Converter low-frequency ripple self-adaptive elimination method

Abstract

The invention discloses a converter low-frequency ripple self-adaptive elimination method, which comprises the following steps: acquiring and processing operation parameters of a fuel cell to estimate a physical state quantity representing internal operation characteristics of the fuel cell, and predicting future voltage ripples by using a pre-configured self-adaptive prediction model based on the physical state quantity to generate a predicted ripple track; calculating a compensation current instruction for counteracting the predicted ripple according to the predicted ripple trajectory; and generating and outputting a control signal for driving the converter based on the compensation current instruction. According to the method, the internal state of the fuel cell can be deeply observed, accurate and prospective prediction and compensation of the ripple waves are realized, and the grid-connected electric energy quality is improved.

Description

Technical Field

[0001] This invention belongs to the field of power electronic control, and in particular to an adaptive elimination method for low-frequency ripple of converters based on predictive compensation in a hydrogen power grid-connected system. Background Technology

[0002] Connecting fuel cell power generation systems to the power grid via grid-connected converters is a crucial step in realizing the large-scale consumption and utilization of hydrogen energy. However, the output characteristics of fuel cells are affected by the complex electrochemical reactions, gas transport, and hydrothermal management processes, resulting in unavoidable low-frequency ripple in their output voltage. If this ripple is injected into the grid without suppression, it will not only affect the power quality but may also induce grid resonance. Therefore, researching efficient and reliable low-frequency ripple suppression technologies for fuel cell grid-connected converters is of significant practical importance for promoting the commercialization of hydrogen power generation technology and ensuring grid security.

[0003] Currently, significant progress has been made in suppressing DC-side voltage ripple in converters. In engineering practice, a common method is passive filtering, which involves connecting large-capacity electrolytic capacitors in parallel or large inductors in series with the DC bus to passively absorb or smooth ripple through their energy storage characteristics. Another widely used technique is active suppression based on feedback control. This method detects voltage or current deviations at the output and uses traditional linear regulators such as PI (proportional-integral) controllers to adjust the converter's switching state, thereby compensating for harmonics at specific frequencies. Furthermore, some more advanced active filtering techniques first perform harmonic analysis on the ripple signal using algorithms such as Fast Fourier Transform (FFT) or Multiple Synchronous Rotating Coordinate System (MSRF) to extract the amplitude and phase information of the main harmonic components, and then generate compensation commands accordingly. These methods all play a role in dealing with harmonic problems caused by traditional power grids or loads with relatively fixed frequencies and amplitudes, forming the mainstream technological foundation in the current field of ripple suppression.

[0004] However, research has found that existing technologies still have profound technical limitations when applied to the specific scenario of hydrogen power grid-connected systems. These limitations mainly stem from insufficient understanding of the unique source-side characteristics of fuel cells, leading to inherent deficiencies in adaptability, accuracy, and foresight of existing methods. Summary of the Invention

[0005] The purpose of this invention is to provide an adaptive method for eliminating low-frequency ripple in converters, so as to solve the above-mentioned problems existing in the prior art.

[0006] Technical solution: An adaptive elimination method for low-frequency ripple in a converter, comprising:

[0007] The operating parameters of the fuel cell are collected and processed to estimate the physical state variables that characterize its internal operating characteristics. Based on the physical state variables, a pre-configured adaptive prediction model is used to predict future voltage ripple and generate a predicted ripple trajectory.

[0008] Based on the predicted ripple trajectory, calculate the compensation current command to counteract the predicted ripple.

[0009] Based on the compensation current command, control signals for driving the converter are generated and output.

[0010] Beneficial effects: This invention can provide in-depth insight into the internal state of fuel cells, enabling accurate and forward-looking prediction and compensation of ripple, thereby improving the quality of grid-connected power. Attached Figure Description

[0011] Figure 1 A flowchart illustrating the steps of an adaptive elimination method for low-frequency ripple in a converter, as provided in this application embodiment.

[0012] Figure 2 A flowchart illustrating the steps for defining a predetermined number of work areas provided in an embodiment of this application.

[0013] Figure 3 A flowchart illustrating the steps for generating a predicted ripple trajectory provided in an embodiment of this application.

[0014] Figure 4 A flowchart illustrating the steps for updating parameters in a local autoregressive integral moving average (ARIMA) model, as provided in this application embodiment. Detailed Implementation

[0015] 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.

[0016] 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.

[0017] The research revealed that existing technologies face at least the following challenges: Current methods generally treat fuel cells as a black box generating disturbances, with control models relying solely on externally measurable electrical quantities such as voltage and current, completely ignoring the fundamental physical mechanism of ripple generation—the dynamic evolution of the fuel cell's internal operating state. The output voltage ripple characteristics of a fuel cell, especially its low-frequency dynamics in the 0.1-10Hz range, are tightly coupled with its internal water management state (i.e., the water content of the proton exchange membrane). When the cell is in a water-deficient (dry) or flooded state, its output impedance and voltage stability undergo significant nonlinear changes, resulting in ripples with different characteristics. Because existing methods cannot perceive and distinguish these internal states, their fixed-parameter models or controllers have poor adaptability to the transitions in the fuel cell's internal state, making it difficult to achieve continuous and efficient ripple suppression. Furthermore, the ripple dynamics of fuel cells exhibit strong nonlinearity and non-stationarity, especially during the transition phases between different operating ranges. Even when employing multi-model switching strategies, existing technologies often experience control abrupt changes or oscillations due to hard switching mechanisms between models. The lack of a mechanism that can smoothly and continuously describe and predict the dynamic characteristics of the ambiguous region of the state boundary leads to hysteresis or overshoot in the prediction model when the operating conditions (such as load and temperature) change and the membrane water content slowly transitions from the optimal region to the dry region or the flooded region. This makes it impossible to predict and respond to the ripple morphology changes that will occur due to the state evolution, thus limiting the suppression accuracy and dynamic response speed that it can achieve.

[0018] like Figure 1 As shown, an adaptive method for eliminating low-frequency ripple in converters is proposed, comprising the following steps:

[0019] The operating parameters of the fuel cell are collected and processed to estimate the physical state variables that characterize the internal operating characteristics of the fuel cell. Based on the physical state variables, a pre-configured adaptive prediction model is used to predict future voltage ripple and generate a predicted ripple trajectory.

[0020] In this embodiment, in order to obtain the basic data for prediction, the system synchronously acquires the DC bus voltage V at the output terminal of the fuel cell stack using multi-channel synchronous signal acquisition technology. dc (t), DC bus current I dc (t), and such as the pile temperature T stack (t), hydrogen inlet pressure P H2 (t) and other operating status parameters. The collected data is interpolated and aligned to generate a synchronous data matrix with a uniform sampling rate (e.g., 1kHz). Through signal processing techniques such as adaptive moving average filters, the voltage ripple sequence v is separated from the voltage signal. r(t). Physical state quantity, preferably, is an estimate of membrane water content that reflects the internal water management state of the fuel cell. The output impedance and voltage stability of the fuel cell are closely related to its internal membrane water content. By estimating the membrane water content, the prediction model can understand and predict the ripple generation mechanism from a symptom-based rather than surface-based perspective, thereby achieving more accurate predictions. The adaptive prediction model can adopt a piecewise modeling method based on autoregressive integral moving average (ARIMA) with fuzzy boundary partitioning. This model dynamically maps the operating state of the fuel cell to different operating ranges (such as dry, optimal, and flooded) according to the membrane water content estimate, and integrates the prediction results of multiple local models for different ranges to adapt to the complex nonlinear characteristics of the fuel cell.

[0021] Based on the predicted ripple trajectory, calculate the compensation current command used to counteract the predicted ripple.

[0022] In this embodiment, after obtaining the predicted ripple trajectory V for a future period of time... pred Subsequently, the compensation current is not simply calculated inversely using Ohm's law. To ensure that the compensation process itself does not cause secondary disturbances to the system or exceed hardware capabilities, a constrained optimization strategy is adopted. Specifically, the process is constructed as a quadratic programming problem, with its objective function J comprehensively considering multiple performance indicators, such as minimizing the predicted residual ripple amplitude, minimizing the rate of change of the compensation current (to reduce the impact on the converter), and minimizing additional switching losses. Simultaneously, this optimization process is subject to strict constraints, including at least: the total output current of the converter (load current + compensation current) must not exceed its rated current; the slope of the compensation current change must not exceed the maximum value allowed by the hardware; and the modulation of the converter must be kept within a reasonable range to avoid over-modulation or under-modulation regions. By solving this optimization problem, the optimal compensation current trajectory I, satisfying all safety and performance constraints, can be obtained. compopt This trajectory represents the compensation current command.

[0023] Based on the compensation current command, control signals for driving the converter are generated and output.

[0024] In this embodiment, in order to calculate the optimal compensation current command I compoptThe system converts the current into actual hardware drive signals and superimposes them onto the original dq-axis current reference values. Specifically, in a synchronous rotating coordinate system, the optimal compensation current command is decomposed or transformed and superimposed onto the reference value of the current loop. The PI controller calculates the three-phase voltage command based on the deviation between the superimposed new reference value and the actual current feedback. To accurately synthesize this compensation component, an improved space vector pulse width modulation (SVPWM) strategy is preferred. In this strategy, in addition to synthesizing the main voltage vector calculated by the PI controller, a compensation voltage vector corresponding to the compensation current command is additionally synthesized and injected during the zero-vector action time of the SVPWM. The compensation command is precisely and efficiently modulated into the final output of six PWM control signals to drive the power switches (such as IGBTs) of the converter, thereby generating current or voltage components at the converter output that are equal in magnitude and opposite in direction to the predicted ripple, thus canceling the low-frequency ripple.

[0025] According to one aspect of this application, the physical state quantity is an estimate of the membrane water content that characterizes the water management state of the fuel cell, obtained by processing operating parameters.

[0026] Furthermore, the calculation process for the estimated membrane water content is as follows:

[0027] The current ripple sequence is isolated from the operating parameters, and the root mean square value of the current ripple sequence is calculated to quantify the intensity of the current fluctuation.

[0028] In this embodiment, the separated current ripple sequence i is obtained. r (t), and calculate its root mean square value (RMS) over a time window. i =mean(i r (t) 2 The internal impedance of a fuel cell changes under different water management conditions (especially during water shortage or flooding), resulting in varying fluctuations in its output current under the same load disturbance. Therefore, the root mean square value of the current ripple can serve as an indirect indicator of the membrane's water content.

[0029] The operating parameters, including the fuel cell stack temperature and hydrogen inlet pressure, are obtained. These operating parameters and the root mean square value are then fed into a preset empirical formula for calculation to obtain an estimated membrane water content.

[0030] To improve the accuracy of the estimation, this embodiment does not use any single parameter in isolation, but rather integrates multiple key operating parameters. In a preferred implementation, the collected reactor temperature T... stack (t) (e.g., range 40-80°C), hydrogen inlet pressure P H2 (t) (e.g., range 1-3 bar) and the calculated root mean square value (RMS) of the current ripple.i After standardization, the dimensionless reactor temperature T is obtained. norm Hydrogen inlet pressure P norm and current ripple root mean square value (RMS) norm Substituting these standardized parameters into the nonlinear empirical relationship obtained through fitting a large amount of experimental data, the water content λ of the base membrane is calculated. base An exemplary empirical relation is: λ base =14+3×tanh(0.5×T norm )-2×exp(-P norm )+4×sigmoid(RMS norm ); where tanh is the hyperbolic tangent function, exp is the natural exponential function, and sigmoid is the sigmoid function. This formula integrates the effects of temperature, pressure, and current fluctuations on membrane water content. For example, increased temperature and pressure generally help increase membrane water content, while abnormal current fluctuations may indicate water shortage or flooding. Furthermore, to capture the dynamic changes of the system, preferably, the membrane water content estimate λ(t) = λ base × f dynamic , where λ base f represents the basic membrane water content obtained through empirical formula calculation; dynamic This is a dynamic correction factor, the value of which is calculated based on the time derivative of the current ripple sequence and the rate of change of the reactor core temperature. Specifically, the dynamic correction factor f... dynamic The calculation formula can be: f dynamic = 1 + 0.1 · tanh((dRMS) i / dt ) / 10) + 0.05 · ((dT stack / dt) / 0.5). Base membrane water content λ base This reflects the steady-state or quasi-steady-state condition. However, when the system experiences transient processes such as load jumps, temperature and current change rapidly. Introducing their rate of change as a correction term can more accurately capture the instantaneous dynamics of the film content and improve the real-time performance of the estimation. Let λ... base with f dynamic Multiplying, and optionally subjecting to amplitude limiting (e.g., [0, 22]) and low-pass filtering, yields a smooth and robust estimate of membrane water content λ(t) for subsequent prediction models.

[0031] This embodiment goes beyond analyzing already occurring voltage and current signals. Instead, it transforms specific input data such as stack temperature, hydrogen pressure, and root mean square current ripple into an output—an estimated membrane water content—that accurately characterizes the internal water management state of the fuel cell. This allows the entire prediction system to anticipate impending state changes due to water shortage or flooding within the fuel cell. In hydrogen power grid-connected scenarios, when sudden load changes or environmental shifts disrupt the internal water balance of the battery, it can predict the future form and amplitude of voltage ripple before it significantly deteriorates. This provides valuable response time for compensation control, improving the timeliness and accuracy of ripple suppression under dynamic operating conditions and ensuring high-quality grid-connected power.

[0032] According to one aspect of this application, prediction is performed using an adaptive prediction model, further comprising:

[0033] Based on the estimated membrane water content, the current operating status is mapped to one of the predefined working intervals to obtain the determination of the current working interval.

[0034] The adaptive prediction model is selected or configured based on the attribution determination to perform voltage ripple prediction.

[0035] In this embodiment, the mapping is not a rigid, either-or division, but a flexible, fuzzy logic-based membership assignment process. Specifically, the system predefines multiple operating intervals directly corresponding to the fuel cell's water management state, such as a dry zone, an optimal zone, and a flooded zone. The membership determination does not select a single interval, but rather calculates the membership degree of the current membrane water content λ(t) to these three intervals, obtaining a fuzzy membership vector. A single linear model is insufficient to capture the complex nonlinear dynamic characteristics of the fuel cell throughout its operating domain. By decomposing the complex nonlinear problem into a quasi-linear problem within multiple local operating intervals and configuring a dedicated prediction model for each interval, the accuracy and robustness of the prediction can be improved.

[0036] Furthermore, such as Figure 2 As shown, the process of defining a predetermined work area includes:

[0037] Historical estimates of membrane water content are compiled to form a historical data sequence;

[0038] K-means clustering analysis was performed on the historical data sequence to identify data clusters belonging to different water management states;

[0039] Based on the analysis results of the data clusters, the statistical characteristic parameters of each working interval are extracted to complete the adaptive definition of the working interval.

[0040] In this embodiment, all estimated membrane water content values ​​λ(t) over a relatively long period (e.g., 24 hours) are read and stored, forming a historical sequence containing tens of thousands of data points. A K-means clustering algorithm is applied to this historical data sequence, with the number of clusters K preset to 3, corresponding to the three states of dryness, optimal conditions, and flooding, respectively. The K-means algorithm automatically divides the data points into three clusters, thereby identifying data clusters representing different water management states. After clustering, statistical analysis is performed on each data cluster to calculate the cluster center λ for each cluster. center (i.e., mean) and standard deviation σ. These statistical parameters are the set P. stat ={λ center σ} constitutes a quantitative description of the intrinsic characteristics of these three operating intervals. For example, the center of the dry zone might be around 2.5, the center of the optimal zone around 9.5, and the center of the flooded zone around 18. This embodiment does not rely on fixed, theoretical thresholds to divide the intervals, but rather learns its unique operating interval distribution from the historical operating data of the fuel cell itself. This allows the model to automatically adapt to different stack aging levels, manufacturing differences, or specific operating environments, improving the model's personalization and accuracy. Optionally, in some implementations, other clustering algorithms, such as Gaussian Mixture Model (GMM) or DBSCAN, can be used instead of the K-means algorithm.

[0041] like Figure 3 As shown, according to one aspect of this application, generating a predicted ripple trajectory includes:

[0042] Configure a local autoregressive integral moving average (ARIMA) model for each working interval defined by statistical characteristic parameters;

[0043] Based on the statistical characteristic parameters, a Gaussian membership function is constructed. Using the Gaussian membership function and the current membrane water content estimate, the fuzzy membership degree of the statistical characteristic parameters to each working interval is calculated to obtain the fuzzy membership vector.

[0044] The autoregressive integral moving average (ARIMA) models of each locality are driven to generate independent predicted values, and fuzzy membership vectors are used to weight and fuse the predicted values ​​to generate predicted ripple trajectories.

[0045] In this embodiment, K-means clustering analysis is performed on the historical membrane water content estimation sequence to reveal the data distribution pattern and extract statistical characteristic parameters for multiple working intervals. Based on the extracted statistical characteristic parameters, a unique Gaussian membership function is constructed for each working interval, thereby establishing a nonlinear mapping relationship between the membrane water content estimate and the fuzzy membership degree. Using this nonlinear mapping relationship and the current membrane water content estimate, the fuzzy membership degree vector is calculated in real time. Specifically, local ARIMA models are configured for the three working intervals (dry, optimal, and flooded). For example, based on historical data analysis, the data in the dry area may be more suitable for the ARIMA(2,1,1) model, the optimal area for the ARIMA(2,0,2) model, and the flooded area for the ARIMA(1,1,2) model. The order (p, d, q) of the model is preferably automatically determined during offline training or periodic updates using the Akaike Information Content Criterion (AIC) to seek the best fitting effect. A Gaussian membership function is constructed based on the statistical characteristic parameters. Its calculation formula is: μ i (λ) = exp(-((λ- c i ) / σ i_fuzzy ) 2 ); where μ i (λ) represents the fuzzy membership degree of the current membrane water content λ with respect to the i-th working interval, with a value range of [0, 1]; λ is the estimated value of the current membrane water content; c i The center point of the i-th membership function is determined by the cluster center λ of the corresponding interval (e.g., c). dry =λ centerdry ); σ i_fuzzy Let be the width parameter of the i-th membership function, whose value can be obtained from the statistical standard deviation σ of the corresponding interval. i Multiplying by an amplification factor (e.g., 1.2 or 1.5) ensures a smooth transition between intervals. After calculating the three original membership values, they need to be normalized so that their sum is always 1. Optionally, to avoid drastic jumps in membership values ​​caused by small fluctuations in the λ value, a time-pass filter can be applied to the normalized membership vector to make its changes smoother. At any time t, the three local ARIMA models are driven to generate their respective independent predictions y* for time t+1 based on the most recent voltage ripple history sequence. dry (t+1), y* opt (t+1) and y* flood (t+1). Using the (preferably smoothed) fuzzy membership vector μ(t) at this moment, the three independent predicted values ​​are weighted and fused, and the calculation formula is as follows: y* fuzzy (t+1) = μ dry (t) · y* dry(t+1) + μ opt (t) · y* opt (t+1) + μ flood (t) · y* flood (t+1); where y* fuzzy (t+1) is the final fused prediction value at time t+1, constituting a data point for the predicted ripple trajectory; μ dry (t), μ opt (t), μ flood (t) represents the membership degree of the three intervals at time t. Soft switching between different local models is achieved. Compared to hard switching, which causes abrupt changes at the boundaries of the working interval, this ensures that the final output predicted trajectory is continuous and stable, effectively avoiding unnecessary control disturbances caused by model switching.

[0046] like Figure 4 As shown, according to one aspect of this application, it also includes updating the parameters in the local autoregressive integral moving average (ARIMA) model, specifically:

[0047] The pre-stored actual voltage ripple sequence is continuously compared with the predicted values ​​generated by the autoregressive integral moving average (ARIMA) model to obtain the prediction error sequence.

[0048] The parameters of the autoregressive integral moving average (ARIMA) model are dynamically adjusted based on the prediction error sequence.

[0049] In a preferred implementation, the online update process for this parameter is triggered once every control cycle (e.g., every 100 milliseconds), and is specifically implemented as follows: Obtain the actual voltage ripple value v at the current time t. r (t), and compare it with the model's prediction of time t from the previous time t-1. fuzzy By comparing (t) with the current prediction error e(t) = v, we can calculate the current prediction error e(t) = v r (t)- y* fuzzy (t); where y* fuzzy (t) is the final predicted value after fuzzy weighted fusion. To reasonably distribute the total prediction error among the three different local ARIMA models, the system uses the fuzzy membership vector μ(t-1) from the previous time step t-1 as weights. Specifically, the error components assigned to the dry zone, optimal zone, and flooded zone models are et, et, and et, respectively. dry = μ dry (t-1) *e(t); e opt = μ opt (t-1) * e(t); e flood = μ flood(t-1) * e(t). Weighted allocation based on membership degrees ensures that the local model contributing the most to the current prediction bears the largest share of the correction responsibility, making parameter updates more targeted and efficient. For each local model, its parameter θ... i The updates to the vector containing the coefficients of the ARIMA model are all implemented using the Recursive Least Squares (RLS) algorithm. The core step of this algorithm is to calculate the Kalman gain K. i (t), its formula is: K i (t) = P i (t-1) * Φ i (t) / (Φ i (t)ᵀ * P i (t-1) * Φ i (t) + R i ); where K i (t) is the Kalman gain matrix of the i-th local model at time t; P i (t-1) is the parameter covariance matrix at time t-1, which represents the uncertainty of the current parameter estimate; Φ i (t) is the regression vector at time t, composed of the inputs used for prediction, such as several recent actual voltage ripple values ​​and historical prediction error values; R i θ represents the preset observation noise variance, reflecting the intensity of the noise in the measurement system itself; ᵀ represents the transpose. After obtaining the Kalman gain, the model parameters can be recursively updated: θ i (t) = θ i (t-1) + K i (t) * e i ;where θ i (t) represents the parameter set after the i-th local model update; θ i (t-1) represents the parameter set before the update; e i Let be the error component belonging to the i-th model. Optionally, to enable the model to better track time-varying characteristics and gradually forget outdated historical data, a forgetting factor λ can be introduced into the update of the covariance matrix. f (A number slightly less than 1, such as 0.98).

[0050] To ensure the stability and robustness of the online update process, after calculating the new parameter set θ... i After (t), the updated model is not immediately adopted, but its effectiveness is first verified. This verification includes at least a stability test, which calculates the eigenvalues ​​of the updated ARIMA model such that their moduli are all less than a safety threshold (e.g., 0.95) to avoid updating the model parameters to an unstable region. If an unstable tendency is detected, the update is abandoned, and the model is rolled back to the parameters θ of the previous time step. i(t-1), and optionally temporarily reduce the Kalman gain of the model (e.g., multiply by 0.5) to adopt a more conservative step size in subsequent updates. Furthermore, the rate of change of parameters can be monitored; if a single update causes excessive parameter changes (e.g., exceeding 20%), the update magnitude is limited to prevent drastic parameter fluctuations due to noise or outliers. Only parameters that pass all validity validations are ultimately adopted and used for subsequent prediction calculations.

[0051] According to one aspect of this application, the method further includes updating the center point and width parameters of the Gaussian membership function through an adaptive adjustment loop, specifically:

[0052] Receive the residual ripple sequence at the converter output;

[0053] Based on the residual ripple sequence, the gradient of the performance index with respect to the center point and width parameter is estimated, and a performance gradient vector is generated.

[0054] The center point and width parameters are corrected using the performance gradient vector.

[0055] In this embodiment, the input to the regulation loop is the residual voltage ripple sequence v actually measured at the converter output after the compensation strategy has been applied. r_res (t). The system calculates the comprehensive performance index J based on this sequence. For example, this index could be the root mean square value (RMS) of the residual ripple over a period of time. r_res Alternatively, it could be a weighted function incorporating the ripple suppression ratio (SR), such as J = w1 * RMS(v r_res ) + w2 * (20 - SR), where w1 and w2 are weighting coefficients. A smaller J value indicates better ripple suppression. To determine the center point c of the membership function... i and width σ i Optimization requires knowing the gradient of J with respect to these parameters. The finite difference method is preferred for estimating this gradient online. For example, to estimate ΨJ / Ψc... i The system can be slightly perturbed by c. i The value of J(c) (i.e., calculate J(c) separately) i + Δc) and J(c i - Δc), then through the formula (J(c) i + Δc) - J(c i The gradient value is approximated by -Δc)) / (2 * Δc). This process is repeated for all parameters that need adjustment to obtain the performance gradient vector. Here, Ψ is the partial derivative, and Δc is the small perturbation in the finite difference method.

[0056] Furthermore, adjustments are made to the center point and width parameters, including calculating the adaptive learning rate, specifically:

[0057] The learning rate is adaptively adjusted by monitoring the direction of the performance gradient vector over time. When consecutive gradients are in the same direction, the learning rate is increased; when consecutive gradients are in the opposite direction, the learning rate is decreased. For example, if the system detects three consecutive gradients in the same direction, it means the optimization direction is correct, and the learning rate can be increased by α(t) = min(1.2 * α(t-1), α... max If the gradient direction is reversed, it means that oscillations may have occurred near the optimal value. In this case, the learning rate α(t) = 0.5 * α(t-1) should be reduced to make the search step size more precise. Here, α is the learning rate. max Maximum learning rate.

[0058] The center point and width parameters are updated using gradient descent with an adaptive learning rate and the current performance gradient vector. Specifically, the update formula is: c i (t+1) = c i (t) - α(t) * (ΨJ / Ψc i ), and σ i (t+1) = σ i (t) -α(t) * (ΨJ / Ψσ i ).

[0059] Furthermore, to ensure that the updated parameters have a reasonable physical meaning and actually bring performance improvements, gradient descent updates include constraint handling and validity verification of the updated parameters, specifically:

[0060] A preset feasible region constraint is applied to the parameters updated by gradient descent, ensuring that the values ​​of the center point and width parameters fall within a preset range. If the center point or width parameters exceed the boundary, they are projected back to the boundary of the feasible region constraint to obtain the parameters after constraint processing. For example, the center point c of the drying area can be set. dry The c in the optimal region must be between [1, 4]. opt Between [7, 12], and all width parameters σ i The value must be between [1, 5]. If the updated value is outside this range, it will be forced to be set to the nearest boundary value.

[0061] The preset test sample set is re-predicted using the constrained parameters to obtain the prediction mean square error. The current constrained parameters are confirmed to be valid only if the current prediction mean square error is less than the previous prediction mean square error by a predetermined threshold; otherwise, they are revoked.

[0062] In this embodiment, the parameter update is only confirmed if the mean square error of the prediction brought by the new parameters is less than the error under the old parameters reaching a predetermined threshold; otherwise, it is revoked. Specifically, after obtaining the new parameter set Φ... new Then, the system can use it to perform retrospective predictions on a small recent segment (e.g., 10) of historical samples, calculating its mean squared error (MSE). new and using the old parameter Φ old Calculated MSE old Comparison. Only when MSE new Compared to MSE old When there is sufficient improvement (e.g., MSE) new < 0.95 * MSE old Only updates that show an improvement of more than 5% are ultimately confirmed and adopted; otherwise, the update is considered invalid and withdrawn, and the system will continue to use the old parameters. The entire prediction model can not only adapt to the short-term dynamic changes of fuel cells, but also optimize its most fundamental operating condition classification logic from the perspective of long-term operating effects, achieving a higher-dimensional adaptive and self-learning capability.

[0063] This embodiment calculates the fuzzy membership vector of the current state to each interval using a Gaussian membership function, and then uses this vector as weights to perform real-time weighted fusion of the prediction results from multiple professional local ARIMA models. By combining abstract fuzzy rules with specific prediction algorithms, the continuity and smoothness of the model output are achieved. In scenarios where the fuel cell state slowly evolves from the optimal region to the flooded region, the prediction results can smoothly transition from the optimal region model to the flooded region model, effectively avoiding prediction errors caused by model switching lag. This ensures the continuity and stability of the compensation current, ultimately suppressing the distortion of the grid-connected current and improving the dynamic response quality of the system.

[0064] According to another aspect of this application, generating the predicted ripple trajectory can also be:

[0065] The predicted values ​​within the future time window are generated point by point in an iterative manner. Each iteration takes the prediction result of the previous step as input to call the adaptive prediction model. In each iteration step, the prediction uncertainty is accumulated and calculated synchronously. It is determined whether the accumulated uncertainty exceeds a preset threshold. If it does, a decay factor is applied to the prediction value generated in the current step for correction. The predicted ripple trajectory is formed by the predicted values ​​that have been corrected in all steps.

[0066] In this embodiment, in order to predict the ripple over a future period (e.g., the next 10 milliseconds, corresponding to 10 sampling points), the system performs the following operations: It retrieves the ripple from the stored voltage ripple sequence v... rIn (t), the most recent historical data segment is extracted as the basis for prediction, such as the most recent 100 sampling points. To improve the robustness of the prediction, this data segment is preferably preprocessed, including: detecting and removing outliers exceeding three times the standard deviation, and using methods such as cubic spline interpolation to fill in any missing data; and performing detrending processing on the data to meet the ARIMA model's requirement for data stationarity. The adaptive ripple prediction model M is then invoked using the historical data window. fuzzy The prediction function is used to obtain the predicted value y*(t+1) and its prediction uncertainty σ(t+1) for the first sampling point t+1 in the future. This predicted value y*(t+1) is added to the end of the data window as pseudo-historical data, and the old data point at the front of the window is removed to form a new input sequence, which is used to predict the ripple value y*(t+2) and uncertainty σ(t+2) for the second sampling point t+2. This process is repeated K times to obtain the prediction sequence for the next K points. A crucial step in this process is the synchronous accumulation of prediction uncertainty. At the k-th prediction step, the total accumulated uncertainty can be calculated as σ. total (k) = sqrt(Σ(σ(i)²)), where i ranges from 1 to k. Prediction error accumulates with increasing prediction step size. The reliability of long-term predictions can be assessed by explicitly quantifying the accumulated uncertainty. A threshold for accumulated uncertainty is set (e.g., 2.0V). During iterative prediction, if σ(i)² at step k... total If (k) exceeds this threshold, it indicates that the reliability of the prediction results from this step onwards decreases. At this point, to prevent the controller from overcompensating based on unreliable predictions, the system applies a decay factor to the prediction value generated in the current step for correction. For example, the corrected prediction value y* adj (t+k) = α k * y*(t+k) + (1 - α k ) * mean(v pre ), where α k It is the decay factor that increases with k (such as α) k = exp(-0.1 * k)), mean(v pre The mean of the historical data window is used. This allows long-term, unreliable predictions to regress to a conservative benchmark (such as zero or the mean), achieving a graceful degradation of the prediction information. The K-point prediction sequence generated by the above steps may contain undesirable jumps due to iterative calculations. Therefore, preferably, a smoothing filter such as Savitzky-Golay is applied before output to obtain the final, smooth prediction ripple trajectory V. pred .

[0067] This embodiment links abstract prediction confidence with specific mathematical operations and control behaviors. In the actual operation of a hydrogen power plant connected to the grid, the control system can be aware of the boundaries of its predictive capabilities. When highly confident in near-term (e.g., 1-3 milliseconds) predictions, it performs precise compensation; however, when less confident in long-term (e.g., 8-10 milliseconds) predictions, it automatically reverts to a more conservative compensation strategy. This enhances the robustness and safety of the entire predictive compensation system under complex operating conditions, avoiding new system disturbances caused by overconfident and erroneous predictions.

[0068] According to one aspect of this application, the compensation current command, i.e. the optimal compensation current trajectory, is calculated by constructing and solving a quadratic programming problem; wherein the objective function of the quadratic programming problem is to minimize the comprehensive performance index, which at least weights and integrates the predicted residual ripple amplitude and the rate of change of the compensation current; the constraints at least include the amplitude constraint that the total output current does not exceed the rated current, and the dynamic constraint that the rate of change of the compensation current does not exceed the slope limit.

[0069] In this embodiment, the ideal compensation current may be unrealistic or prohibitively expensive in practice. Therefore, an optimal balance must be sought between the ripple suppression effect and the system's costs and risks. Specifically, the quadratic programming (QP) problem is constructed as the standard QP form min(0.5 * x T Hx + f T x); where the decision variable x is the compensation current i at the future K sampling points. comp The vector is composed of (t+k). The terms of the objective function J are integrated into the Hessian matrix H and the linear term f. For example, the minimization of the residual ripple term in the objective function (i.e., w1 * ||V) pred + Z out * i comp || 2 ) and minimizing the rate of change of the compensation current (i.e., w2 * ||di) comp / dt|| 2 All of these can be expressed as quadratic forms with respect to the decision variable x. The physical and operational constraints of the system are transformed into linear inequality forms Ax ≤ b that can be recognized by the QP solver, where A is the linear inequality constraint matrix, x is the decision variable vector, and b is the linear inequality constraint vector. Amplitude constraint: |I load (t) + i comp (t+k)| ≤ 0.95 *I rated This constraint ensures that the total output current of the converter will never exceed its rated safe range at any time; where I load (t) represents the system load current at the current moment, I rated This represents the rated current value of the converter. Dynamic constraint: |icomp (t+k) - i comp (t+k-1)| / Δt≤50A / ms, this constraint limits the rate of change of the compensation current to protect the power switching devices from excessive current surges; where Δt is the sampling period of the control system. Modulation constraint: m = V out / (V dc / 2)∈[0.1,0.9], this constraint ensures that the converter operates in the linear modulation region; where V out This is the output voltage of the converter. It is connected through the converter's output impedance Z. out This voltage constraint can be converted into a constraint on the compensation current i comp The constraints are significant. In real-time control systems, this QP problem must be solved quickly. Optional solution algorithms include interior-point methods or active-set methods, which can provide optimal solutions in milliseconds.

[0070] Furthermore, the optimal compensation current trajectory obtained from solving the quadratic programming problem also includes a safety check of the compensation strategy before execution, specifically:

[0071] The evaluation should assess whether the optimal compensation current trajectory will cause system instability. The evaluation should include at least the following: whether the trajectory will cause the DC bus voltage to drop below the safety threshold, and whether it may cause false triggering of the protection device.

[0072] When a risk is detected, the preset compensation gain is actively reduced, and the calculation of the compensation current requirement is returned to the calculation step based on the reduced compensation gain for recalculation until the generated trajectory passes the safety verification.

[0073] The verified compensation current trajectory will be output as a safety compensation current command.

[0074] In this embodiment, after obtaining the optimal compensation current trajectory I compopt Subsequently, the system uses a simplified system model to simulate the potential impact of this current trajectory. For example, it assesses whether the current trajectory will cause a transient voltage drop exceeding 5% on the DC bus, or whether its high-frequency components might cross-link with the resonant frequencies of other parts of the system, causing oscillations or triggering hardware protection. If the pre-assessment detects a risk, the system will not directly abandon compensation but will instead adopt an iterative risk avoidance strategy. The compensation gain G used in QP optimization will be... comp Actively reduce by a fixed percentage (e.g., G) compnew = 0.8 *G compoldThen, return to the step of calculating the ideal compensation current and perform QP optimization again. This cycle of reducing the objective, re-optimizing, and re-verifying will be repeated until a compensation current trajectory that is both QP-optimal and passes the final safety check is generated. This generates a smart and safe compensation current command I. comp_safe This provides a solid foundation for achieving efficient and stable ripple suppression.

[0075] This embodiment constructs a multi-objective, multi-constraint quadratic programming (QP) optimization problem and supplements it with final security verification. This ensures that the generated compensation instructions not only achieve optimal ripple suppression but also meet the stringent requirements of hardware security and stable operation in engineering practice. It solves the practical problems that traditional compensation methods may suffer from inconsistencies or even threaten system security. This guarantees that the final output compensation instructions can efficiently complete the compensation task without violating any system security red lines, thus improving the overall practical usability and engineering reliability of the solution.

[0076] According to one aspect of this application, generating a control signal includes:

[0077] The compensation current command is superimposed on the original dq axis current reference value, and the three-phase voltage command is calculated by the PI controller based on the deviation between the superimposed current reference value and the actual current feedback.

[0078] SVPWM is executed to synthesize three-phase voltage commands, and during the zero-vector action time of SVPWM, an additional compensation voltage vector corresponding to the compensation current command is synthesized and injected to generate control signals.

[0079] This embodiment is preferably implemented within a vector control framework based on a synchronously rotating dq coordinate system. Specifically, the safety compensation current command I, represented in the time domain and containing K future sampling points, is... comp_safe By transforming to the dq synchronous rotating coordinate system using the Park transformation, the d-axis component i of the compensation current is obtained. d_comp and q-axis component i q_comp Simultaneously, the system has a raw dq-axis current reference value I generated by an outer loop (such as a power loop or voltage loop) for controlling active and reactive power. d_ref_orig and I q_ref_orig The compensation component is superimposed on the original reference value to form the corrected current reference value: I d_refnew =I d_ref_orig + i d_comp , and I q_refnew = I q_ref_orig + i q_compSuperimposing the compensation command onto the current loop reference value allows full utilization of the carefully tuned, high-bandwidth, and dynamically responsive inner current loop in the existing control system. This enables rapid and accurate tracking of the compensation current without requiring a completely independent, parallel compensation control channel, thus simplifying the control system structure. The corrected current reference value I d_refnew and I q_refnew It is then fed into the PI controller and compared with the actual feedback current I obtained through coordinate transformation. d_actual and I q_actual By comparing the values, the PI controller outputs the required dq-axis voltage command V based on the deviation. d_cmd and V q_cmd After receiving the dq-axis voltage command, the reference voltage vector V in the two-phase stationary coordinate system is obtained through inverse Park transformation. ref By utilizing space vector pulse width modulation (SVPWM) technology, the switching on and off of six power switches in a three-phase bridge is controlled to achieve the desired effect in each switching cycle T. s V was synthesized on an average basis within the interior. ref Standard SVPWM technology synthesizes V0 within one switching cycle using two adjacent effective voltage vectors (e.g., V1, V2) and a zero vector (V0 or V7). ref The zero vector's duration T0 is typically used to provide a switching dead zone, or it is evenly distributed at the beginning and end of the cycle. This embodiment does not allow the converter to be completely inactive in a zero state during the T0 time period; instead, during this time, an additional compensation voltage vector corresponding to the compensation current command is synthesized and injected. Specifically, the compensation current command I... comp_safe By using the dynamic model of the converter, the required small, instantaneous compensation voltage vector V is calculated. comp Only then can it be generated. During modulation, the SVPWM algorithm divides the T0 time period into a series of extremely short combinations of different effective vectors, such that the average voltage effect of these additional switching actions over the entire T0 time period is exactly equivalent to the desired compensation voltage vector V. compThe zero-vector time injection strategy decouples the main voltage vector synthesis used to maintain basic power output from the compensation voltage vector synthesis used to eliminate high-frequency ripple in time. The compensation behavior is cleverly embedded into the intermittent periods of main power control, avoiding interference with the fundamental voltage. Simultaneously, it provides a fast track for the compensation signal to bypass the bandwidth limitations of traditional PI controllers and be directly injected into the PWM modulation stage, enabling the system to respond to and compensate for higher-frequency ripple components. No hardware modifications to the converter main circuit are required; implementation is achieved solely through software modifications to the SVPWM algorithm, making it practically applicable in engineering. The generated six PWM pulse signals are sent to the drive circuit to precisely control the switching action of the converter, thereby generating a current at the output that cancels out the predicted ripple, completing the closed loop of the adaptive elimination process.

[0080] In summary, this application discloses an adaptive method for eliminating low-frequency ripple in a converter, comprising: collecting and processing operating parameters of the fuel cell to estimate the membrane water content, which characterizes its internal water management state; based on the membrane water content, predicting future voltage ripple using an adaptive prediction model to generate a predicted ripple trajectory, wherein the adaptive prediction model weights and fuses multiple local ARIMA models corresponding to different operating ranges based on the fuzzy membership degrees calculated by the membrane water content; calculating the optimal compensation current command by constructing and solving a quadratic programming problem based on the predicted ripple trajectory and the dynamic operating constraints of the converter; and generating and outputting a control signal for driving the converter based on the optimal compensation current command to actively cancel low-frequency ripple. This method can provide in-depth insight into the internal state of the fuel cell, achieving accurate and forward-looking prediction and compensation of ripple, and improving the quality of grid-connected power.

[0081] In a specific embodiment, assuming that the system has completed all preparatory work during a certain control period t, the specific input conditions are as follows: the estimated membrane water content at the current moment is λ(t) = 8.0. Through historical data clustering, the system has determined the Gaussian membership function parameters for three working intervals: Dry zone: center point c dry = 2.5, width parameter σ dry_fuzzy = 2.0; Optimal region: center point c opt = 9.5, width parameter σ opt_fuzzy = 3.0; Flooded area: center point c flood = 18.0, width parameter σ flood_fuzzy = 2.0. Three parallel local ARIMA models, based on their respective historical data and current state, have independently provided voltage ripple predictions for the next time step t+1: Dry zone model prediction: y* dry (t+1) = 0.50 V; Optimal region model prediction: y* opt(t+1) = -0.20 V; Flooded area model prediction: y* flood (t+1) = 0.80 V. Substituting the current membrane water content λ = 8.0 into the three Gaussian membership function formulas μ... i (λ) = exp(-((λ -c i ) / σ i_fuzzy )2) In the dry zone: original membership degree: μ dry_raw = exp(-((8.0 - 2.5) / 2.0) 2 ) = exp(-(2.75) 2 ) = exp(-7.5625) ≈ 0.00052; Original membership degree of the optimal region: μ opt_raw = exp(-((8.0 - 9.5) / 3.0) 2 ) = exp(-(-0.5) 2 ) = exp(-0.25) ≈ 0.77880; Initial membership degree of the flooded area: μ flood_raw =exp(-((8.0 - 18.0) / 2.0) 2 ) = exp(-(-5.0) 2 = exp(-25) ≈ 1.389e -11 (This value is extremely small, close to 0). Calculate the sum: S = μ dry_raw + μ opt_raw + μ flood_raw ≈ 0.00052 + 0.77880 + 0 ≈0.77932. Calculate the normalized membership degree: μ dry_norm = μ dry_raw / S ≈ 0.00052 / 0.77932 ≈ 0.00067; μ opt_norm = μ opt_raw / S ≈ 0.77880 / 0.77932 ≈ 0.99933; μ flood_norm = μ flood_raw / S ≈ 0 / 0.77932 ≈ 0. This yields the normalized membership vector μ. norm (t)≈[0.00067, 0.99933, 0]. It can be seen that the current state almost entirely belongs to the optimal region. Using the normalized membership degrees, the independent predictions of the three local models are weighted and summed to obtain the final fused prediction: y* fuzzy (t+1) = μ dry_norm · y* dry (t+1) +μ opt_norm · y*opt (t+1) + μ flood_norm · y* flood (t+1)≈ 0.00067 · (0.50 V) + 0.99933 · (-0.20 V) + 0 · (0.80 V)≈ 0.000335 V - 0.199866 V + 0 V≈ -0.1995 V. Through the above calculations, the adaptive prediction model gives a voltage ripple prediction of approximately -0.1995 V for time t+1. This value is very close to the prediction of the optimal region model, which is -0.20 V. This is also very close to the current membrane water content λ=8.0, which is extremely close to the center of the optimal region. opt The prediction of -0.1995 V is consistent with the physical fact, validating the rationality of the fuzzy fusion mechanism. This predicted value will then be used as the predicted ripple trajectory V. pred One of the key data points is sent to the compensation strategy optimization module.

[0082] Optionally, an adaptive method for eliminating low-frequency ripple in a converter may further include: acquiring the operating parameters of the fuel cell, separating the voltage ripple sequence from the operating parameters, and extracting operating parameters characterizing the current operating condition; estimating the internal physical state variables of the fuel cell based on the operating parameters, and constructing or adjusting an adaptive ripple prediction model based on these internal physical state variables; generating a predicted ripple trajectory using the adaptive ripple prediction model and historical voltage ripple sequences; solving for the optimal compensation current trajectory based on the predicted ripple trajectory and the preset operating constraints of the converter; superimposing the optimal compensation current trajectory onto the original current reference value of the converter to generate a corrected current reference value; and generating a pulse width modulation (PWM) control signal for driving the converter based on the corrected current reference value.

[0083] In one embodiment of this application, before predicting future voltage ripple, an adaptive dual-window analysis strategy can more accurately separate and quantify the characteristic components at different time scales from the complex voltage ripple signal, providing higher-quality input for subsequent prediction models. Specifically, during the operation of a fuel cell, its voltage ripple typically contains two typical components: low-frequency fluctuations of 0.1-1Hz related to slow dynamic processes such as gas and water, and mid-frequency oscillations of 1-10Hz related to the operation of auxiliary systems (such as air compressors) and electrochemical reactions. The optimal analysis methods for these two components differ; low-frequency analysis requires a long time window to ensure frequency resolution, while mid-frequency analysis requires a short time window to ensure time resolution. To resolve this contradiction, the following method is adopted: The input voltage ripple sequence v... r(t) Perform rapid instantaneous frequency prediction. Optionally, this prediction can be achieved by analyzing the zero-crossing rate of the signal or by using methods such as Hilbert transform. Based on the predicted dominant frequency range, dynamically determine the length of a pair of long and short analysis windows. For example, when the predicted dominant frequency is below 1Hz, the system determines that the current ripple is dominated by low-frequency fluctuations. In this case, in order to finely distinguish the specific low-frequency points, a longer analysis window, such as W, is used. long = 2048 sampling points (corresponding to 2.048 seconds at a 1kHz sampling rate); simultaneously used with a short window, such as W, to capture potential mid-frequency disturbances. short = 128 points (128 milliseconds). Correspondingly, when the estimated main frequency is higher than 5Hz, the system determines that the intermediate frequency oscillation is dominant. In this case, to accurately capture the timing of the oscillation, a shorter analysis window is used, such as W... short = 64 points (64 milliseconds), paired with a moderately long window, such as W long = 1024 points (1.024 seconds).

[0084] After selecting the window parameter pair, the system processes the voltage ripple sequence v. r (t) Perform two STFT calculations in parallel. To reduce spectral leakage caused by window truncation, a smooth window function, such as the Hanning window, is preferred in the calculation. Long window STFT: Calculations are performed with a high overlap ratio (e.g., 90%). A high overlap ratio means a larger computational cost, but it allows for very fine-step sliding in time, without missing any small changes in low-frequency details, ultimately resulting in a spectral matrix S with high frequency resolution (e.g., approximately 0.5 Hz). long (f, t). Short-window STFT: Calculations are performed using a lower overlap ratio (e.g., 50%) to balance time resolution and computational load, ultimately resulting in a spectral matrix S with high time resolution (e.g., 64 milliseconds). short (f, t). After obtaining the two-dimensional spectrum matrix, the system performs energy analysis within the preset frequency bands. In the long window spectrum matrix S... long In the middle, calculate the total energy density E in the low-frequency band (0.1-1Hz). low And identify the frequency f corresponding to the energy peak. low and amplitude A low In the short window spectrum matrix S short In the middle, calculate the total energy density E in the mid-frequency band (1-10Hz). mid And identify the peak frequency f mid and amplitude A mid The system further compares the energy ratio E. low / E midIt determines the dominant mode of the current ripple and outputs the characteristics (amplitude and frequency) of one or two identified dominant components. It can see the texture details of low-frequency fluctuations and capture the instantaneous dynamics of mid-frequency oscillations, thus providing the prediction model with much richer and more accurate input information than traditional single-window FFT.

[0085] In an optional embodiment, when the operating state of the fuel cell (i.e., the membrane water content λ(t)) rapidly crosses the boundaries of different operating ranges, the standard fuzzy fusion model may exhibit prediction lag due to its quasi-static characteristics. Therefore, a dynamic compensation term is introduced during the fuzzy weighted model fusion process. Specifically, the system continuously monitors the current membrane water content estimate λ(t). This compensation mechanism is activated when λ(t) enters any preset boundary region. The boundary region can be defined as a certain range on both sides of the center point of the membership function, for example, |λ(t) - c i | < 1.5 * σ i , where c i and σ i It represents the center point and width of the i-th membership function. Once the system is detected to be in a boundary transition state, the system calculates the dynamic compensation term Δy. trans The preferred calculation formula is: Δy trans = k trans * (dλ / dt) * (y* i - y* j ); where Δy trans The calculated boundary transition compensation term is expressed in volts (V); k trans is a dimensionless compensation coefficient, the value of which is calibrated experimentally, for example, it can be set to 0.05; dλ / dt is the time derivative of the estimated membrane water content, i.e., the rate of change of λ(t), which quantifies the speed and direction of the state crossing the boundary; y* i and y* j This represents the independent predictions of two local ARIMA models adjacent to the current boundary region where λ(t) is located. For example, if λ(t) is transitioning from an optimal region to a dry region, then y* i and y* j These are the predicted values ​​from the models for the optimal and dry regions, respectively. The calculated dynamic compensation term Δy... trans It is directly added to the fuzzy fusion prediction value y* fuzzy Above this, we obtain the final compensated predicted value y* comp = y* fuzzy + Δy trans .

[0086] This embodiment introduces information about the state velocity (dλ / dt) into the original fuzzy model based on the state position (i.e., the value of λ). When the fuel cell state is stable or changes slowly, dλ / dt is close to zero, and the compensation term also approaches zero, having no impact on the system. However, when a sudden change in operating conditions occurs and λ(t) changes rapidly, the value of dλ / dt becomes large, and the compensation term Δy becomes significant. trans It can adjust the prediction results in advance based on the direction and speed of state changes, as well as the difference between the model prediction values ​​of the target interval and the source interval. This can effectively compensate for prediction lag or overshoot that may be caused by untimely model switching, and improve the prediction accuracy of the model in dynamic processes.

[0087] In a specific implementation, after the control signal is generated, a global monitoring and protection mechanism runs in the background. Specifically, the system not only focuses on ripple-related signals but also continuously monitors key global parameters that reflect the overall health of the grid-connected system. These parameters preferably include: DC bus voltage fluctuation: monitoring whether the DC bus voltage exceeds ±10% of its nominal value; excessive bus voltage fluctuation is a core indicator of system instability; total harmonic distortion (THD) of the grid-connected current: monitoring whether the THD of the grid-connected current exceeds the upper limit specified by the grid standard (e.g., 5%); excessive THD not only pollutes the grid but may also indicate oscillations in the control system; converter temperature rise: real-time monitoring of the temperature of key power devices (such as IGBT modules) in the converter using temperature sensors to ensure that it does not exceed the safe operating upper limit (e.g., 85°C). When any of the above-monitored parameters exceeds its preset safety threshold, the protection mechanism is triggered. Unlike traditional direct tripping or shutdown, this embodiment adopts a flexible exit or graceful degradation strategy. Specifically, the system gradually and smoothly reduces the compensation gain G. comp For example, every control cycle, G will be... comp Multiply by a coefficient less than 1 (e.g., 0.9) until it decreases to 0. When the compensation gain G... comp Once reduced to zero, the entire predictive compensation-based adaptive ripple cancellation function is safely and temporarily suspended. At this point, the converter's control system automatically and seamlessly switches back to its basic, conventional control mode that does not include any active ripple compensation (e.g., standard dq decoupled PI control).

[0088] This embodiment ensures that regardless of unforeseen circumstances within the algorithm or severe disturbances in the external power grid environment, as long as the overall stability of the system is threatened, the compensation function will be safely and smoothly withdrawn, while the basic grid connection function is maintained. This improves the practical usability and reliability of this application in critical power applications.

[0089] In an optional embodiment, two functional enhancement schemes are provided for the adaptive prediction model. Specifically: In practical control applications, simply providing a prediction value for future points is insufficient; understanding the reliability of that prediction value is equally important. Therefore, the adaptive prediction model M... fuzzy Along with outputting the predicted value, the uncertainty σ of the prediction is also quantified and output simultaneously. fuzzy The uncertainty of the fusion prediction stems from two aspects: firstly, the inherent prediction variance σ of each local ARIMA model itself. 2 i Secondly, the uncertainty stems from the model selection itself, as the current state lies on a fuzzy boundary. A weighted method incorporating information entropy can be used to calculate the total uncertainty. Specifically, the Shannon entropy formula H = -Σ(μ) can be applied. i * log(μ i Calculate the entropy of the current fuzzy membership vector μ(t). The magnitude of the H value intuitively reflects the degree of fuzziness or confusion in the current decision. When the system is clearly within a certain working area (a certain μ... i Approaching 1), H value approaching 0; when the system is at the transition boundary of two or more operating regions (multiple μ i The values ​​are all quite significant), and the H value is high. The final fusion prediction variance σ 2 fuzzy σ is calculated using the following formula: 2 fuzzy = Σ(μ i * (σ 2 i + (y* i - y* fuzzy ) 2 )); where Σ represents the summation over all working intervals (dry, optimal, flooded); μ i σ is the membership degree of the i-th interval; 2 i y* is the prediction variance of the i-th local model; i It is its independent predicted value; y* fuzzy This is the final fused prediction. This formula calculates the weighted average of the variances of each local model (Σ(μ)). i * σ 2 i ), and the weighted dispersion of each local model prediction relative to the fused value (Σ(μ) i * (y* i -y* fuzzy ) 2 The combination of these two methods comprehensively reflects the sources of uncertainty. Preferably, when the entropy H value is detected to exceed a threshold (e.g., H > 1.0), the system can calculate σ...2 fuzzy An additional penalty factor (e.g., increasing it by 50%) is applied to more conservatively reflect the predicted risks in the border area.

[0090] To ensure optimal performance of the prediction model throughout long-term operation, the system integrates a performance self-evaluation and structural self-adjustment mechanism. The system continuously compares the model's predicted output with the actual voltage ripple value and continuously calculates prediction performance metrics over a period of time, primarily including root mean square error (RMSE) and mean absolute percentage error (MAPE). Preset performance degradation thresholds are established, for example, RMSE > 0.1V or MAPE > 10%. If the system detects that the model's performance metrics consistently exceed these thresholds over a continuous period, it considers the current model structure potentially unsuitable for the latest state of the fuel cell. In this case, the system automatically triggers a model structure adjustment. This adjustment process involves re-executing the automatic order selection step for the local ARIMA model. Using the latest historical data and based on statistical methods such as the Akaike Information Criterion (AIC) or Bayesian Information Criterion (BIC), the system automatically re-searches for and determines the optimal ARIMA order (p, d, q) for one or more poorly performing local models. This allows the adaptive model to transcend simple parameter adaptation and reach a level of structural adaptation. This ensures that the model can not only cope with chronic parameter drift, but also actively reconstruct itself when the fuel cell undergoes more profound characteristic changes, thereby maintaining the highest prediction accuracy throughout its entire life cycle.

[0091] In another embodiment of this application, the abstract quadratic programming (QP) problem is solved efficiently and robustly while meeting real-time requirements, and the quality of the solution is further verified and adjusted. Specifically, considering that the control cycle of the converter is typically in the millisecond range, the QP problem must be solved in an extremely short time. To achieve this goal, modern optimization algorithms with high computational efficiency and good convergence performance are preferred. In this embodiment, the interior-point method is used as the solver for the QP problem. To accelerate the convergence of the interior-point method iteration process, the selection of the initial point x0 of the solver is crucial. A preferred strategy is to use the ideal compensation current trajectory I without considering any constraints. comp_ideal Multiply by a discount factor less than 1 (e.g., 0.8) as the initial point for the solution, i.e., x0 = 0.8 * I comp_ideal This provides the solver with a good and safe starting point close to the optimal solution. The termination condition for the iterative process is set when the degree to which the solution satisfies the KKT (Karush-Kuhn-Tucker) conditions is below a very small threshold (e.g., the norm of the gradient is less than 1e^(-1 / 2)). -6The iteration terminates and outputs the current solution when the number of iterations reaches a preset limit (e.g., 50). The QP solver returns a mathematically optimal solution x. opt Afterwards, the system will perform final verification and adjustments to ensure the solution's suitability for engineering practice. Specifically, it will check x. opt Does the solution strictly satisfy all linear inequality constraints Ax≤b? Due to numerical computation errors, the solution returned by the solver may contain minor out-of-bounds errors. If any constraint violation is detected, a projection operation is performed to project the solution back onto the boundary of the feasible region. In some extreme cases, strict constraints may lead to a situation where the QP problem has a solution, but the quality of the solution (i.e., the objective function J) is poor. opt The value of J is significantly worse than the ideal situation (without considering constraints). For example, if J... opt The value is more than 20% worse than ideal, which may mean that too much ripple suppression performance has been sacrificed to meet a certain constraint. In this case, the system can implement a constraint adaptive relaxation strategy. For example, the system can temporarily and slightly relax a non-critical constraint, such as reducing the current amplitude constraint from no more than 95% of the rated current (0.95 * I) rated The constraints are relaxed to 90%, and then the QP problem is solved again based on this relaxed constraint set. This gives the system the intelligence to balance strict adherence to rules and the pursuit of optimal results. By appropriately relaxing secondary constraints when necessary, the system can avoid a severe decline in overall performance due to an overly strict restriction, thereby finding a more effective and balanced compensation strategy in reality.

[0092] According to one aspect of this application, a converter low-frequency ripple adaptive cancellation system based on predictive compensation in a hydrogen power grid-connected system includes:

[0093] Signal acquisition device, used to collect operating parameters of fuel cell;

[0094] The processor communicates with the signal acquisition device.

[0095] And the converter controlled by the processor;

[0096] The processor is configured to perform the following steps:

[0097] The operating parameters acquired by the signal acquisition device are processed to estimate the physical state quantities characterizing the internal operating characteristics of the fuel cell. Based on the physical state quantities, an adaptive prediction model is used to predict the future voltage ripple and generate a predicted ripple trajectory.

[0098] Based on the predicted ripple trajectory, calculate the compensation current command to counteract the predicted ripple.

[0099] Based on the compensation current command, control signals for driving the converter are generated.

[0100] This invention proposes a prediction mechanism based on internal physical state quantities. Specifically, instead of relying solely on external electrical quantities, it uses readily available operating parameters such as fuel cell stack temperature, hydrogen pressure, and current ripple intensity to estimate in real time the core physical quantity that profoundly characterizes the internal water management state—membrane water content. This internal state quantity is used as a key input to the prediction model, enabling the model to understand the causes of ripple at a mechanistic level. When the fuel cell transitions from an optimal state to a dry or flooded state, the model can detect this in advance and adjust its prediction strategy accordingly, solving the problems of poor adaptability and model failure caused by the inability to perceive the internal state in traditional methods. An adaptive prediction model based on fuzzy boundary delineation and weighted fusion is constructed. This model first adaptively defines the operating intervals under different water management states from historical data using methods such as K-means clustering, and then configures a specialized local ARIMA model for each interval. During prediction, instead of abruptly switching between different models, the model uses a Gaussian membership function to calculate the fuzzy membership degree of the current state to each interval, and uses this as a weight to smoothly weight and fuse the prediction results of each local model. It accurately captures the nonlinear transition dynamics at the boundary of the working range, solving the problems of decreased prediction accuracy and control oscillation caused by model switching lag or abrupt changes when the working conditions change, thus ensuring the continuity and accuracy of prediction.

[0101] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details of 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 method for adaptive elimination of low-frequency ripple in a converter, characterized in that, include: The operating parameters of the fuel cell are collected and processed to estimate the physical state variables that characterize its internal operating characteristics. Based on the physical state variables, a pre-configured adaptive prediction model is used to predict future voltage ripple and generate a predicted ripple trajectory. Based on the predicted ripple trajectory, calculate the compensation current command to counteract the predicted ripple. Based on the compensation current command, control signals for driving the converter are generated and output.

2. The method according to claim 1, characterized in that, The physical state quantity is the membrane water content estimate, which characterizes the water management state of the fuel cell, obtained by processing operating parameters; Using adaptive prediction models for prediction further includes: Based on the estimated membrane water content, the current operating status is mapped to one of the predefined working intervals to obtain the assignment of the current working interval; The adaptive prediction model is selected or configured based on the attribution determination to perform voltage ripple prediction.

3. The method according to claim 2, characterized in that, The process of defining a pre-defined work area includes: Historical estimates of membrane water content are compiled to form a historical data sequence; K-means clustering analysis was performed on the historical data sequence to identify data clusters belonging to different water management states; Based on the analysis results of the data clusters, the statistical characteristic parameters of each working interval are extracted to complete the adaptive definition of the working interval.

4. The method according to claim 3, characterized in that, Generate the predicted ripple trajectory, including: Configure a local autoregressive integral moving average (ARIMA) model for each working interval defined by statistical characteristic parameters; A Gaussian membership function is constructed based on statistical characteristic parameters. Using the Gaussian membership function and the current membrane water content estimate, the fuzzy membership degree of the function to each working interval is calculated to obtain the fuzzy membership vector. The autoregressive integral moving average (ARIMA) models of each locality are driven to generate independent predicted values, and fuzzy membership vectors are used to weight and fuse the predicted values ​​to generate predicted ripple trajectories.

5. The method according to claim 4, characterized in that, This also includes updating the parameters in the local autoregressive integral moving average (ARIMA) model, specifically: The pre-stored actual voltage ripple sequence is continuously compared with the predicted values ​​generated by the autoregressive integral moving average (ARIMA) model to obtain the prediction error sequence. The parameters of the autoregressive integral moving average (ARIMA) model are dynamically adjusted based on the prediction error sequence.

6. The method according to claim 4, characterized in that, It also includes updating the center point and width parameters of the Gaussian membership function through an adaptive adjustment loop, specifically: Receive the residual ripple sequence at the converter output; Based on the residual ripple sequence, the gradient of the performance index with respect to the center point and width parameter is estimated, and a performance gradient vector is generated. The center point and width parameters are corrected using the performance gradient vector.

7. The method according to claim 6, characterized in that, The center point and width parameters are adjusted, including the calculation of the adaptive learning rate, specifically as follows: Monitor the direction of change of the performance gradient vector over time. Increase the learning rate when consecutive gradients are in the same direction and decrease the learning rate when consecutive gradients are in opposite directions to obtain an adaptive learning rate. The center point and width parameters are updated using gradient descent with an adaptive learning rate and the current performance gradient vector.

8. The method according to claim 7, characterized in that, Gradient descent updates include constraint handling and validity verification of the updated parameters, specifically: Apply a preset feasible region constraint to the parameters updated by gradient descent, so that the values ​​of the center point and width parameter fall within the preset range. If the center point and width parameter exceed the limit, project them back to the boundary of the feasible region constraint to obtain the parameters after constraint processing. The preset test sample set is re-predicted using the constrained parameters to obtain the prediction mean square error. The current constrained parameters are confirmed to be valid only if the current prediction mean square error is less than the previous prediction mean square error by a predetermined threshold; otherwise, they are revoked.

9. The method according to claim 2, characterized in that, The calculation process for the estimated membrane water content is as follows: The current ripple sequence is separated from the operating parameters, and the root mean square value of the current ripple sequence is calculated. The operating parameters, including the fuel cell stack temperature and hydrogen inlet pressure, are obtained and fed into a preset empirical formula along with the root mean square value to obtain an estimated membrane water content.

10. The method according to claim 1, characterized in that, Generate control signals, including: The compensation current command is superimposed on the original dq axis current reference value, and the three-phase voltage command is calculated by the PI controller based on the deviation between the superimposed current reference value and the actual current feedback. SVPWM is executed to synthesize three-phase voltage commands, and during the zero-vector action time of SVPWM, an additional compensation voltage vector corresponding to the compensation current command is synthesized and injected to generate control signals.

Images