allocation strategy, significantly reducing current overshoot and power imbalance in each unit. Through a closed-loop iterative mechanism, the system possesses self-learning capabilities, and the parameter identification accuracy continuously improves over time. The overall solution significantly improves parameter identification accuracy and system response speed without increasing hardware costs, providing a breakthrough solution for high-performance control of large-scale cascaded H-bridge systems.

[0213] This invention fundamentally solves the technical problem of low parameter identification accuracy in cascaded H-bridge frequency converters under high-power transient conditions by creating an innovative method of quasi-zero energy points through spiral phase encoding. This is achieved by carefully designing each H-bridge unit to follow θ... i (t)=φ i +ω i t+β i • Determine the spiral phase trajectory following the sin(2πt / Tsweep) pattern and solve the constraint equation |Σ i A i •exp(jθ i (t))|<0.01•Σ i |A i This system achieves precise cancellation of disturbance signals at predetermined times, creating a high signal-to-noise ratio sampling window with noise suppression exceeding 40dB. Combined with a distributed Kuramoto synchronization mechanism featuring adaptive phase elasticity coefficients, it maintains phase synchronization accuracy within 5° even in complex systems with a communication delay of 10ms and more than 20 units. Based on highly accurate transient parameters, the system dynamically optimizes the power allocation strategy, reducing current overshoot in each unit from ±20% to ±8% and power imbalance from 15% to 5%. Through a closed-loop iterative mechanism, the system possesses self-learning capabilities, gradually improving parameter identification accuracy from an initial 5% to 3%. The overall solution improves parameter identification accuracy by more than 4 times and response speed by 50 times without increasing hardware costs, providing a breakthrough solution for high-performance control of large-scale cascaded H-bridge systems.

[0214] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. A hierarchical collaborative control method for a cascaded H-bridge frequency converter, wherein the cascaded H-bridge frequency converter comprises at least three H-bridge units, characterized in that, include: Each H-bridge unit is assigned a perturbation signal with a predetermined phase relationship, so that the vector sum of the perturbation signals of each unit satisfies the minimum condition at a predetermined time, and multiple energy quasi-zero time points are generated within a preset perturbation period. At the near-zero energy point, multiple H-bridge units are simultaneously triggered to acquire data with high signal-to-noise ratio, and transient parameters characterizing the dynamic characteristics of each unit are identified based on the acquired data. Based on transient parameters, the multi-level control law of the cascaded H-bridge frequency converter is adjusted in a closed loop to generate and execute optimized control commands.

2. The method according to claim 1, characterized in that, Coordinated planning of disturbances across the spatiotemporal dimensions includes: Based on the initial system state, a phase evolution equation following the spiral phase coding law is established for each H-bridge unit to describe the dynamic trajectory of the perturbation phase of each unit in the time dimension; Based on this set of phase evolution equations, an energy quasi-zero constraint equation is constructed, which is valid under the condition that the magnitude of the vector sum of the disturbance signals of all H-bridge units in the complex domain is less than a preset threshold. Solve the energy quasi-zero constraint equations to determine the time of the energy quasi-zero point.

3. The method according to claim 2, characterized in that, Establish phase evolution equations for each H-bridge element, including: The voltage imbalance, which characterizes the differences in electrical state of each unit, is extracted from the initial system state, and an initial phase compensation term is determined for each unit accordingly to correct the initial value of the disturbance phase of each unit in order to accelerate the energy balance of the system. Extract the system dynamic characteristics that characterize the response speed of each unit from the initial system state, and combine them with the voltage imbalance to calculate the adaptive contraction factor for each unit, which is used to regulate the convergence pattern of the disturbance energy in the spatiotemporal dimension. The initial phase compensation term and the adaptive shrinkage factor are used as key parameters to construct the phase evolution equation.

4. The method according to claim 2, characterized in that, Solving the energy quasi-zero constraint equations includes: Numerical iterative solutions are performed on the energy quasi-zero constraint equations to obtain a set of candidate zeros. For each candidate zero in the candidate zero set, the validity is evaluated. The evaluation process verifies at least whether the candidate zero simultaneously meets the preset energy suppression ratio threshold and the phase space dispersion threshold. From the candidate zero points that have passed the validity screening, the final energy quasi-zero point timetable is selected and constructed.

5. The method according to claim 2, characterized in that, Before constructing the energy quasi-zero constraint equations, the following is also included: The theoretical phase trajectory is generated based on the established phase evolution equation; A predictive model capable of quantifying phase distortion caused by the nonlinear characteristics of power devices is established, and the phase pre-compensation amount is calculated based on the model. The phase pre-compensation amount is superimposed onto the theoretical phase trajectory to form the compensated phase trajectory, and the energy quasi-zero constraint equation is constructed using the compensated phase trajectory.

6. The method according to claim 1, characterized in that, Before simultaneously triggering multiple H-bridge units to acquire high signal-to-noise ratio data, the following steps are also included: Each H-bridge unit performs a distributed phase coordination process, which generates a phase correction based on the deviation between the actual phase obtained from neighboring units and the theoretical phase it follows. The theoretical phase and the phase correction are combined to form the injected phase, and the injected phase is used to drive the perturbation of the H-bridge unit so that the actual phase of each unit remains synchronized during the perturbation period.

7. The method according to claim 6, characterized in that, The distributed phase coordination process includes: Real-time assessment of the instantaneous synchronization error of each H-bridge unit; Based on instantaneous synchronization error, the phase elasticity coefficient of each unit is dynamically updated through a preset fuzzy adaptive rule, which is used to adjust the response characteristics of phase correction between strong coupling and weak coupling. An improved Kuramoto synchronization model that takes into account communication delays is applied, and the phase correction is calculated by combining dynamically updated phase elasticity coefficients.

8. The method according to claim 7, characterized in that, The application of the improved Kuramoto synchronization model also includes: A momentum term that increases with the number of iterations is introduced when calculating the phase correction. An event-triggered mechanism is adopted. When the instantaneous synchronization error exceeds the preset trigger threshold, the phase correction amount is updated and broadcast to reduce communication overhead.

9. The method according to claim 1, characterized in that, The multi-level control law for a closed-loop cascaded H-bridge frequency converter includes: Based on the equivalent capacitance and equivalent series resistance of each unit contained in the transient parameters, the upper-level power allocation strategy is optimized to compensate for the dynamic response differences between units. Based on transient parameters, the underlying current loop parameters of each unit are synchronously tuned to match the updated equivalent impedance characteristics of each unit. After executing the optimized control commands, the power balance index of the system is evaluated, and the initial parameters used for spatiotemporal collaborative planning are corrected based on the feedback of the index, so as to achieve closed-loop optimization across cycles.

10. The method according to claim 9, characterized in that, Optimize upper-layer power allocation strategies, including: The dynamic response time constant of each H-bridge element is calculated using the equivalent capacitance and equivalent series resistance in the transient parameters. Determine a set of power allocation weights that are inversely proportional to the dynamic response time constant; Update the upper-layer power allocation strategy based on the group power allocation weight.

Images