Hydrogen power supply DCDC converter control method

Abstract

The invention discloses a hydrogen power supply DCDC converter control method comprising the following steps: carrying out signal decomposition processing on a converter output current signal to generate signal characteristics representing transient and steady state characteristics; performing phase tracking and control according to the signal characteristics to generate a target phase; based on the signal characteristics and the target phase, performing self-learning optimization on a pre-configured control parameter model, and generating control parameters adaptive to the system state; and generating a PWM control signal in combination with the target phase and the control parameter. The technical problems such as phase jitter caused by bubble burst transient interference and parameter space non-uniformity under the temperature gradient are solved, and the efficiency and stability of a hydrogen production system are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of hydrogen energy control, and particularly relates to a hydrogen power DCDC converter control method. BACKGROUND

[0002] Water electrolysis is a main technical path for realizing large-scale hydrogen energy production. A DCDC (direct current-direct current) converter is a core component of an electrolytic cell power supply system, and its control performance directly affects hydrogen production efficiency and equipment life. With the expansion of the scale of renewable energy grid connection, the hydrogen power system needs to frequently cope with power fluctuations, which puts higher requirements on the dynamic response capability and stability of the DCDC converter. In particular, in a megawatt-level high-power electrolysis system, the annual hydrogen production cost can be reduced by hundreds of thousands of yuan for every 1% improvement in control accuracy. Therefore, developing a high-performance DCDC converter control method has important engineering value and economic significance.

[0003] Currently, the hydrogen power DCDC converter control mainly adopts a classical control strategy based on proportional-integral (PI / PID), and a single fast Fourier transform (FFT) spectrum analysis is used for current harmonic detection. In terms of phase control, a standard phase-locked loop (PLL) technology is usually used to track the fundamental frequency phase, and some advanced systems introduce parallel multi-channel PLLs to track each harmonic. Parameter adaptation mainly relies on the traditional recursive least squares (RLS) algorithm, and the control parameters are adjusted according to the hydrogen production efficiency feedback. Temperature compensation mostly uses a lookup table method or a simple linear compensation model. For transient disturbances, the dynamic performance is improved by increasing the controller bandwidth or introducing feedforward compensation. These methods can meet the basic control requirements under steady-state conditions and lay a technical foundation for the industrialization application of the hydrogen power system.

[0004] However, the existing control methods have obvious deficiencies in dealing with the complex dynamic process specific to the electrolytic cell. First, the random bursting of hydrogen bubbles in the electrolysis process produces a wideband impact signal lasting for 20-50 microseconds. This ultra-short transient state causes serious aliasing in the FFT spectrum analysis, resulting in a ±30% instantaneous deviation of the fundamental frequency amplitude and a harmonic phase calculation error of more than ±40°. Moreover, the aliasing effect lasts for 5-8 calculation periods, and the traditional window function cannot effectively suppress such impacts. Second, there is a significant temperature gradient inside the large electrolytic cell, and the temperature difference between the inlet and outlet can be as high as 15°C, resulting in significant differences in electrochemical characteristics in different regions. However, the standard RLS algorithm assumes that the parameters change uniformly and slowly over the entire control object, and cannot adapt to such spatial non-uniformity, resulting in a high parameter estimation error rate at the temperature gradient boundary. SUMMARY

[0005] The application aims to provide a hydrogen power DCDC converter control method to solve at least one technical problem in the prior art.

[0006] Technical solution: A control method for a hydrogen power source DC-DC converter, comprising:

[0007] The converter output current signal is acquired and processed by signal decomposition to generate signal features characterizing transient and steady-state characteristics.

[0008] Based on signal characteristics, phase tracking and control are performed to generate the target phase;

[0009] Based on signal characteristics and target phase, the pre-configured control parameter model is self-learned and optimized to generate control parameters that adapt to the system state.

[0010] A PWM control signal is generated by combining the target phase and control parameters.

[0011] Beneficial effects: This invention solves technical problems such as transient interference from bubble bursting and phase jitter caused by non-uniformity of parameter space under temperature gradient, thereby improving the stability of hydrogen production systems. Attached Figure Description

[0012] Figure 1 A flowchart illustrating the steps of a hydrogen power source DC-DC converter control method provided in this application embodiment.

[0013] Figure 2 A flowchart illustrating the steps for generating signal features provided in this application embodiment.

[0014] Figure 3 A flowchart illustrating the steps for generating control parameters provided in this application embodiment.

[0015] Figure 4 A flowchart illustrating the steps for generating a target phase provided in an embodiment of this application. Detailed Implementation

[0016] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0017] It should be noted that the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or device.

[0018] The study found that when the electrolyte flows at high speed (flow velocity > 2 m / s), the turbulent eddies generated will cause micro-fluctuations in current density of 100-500 Hz. Although the amplitude is small, it will cause continuous phase jitter. The fixed loop bandwidth of the standard PLL cannot simultaneously meet the contradictory requirements of noise suppression and dynamic tracking, resulting in a phase jitter standard deviation of 0.8°, which seriously affects the efficiency of the multiphase rectifier system.

[0019] like Figure 1 As shown, a control method for a hydrogen power source DC-DC converter is proposed, including the following steps:

[0020] The converter output current signal is acquired, and the converter output current signal is processed by signal decomposition to generate signal features that characterize transient and steady-state characteristics.

[0021] According to one aspect of this application, the signal decomposition process includes: calculating the rate of change of the current of the converter output current signal; determining a transient detection function characterizing the transient intensity of the signal based on the rate of change of the current; and using the transient detection function to separate the converter output current signal into a steady-state path signal and a transient path signal.

[0022] Furthermore, signal features are generated, specifically by separating the output current signal into steady-state path signals and transient path signals through a transient detection function, performing matching and identification of the transient path signals based on a preset physical model library, and generating signal features that include the amplitude, phase, and transient event type of the fundamental frequency and harmonics.

[0023] Specifically, the output current signal i(t) of the DC-DC converter is acquired at a sampling rate of 10kHz. This 10kHz current signal sampling is typically implemented by a high-speed, high-precision analog-to-digital converter (ADC) module located in the core controller (such as a DSP- or FPGA-based digital controller). To overcome the spectral leakage and aliasing problems caused by traditional Fourier transforms when processing non-stationary impact signals (such as current pulses caused by bubble bursts during electrolysis), a signal separation strategy based on transient feature detection is adopted: the first derivative of the current signal, i.e., the rate of change of current di / dt, is calculated; and the transient intensity of the signal is evaluated in real time using the transient detection function D(t). In this embodiment, D(t) is modified to a form more consistent with physical logic: D(t) = 1 - exp(-|di / dt| 2 / σ 2 ); where D(t) is the transient detection function, with a range of [0, 1], and a value closer to 1 indicates a more significant transient characteristic; di / dt is the rate of change of current; σ is the transient sensitivity parameter, used to adjust the sensitivity of the function to the rate of change of current. The value of σ can be adaptively adjusted according to the statistical characteristics (such as standard deviation) of historical current data, for example, σ = 0.5 + 0.3 • std(ihistory Optionally, to obtain smoother transition characteristics, the transient detection function can also adopt the hyperbolic tangent function D(t) = tanh(|di / dt| / σ). Based on the calculated D(t), the original current signal i(t) is intelligently separated into two paths: the steady-state path signal i... s (t)=i(t)•(1-D(t)); Transient path signal: i t (t) = i(t)•D(t). The signals from the two paths are processed separately. For the steady-state path signal i... s (t) A refined spectral analysis is performed using a frequency-adaptive piecewise FFT algorithm. For the transient path signal i... t (t) employs a specific compensation algorithm. The analysis results from both methods are fused to generate high-precision signal characteristics. These characteristics include at least the amplitudes A1, A2, A3 and phases Φ1, Φ2, Φ3 of the fundamental frequency and higher harmonics, as well as possible transient event types. In an optional embodiment, the value of σ is crucial to ensuring the accuracy of transient detection. Its determination method includes: offline calibration: collecting current data of the electrolytic cell under various typical operating conditions (e.g., 1 hour), calculating the probability distribution of its rate of change di / dt, and taking its 99th percentile as (di / dt). 99 Then σ can be set as (di / dt). 99 / 3, so that the vast majority of normal fluctuations are not misjudged as transients. Online adaptive: σ(n)=σ(n-1)+μ•[std(di / dt)] history [)-σ(n-1)]; where μ is the adaptive step size (e.g., 0.01), std(di / dt) history σ is the standard deviation of the historical derivative sequence. This allows σ to automatically adapt to changing operating conditions.

[0024] Based on signal characteristics, phase tracking and control are performed to generate the target phase.

[0025] In this embodiment, stable and reliable phase information is extracted from the characteristics of a signal containing noise and disturbances, and necessary control is applied. This is achieved through a hierarchical phase-locked loop (PLL) architecture. Specifically, a dual-loop main PLL is used to accurately track the fundamental frequency phase Φ1. This main PLL includes a high-bandwidth (e.g., 500Hz) fast loop for rapid response to dynamic phase changes and a low-bandwidth (e.g., 50Hz) slow loop for extracting the long-term stable trend of the phase. The outputs of the two loops are weighted and averaged using an adaptive fusion algorithm to generate a stable and reliable target fundamental frequency phase Φ1. 1out Multiple phase-locked loops (PLLs) are used to control the relative phase of higher harmonics (such as second and third harmonics). The control target of these PLLs is not the absolute value of the harmonic phase, but rather its phase Φ relative to the target fundamental frequency. 1outThe relative phase error between integer multiples enables precise coordinated control of harmonics, generating the second harmonic phase control quantity θ2 and the third harmonic phase control quantity θ3.

[0026] Based on signal characteristics and target phase, the pre-configured control parameter model is self-learned and optimized to generate control parameters that adapt to the system state.

[0027] In this embodiment, to enable the control system to autonomously adapt to changes in the internal state of the electrolyzer (especially uneven temperature distribution), a spatially aware partitioned parameter self-learning optimization strategy is adopted. Specifically, based on data from multiple temperature measurement points within the electrolyzer, its temperature gradient distribution is analyzed, and the electrolyzer is dynamically divided into multiple control regions (such as high-temperature, medium-temperature, and low-temperature regions). For each region, a recursive least squares (RLS) algorithm with boundary smoothing constraints is independently used for parameter self-learning, where the hydrogen production efficiency η... H2 As a feedback signal, the boundary smoothing constraint ensures that the control parameters between adjacent regions do not change abruptly, maintaining the continuity of spatial control. The local control parameters θ learned from each partition are then used as feedback signals. k The optimal global control parameter θ is generated through spatial weighted fusion. global .

[0028] A PWM control signal is generated by combining the target phase and control parameters. This signal is used to drive the DC-DC converter.

[0029] In this embodiment, the target phase Φ 1out Phase control quantities θ2, θ3, and global control parameter θ global Information such as these is integrated. The basic control quantity u is calculated using a multi-objective optimization algorithm. base This optimization algorithm aims to synergistically optimize multiple objectives, including phase tracking accuracy, control parameter stability, and system temperature. It also includes an alternative predictive-correction dual-mode control mechanism. When integral saturation is detected in the control loop (usually caused by strong nonlinear events such as excessive bubble adhesion), the system automatically switches to this mode. In this mode, the system determines the bubble growth and detachment state based on the real-time impedance change rate of the electrolyzer and calculates the corrected control quantity u. corr To proactively respond to such strong disturbances. The final control voltage u final (u in normal mode) base Under dual-mode control, it is u corr (or its smooth transition value) is converted into a pulse width modulation (PWM) duty cycle signal D. compAfter safety limiting and dead-time compensation, the power switching transistors of the DC-DC converter are driven. Optionally, the PWM duty cycle digital signal generated by the controller does not directly drive the power switching transistors (such as MOSFETs or IGBTs). It is first sent to a dedicated gate driver IC. This driver IC is responsible for converting the logic level signal output by the controller (such as 3.3V) into a drive signal with sufficient current and voltage swing (such as 0-15V) to quickly and reliably turn the power switching transistors on and off, while providing electrical isolation and various protection functions.

[0030] This embodiment monitors the rate of change of current in real time using a transient detection function. When transient events such as hydrogen bubble bursting are detected, the original current signal is intelligently separated into a steady-state path and a transient path. The steady-state path uses a long-window FFT to ensure frequency resolution, while the transient path is decomposed on an overcomplete dictionary using a sparse reconstruction algorithm, avoiding spectral aliasing caused by transient impacts in traditional single FFT processing. This reduces the fundamental frequency amplitude detection error and harmonic phase calculation error, eliminates the continuous influence of aliased signals on subsequent calculation cycles, solves the problem of ultra-short-time transient interference unique to electrolytic cells, and improves the accuracy and reliability of spectral analysis.

[0031] In an optional embodiment, the signal decomposition process further includes: separating the transient path signal from the converter output current signal; using a sparse reconstruction algorithm to decompose the transient path signal on a preset overcomplete dictionary to obtain sparse coefficients; reconstructing the transient path signal based on the sparse coefficients and the overcomplete dictionary to generate a transient compensation spectrum, wherein the transient compensation spectrum is used to constitute signal features.

[0032] In this embodiment, any complex transient signal i t (t) can be approximated by a linear combination of a very small number of atoms in a large library of signal atoms (i.e., an overcomplete dictionary Ψ). Specifically, the overcomplete dictionary Ψ is a matrix whose column vectors ψ j The dictionaries are called atoms, and each atom represents a basic, typical signal waveform. For example, the dictionary could contain exponentially decaying atoms exp(-at)u(t), sinusoidal oscillation atoms of different frequencies sin(2πft), and Gaussian pulse atoms exp(-bt). 2The overcomplete dictionary Ψ can be a set of atoms, where a is the attenuation factor, u(t) is the unit step function, and b is the control factor. Theoretically, arbitrarily complex transient waveforms can be synthesized by combining these atoms. Optionally, to improve the efficiency of representing transient signals of a specific system, the overcomplete dictionary Ψ can not be a fixed standard dictionary, but rather an adaptive dictionary trained from a large number of historical transient signal samples using a dictionary learning algorithm (such as K-SVD). The atoms of this adaptive dictionary can better match the transient characteristics of a specific electrolytic cell system, thus allowing the signal to be reconstructed with fewer atoms (i.e., sparser coefficients).

[0033] The goal of the sparse reconstruction algorithm is to solve the following optimization problem: min ||s||0 subject to ||i t -Ψs||2 2 <ε; where s is the sparse coefficient vector to be solved, with most of its elements being zero; ||s||0 represents the number of non-zero elements in s, i.e., the L0 norm; Ψ is an overcomplete dictionary; i t ε represents the transient path signal; ||•||2 represents the L2 norm, i.e., the Euclidean distance; ε is a minimal error tolerance used to balance reconstruction accuracy and sparsity. Since directly solving the L0 norm minimization problem is NP-hard, approximate algorithms are typically used in practical engineering. In this embodiment, the Orthogonal Matching Pursuit (OMP) algorithm is used to efficiently solve for the sparse coefficients s. The iterative steps of the OMP algorithm are as follows: Set the residual r0 = i t ; Sparse coefficients s0 = 0; Selected atom index set ∧0 = empty set; Iteration count n = 1. Calculate the current residual r. (n-1) The inner product of k with all atoms in the dictionary Ψ n =argmax j |<r (n-1) , ψ j >|, find the index k of the atom that best matches the residual. n , where <> is the inner product symbol, ψ j For the column vectors in the overcomplete dictionary. Update the index set and atom set: ∧ n =∧ (n-1) ∪{k n Construct a matrix Ψ consisting of the selected atoms. n =[ψ k ], where k∈∧ n The solution for s is obtained using the least squares method. n =argmin x ||i t -Ψ n x||2 2 This yields the optimal coefficients on the currently selected set of atoms. Update the residual: r n =i t -Ψn s n If ||r n ||2 2 If the iteration count is less than ε or the preset maximum number of iterations is reached, the iteration stops; otherwise, let n = n + 1 and return to the inner product calculation. Alternatively, as an alternative to the OMP algorithm, other sparse recovery algorithms can be used, such as the Basis Pursuit (BP) algorithm, which relaxes the L0 norm to the L1 norm for convex optimization; or the Compact Sample Matching Pursuit (CoSaMP) algorithm, which selects multiple atoms in each iteration and has a faster convergence speed. After iteration, a sparse coefficient vector s is obtained. Based on the sparse coefficients s and the overcomplete dictionary Ψ, the transient signal is reconstructed to obtain a denoised and structured reconstructed transient signal i. t_sparse =Ψs. This is the reconstructed signal i. t_sparse Performing a Fast Fourier Transform (FFT) generates a very clean, alias-free transient compensated spectrum X. tcomp (f) This compensated spectrum will be used to fuse with the spectrum of the steady-state path, thereby forming the final, high-precision signal characteristics. This embodiment can effectively overcome the limitations of the uncertainty principle in traditional FFT analysis of transient signals, accurately attributing the energy of transient impulses to their corresponding spectral components, avoiding contamination of the steady-state spectrum, and improving the fidelity of signal decomposition.

[0034] like Figure 2 As shown, according to one aspect of this application, generating signal features includes:

[0035] The converter output current signal is deconstructed to extract its transient components;

[0036] Using a pre-defined transient fingerprint database, transient components are matched and identified to determine the event type of the transient component and generate signal features. The transient fingerprint database stores a predetermined number of mathematical models used to characterize specific physical transient events in the electrolytic cell.

[0037] In this embodiment, the transient path signal i can be separated from the original current signal i(t) using the transient detection function D(t). t(t), this signal is the transient component. To intelligently identify this transient component, a transient fingerprint database Ψ is pre-constructed. This database stores mathematical models of various typical transient events, which are based on a deep understanding of the microscopic physical processes inside the electrolyzer. The mathematical models stored in the transient fingerprint database include at least one physical process model selected from the following: a bubble bursting model to characterize the dynamic process of hydrogen bubbles growing, adhering, and eventually bursting on the electrode surface; and a micro-arc model to depict the microscopic discharge phenomenon caused by abnormally high local current density. Specifically, the model can be exemplarily defined as follows: Bubble bursting model: ψ bubble (t)=A b •t•exp(-t / τ b )•sin(2πf b •t); where ψ bubble (t) is a model characterizing the oscillating damped current impact generated when the bubble bursts; A b τ is the impact amplitude coefficient; t is time; b The decay time constant is related to factors such as electrolyte viscosity; f b ψ is the characteristic oscillation frequency generated when the bubble bursts. Micro-arc model: ψ arc (t)=A a •exp(-t / τ arc )•[1-exp(-t / τ rise )]; where ψ arc (t) is a model characterizing the microscopic discharge phenomenon; A a τ is the peak value of the arc current. arc τ is the decay time constant for arc extinction; rise The rise time constant established for the electric arc. The transient component i separated in real time. t (t) and each model ψ in the fingerprint database Ψ k (t) is used for matching. A direct matching method is to calculate the normalized correlation coefficient: ρ k =<i t (t), ψ k (t)> / (||i t (t)||•||ψ k (t)||); where ρ k For i t The correlation coefficient between (t) and the k-th fingerprint model has a range of [-1, 1]; <> represents the inner product operation; ||•|| represents the norm operation. The closer the absolute value of the correlation coefficient is to 1, the stronger the correlation coefficient. t The more similar the waveform of (t) is to the fingerprint model, the better. By calculating i t The correlation coefficient between (t) and all models in the library can be used to find the best-matching model, whose index k*=argmax(|ρk The event type corresponding to k* (such as bubble bursting or micro-arc) is identified as the type of the current transient event. This event type information will be passed to subsequent processing modules as part of the signal characteristics.

[0038] Optionally, to improve recognition accuracy, especially when multiple transient features are superimposed, more advanced pattern recognition algorithms can be used, such as a Bayesian classifier: P(k|i t )=(P(i t |k)•P(k)) / P(i t ); where P(k|i t ) is the transient signal i observed t Under the given conditions, the posterior probability that the transient belongs to the k-th type of event; P(i t |k) represents the signal i observed at the time of observation. t Then, the prior probability that the transient belongs to the k-th type of event; P(i t For all possible events, the transient signal i is observed. t The overall probability. By maximizing the posterior probability to determine the event type, prior knowledge (such as the probability of different events P(k)) can be incorporated, thereby obtaining more robust recognition results.

[0039] Furthermore, it also includes: in response to the event type identified by matching and identification, applying targeted compensation processing to the signal features; wherein, when the event type is identified as bubble bursting, the compensation processing is manifested as a preset correction of the phase information in the signal features; when the event type is identified as micro-arc, the compensation processing is manifested as a preset adjustment of the amplitude information in the signal features.

[0040] In this embodiment, different physical transients have different effects on the signal spectrum. Bubble bursting typically causes high-frequency oscillations, primarily interfering with the accuracy of phase calculation; while micro-arcs exhibit a one-time energy release, mainly affecting the accuracy of amplitude calculation. Therefore, after identifying the specific event type, asymmetric, targeted compensation can be applied. Specifically, the compensation process is as follows: when the event type is identified as bubble bursting, the phase in the signal characteristics is corrected: Φ comp (f)=Φ raw (f)+ΔΦ; where ΔΦ=-k b •ρ bubble •f;Φ comp (f) represents the compensated phase; Φ raw (f) represents the original calculated phase; ΔΦ is the phase compensation amount; k b The phase compensation coefficient for bubble bursting can be calibrated through offline experiments; ρ bubbleThe correlation coefficient is used to match the bubble burst model; f is the frequency. This compensation aims to counteract the disturbance to the high-frequency phase caused by bubble burst. When the event type is calibrated as micro-arc, the amplitude in the signal characteristics is adjusted: A comp (f)=A raw (f)+ΔA; where ΔA=-k a •ρ arc •A raw (f); A comp (f) represents the compensated amplitude; A raw (f) represents the original calculated amplitude; ΔA is the amplitude adjustment amount; k a The amplitude attenuation coefficient of the micro-arc can also be determined experimentally; ρ arc The correlation coefficient is used to match the micro-arc model. This compensation aims to correct the amplitude calculation deviation caused by the micro-arc. It can improve the accuracy and fidelity of signal characteristics, provide high-quality input for subsequent precision control (such as phase tracking and parameter self-learning), and improve the control performance and operating efficiency of the entire hydrogen power supply DC-DC converter.

[0041] This embodiment establishes a transient fingerprint database containing bubble burst models and micro-arc models, and performs intelligent matching by calculating the normalized correlation coefficient between the transient signal and the models in the database. When a bubble burst is identified, frequency-related compensation is applied to the phase information; when a micro-arc is identified, amplitude attenuation compensation is applied. This enables the system to distinguish transients with different physical causes, rather than treating them all the same. Compared with traditional unified filtering methods, this avoids over-compensation or under-compensation, improves steady-state control accuracy, shortens transient recovery time, and enhances the system's adaptability to different types of disturbances.

[0042] According to one aspect of this application, self-learning optimization includes:

[0043] Real-time temperature data of a predetermined set of temperature measurement points within the electrolytic cell are collected; based on the real-time temperature data, the temperature gradient distribution characterizing the overall thermodynamic state of the electrolytic cell is analyzed; and according to the temperature gradient distribution, the electrolytic cell is dynamically divided into predetermined control regions in space.

[0044] Specifically, multiple (e.g., nine) temperature sensors are arranged at different locations within the electrolyzer (e.g., the inlet, middle, and outlet areas). Optionally, the nine temperature sensors can be PT100 platinum resistance thermometers or thermocouples. They are strategically arranged at different locations within the electrolyzer; for example, three each are arranged inside the inlet and outlet manifolds of the electrolyte, and three are arranged on the outer wall of the middle section of the electrolyzer stack to comprehensively reflect its three-dimensional temperature field. The sensor signals, after passing through a signal conditioning circuit, are also input to the controller's ADC module. The controller collects the temperature data T from these measurement points in real time. iBased on this data, the weighted average temperature T of the entire electrolytic cell can be calculated. avg And the temperature gradient characterizing its spatial inhomogeneity (such as the axial temperature gradient ▽T) axial and radial temperature gradient ▽T radial Furthermore, when a sensor reading is significantly abnormal (e.g., outside the physical range or significantly different from neighboring points), the data for that point will be discarded and replaced with the interpolated or averaged readings of several spatially adjacent sensors. When all sensors in a certain area (e.g., the inlet area) fail, the system will switch to a model-based temperature prediction mode, using historical data and current operating conditions to estimate the temperature of that area. When all critical sensors (e.g., current) or controllers fail, the system will trigger the highest level of protection, immediately shutting down the PWM output, putting the converter into a safe shutdown state, and reporting a serious fault to the host computer. Based on the temperature gradient distribution, the electrolytic cell is dynamically divided into multiple (e.g., three) control zones in space. For example, two temperature thresholds T can be set. high and T low To divide the region: High temperature region: Z1(x)=1, if T(x)>T high In the intermediate temperature region: Z2(x) = 1, if T low ≤T(x)≤T high Low temperature region: Z3(x)=1, if T(x)<T low Where T(x) is the temperature at position x; Z k (x) is the region partitioning identifier function.

[0045] Furthermore, to enable the zoning to adapt to different workloads and ambient temperatures, the threshold T... high and T low It can be dynamic. For example, T high =T avg +ΔT base •(1+k grad •|▽T axial |);T low =T avg -ΔT base •(1+k grad •|▽T axial |); where ΔT base Based on the temperature difference (e.g., 5°C), k grad This represents the gradient influence coefficient. It means that when the temperature difference inside the electrolytic cell is large, the range of each temperature zone will also expand accordingly, making the division more physically meaningful.

[0046] Furthermore, self-learning optimization also includes:

[0047] For each segmented control region, the recursive least squares algorithm is executed independently to iteratively generate the local control parameters for that control region;

[0048] Between each pair of adjacent control regions, a preset boundary smoothing constraint is applied to the generated local control parameters to obtain smoothed local control parameters. This ensures the continuity of the parameters in spatial distribution.

[0049] In this embodiment, for each divided region k, a local observation vector Φ is constructed. k This vector contains key state variables that affect the hydrogen production efficiency in this region, such as the local current density i. k Local voltage v k Regional temperature T k Etc. Independently execute the Recursive Least Squares (RLS) algorithm to iteratively update the local control parameter vector θ of the region. k θ k (n+1)=θ k (n)+K k (n)[η H2 -Φ k T (n)θ k [(n)]; where n is the iteration number; θ k (n) is the parameter vector of region k in the nth iteration; K k (n) is the Kalman gain; η H2 This is a real-time feedback value for hydrogen production efficiency; Φ k (n) is the observation vector of the nth iteration. T This is a transpose. Preferably, the forgetting factor λ in the RLS algorithm... k It is adaptive, not a fixed value. For example, λ k =clip(λ base +Δλ trans +Δλ sat , 0.90, 0.99); where λ base The basal forgetting factor (e.g., 0.95); Δλ trans It is an adjustment term related to the number of transient events; the more transients, the higher the λ value. k Reduce to speed up tracking; Δλ sat This is a penalty term related to integral saturation; if saturation occurs, then λ... k Reduce to quickly correct the model.

[0050] Independent RLS algorithms may lead to parameter θ in adjacent regions i and θ jExcessive differences create parameter cliffs at the region boundaries, leading to control instability. To address this, boundary smoothing constraints are introduced. These constraints are determined based on a thermal coupling matrix, and the process includes constructing the thermal coupling matrix before applying the constraints. The matrix elements of the thermal coupling matrix quantitatively characterize the thermal conduction coupling relationship between the control regions in physical space. Specifically, this constraint is determined based on a thermal coupling matrix C that quantifies the physical coupling relationship between the regions. The matrix elements c of the thermal coupling matrix C... ij The thermal conduction coupling relationship between regions i and j was characterized, c ij The larger the value of c, the closer the heat exchange between the two. For example, c ij =exp(-d ij / λ thermal ), where d ij Let λ be the distance between the centroids of the region. thermal Let ||θ be the thermal diffusion length. The boundary smoothing constraint is specifically manifested as: ||θ i -θ j ||2≤ε•c ij Where ε is the constraint strength coefficient, and c ij This represents the corresponding element in the thermal coupling matrix. The physical meaning of this constraint is that the stronger the thermal coupling between two adjacent regions, the more similar their optimal control parameters should be. After the RLS iterative update, this constraint can be applied to the parameter θ through a correction step (e.g., based on the Lagrange multiplier method). k Fine-tuning is performed to obtain the smoothed parameter θ. ksmooth .

[0051] Furthermore, such as Figure 3 As shown, the control parameters are generated, including:

[0052] Calculate the regional fusion weights; wherein the determination of the regional fusion weights is based at least in part on the parameter reliability of the recursive least squares algorithm or the physical coupling relationship related to the control region; using the regional fusion weights, perform weighted fusion processing on the smoothed local control parameters of each control region to generate control parameters.

[0053] In this embodiment, after obtaining the smoothed local control parameters θ for each region... ksmooth Next, they need to be merged into a single global control parameter θ. global This is used to drive the entire converter. In one implementation, the weight w k It can be simply determined based on the geometric area or effective electrode area ratio β of each region. k This is determined by the principle that the larger the area, the greater its influence. Preferably, to obtain a more intelligent and robust fusion result, the region fusion weight w... kThe determination is multi-dimensional. It integrates two aspects of information: one is the convergence state of the RLS algorithm itself, i.e., the parameter confidence level (conf). k conf k The covariance matrix P can be based on the RLS algorithm. k The trace is used to calculate the parameters; a smaller trace indicates lower uncertainty and higher reliability in parameter estimation. Secondly, the physical coupling relationship between regions can be determined by the principal eigenvector q of the thermal coupling matrix C. max To characterize this, larger components in the eigenvector correspond to regions that play a dominant role in the overall heat transfer of the system. A specific formula for calculating the fusion weights can be: w k =(w eigen_k •conf k ) / Σ(w eigen_j •conf j ); where w eigen_k It is composed of the principal eigenvector q max Weights obtained through normalization; conf k This represents the reliability of the parameters in region k. Using regional fusion weights, a weighted fusion process is performed on the local control parameters of each control region to generate the global control parameters: θ. global =Σw k •θ ksmooth θ global This refers to the final output, which is the optimal control parameter that can adapt to the non-uniformity of the internal space of the electrolytic cell.

[0054] This embodiment calculates the temperature gradient based on multi-point temperature data within the electrolyzer, dynamically dividing the cell into high-temperature, medium-temperature, and low-temperature control zones. Each zone independently executes the RLS algorithm to update local parameters. The thermal coupling matrix quantifies the heat conduction relationship between zones, and boundary smoothing constraints are applied to ensure parameter continuity between adjacent zones. Based on parameter reliability and physical coupling relationships, fusion weights are calculated to generate global control parameters. This approach overcomes the limitation of traditional RLS's assumption of globally uniform parameters, reduces parameter estimation errors at temperature gradient boundaries, improves overall hydrogen production efficiency, and effectively solves the control challenges caused by spatial non-uniformity in large electrolyzers.

[0055] According to one aspect of this application, to address the shortcomings of standard linear controllers (such as PI controllers) in exhibiting slow dynamic response and susceptibility to overshoot when faced with integral saturation problems caused by strong nonlinearity and time-varying disturbances such as hydrogen bubble growth-decomposition cycles, a smart-switching dual-mode control strategy is provided to enhance the robustness and dynamic performance of the system. Specifically, a PWM control signal is generated, including operation in at least two switchable control modes: in the default basic control mode, a basic control quantity is calculated based on the target phase and control parameters; and when a preset integral saturation condition is detected in the control loop, the system switches from the basic control mode to a predictive-corrective dual-mode control mode to generate a corrective control quantity.

[0056] In this embodiment, under the basic control mode, the controller calculates the basic control quantity u based on a multi-objective optimization function. base However, when violent physical processes occur within the electrolytic cell, such as a sudden change in system impedance caused by a large number of bubbles instantly covering the electrode surface, the integral term of the PI or PID loop in the basic control mode accumulates rapidly until it reaches the saturation limit. This integral saturation causes the controller to malfunction for a long period (up to hundreds of milliseconds) after the disturbance disappears, a phenomenon known as integrator windup, which severely affects the dynamic response of the system. To address this problem, an integral saturation detection and switching logic is introduced. Specifically, the controller monitors the integral term saturation flags of multiple control loops (such as the main control loop, slave phase-locked loops, etc.). For example, a comprehensive saturation index I is defined. sattotal =I sat_main +I sat2 +I sat3 ; where I sat_main Main loop integral saturation indicator; I sat2 I sat3 This is the integral saturation flag for the harmonic phase control loop. The system maintains a saturation counter; when I... sattotal When the current control cycle is greater than or equal to a preset value (e.g., 2), the counter increments by 1; otherwise, the counter decrements at a certain rate. When the saturation counter continues to accumulate and reaches the trigger threshold (e.g., 3), the system is considered to have entered a continuous integral saturation state. At this time, the control mode will automatically switch from the basic control mode to the predictive-correction dual-mode control mode.

[0057] Furthermore, in the predictive-correction dual-mode control mode, a correction control quantity is generated, including: calculating the real-time impedance of the electrolyzer; deriving the impedance change rate based on the real-time impedance; and identifying the current growth or detachment state of hydrogen bubbles on the electrode surface based on the characteristics of the impedance change rate, and determining the correction control quantity based on the identified bubble state.

[0058] Specifically, upon entering the predictive-corrective dual-mode control mode, the controller's behavior shifts from passive response to active prediction and correction. Based on the fundamental frequency amplitude A1 (from signal decomposition processing) and the voltage measurement V(t) output by the DC-DC converter, the controller calculates the equivalent impedance Z(t) = V(t) / A1 of the electrolytic cell in real time. To obtain the dynamic trend of the impedance, a higher-order differential formula (such as five-point differential) is used to calculate its rate of change dZ / dt, and the result is filtered using a moving average to suppress noise. The characteristics of the impedance rate of change dZ / dt are highly correlated with the bubble dynamics on the electrode surface. When bubbles gradually grow and accumulate on the electrode surface, covering the effective reaction area, it leads to an increase in the equivalent resistance of the electrolytic cell. At this time, dZ / dt exhibits a positive value, and its amplitude is large (e.g., |dZ / dt| > 0.5Ω / s), which can be identified as the bubble growth period. When a bubble grows large enough and suddenly detaches from the electrode surface, the effective reaction area recovers instantaneously, causing a sharp decrease in the equivalent resistance. At this moment, dZ / dt changes from positive to negative and exhibits a sharp negative peak (e.g., dZ / dt < -1.0 Ω / s), which can be identified as the instant of bubble detachment. Optionally, to improve the accuracy of identification, spectral analysis of the current or voltage signal can be combined. If an energy enhancement in a specific frequency band (e.g., 20-50 Hz) is detected when dZ / dt is positive, and this frequency band is related to fluid instability associated with the bubble, then the entry into the bubble growth phase can be more reliably confirmed. Based on the identified bubble state, the control law is nonlinearly modified. A simplified prediction model is constructed to predict future error trends, such as the predicted error e. predicted (t+Δt)=e(t)+e'(t)·Δt+0.5·e"(t)·Δt 2 Based on this, the predictive control quantity u with a lower integral gain is calculated. pred =Kp•e(t)+Ki pred •∫e predicted (τ)dτ, where e(t) is the current error, e'(t) is the rate of change of the current error, e"(t) is the acceleration of the current error, Δt is the time interval, Kp is the response strength of the controller to the current error, τ is the time variable in the integration process, and Ki pred This is the integral gain used in predictive control. The dynamic integral gain correction ΔKi is calculated based on the bubble state. During the bubble growth phase, impedance increases; to suppress excessive accumulation in the integrator, the integral action should be reduced. At this time, ΔKi = -0.3•Ki normal •(1-exp(-|dZ / dt|)), meaning the larger dZ / dt is, the more the integral is reduced; where Ki normal This is the reference value for the integral gain of the controller under normal conditions. During the bubble release period, the impedance drops suddenly; to help the controller respond quickly, the integral action should be enhanced. At this time, ΔKi = 0.5•Ki normal•sign(dZ / dt) (Since dZ / dt is negative at this point, it actually reduces the inverse of the integral action, i.e., rapidly increases the control output). The final corrected control quantity u corr =u pred +ΔKi•e(t). When the saturation condition is removed (e.g., no integral saturation is detected for 10 consecutive control cycles), the controller will smoothly transition the control quantity from u to u using a smooth transition function. corr Smoothly switch back to the basic control variable u base This ensures smooth and uninterrupted switching between control modes. It effectively overcomes integral saturation problems under strong nonlinear disturbances, improving the system's dynamic response speed and stability.

[0059] This embodiment avoids the integral saturation problem of traditional PI controllers during the bubble cycle. Under conditions of dense bubble generation, it reduces control overshoot, shortens settling time, effectively solves the strong nonlinear disturbance problem unique to hydrogen power systems, and improves the system's robustness and dynamic response capability under varying operating conditions.

[0060] In hydrogen power applications, complex fluid dynamics phenomena such as high-speed turbulence and bubble evolution in the electrolyte cause rapid random fluctuations in the load characteristics of the DC-DC converter, resulting in small but frequent jitter in the fundamental frequency phase of the output current signal. Traditional single-loop phase-locked loops (PLLs) struggle to simultaneously guarantee fast dynamic tracking and high steady-state accuracy under such broadband noise conditions. Therefore, a dual-loop main PLL is proposed for phase tracking and control, comprising: a high-bandwidth fast loop for rapid dynamic tracking of the fundamental frequency phase in the signal characteristics, obtaining the fast loop phase; and a low-bandwidth slow loop for filtering the fundamental frequency phase to extract its core trend, obtaining the slow loop phase trend.

[0061] Specifically, the fundamental frequency phase Φ1 obtained from the signal decomposition processing step is simultaneously input into two parallel loops. The fast loop is constructed as a high-bandwidth (e.g., natural frequency ωn = 2π•500 rad / s) second-order phase-locked loop, providing the fastest response to instantaneous phase changes. Its loop filter can be designed as H... fast (s)=(1+s / ω n ) / (s 2 / ω n 2 +2ζs / ω n+1 ); where s is the Laplace operator; ω n The loop's natural angular frequency determines the tracking speed; ζ is the damping coefficient (usually taken as 0.707 to obtain critical damping characteristics), which determines the system's overshoot and stability. The loop output can closely follow the fast loop phase Φ of the input phase dynamics. fastThe slow loop is constructed with an extremely low bandwidth (e.g., cutoff frequency ω). c A first- or second-order low-pass filter (2π•50rad / s) is used to filter out high-frequency noise and random jitter in the phase signal, extracting the long-term, stable evolution trend of the phase. Its loop filter can be designed as H... slow (s)=1 / (1+s / ω c The loop output has a very smooth slow loop phase trend Φ. trend Furthermore, the slow loop can also be used to calculate second-order phase statistics, such as the mean μ. Φ and variance σ 2 Φ .

[0062] Furthermore, such as Figure 4 As shown, generating the target phase includes:

[0063] Based on the phase trends of the fast loop and slow loop, the stability of the current fundamental frequency phase is evaluated, and a phase stability index is generated. Based on the phase stability index, an adaptive fusion weight is dynamically determined. Using the adaptive fusion weight, the phase trends of the fast loop and slow loop are weighted and fused to generate the target phase.

[0064] In this embodiment, the fast loop phase Φ was obtained. fast and slow loop phase trend Φ trend Subsequently, simply averaging the results for fusion fails to leverage the advantages of the dual-loop design. Therefore, an adaptive fusion mechanism based on phase stability is introduced. Specifically, phase stability can be evaluated through multiple dimensions. For example, the phase variance σ calculated in the slow loop can be used as an example. 2 Φ As an indicator, a larger variance indicates greater phase instability. Simultaneously, the rate of change of the fast loop phase, dΦ, can be used as a metric. fast The amplitude of / dt serves as another indicator; a larger rate of change indicates that the phase is undergoing a rapid dynamic process. A comprehensive transient indicator could be I... trans =|dΦ fast / dt| / ω nom ;where ω nom The angular frequency is the rated frequency. The fusion weight α is a value between 0 and 1, used to determine the dependence of the final output on the fast and slow loops. α is dynamically changing, and its calculation is based on the phase stability index. Preferably, α is constructed as a frequency-dependent sigmoid function: α(f) = 1 / (1 + exp(-k(f - fc))); where k is the slope of the function; and fc is the center frequency. The parameters k and fc of the sigmoid function are adaptively adjusted according to the phase stability index: k = 0.1 + 0.05•I trans When transient index I transAs the slope k increases, the weight switching becomes steeper; fc = 250 + 50•σ Φ When the phase variance σ 2 Φ As the frequency increases, the center frequency fc shifts to higher frequencies, meaning that the system has greater confidence in the slow loop that can suppress noise over a wider low-frequency range. The final target fundamental frequency phase Φ 1out The weighted fusion formula yields: Φ 1out =(1-α avg )•Φ fast +α avg •Φ trend ;where α avg This represents the average value of the fusion weighting function α(f) within the operating frequency band.

[0065] As an alternative implementation, the adaptive fusion weight α can also be determined using a fuzzy logic controller. This fuzzy logic controller takes the phase variance and phase change rate as input linguistic variables (e.g., large, medium, small), and outputs a determined fusion weight value through a series of preset fuzzy rules (e.g., if the phase variance is large and the phase change rate is large, then the weight is small). It does not require a precise mathematical model and has stronger robustness.

[0066] This embodiment effectively suppresses noise while maintaining a fast response when turbulent disturbances occur due to high-speed electrolyte flow. Compared to traditional single-loop PLLs, it reduces phase jitter, improves phase tracking bandwidth and noise suppression capability, resolves the inherent contradiction in bandwidth selection of traditional PLLs, and improves the system's power factor and operating efficiency. Through a dual-loop adaptive fusion structure, it can rely primarily on the slow loop output to achieve extremely high phase accuracy and noise suppression capability when the signal is stable; while when the signal undergoes transient changes, it automatically and smoothly switches to rely primarily on the fast loop output, ensuring optimal dynamic tracking performance. The final output target fundamental frequency phase combines the advantages of high steady-state accuracy and high dynamic response.

[0067] According to one aspect of this application, phase tracking and control further includes: enabling one or more slave phase-locked loops to regulate the higher harmonic phase in the signal characteristics; wherein the control target of the slave phase-locked loop is not the absolute value of the higher harmonic phase, but the relative phase error between the higher harmonic phase and a preset integer multiple of the fundamental frequency phase output by the master phase-locked loop.

[0068] Specifically, a stable and reliable target fundamental frequency phase Φ is obtained through a dual-loop main phase-locked loop. 1outNext, precise phase control is required for key higher harmonics such as the second and third harmonics, as their phase directly affects the current waveform, and consequently, the efficiency and uniformity of the electrochemical reaction. Therefore, an independent slave phase-locked loop (PLL) is used for each higher harmonic requiring control (e.g., the second harmonic). Unlike the master PLL, which tracks the absolute phase, the slave PLL controls the relative phase error. Taking the second harmonic as an example, its relative phase error Φ... e2 Defined as: Φ e2 =mod((Φ2-2Φ 1out )-Φ 2ref +π, 2π)-π; where Φ2 is the real-time phase of the second harmonic extracted from the signal features; 2Φ 1out The theoretical second harmonic phase of the main phase-locked loop output fundamental frequency phase; Φ 2ref A preset second-harmonic relative phase reference value (e.g., -30°) is used, which is the optimal relative phase obtained experimentally or through simulation to maximize hydrogen production efficiency; the mod(•, 2π) operation is used to normalize the error to the [-π, π] interval. The relative phase error Φ is typically controlled by a gradient-limited PI controller inside the phase-locked loop. e2 Adjustments are made to output a 2-harmonic phase control quantity θ2. Gradient constraint |dθ2 / dt|≤ω max This prevents the phase control quantity from changing too rapidly when the system is subjected to large disturbances, ensuring smooth control; where ω max This represents the maximum permissible rate of change of the phase control quantity output from the phase-locked loop (PLL). Optionally, when the phase of the PLL output undergoes a jump greater than π, it is considered a phase unwrapping error. In this case, correction is required: if Φ new -Φ old If ∀π, then let Φ new =Φ new -2π; if Φ new -Φ old If <-π, then let Φ new =Φ new +2π. To avoid introducing abrupt changes during correction, the actual update value can be smoothly transitioned: Φ smooth =Φ old +sign(Φ new -Φ old )•min(|Φ new -Φ old |,ΔΦ max ); where ΔΦ max Φ is the maximum permissible phase change in a single cycle. new Φ represents the latest phase value extracted from the phase-locked loop or signal processing at the current moment. old This is the phase value recorded at the previous moment.

[0069] According to one aspect of this application, it further includes: calculating an effective temperature, which is a result of correcting an actual measured temperature, the correction being intended to quantify the thermal effect of the higher harmonic phase control quantity on the electrolytic cell, thereby forming a phase-to-temperature coupled feedback in the control model; wherein the calculation of the effective temperature integrates the actual measured temperature and a temperature rise compensation term related to the magnitude of the higher harmonic phase control quantity.

[0070] In this embodiment, controlling the phase of higher harmonics (i.e., outputting control quantities θ2, θ3, etc.) changes the magnitude and phase of the harmonic current. This portion of the harmonic current generates additional power loss (P) when flowing through the equivalent resistance of the electrolytic cell. loss =I harmonic 2 •R eq , where I harmonic R is the current component caused by higher harmonics. eq (This refers to the equivalent impedance of the electrolyzer at harmonic frequencies). This loss is ultimately dissipated as heat, causing the electrolyzer temperature to rise. In other words, the phase control behavior itself is an internal heat source of the system. To accurately account for this effect in the control model, an effective temperature T is proposed. eff The concept. It does not directly use the temperature T measured by a physical sensor. avg Instead, it was modified to include physical meaning: T eff =T avg +ΔT phase +ΔT trans ; where T eff Effective temperature; T avg The actual measured temperature is obtained by weighted averaging from multiple sensors; ΔT phase This is the temperature rise compensation term, which quantifies the thermal effect caused by phase control behavior and can be calculated as: ΔT phase =γ•(θ2 2 +θ3 2 ); where γ is the phase-temperature coupling coefficient, its physical meaning is the steady-state temperature rise corresponding to the square of a unit phase control quantity, which can be calibrated experimentally; θ2 and θ3 are the second and third harmonic phase control quantities, respectively. ΔT trans This is an optional additional compensation term used to quantify the cumulative effect of transient thermal shock caused by transient events (such as bubble bursting), and can be calculated as: ΔT trans =δ•N transient Where δ is the transient influence coefficient; N transient This counts transient events per unit time. This is achieved by introducing an effective temperature T. effA complete phase-temperature bidirectional coupling closed loop is constructed: temperature T affects the electrochemical reaction, which in turn affects the current signal characteristic Φ; the controller adjusts the phase according to Φ and outputs a control quantity θ; the control quantity θ generates harmonics and loses heat, which in turn affects the system temperature T. This can more profoundly reflect the internal physicochemical processes of the system, thereby achieving more precise and efficient control. In an optional embodiment, the determination of the phase-temperature coupling coefficient γ includes: under stable operating conditions, keeping other parameters constant, actively injecting different phase control quantities (such as θ2 gradually increasing from 0° to 30°) into the phase-locked loop, and measuring the final steady-state temperature rise ΔT of the electrolyzer. By comparing ΔT with θ2... 2 By performing least-squares fitting on the data points, the coefficient γ (unit: °C / rad) can be obtained. 2 ).

[0071] This embodiment introduces a reverse influence mechanism of phase control on temperature, building upon traditional unidirectional temperature compensation. By calculating the power loss generated by the harmonic phase control quantity, it is converted into a temperature rise compensation term, forming an effective temperature for control. This accurately reflects the physical fact that the control behavior itself is an internal heat source. The introduction of bidirectional coupling improves temperature control accuracy, avoids system overheating caused by neglecting the control's thermal effect, enhances the physical consistency of the control system, and extends the service life of the electrolytic cell.

[0072] In summary, this application discloses a control method for a hydrogen power source DC-DC converter, comprising: intelligently separating the converter output current signal into steady-state and transient dual paths using a transient detection function, processing them separately using adaptive piecewise FFT and sparse reconstruction algorithms, and then fusing them to generate high-precision signal features; employing a fast-slow dual-loop main phase-locked loop to track the fundamental frequency phase, and multiple slave phase-locked loops to control the relative phase of higher harmonics, achieving disturbance-resistant hierarchical phase control; dividing the electrolyzer into multiple control regions based on the temperature gradient, with each region independently performing RLS parameter self-learning with boundary smoothing constraints, and introducing phase-temperature bidirectional coupling compensation; combining the above results to generate a PWM control signal through multi-objective optimization, and automatically switching to predictive-correction dual-mode control when integral saturation is detected. This method solves technical challenges such as transient interference from bubble bursting, non-uniform parameter space under temperature gradients, and phase jitter caused by turbulence, thereby improving the efficiency and stability of the hydrogen production system.

[0073] In one specific embodiment, the scenario and parameters are set as follows: a 1MW PEM electrolytic cell system operating at 80% of its rated power. The controller collects real-time data (unit: °C) from the following nine temperature measurement points: Inlet zone (k=3, low temperature zone): T in =[70.1, 70.5, 70.3]; Central region (k=2, medium temperature region): T mid =[75.2, 75.8, 75.5]; Export zone (k=1, high temperature zone): Tout =[80.3, 80.9, 80.6]. Constraint strength coefficient ε=0.5; Intercentroid distance (estimated): d 12 =0.4m, d 23 =0.4m, d 13 =0.8m. (Corrected parameter) Thermal diffusion length λ thermal =0.2m, determined based on the structure and material properties of the electrolytic cell; here, a value that better reflects the inter-regional heat conduction effect is selected. Calculate the average temperature of each zone: T in_avg =70.3°C; T midavg =75.5°C; T outavg =80.6°C. Calculate the weighted average temperature of the electrolytic cell: T avg =0.3•T in_avg +0.4•T midavg +0.3•T outavg =75.5°C. Determine the temperature threshold: T high =T avg +2.5°C = 78.0°C; T low =T avg -2.5°C = 73.0°C. The exit area is divided into a high-temperature zone (k=1); the middle area is a medium-temperature zone (k=2); and the entrance area is a low-temperature zone (k=3). Calculate the matrix element c. ij =exp(-d ij / λ thermal ). c 12 =exp(-0.4 / 0.2)=exp(-2)≈0.135; c 23 =exp(-0.4 / 0.2)=exp(-2)≈0.135; c 13 =exp(-0.8 / 0.2)=exp(-4)≈0.018. The thermal coupling matrix C (symmetric matrix, diagonal is 1) is obtained: C=|1.000, 0.135, 0.018; 0.135, 1.000, 0.135; 0.018, 0.135, 1.000|. This matrix indicates that there is moderate thermal coupling between adjacent regions, and weak coupling between non-adjacent regions. Taking the medium-temperature region k=2 as an example, assuming at the current time n, the parameter vector θ2(n-1)=[0.85, -0.22, 0.11]ᵀ; the observation vector Φ2(n)=[150A, 55V, 75.5°C]ᵀ and the hydrogen production efficiency feedback η are obtained. H2 (n)=0.92. After RLS iterative calculation, the updated parameter vector is assumed to be θ2(n)=[0.81, -0.25, 0.14]ᵀ. To ensure the correct physical meaning of the constraint, the modified constraint formula is adopted: the stronger the thermal coupling between two regions (c ijThe larger the value of ||θ|, the more similar their control parameters should be; that is, the smaller the allowable difference. The constraint is defined as: ||θ| i -θ j ||2≤ε•(1-c ij Taking the high-temperature region (k=1) and the medium-temperature region (k=2) as an example for verification: Assume the parameter vector of the high-temperature region after RLS update is θ1(n)=[0.71, -0.45, 0.22]ᵀ. Calculate the Euclidean distance between the two parameter vectors: ||θ1(n)-θ2(n)||2=||[0.10, 0.20, 0.08]ᵀ||2=sqrt(0.10) 2 +0.20 2 +0.08 2 )≈0.237. Calculate the maximum allowable parameter difference (constraint threshold) at this boundary: Threshold 12 =ε•(1-c 12 =0.5•(1-0.135)=0.4325. Judgment: 0.237<0.4325. In this case, the parameter difference between the two regions does not exceed the allowable threshold, therefore the boundary smoothing constraint is not violated, and no correction procedure needs to be initiated. This indicates that although there is a difference in the parameters between the two regions, this difference is within a reasonable range determined by their thermal coupling strength. If the calculated parameter distance is greater than 0.4325, a correction mechanism will be triggered, pulling θ1(n) and θ2(n) closer to each other until their distance satisfies the constraint condition. Assume that the local parameter vectors of the three regions after RLS update (and possibly boundary smoothing correction) are θ1(n) and θ2(n). 1smooth θ 2smooth θ 3smooth Based on the convergence and reliability of the parameters in each region (conf) k Based on the physical importance determined by the eigenvectors of the thermal coupling matrix, the final region fusion weight is calculated as w = [0.3, 0.4, 0.3] (assuming the importance of the mid-temperature region and the inlet / outlet region is similar). The final global control parameter is then calculated: θ. global =w1•θ 1smooth +w2•θ 2smooth +w3•θ 3smooth Through the above steps, a single global control parameter θ is generated, which embodies a profound understanding of the spatial non-uniformity of the electrolytic cell. global .

[0074] In another detailed embodiment, the controller acquires the current signal i(t) at a sampling rate of 10 kHz. To balance the efficiency of real-time computation with the needs of historical data analysis, the controller internally maintains a circular buffer with a length of 512 sampling points. This buffer is logically divided into two parts: the most recent 256 sampling points constitute the real-time computation sequence i. bufferThis is used to calculate instantaneous quantities such as the derivative at the current moment; the 256 sampling points immediately preceding it constitute the historical reference sequence i. history This is used to calculate statistical parameters, such as standard deviation and median, thus providing a benchmark for adaptive algorithms. To accurately capture the dynamic characteristics of the signal, a higher-order difference formula is used to calculate the derivative. The first derivative (rate of change of current) is calculated using a five-point center difference formula: di / dt≈[i(n+2)-8i(n+1)+8i(n-1)-i(n-2)] / (12Δt); where i(n) is the current value at the current sampling point; i(n±1), i(n±2) are the values ​​at adjacent sampling points; and Δt is the sampling period, i.e., 0.1ms. Compared to a simple two-point difference, this has higher accuracy and better noise suppression characteristics. Based on this, the second derivative can be calculated by performing a three-point difference again on the first derivative sequence: d 2 i / dt 2 ≈[di / dt(n+1)-2di / dt(n)+di / dt(n-1)] / Δt 2 To further eliminate the influence of measurement and quantization noise on derivative calculation, the obtained first and second derivative sequences need to be smoothed using a three-point moving average filter to obtain the smoothed first derivative di / dt. smooth and smooth second derivative d 2 i / dt 2 smooth .

[0075] The transient detection function D(t) is calculated based on the sensitivity σ and the threshold θ. The adaptive capability of the threshold θ is crucial for avoiding false positives and false negatives. The threshold θ is dynamically set based on the statistical characteristics of the historical reference sequence: θ = 2.5•median(|di / dt) history |); where median(|di / dt) history |) represents the median of the historical derivative absolute value sequence |di / dt|. Using the median instead of the average effectively resists interference from occasional strong transient shocks in historical data, allowing the threshold θ to more stably reflect the system's fundamental noise level under current operating conditions. This ensures the transient detection logic is robust across electrolytic cells with different loads and aging levels. The controller continuously monitors the output of the transient detection function D(t). When a rising edge of D(t) is detected (e.g., D(t) > 0.8 and the previous time step D(t-1) < 0.8), the transient event counter N... transient Add 1. Simultaneously, the system will accumulate the transient duration τ. trans =Σ(D(t)>0.5)•Δt. The controller maintains a queue containing the durations of the most recent 100 transient events, through which the average transient duration τ can be calculated. avg and transient occurrence frequency f trans. τ avg It can be used for the calculation of subsequent spectrum fusion weights, and N transient It can then be used to calculate the effective temperature and adjust the partitioned RLS forgetting factor, thereby closely linking microscopic transient behavior with macroscopic control strategies.

[0076] According to another aspect of this application, in signal decomposition processing, the process of performing frequency domain analysis on steady-state and transient signals and finally fusing them into a unified spectrum specifically involves: the original current signal being separated into steady-state path signals i s (t) and transient path signal i t (t). Before performing the FFT, different window functions are applied to the signals of the two paths for preprocessing: for the steady-state path signal i s (t) is processed using the Hanning window. The Hanning window has good main lobe width and side lobe suppression capability, which can effectively reduce spectral leakage caused by signal truncation, and is suitable for steady-state signal analysis with high frequency resolution requirements. For transient path signal i t For (t), a rectangular window is used because it does not cause any attenuation in the time domain and can most completely preserve the edge information and impulse pattern of the transient signal. To achieve optimal analysis of different frequency bands of the steady-state signal, a piecewise FFT strategy is adopted: Low frequency band (0-200Hz): This band contains the fundamental wave and the main low-order harmonics, requiring the highest frequency resolution. Therefore, a 4096-point FFT is used for calculation, achieving a frequency resolution of approximately 2.44Hz. To further improve accuracy, a modulated Hanning window w is preferred here. l (n) = 0.5 - 0.5cos(2πn / N) l )•(1-0.1•f trans / f max Its shape will vary depending on the transient frequency f. trans Fine-tuning is performed; the more transients, the slightly reduced window smoothness is used to accommodate the slight non-stationarity of the signal; where N l f represents the sampling points used when performing Fast Fourier Transform in the low-frequency range (e.g., 0–200 Hz). max This represents the upper limit of the frequency range covered by the spectrum analysis. Mid-frequency band (200Hz-2kHz): Calculated using a 1024-point FFT, with a frequency resolution of 9.77Hz. The Blackman window is preferred for this band due to its superior sidelobe suppression compared to the Hanning window, making it suitable for analyzing mid-frequency harmonics with lower amplitudes. High-frequency band (>2kHz): Calculated using a 256-point FFT, with a frequency resolution of 39.06Hz. This band requires high amplitude accuracy; a flat-top window is preferred, as this window function sacrifices some frequency resolution for extremely high amplitude measurement accuracy. The low, mid, and high-frequency spectrum X segments of the steady-state path are obtained. l(f), X m (f), X h (f) and the compensated spectrum X of the transient path tcomp After (f), they need to be seamlessly fused into a complete final spectrum X(f) covering the entire frequency band. This can be achieved through adaptive weighting, which solves the problem of the contradiction in the design of the weighting function in the original scheme: X(f) = W l (f)•X l (f)+W m (f)•X m (f)+W h (f)•X h (f)+W t (f)•X tcomp (f); where the weighting function is carefully designed to ensure a smooth transition in the transition zone and a sum of 1: W l (f)=0.5•[1-tanh(10•(ff c1 ) / Δf1)];where f c1 This represents the low-to-mid-frequency transition center (e.g., 190Hz), and Δf1 is the transition bandwidth. W m (f) Use a similar tanh function in the transition band between low and high frequency bands to ensure consistency with W. l (f) and W h (f) Smooth transitions are achieved, with a weight of 1 in the intermediate frequency band, ensuring seamless spectrum splicing. W t (f)=0.3•(1-exp(-f•τ avg It was clarified that the contribution of the transient compensation spectrum increases with frequency, and its growth rate is related to the average transient duration τ. avg This aligns with the physical law that transient impact energy is primarily distributed at high frequencies. From the fused complete spectrum X(f), the frequencies of the fundamental and harmonics are precisely located by searching for local maxima within a preset frequency tolerance range (e.g., fundamental frequency 50Hz ± 0.5Hz, second harmonic (2f1 ± 1Hz)). Once located, their complex values ​​X(f) can be extracted. n )=A n •exp(jΦ n This allows us to obtain the precise amplitude A. n and phase Φ n Based on the extracted harmonic amplitudes, calculate the total harmonic distortion (THD): THD = sqrt(A2) 2 +A3 2 +...) / A1.

[0077] In another embodiment of this application, the process of generating the PWM control signal is specifically as follows: In the basic control mode, the controller does not simply track a target, but calculates the basic control quantity u through a multi-objective optimization function (or cost function) J. base An example optimization function is as follows: J = w1•(Φ 1out -Φ ref_T ) 2 +w2•||θ global || 2 +w3•(T eff -T opt ) 2 The first term, w1•(Φ), is... 1out -Φ ref_T ) 2 Represents phase tracking error, the goal is to make the actual output target phase Φ 1out A phase reference Φ that closely follows temperature adaptation ref_T The second term w2•||θ global || 2 This represents controlling energy consumption, with the goal of penalizing excessive control parameters, preventing overly aggressive controller actions, and ensuring stable system operation. The third term, w3•(T) eff -T opt ) 2 This represents the temperature control error, and its goal is to reduce the effective temperature T of the system. eff Maintain at the optimal operating temperature T opt Nearby. w1, w2, and w3 are the weighting coefficients for each item. These weighting coefficients are not fixed but can be dynamically adjusted according to the system's macro-operation strategy. For example, in an economic mode that pursues the highest hydrogen production efficiency, the weight of w3 can be increased; in a dynamic response mode that needs to cope with rapid load changes, the weight of w1 can be increased. The controller obtains the optimal basic control quantity u by minimizing the cost function J. base In the above optimization function, the phase reference Φ ref_T It is also dynamic and nonlinear. In addition to the quadratic relationship with temperature T, a nonlinear compensation term f(dT / dt) related to the rate of temperature change dT / dt is introduced to cope with rapid temperature changes: Φ ref_T =[Φ0+α1(T-T0)+α2(T-T0) 2The coefficients α1 and α2 are obtained by establishing a temperature-controllable experimental platform. While maintaining constant parameters such as current and flow rate, the overall temperature T of the electrolytic cell is slowly changed (e.g., from 25°C to 65°C), and the change in fundamental frequency phase Φ is measured using a high-precision phase-locked loop device. The coefficients α1 and α2 are obtained by performing a quadratic polynomial fitting on the data points of Φ and (T-T0). The nonlinear compensation function is f(dT / dt) = tanh(β•dT / dt)•exp(-|dT / dt| / τ); where β and τ are coefficients. The tanh term provides fast feedforward compensation proportional to the direction of temperature change, while the exponential decay term ensures that the compensation effect is suppressed when the temperature change is too drastic (potentially indicating sensor failure or abnormal operating conditions), thus guaranteeing safety. The final control voltage u is obtained... final (u from the basic mode) base or dual-mode controlled u corr After that, the physical PWM signal needs to be generated through the following steps: linearly mapping the voltage value to the PWM duty cycle D. D=(u final +V dc / 2) / V dc ;where V dc This is the DC bus voltage of the DC-DC converter. To prevent damage to power devices or control failure due to extreme duty cycles (0% or 100%), the duty cycle D is limited: D safe =max(D min ,min(D max ,D));where D min and D max The preset safety upper and lower limits are set, for example, 0.1 and 0.9. In half-bridge or full-bridge topologies, a small dead time t is set to prevent short circuits caused by the simultaneous conduction of two switches in the same bridge arm. dead Dead time causes a deviation between the actual output voltage and the theoretical value. This embodiment compensates for this: D comp =D safe +t dead •f sw / (1-D safe ); where f sw The switching frequency. The final compensated duty cycle D. comp The PWM peripheral is loaded into the digital controller (such as a DSP or FPGA) to generate a precise gate signal to drive the power switching transistor, thus completing the entire control closed loop.

[0078] 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 control method for a hydrogen power source DC-DC converter, characterized in that, include: The converter output current signal is acquired and processed by signal decomposition to generate signal features characterizing transient and steady-state characteristics. Based on signal characteristics, phase tracking and control are performed to generate the target phase; Based on signal characteristics and target phase, the pre-configured control parameter model is self-learned and optimized to generate control parameters that adapt to the system state. A PWM control signal is generated by combining the target phase and control parameters.

2. The method according to claim 1, characterized in that, Generated signal features include: The output current signal of the converter is deconstructed to extract the transient components. Using a pre-defined transient fingerprint database, transient components are matched and identified to determine the event type of the transient component and generate signal features. The transient fingerprint database stores a predetermined number of mathematical models used to characterize specific physical transient events in the electrolytic cell.

3. The method according to claim 2, characterized in that, The mathematical models stored in the transient fingerprint database include at least one physical process model selected from the following: A bubble bursting model used to characterize the dynamic process of hydrogen bubbles growing, adhering, and eventually bursting on the electrode surface; And a micro-arc model used to describe the micro-discharge phenomenon caused by abnormally high local current density.

4. The method according to claim 2, characterized in that, Also includes: In response to the event type identified by the matching and identification, targeted compensation processing is applied to the signal characteristics; When the event type is identified as bubble bursting, the compensation process is manifested as a preset correction of the phase information in the signal characteristics; When the event type is identified as a micro-arc, the compensation process involves pre-setting and adjusting the amplitude information in the signal characteristics.

5. The method according to claim 1, characterized in that, Self-learning optimization includes: Collect real-time temperature data at a predetermined set of temperature measurement points within the electrolytic cell; Based on real-time temperature data, the temperature gradient distribution characterizing the overall thermodynamic state of the electrolyzer is analyzed. Based on the temperature gradient distribution, the electrolytic cell is dynamically divided into predetermined control zones in space.

6. The method according to claim 5, characterized in that, Self-learning optimization also includes: For each segmented control region, the recursive least squares algorithm is executed independently to iteratively generate the local control parameters for that control region; Between two adjacent control regions, a preset boundary smoothing constraint is applied to the generated local control parameters to obtain smoothed local control parameters.

7. The method according to claim 6, characterized in that, Generate control parameters, including: Calculate the region fusion weights; wherein the determination of the region fusion weights is based at least in part on the reliability of the parameters of the recursive least squares algorithm or the physical coupling relationship related to the control region; By employing regional fusion weights, the smoothed local control parameters of each control region are weighted and fused to generate control parameters.

8. The method according to claim 1, characterized in that, Generate PWM control signals, including operation in at least two switchable control modes: In the default basic control mode, the basic control quantity is calculated based on the target phase and control parameters; Furthermore, when the control loop is detected to meet the preset integral saturation condition, the system switches from the basic control mode to the predictive-correction dual-mode control mode to generate a corrected control quantity.

9. The method according to claim 1, characterized in that, Phase tracking and control are achieved through a dual-loop main phase-locked loop, including: A high-bandwidth fast loop is used to quickly and dynamically track the fundamental frequency phase in the signal characteristics to obtain the fast loop phase; Additionally, a low-bandwidth slow loop is used to filter the fundamental frequency phase to extract its core trend and obtain the slow loop phase trend.

10. The method according to claim 9, characterized in that, Generate the target phase, including: Based on the fast loop phase and slow loop phase trends, the stability of the current fundamental frequency phase is evaluated, and a phase stability index is generated. The adaptive fusion weights are dynamically determined based on the phase stability index. An adaptive fusion weight is used to weight and fuse the trends of the fast loop phase and the slow loop phase to generate the target phase.

Images