Adaptive Control Method for Hydrogen Production Systems Based on Real-Time Parameter Feedback

Abstract

This invention proposes an adaptive control method for a hydrogen production system based on real-time parameter feedback, relating to the field of process control technology. The method involves acquiring the input command signal and controlled state variable signal of the hydrogen production system, comparing the fluctuating energy index with a preset identification dead zone threshold; if the fluctuating energy index is lower than the identification dead zone threshold, the current controller parameters remain unchanged; if the fluctuating energy index is higher than the identification dead zone threshold, the online parameter identification step is activated; the dynamic confidence interval of the process model parameters is calculated; it is determined whether the real-time estimated process model parameters fall within the dynamic confidence interval; if the process model parameters fall within the dynamic confidence interval, the real-time estimated process model parameters are converted into control parameters for the PID controller and updated in the control loop of the hydrogen production system; if the process model parameters exceed the dynamic confidence interval, the internal state vector of the recursive algorithm in the online parameter identification step is reset using boundary values.

Description

Technical Field

[0001] This invention relates to an adaptive control method for a hydrogen production system based on real-time parameter feedback, belonging to the field of process control technology. Background Technology

[0002] In the current automated control of alkaline or proton exchange membrane water electrolysis hydrogen production systems, the PID control strategy is widely used due to its simple structure. This strategy is based on the fact that the controlled object has linear time-invariant characteristics near its rated operating point. By tuning fixed proportional, integral and derivative parameters, it maintains the pressure or temperature balance under steady-state conditions. In chemical production scenarios where the input power is relatively constant, this mode meets the basic continuity requirements.

[0003] In renewable energy coupled hydrogen production scenarios, the operating conditions of electrolyzers fluctuate randomly over a wide power range due to the characteristics of wind and solar resources. The gas-liquid two-phase flow state and electrochemical polarization impedance inside the electrolyzer exhibit nonlinear changes with current density, leading to significant drift in the dynamic gain and time constant of the controlled object. Fixed parameter control strategies struggle to simultaneously address the high impedance hysteresis in low-load areas and the rapid response requirements in high-load areas in such variable parameter objects. Although online identification and adaptive control have been introduced to solve the parameter mismatch problem, the practical application of hydrogen production under wide power fluctuations still faces limitations. For example, Chinese invention patent CN120237793B discloses an energy adaptive regulation method for a hydrogen fuel cell flywheel UPS system, which relies on load prediction and dynamic hierarchical priority coordination to regulate the hydrogen fuel cell unit and flywheel. This approach is essentially based on predictive energy scheduling and does not fundamentally solve the problem of drastic drift in the dynamic gain and time constant of the hydrogen production unit itself under wide power fluctuations.

[0004] Therefore, the technical problem to be solved by this invention is how to construct a system that can capture the characteristics of the object by utilizing the fluctuations in operating conditions, avoid the divergence of steady-state parameters and the lag in transient response, and have an adaptive control mechanism with physical boundary constraints. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of the present invention is as follows: an adaptive control method for a hydrogen production system based on real-time parameter feedback, wherein the adaptive control method for the hydrogen production system operates in the control unit of the hydrogen production system, comprising:

[0006] The system collects input command signals and controlled state variable signals from the hydrogen production system and calculates the fluctuation energy index of the input command signal within a preset time window. The fluctuation energy index is compared with a preset identification dead zone threshold. If the fluctuation energy index is lower than the identification dead zone threshold, the current controller parameters are kept unchanged.

[0007] If the fluctuation energy index is higher than the identification dead zone threshold, the online parameter identification step is activated.

[0008] The online parameter identification step uses a recursive algorithm to estimate the current process model parameters of the hydrogen production system in real time based on the collected input command signals and controlled state variable signals. The process model parameters include at least the process gain and the process time constant. The dynamic confidence interval of the process model parameters is calculated based on the preset parameter reference function and the current input command signal. The parameter reference function defines the nonlinear correspondence between the process model parameters and the input command signal.

[0009] Determine whether the process model parameters for real-time estimation fall within the dynamic confidence interval;

[0010] If the process model parameters fall within the dynamic confidence interval, the real-time estimated process model parameters are converted into control parameters of the PID controller according to the preset internal model control mapping rules, and updated in the control loop of the hydrogen production system; if the process model parameters exceed the dynamic confidence interval, the process model parameters are constrained to the boundary values ​​of the dynamic confidence interval, and the internal state vector of the recursive algorithm in the online parameter identification step is reset using the boundary values.

[0011] Furthermore, it also includes a feedforward compensation step; the feedforward compensation step is executed in parallel with the state monitoring step, including: calculating the time change rate of the input command signal; obtaining the process gain and process time constant estimated in real time by the online parameter identification step; and calculating the feedforward compensation gain at the current moment based on the ratio of the process time constant to the process gain. The calculation rules satisfy the following relationship:

[0012] ,

[0013] in, The process time constant is estimated in real time for the online parameter identification step. The process gain is estimated in real time for the online parameter identification step; the product of the time change rate and the feedforward compensation gain is used as the feedforward control quantity and superimposed on the control loop of the hydrogen production system.

[0014] Furthermore, the adaptive control method for the hydrogen production system is also configured with a pre-constructed static parameter fingerprint database, which stores the benchmark model parameters of the hydrogen production system under different steady-state operating conditions. After comparing the fluctuation energy index with a preset identification dead zone threshold, and before activating the online parameter identification step, the method further includes: calculating the instantaneous change amplitude of the input command signal; determining whether the instantaneous change amplitude exceeds a preset step judgment threshold; if the instantaneous change amplitude exceeds the step judgment threshold, then based on the current input command signal, the corresponding benchmark model parameters are retrieved from the static parameter fingerprint database, and the benchmark model parameters are directly assigned to the parameter estimation algorithm in the online parameter identification step as the initial parameter values ​​at the current moment, while resetting the covariance matrix of the parameter estimation algorithm.

[0015] Furthermore, in the parameter mapping step, the internal model control mapping rules include: setting the closed-loop time constant, which is used to define the target response speed of the control system; determining the integral time parameter of the PID controller using the process time constant; and determining the proportional gain parameter of the PID controller through preset algebraic operations using the process time constant, process gain, and closed-loop time constant.

[0016] Furthermore, the calculation process of the dynamic confidence interval includes: substituting the current input command signal into the parameter reference function to obtain the reference parameter value under the current operating condition; setting the allowable tolerance coefficient, and using the product of the reference parameter value and the allowable tolerance coefficient to determine the upper and lower limits of the dynamic confidence interval; the upper and lower limits are dynamically adjusted in real time according to the changes in the input command signal to form a dynamic allowable range that envelops the parameter reference function.

[0017] Furthermore, the step of resetting the internal state vector of the recursive algorithm in the online parameter identification step using the boundary value includes: forcibly setting the process model parameters to the boundary value of the dynamic confidence interval; using the forcibly set process model parameters to reverse correct the parameter estimation vector in the recursive algorithm; resetting the covariance matrix in the recursive algorithm to the preset initial diagonal matrix state, eliminating the cumulative influence of historical data on the subsequent iteration process, and enabling the recursive algorithm to re-converge from the physical state defined by the boundary value.

[0018] Furthermore, the step of calculating the fluctuation energy index of the input command signal within a preset time window includes: establishing a sliding time window with a fixed length; collecting a series of continuous input command signal sample values ​​within the sliding time window; and calculating the variance or absolute value integral of the rate of change of the series of continuous input command signal sample values ​​as a fluctuation energy index to characterize the dynamic excitation intensity of the input command signal.

[0019] Furthermore, the adaptive control method for the hydrogen production system also includes a circuit breaker protection step. The circuit breaker protection step is performed before updating the control parameters of the PID controller to the control loop of the hydrogen production system, and includes: calculating the change range between the newly generated control parameters of the PID controller and the control parameters currently in use; comparing the change range with a preset single-step adjustment limit threshold; if the change range exceeds the single-step adjustment limit threshold, then the control parameters are adjusted only according to the single-step adjustment limit threshold to limit the rate of change of the control parameters within a single control cycle.

[0020] Furthermore, the controlled state variable signal of the hydrogen production system is selected from one or more of the following: DC bus voltage of the hydrogen production system, electrolyzer operating temperature, or gas-liquid separator pressure; the input command signal is the target power command or target current command of the hydrogen production system, and the adaptive control method of the hydrogen production system realizes the adaptive control of the hydrogen production system by adjusting the output characteristics of the rectifier power supply.

[0021] Compared with the prior art, the beneficial effects of the present invention are:

[0022] 1. This invention uses the energy index of input command fluctuation as the activation gate condition of the identification algorithm, which solves the problem of data collinearity and parameter drift caused by insufficient excitation signal during the steady-state operation of the conventional recursive least squares algorithm. By real-time monitoring and releasing parameter iteration permission only when the input signal naturally changes with sufficient dynamic information, the control system can accurately capture the dynamic characteristics of the controlled object without actively injecting interference signals to affect production. Based on the excitation-triggered intermittent operation logic, it avoids the dilution of model accuracy by invalid data, ensuring that the dynamic adjustment capability of the adaptive controller does not degrade after long-term steady-state operation, and maintaining the physical authenticity and high confidence of the control parameters in the entire time domain.

[0023] 2. This invention establishes an instantaneous initial value projection mechanism based on the collaborative operation of a static polarization fingerprint database and online recursive identification. This avoids the inherent parameter convergence lag and adjustment oscillations of traditional adaptive control when facing large jumps in operating conditions. When a large step change in the input command is detected, the identification algorithm's state vector is forcibly overwritten and the covariance matrix is ​​reset by directly retrieving the reference physical parameters that match the current load. During the transient window period before the algorithm's mathematical iteration has completed convergence, the controller is given near-true initial control parameters. Prior physical laws are used to fill the learning lag of the posterior algorithm, achieving smooth transition and rapid stabilization of the control system during wide load and large-span switching processes, eliminating the risk of transient shocks caused by insufficient parameter tracking.

[0024] 3. This invention constructs a dynamic inverse model compensation architecture where the feedforward control gain and feedback identification parameters are from the same source and synchronized. This solves the technical problem of compensation failure in traditional fixed feedforward strategies after the characteristics of the controlled object drift over time or in the environment. It directly reuses the process gain and time constant identified in real time in the feedback loop to synthesize the inverse dynamic feedforward gain, so that the compensation strength of the feedforward path automatically follows the physical factors of equipment aging and temperature changes for real-time self-calibration. The parameter reuse mechanism ensures that the feedforward compensation amount always maintains a precise match with the actual physical response characteristics of the current equipment in whatever load range the control system is in throughout its entire life cycle, eliminating the inherent phase lag when the feedback control responds to rapid command changes. Attached Figure Description

[0025] Figure 1 This is a flowchart of the adaptive control process of the hydrogen production system based on real-time parameter feedback and dynamic constraints according to the present invention.

[0026] Figure 2 This is a comparison diagram of the pressure response of adaptive regulation and conventional control under wide power fluctuation conditions according to the present invention;

[0027] Figure 3 Fishbone diagram of key technical elements for ensuring the stability of the hydrogen production system with wide power regulation in this invention. Detailed Implementation

[0028] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0029] This invention discloses an adaptive control method for a hydrogen production system based on real-time parameter feedback. Deployed in the central control unit of the hydrogen production system, this unit can be a programmable logic controller (PLC) or a distributed control system (DCS). The overall control logic architecture consists of a state monitoring and identification activation module, a parameter initial value projection module, an online parameter identification module, a parameter constraint and mapping module, and a feedforward compensation module. The state monitoring and identification activation module determines whether the system is in an effective identification zone rich in dynamic information. The parameter initial value projection module provides initial values ​​close to the actual operating conditions for the identification algorithm when the system operating conditions change drastically. The online parameter identification module estimates the process model parameters under the current operating conditions in real time. The parameter constraint and mapping module constrains the identified parameters to a physically reasonable range and resolves them into control parameters for the PID controller. The feedforward compensation module uses the identified model parameters to pre-compensate for changes in input commands. All modules work together to achieve stable and rapid control of the hydrogen production system under wide power fluctuations. The control unit continuously collects data. The input command signal of the hydrogen production system can be a target current command or a controlled state variable signal, which can be the electrolyzer operating temperature or the gas-liquid separator pressure. The system establishes a fixed-length sliding time window, taking 5 seconds as an example, and calculates the variance of the sampled values ​​of the input command signal within this time window as a fluctuation energy index characterizing the dynamic excitation intensity of the signal. The system compares this fluctuation energy index with a preset identification dead zone threshold. It is calibrated by analyzing the noise floor variance of the input command signal under typical steady-state conditions, such as constant operation at 50% rated power, which can be set to 3 to 5 times the noise floor variance to ensure that the identification algorithm is only triggered when the natural variation of the input signal exceeds the pure noise level. If the fluctuation energy index is lower than the identification dead zone threshold, it indicates that the system is in a steady state or the excitation is insufficient. At this time, the system will force the current controller parameters to remain unchanged, putting the subsequent identification module into a dormant state. This avoids parameter drift caused by data collinearity in the recursive algorithm when there is a lack of dynamic excitation.

[0030] The specific method for measuring the noise floor variance in the hydrogen production system control unit is as follows:

[0031] The hydrogen production system should be operated under a typical steady-state condition, defined as the system operating at 40%-60% of its rated power, with key controlled state variables, such as electrolyzer operating temperature and gas-liquid separator pressure, fluctuating within ±1% of their set values ​​over a 10-minute period. During this period, interference from external renewable energy power fluctuations should be eliminated to ensure a constant input command signal. The control unit should continuously acquire the input command signal (current) at a sampling frequency of at least 100Hz for at least 30 minutes. For the acquired discrete-time series data, obvious outliers caused by sensor malfunctions should first be removed. Then, the variance of all sampled values ​​within the time window should be calculated; this value is considered the noise floor variance under this steady-state condition. The dead zone threshold was ultimately set to be 3 to 5 times the variance of the noise floor. This range was verified through Monte Carlo simulations under typical noise distributions, demonstrating its ability to effectively distinguish between inherent system noise and command fluctuations containing dynamic information.

[0032] If the fluctuation energy index exceeds the identification dead zone threshold, it indicates that the current data contains sufficient dynamic information, and the system activates the subsequent identification and update process. Before activating the identification step, the system performs an additional graded arbitration of the disturbance amplitude. The system calculates the instantaneous change amplitude of the current input command signal and compares it with a preset step judgment threshold, which can be set to 20% of the rated command. It should be noted that the choice of 20% as the threshold is not arbitrary, but based on the nonlinear characteristics of the electrolytic cell's VI curve. The VI curve of the electrolytic cell has a clear inflection point in the low current density region. Below this inflection point, the electrochemical overpotential dominates, the system dynamic gain is low, the time constant is large, and the response is slow. Above this inflection point, the ohmic overpotential dominates, the dynamic gain increases, the time constant decreases, and the response is faster. Therefore, setting a 20% step judgment threshold aims to capture command changes that cross this nonlinear critical region. When the command change amplitude exceeds this threshold, it means that the system operating point is switching between the low gain region and the high gain region. At this time, the process model parameters (K) est ,T est Dramatic changes will occur.

[0033] If the instantaneous change does not exceed the step threshold, it indicates that the operating condition change is relatively gentle. In this case, the online parameter identification step uses the parameter estimate from the previous moment as the initial parameter value for the current moment. If the instantaneous change exceeds the step threshold, it indicates that the system is undergoing a significant operating condition transition. At this time, the system triggers the parameter initial value projection mechanism. Based on the current input command signal, the system retrieves the corresponding baseline model parameters from the pre-built static parameter fingerprint database. This fingerprint database is established through the step test during system factory commissioning and stores the process gain and process time constant corresponding to different steady-state operating points. The retrieved baseline model parameters are directly assigned to the parameters in the online parameter identification step. The estimation algorithm uses the initial parameter values ​​for the current moment and forcibly resets the covariance matrix of the parameter estimation algorithm to a pre-set diagonal matrix with a large value. This operation is used to clear the cumulative influence of historical data, improve the algorithm's tracking sensitivity to new operating conditions, and use prior physical laws to fill the convergence lag of the posterior algorithm in the early stage of sudden changes in operating conditions, eliminating transient shocks caused by parameter tracking lag. The system executes an online parameter identification step, based on the input command signal and controlled state variable signal acquired after activation, and uses a recursive algorithm, which can be the recursive least squares (RLS) method, to estimate the current process model parameters of the hydrogen production system in real time. The process model parameters include at least the process gain. and process time constant .

[0034] After obtaining the real-time estimated process model parameters, the system imposes physical reality constraints on them. Based on a preset parameter baseline function and the current input command signal, the system calculates the dynamic confidence interval of the process model parameters. The parameter baseline function defines the nonlinear correspondence between the process model parameters and the input command signal. Preferably, the parameter baseline function is a cubic polynomial, which can effectively fit the common nonlinear S-shaped variation trend of electrolytic cell parameters. The data required for fitting comes from the stepped test described in Example 4: discrete parameter values ​​identified by the system at each steady-state operating point. Using the least squares method to perform polynomial regression on this discrete dataset yields a continuous function expression, which is used for the real-time calculation of the dynamic confidence interval. The calculation process of the dynamic confidence interval includes: substituting the current input command signal into the parameter reference function to obtain the reference parameter value under the current operating condition; setting an allowable tolerance coefficient (taking 30% as an example); using the reference parameter value and the tolerance coefficient to determine the upper and lower limits of the interval, forming an envelope range that is dynamically adjusted in real time according to the input command; the system determines whether the real-time estimated process model parameters fall within the dynamic confidence interval; if they fall within the interval, the parameters are deemed valid, and the subsequent parameter mapping steps are initiated; if the process model parameters exceed the dynamic confidence interval, the identification result is determined to be contaminated by strong noise or atypical disturbances, and the system forcibly constrains the process model parameters to the boundary value of the dynamic confidence interval, using this boundary value to reset the online parameters. The internal state vector of the recursive algorithm in the parameter identification step specifically includes reversing the parameter estimation vector in the algorithm to the boundary value and resetting the covariance matrix to the initial diagonal matrix state. This ensures that the identification algorithm is always constrained to operate within a physically reasonable space, avoiding control instability caused by parameter divergence. The step of resetting the internal state vector of the recursive algorithm in the online parameter identification step using boundary values ​​includes: directly setting the parameter estimation vector of the recursive algorithm to the boundary value of the dynamic confidence interval in the current control cycle, and resetting the covariance matrix to the preset initial diagonal matrix state; the diagonal elements of the initial diagonal matrix are set to 100 to 1000 times the square of the benchmark value under the corresponding parameter's rated operating condition, clearing the accumulated influence of historical data. It should be noted that 100 times represents a moderate level of initial uncertainty, allowing the algorithm to converge relatively quickly, suitable for systems with relatively gentle dynamic changes. 1000 times represents a very large initial uncertainty, giving new data extremely high weight, enabling the algorithm to reconverge very quickly after drastic changes in system characteristics, which is precisely the expected behavior during parameter reset.

[0035] The system performs parameter mapping and updating, converting the real-time estimated or boundary-constrained process model parameters into PID controller control parameters according to preset internal model control mapping rules. The internal model control mapping rules include: setting the closed-loop time constant. Used to define the target response speed of the control system; utilizing the process time constant. Determine the integral time parameter of the PID controller ,by For example; and using Process gain and closed-loop time constant The proportional gain parameter of the PID controller is determined through preset algebraic operations. ,by For example, among which This is the pure time delay estimate of the system. Before updating to the control loop, the system can also perform a circuit breaker protection step to calculate the change range between the newly generated PID control parameters and the currently used control parameters. If the change range exceeds the preset single-step adjustment limit threshold (e.g., 10%), the control parameters are gradually adjusted according to the single-step adjustment limit threshold to limit the rate of change of the control parameters within a single control cycle and prevent drastic jumps in the control output. The PID control parameters after the limit adjustment are then updated to the control loop of the hydrogen production system. In addition, this method can also execute a feedforward compensation step in parallel; this step calculates the time change rate of the input command signal in real time. When calculating the time change rate of the input command signal in the feedforward compensation step, to avoid noise amplification caused by direct differential processing of the original command signal, the following steps are taken: applying a first-order low-pass filter to the input command signal, with the filter time constant set to the closed-loop time constant in the internal model control mapping rule. 1 / 5 to 1 / 10; calculate the difference between the two sampling periods of the filtered signal, and use it as the smoothed estimate of the time rate of change to calculate the final feedforward control quantity, thus obtaining the process gain estimated in real time during the online parameter identification step. and process time constant ; Calculate the feedforward compensation gain at the current moment based on the ratio of the process time constant to the process gain. The calculation rules satisfy the following relationship: ,in, The process time constant is estimated in real time for the online parameter identification step. The process gain is estimated in real time for the online parameter identification step; The dynamic inverse model gain constitutes the controlled object; the system will change the rate of change over time. With feedforward compensation gain The product of these two values ​​is used as the feedforward control quantity and superimposed on the control loop of the hydrogen production system. It can be directly superimposed on the output signal of the PID controller; due to the feedforward gain... With feedback identification parameters , Synchronization from the same source enables the feedforward compensation to automatically self-calibrate as the equipment ages or its operating conditions change.

[0036] It should be noted that the choice of closed-loop time constant... The ratio of 1 / 5 to 1 / 10 is used to balance noise suppression and phase lag. This ratio ensures that the response speed of the feedforward channel is much faster than that of the main feedback loop, effectively filtering out high-frequency noise, while the phase lag it introduces is small enough not to affect the timeliness of feedforward compensation.

[0037] Example 1

[0038] In hydrogen production systems coupled with renewable energy, the input power command exhibits both significant step changes and high-frequency random fluctuations. The input command signal for a PEM electrolyzer system originates directly from a large photovoltaic array. At midday, the system operates stably at 90% of its rated power. Under clear skies, the fluctuation energy of the input command signal is below the preset dead-zone threshold, and the online parameter identification process is in a dormant state. The controller uses PID control parameters tuned near this operating point, along with feedforward compensation gain. To maintain stable hydrogen production pressure, at a certain moment, a cumulonimbus cloud rapidly obscures the photovoltaic array, causing the input command signal to plummet from 90% of its rated value to 20% within seconds. The control unit's status monitoring step calculates that the instantaneous change in the input command signal is 70%, which exceeds the preset 20% step threshold. The system immediately triggers the parameter initialization projection mechanism. Based on the current 20% target command signal, it retrieves the corresponding baseline model parameters from the static parameter fingerprint database. These baseline model parameters reflect the high polarization impedance and low process gain characteristics of the electrolyzer under low current density. The system directly assigns these baseline model parameters to the parameter estimation algorithm in the online parameter identification step as the initial parameter values ​​for the current moment and resets its covariance matrix. The parallel feedforward compensation step also instantly acquires these projected baseline model parameters. and Based on the time change rate of the command signal, a large negative feedforward control quantity is calculated. This feedforward control quantity is superimposed on the output of the PID loop, and before the controlled state variable, i.e. the pressure, deviates, the control output applied to the rectifier power supply is actively reduced.

[0039] After the system enters a 20% low-load operating state, the scattering effect at the cloud edge causes the input command signal to exhibit continuous high-frequency random fluctuations around the 20% mean. The fluctuation energy index calculated by the state monitoring step remains higher than the identification dead zone threshold, thus activating the online parameter identification step and keeping it running continuously. This identification step utilizes these dynamic excitation signals, using the previously projected baseline model parameters as initial values, and employs a recursive algorithm with a variable forgetting factor for convergence, to estimate the process model parameters in real time. and It begins to approximate the actual physical characteristics under the current 20% load, which already includes the effects of equipment temperature and aging; these real-time updated parameters are fed into the parameter mapping and update process, continuously fine-tuning the control parameters of the PID controller and synchronously updating the feedforward compensation gain. During this period, a sensor signal interference caused the online parameter identification step to calculate an abnormal process gain value, which exceeded the dynamic confidence interval calculated by the parameter reference function under 20% load. The parameter constraint mechanism immediately took effect, forcibly constraining the process model parameters to the boundary value of the dynamic confidence interval. This boundary value was used to reset the internal state vector of the recursive algorithm, which enabled the identification algorithm to return to a physically reasonable convergence trajectory in subsequent control cycles, maintaining the stability of the control system. After the cumulonimbus cloud passed, the photovoltaic array output jumped rapidly from 20% to 90% within a few seconds, triggering the step judgment threshold again. The parameter initial value projection mechanism was then activated, and the system retrieved the reference model parameters corresponding to 90% rated operating conditions from the static parameter fingerprint library and overwrote the initial values ​​of the identification algorithm. The feedforward compensation step then calculated a large positive feedforward control quantity based on the model parameters of this high-load area, driving the system to quickly track the power jump, enabling the hydrogen production system to quickly recover to a stable operating state under high load after drastic fluctuations across a very large power range.

[0040] Example 2

[0041] This embodiment constructs a semi-physical simulation test platform to verify the control effect of the adaptive control method in dealing with the wide range of dynamic and nonlinear characteristics of the hydrogen production system. The platform consists of an industrial programmable logic controller (PLC) running the adaptive control method as the control unit, and a dynamic simulation model of the hydrogen production system running in real time on a host computer as the controlled object. The simulation model is established based on the electrochemical and thermodynamic characteristics of the PEM electrolyzer. The process model parameters, namely process gain and process time constant, exhibit nonlinear characteristics with changes in the input command signal (current). Specifically, in the low-load region, such as 20% of the rated current, the process gain is low while the time constant is large; in the high-load region, such as 90% of the rated current, the process gain increases while the time constant decreases. The experiment includes a control group and the prototype group of this invention. The control group uses a fixed-parameter PID controller tuned and optimized at 75% of the rated operating point, while the prototype group of this invention uses the complete adaptive control method, which includes fluctuation energy index activation, initial value projection of the static parameter fingerprint database, online parameter identification with a variable forgetting factor, dynamic confidence interval constraints, and feedforward compensation based on parameter homology. Mechanism: The experiment used a standardized input command signal sequence, which was applied simultaneously to the control group and the sample group of the present invention. The sequence included three typical stages: In the first stage, the command signal dropped from 90% of the rated value to 20% of the rated value; in the second stage, Gaussian white noise with a root mean square deviation of 5% was superimposed on the command signal near 20% of the rated value to simulate random fluctuations under low load; in the third stage, the command signal rose from 20% of the rated value to 90% of the rated value. The controlled state variable, i.e., the dynamic response of the gas-liquid separator pressure, was monitored and recorded throughout the experiment. Key performance indicators included the maximum deviation of the step response, i.e., undershoot or overshoot, the settling time, defined as the time required for the pressure to enter and remain within ±2% of the set value, and the integral absolute error (IAE) under fluctuating conditions. The experimental data are shown in Table 1.

[0042] Table 1: Performance Comparison of Two Control Methods under Typical Operating Conditions

[0043]

[0044] Experimental data shows that in the first stage of the step descent condition, the fixed PID parameters of the control group responded slowly in the 20% low load range, resulting in a pressure downsurge of -0.38 MPa and a settling time of 48.5 s. In contrast, the sample group of this invention, upon detecting a large step jump, triggered the parameter initial value projection mechanism, immediately retrieving the baseline model parameters applicable to the 20% condition from the static parameter fingerprint database, and simultaneously updating the PID parameters and feedforward compensation gain. The pressure downsurge was suppressed to -0.07 MPa, and the stabilization time was shortened to 8.8 s. In the second stage of low-load fluctuation conditions, the controller of the control group had poor ability to suppress fluctuations due to parameter mismatch, and the integral absolute error (IAE) accumulated to 135.8. The fluctuation energy index of the sample group of this invention was consistently higher than the identification dead zone threshold, and the online parameter identification step was activated. Under the constraint of the dynamic confidence interval, the model parameters were identified and updated online, enabling the controller to closely track pressure changes, with an IAE of only 24.1. In the third stage of step-up conditions, the fixed parameters of the control group were adjusted too aggressively for the 90% high-load area, resulting in a pressure overshoot of +0.42 MPa. The sample group of this invention once again switched to the benchmark model parameters of the high-load area instantaneously through the parameter initial value projection mechanism, achieving a smooth response with a maximum overshoot of only +0.06 MPa.

[0045] Example 3

[0046] This embodiment combines Figures 1 to 3 The adaptive control method for hydrogen production systems based on real-time parameter feedback is explained, such as... Figure 1As shown, the process begins with signal acquisition, which involves acquiring the input command signal and the controlled state variable signal. It then proceeds to calculate the fluctuation energy index, using the variance or rate of change within a sliding window to characterize the dynamic excitation intensity. This index is compared to an identification dead zone threshold. If the threshold is not exceeded, the system is determined to be in a steady state or a low-excitation state, and the current controller parameters are kept constant. If the threshold is exceeded, the identification process is activated, and the instantaneous amplitude is further determined to be above the step threshold. If the threshold is exceeded, a transition condition is triggered, and the parameter initialization projection mechanism is activated to retrieve steady-state baseline model parameters from the static parameter fingerprint database and provide them. Initial values ​​are set, and the covariance matrix is ​​reset. If the change does not exceed the threshold, the online parameter identification step is initiated. A recursive algorithm with a variable forgetting factor is used to estimate the process gain and time constant. The dynamic confidence interval is calculated based on the parameter benchmark function, and it is determined whether the real-time estimated parameters fall within this interval. If not (parameter anomaly), the parameter constraint and reset step is executed to constrain the parameters to boundary values ​​and reset the algorithm state. If yes (valid), the model parameters are converted into PID control parameters through the internal model control mapping rule. The feedforward compensation step is executed in parallel, and the feedforward compensation gain is calculated based on the process time constant and the gain ratio. The PID parameters and feedforward values ​​generated above are used to update the control loop after being adjusted and limited in a single step during the circuit breaker protection process.

[0047] like Figure 2 As shown, the graph establishes a coordinate system with time (s) as the horizontal axis, the left vertical axis representing the gas-liquid separator pressure (MPa), and the right vertical axis representing the percentage of the input command signal. The graph contains three curves: the dashed line represents the input command signal, the solid line represents the control group pressure, and the dotted line represents the pressure of the present invention's sample group. The graph shows that when the input command signal drops from 90% to 20% at 12s, the control group pressure curve experiences a significant downward plunge, falling to a minimum of 1.60MPa, while the present invention's sample group pressure curve remains stable with only minor fluctuations. In the low-load fluctuation range from 13.2s to 21.6s, the control group exhibits obvious oscillations, while the present invention's sample group maintains a pressure near the 2.00MPa setpoint. When the input command signal rises from 20% to 90% at 22.8s, the control group pressure surges to 2.40MPa, resulting in overshoot, while the present invention's sample group pressure peaks at only 2.06MPa, quickly returning to steady state. Figure 3As shown, this diagram uses a fishbone diagram format. The fish head on the right represents the ultimate goal, namely the regulation stability of the hydrogen production system under wide power fluctuations. The four main branches pointing to the backbone are identification, activation, and monitoring; transient impact suppression; parameter estimation and constraint; and closed-loop control strategy. The specific technical modules under each branch constitute the fishbone elements: identification, activation, and monitoring includes fluctuation energy indicators, identification dead zone thresholds, and sliding time windows; transient impact suppression covers static parameter fingerprint databases, parameter initial value projection, and step judgment thresholds; parameter estimation and constraint integrates parameter benchmark functions, dynamic confidence intervals, and variable forgetting factors; and the closed-loop control strategy consists of circuit breaker protection steps and feedforward compensation steps. It consists of internal model control mapping rules, and these elements together support the stable operation of the system under complex working conditions.

[0048] Example 4

[0049] This embodiment discloses a standardized engineering calibration procedure required before deploying the adaptive control method in a specific 100MW PEM hydrogen production system. This procedure provides reproducible benchmark parameters and functions that match the specific physical object for the method's identification activation determination, parameter initialization projection, and dynamic confidence interval constraint modules. To determine the identification dead zone threshold, the 100MW hydrogen production system controller is placed in a fixed-parameter PID control mode and runs stably for 30 minutes at 50% of rated power. During this period, the control unit continuously acquires the input command signal with a sampling period of 100ms and calculates the variance of the signal in steady state, denoted as . The dead zone threshold is set to a preset multiple of the base noise level; for example, 4 times. This configuration ensures that the online parameter identification step is activated only when the signal fluctuation energy is greater than the random noise. To construct a static parameter fingerprint library and parameter reference function, a stepped test covering the electrolytic cell's operating range is performed. The input command signal starts from the lowest load of 10% and gradually increases in 10% increments to the highest load of 100%. At each load point (10%, 20%, up to 100%), the system operates stably for 20 minutes. During the last 10 minutes of each stable operating point, a pseudo-random binary sequence PRBS excitation signal with an amplitude of 2% of the rated value is applied. Using an offline system identification tool, the steady-state process gain at that specific operating point is identified based on the input and output data during this phase. and process time constant ; these discrete operating points and its corresponding Data pairs are stored in the control unit, forming a static parameter fingerprint database. Simultaneously, a polynomial fitting method is used to fit these discrete data points, yielding the process gain and process time constant as a function of the input command signal. Continuous nonlinear function of change and These two functions constitute the parameter baseline function. After obtaining the parameter baseline function, all stepped test conditions are calculated. Actual identification parameters on Compared with the baseline function prediction value The relative errors between them are calculated, and the standard deviations of these relative errors are also calculated. Tolerance factor for dynamic confidence intervals That is, it is determined based on the statistical standard deviation: set as Or a fixed percentage covering 99.7% of data points, taking 30% as an example, this tolerance coefficient defines the statistically reasonable range of parameter fluctuations enveloping the reference function; the step judgment threshold is set according to the actual operating constraints of the system, and in this 100MW PEM hydrogen production system, it is set to 20% of the rated power; the closed-loop time constant in the internal model control mapping rule As a tuning parameter, a tradeoff is set between the system's response speed and its adaptability to model mismatch. In this system, When set to 0.5 times the average process time constant, the system achieves a balance between tracking speed and overshoot suppression. By executing the above calibration procedure, the dead zone threshold, static parameter fingerprint library, parameter reference function, allowable tolerance coefficient, step judgment threshold, and closed-loop time constant are all assigned explicit engineering values ​​or functional relationships based on the physical characteristics test of this specially customized hydrogen system. This completes the algorithm framework of the adaptive control method, making it ready for online operation.

[0050] It should be noted that the PRBS signal is used to inject sufficient dynamic excitation at each steady-state operating point to identify process model parameters. The amplitude of the PRBS is typically set to 1%-2% of the input command value at the current steady-state point. Preferably, at 50% rated load, with a steady-state command of 50, the PRBS amplitude is between ±0.5 and ±1.0. The switching speed of the PRBS signal is based on the system's dominant time constant T. dominant To set it. T is estimated through preliminary rough step response tests. dominant This ensures that the excitation signal effectively covers the main dynamic frequency band of the system. The PRBS sequence should be long enough to continuously excite the system and capture its dynamics; the sequence length should ensure that the total duration of the sequence is ≥5×T. dominant .

[0051] Preferably, the standard practice in the field of system identification, namely offline least squares method, is used to identify process model parameters.

[0052] In a preferred embodiment, the stepped testing is designed to cover the entire operating range of the hydrogen production system, and the procedure is as follows:

[0053] Starting from the minimum allowable load of the system, the load is gradually increased in 10% increments to the maximum load, i.e., 100%. After each step, the load is gradually reduced back to the minimum load to examine any potential hysteresis characteristics.

[0054] After a new step command is applied, the controlled state variable, such as the gas-liquid separator pressure, is continuously monitored. When this variable enters and remains within ±1% of its new steady-state value for at least 5 minutes, the system is considered to have reached steady state.

[0055] After the system stabilizes, firstly, input and output data for one minute without excitation is collected at a sampling frequency of 100Hz to calculate the noise floor at that point. Then, the previously set PRBS excitation signal is applied. Input and output data for the entire duration of the PRBS sequence is continuously collected at the same sampling frequency of 100Hz. The amount of data should be sufficient for offline least-squares identification.

[0056] Example 5

[0057] This embodiment discloses an online repair procedure for the static parameter fingerprint database mismatch problem caused by aging after long-term operation of electrolytic cell equipment in the adaptive control method; static parameter fingerprint database Calibration is completed during the initial equipment commissioning, and all operating points are stored. Corresponding baseline model parameters As the electrolyzer operates for several years, the decrease in electrode activity and the increase in diaphragm resistance will cause its actual physical characteristics to drift, meaning that at the same operating point... Actual process gain Decrease, while the time constant To address this slow time-varying characteristic, this method adds a background fingerprint database adaptive maintenance module to the online parameter identification step; it continuously monitors real-time estimated parameters that meet the identification dead zone threshold, are output by the online parameter identification step, and are determined to be valid after dynamic confidence interval constraints. When the system is at a certain operating point After stable operation in the vicinity and accumulation of sufficient identification data, specifically defined as an accumulated identification activation time exceeding 1 hour within ±5% of the operating point, this module calculates these... and Statistical average value in this operating range The system calculates the statistical average. Compared with the initial baseline value stored in the fingerprint database Deviation between If the norm of this deviation exceeds a preset maintenance threshold, to avoid frequent corrections, the system determines that the fingerprint database is mismatched at this operating point and utilizes this deviation. For fingerprint database and Update the system; utilize the incentive data provided by renewable energy fluctuations during long-term operation to perform online calibration of the prior knowledge model of the static parameter fingerprint database on a slow time scale, ensuring that the parameter initial value projection mechanism can provide initial values ​​close to the current physical characteristics throughout the entire life cycle of the equipment, and maintain the system's ability to respond quickly to large step disturbances.

[0058] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0059] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A real-time parameter feedback-based adaptive control method for a hydrogen production system, characterized in that, The hydrogen production system adaptive regulation method is run in a control unit of the hydrogen production system, and comprises the following steps: The input instruction signal and the controlled state variable signal of the hydrogen production system are collected, and the fluctuation energy index of the input instruction signal in a preset time window is calculated; the fluctuation energy index is compared with a preset identification dead zone threshold value; if the fluctuation energy index is lower than the identification dead zone threshold value, the current controller parameter is kept unchanged; If the fluctuation energy index is higher than the identification dead zone threshold value, an online parameter identification step is activated; The online parameter identification step uses a recursive algorithm to estimate the current process model parameter of the hydrogen production system in real time based on the collected input instruction signal and the controlled state variable signal; the process model parameter at least comprises a process gain and a process time constant; a dynamic confidence interval of the process model parameter is calculated according to a preset parameter reference function and the current input instruction signal; the parameter reference function defines a nonlinear corresponding relationship between the process model parameter and the input instruction signal; It is judged whether the real-time estimated process model parameter falls within the dynamic confidence interval; If the process model parameter falls within the dynamic confidence interval, the real-time estimated process model parameter is converted into the control parameter of the PID controller according to a preset internal model control mapping rule, and is updated to the control loop of the hydrogen production system; if the process model parameter exceeds the dynamic confidence interval, the process model parameter is constrained to the boundary value of the dynamic confidence interval, and the internal state vector of the recursive algorithm in the online parameter identification step is reset using the boundary value.

2. The adaptive control method for hydrogen production system based on real-time parameter feedback according to claim 1, characterized in that, The adaptive regulation method of the hydrogen production system further comprises a feedforward compensation step, including: calculating the time variation rate of the input instruction signal; obtaining the process gain and the process time constant estimated in real time in the online parameter identification step; calculating the feedforward compensation gain at the current time based on the ratio of the process time constant to the process gain , and the calculation rule satisfies the following relationship: ， wherein, a process time constant estimated in real time for the online parameter identification step, a process gain estimated in real time for the online parameter identification step; a product of the time variation rate and the feedforward compensation gain is taken as a feedforward control amount, and superimposed into the control loop of the hydrogen production system.

3. The adaptive control method for a hydrogen production system based on real-time parameter feedback according to claim 1, characterized in that, The hydrogen production system adaptive regulation method is further configured with a pre-constructed static parameter fingerprint library, and the static parameter fingerprint library stores the reference model parameters of the hydrogen production system under different steady-state working conditions; After the fluctuation energy index is compared with the preset identification dead zone threshold value, and before the online parameter identification step is activated, the following steps are further included: the instantaneous change amplitude of the input instruction signal is calculated; it is judged whether the instantaneous change amplitude exceeds a preset step judgment threshold value; if the instantaneous change amplitude exceeds the step judgment threshold value, the corresponding reference model parameter is retrieved from the static parameter fingerprint library according to the current input instruction signal, and the reference model parameter is directly assigned to the parameter estimation algorithm in the online parameter identification step as the parameter initial value at the current time, and the covariance matrix of the parameter estimation algorithm is reset.

4. The adaptive control method for a hydrogen production system based on real-time parameter feedback according to claim 1, characterized in that, The internal model control mapping rule comprises the following steps: a closed-loop time constant is set to define the target response speed of the control system; the integral time parameter of the PID controller is determined using the process time constant; the proportional gain parameter of the PID controller is determined through a preset algebraic operation using the process time constant, the process gain and the closed-loop time constant.

5. The adaptive control method for hydrogen production system based on real-time parameter feedback according to claim 1, characterized in that, The calculation process of the dynamic confidence interval comprises the following steps: the current input instruction signal is substituted into the parameter reference function to obtain the reference parameter value under the current working condition; an allowable tolerance coefficient is set, and the upper limit value and the lower limit value of the dynamic confidence interval are determined using the product of the reference parameter value and the allowable tolerance coefficient; the upper limit value and the lower limit value are dynamically adjusted in real time with the change of the input instruction signal to form a dynamic allowable range enveloping the outside of the parameter reference function.

6. The adaptive control method for a hydrogen production system based on real-time parameter feedback according to claim 1, characterized in that, The step of resetting the internal state vector of the recursive algorithm in the online parameter identification step using the boundary value includes: forcibly setting the process model parameters to the boundary value of the dynamic confidence interval; using the forcibly set process model parameters to correct the parameter estimation vector in the recursive algorithm in reverse; resetting the covariance matrix in the recursive algorithm to a preset initial diagonal matrix state to eliminate the cumulative effect of historical data on the subsequent iteration process, so that the recursive algorithm starts to converge from the physical state defined by the boundary value.

7. The adaptive control method for a hydrogen production system based on real-time parameter feedback according to claim 1, characterized in that, The step of calculating the fluctuation energy index of the input instruction signal within the preset time window includes: establishing a sliding time window with a fixed length; collecting a series of continuous input instruction signal sample values within the sliding time window; calculating the variance or the absolute value integral of the rate of change of the series of continuous input instruction signal sample values as the fluctuation energy index, representing the dynamic excitation strength of the input instruction signal.

8. The adaptive control method for a hydrogen production system based on real-time parameter feedback according to claim 1, characterized in that, The hydrogen production system adaptive regulation method further includes a fuse protection step; The fuse protection step is performed before updating the control parameters of the PID controller to the control loop of the hydrogen production system, including: calculating the change amplitude between the newly generated control parameters of the PID controller and the control parameters currently being used at the current time; comparing the change amplitude with a preset single-step adjustment amplitude threshold; if the change amplitude exceeds the single-step adjustment amplitude threshold, only adjusting the control parameters according to the single-step adjustment amplitude threshold to limit the change rate of the control parameters within a single control period.

9. The adaptive control method for a hydrogen production system based on real-time parameter feedback according to claim 1, wherein, The controlled state variable signal of the hydrogen production system is selected from one or more of the DC bus voltage, electrolytic cell operating temperature, or gas-liquid separator pressure of the hydrogen production system; the input instruction signal is a target power instruction or a target current instruction of the hydrogen production system, and the hydrogen production system adaptive regulation method realizes adaptive regulation of the hydrogen production system by adjusting the output characteristics of the rectifier power supply.

Images