Intelligent hierarchical control system and method for hydrogen production through alkaline electrolysis of water

Abstract

The invention discloses an intelligent hierarchical control system and method for hydrogen production through alkaline electrolysis of water, and belongs to the technical field of industrial system control, and the intelligent hierarchical control system comprises an L3 adaptive power predictor, an L2 table look-up-hybrid constraint optimizer, an L1 life integrator-dynamic hot standby scheduler and an L0 dynamic arbitration bus; the power prediction precision and the load adaptability are improved, and meanwhile the hydrogen yield and the system energy efficiency are optimized; by optimizing a start-stop sequence and hot standby scheduling, the start-stop frequency of the electrolytic cell is reduced, and the service life of equipment is remarkably prolonged; multi-layer instruction conflicts are eliminated by means of a dynamic arbitration mechanism, and the stability of the system under complex working conditions and multi-constraint scenes is improved; and embedded real-time operation is supported, and operation requirements of fixed or mobile AWE equipment can be flexibly adapted.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of industrial system control, and particularly relates to an intelligent hierarchical control system and method for hydrogen production by alkaline electrolysis of water. BACKGROUND

[0002] In the field of hydrogen production by alkaline electrolysis of water (AWE), the existing technology has the following technical defects: insufficient power prediction accuracy: renewable energy power prediction relies on a single physical model or a large-scale data-driven model, which cannot adapt to the rapid power fluctuation characteristics of the AWE scene, and has poor adaptability to complex load characteristics, resulting in a large deviation between the predicted results and the actual demand. Weak system coordination ability: the multi-level control system lacks a dynamic coordination mechanism, and conflicts may occur when transmitting instructions at each level, and there is a response delay problem, which affects the efficiency of the system operation. Poor electrolytic cell durability: the electrolytic cell start-stop strategy does not integrate multi-dimensional life loss parameters (such as electrode loss, electrolyte aging, etc.) and real-time working condition constraints (such as temperature, pressure, etc.), and unreasonable start-stop operation shortens the service life of the electrolytic cell. Low control rule adaptability: existing control rules are mostly statically configured and cannot be adaptively adjusted according to dynamic working condition changes (such as renewable energy output fluctuation, load change, etc.) and complex constraint scenarios (such as multi-objective optimization requirements).

[0003] Based on the above defects, the existing technology cannot meet the demand for improving yield, energy efficiency and equipment life of the AWE system, and therefore there is an urgent need for an efficient and dynamic adaptive AWE control method. SUMMARY

[0004] The purpose of the present application is to solve the problems of the prior art and provide an intelligent hierarchical control system and method for hydrogen production by alkaline electrolysis of water, which improves the power prediction accuracy and load adaptability, and optimizes the hydrogen production rate and system energy efficiency; by optimizing the start-stop sequence and hot standby scheduling, the number of electrolytic cell start-stops is reduced, significantly prolonging the service life of the equipment; with the help of a dynamic arbitration mechanism, the multi-level instruction conflicts are eliminated, and the stability of the system under complex working conditions and multi-constraint scenarios is improved; embedded real-time operation is supported, which can flexibly adapt to the operation requirements of fixed or mobile AWE equipment.

[0005] The present application is realized by the following technical solutions: An intelligent hierarchical control system for hydrogen production by alkaline electrolysis of water, comprising an L3 adaptive power predictor, an L2 lookup table-mixed constraint optimizer, an L1 life integrator-dynamic hot standby scheduler, and an L0 dynamic arbitration bus; L3 adaptive power predictor: generating a power prediction curve through an adaptive power predictor, based on multi-scale wavelet transform and adaptive kernel regression, adapting to rapid load fluctuations; L2 Look-up table-mixed constraint optimizer: Through efficiency and loss look-up table, mixed constraint objective function and dynamic weight optimization algorithm, real-time adjustment of current density and electrolyte flow, realization of multi-objective balance of hydrogen yield, energy efficiency and thermal stability; L1 Life integrator-dynamic thermal standby scheduler: Based on multi-dimensional life weight calculation, dynamic programming model and adaptive temperature threshold, the start-stop sequence of electrolytic cell is optimized; when potential life loss risk is detected, the thermal standby mode is triggered to protect the equipment; L0 Dynamic arbitration bus: Using fixed priority mechanism and dynamically generated JSON interlocking rules, real-time coordination of control instructions at each level, output of conflict-free control.

[0006] An intelligent layered control method for hydrogen production by alkaline electrolysis of water, comprising the following steps: Step one, generating a power prediction curve through an adaptive power predictor, based on multi-scale wavelet transform and adaptive kernel regression, adapting to rapid load fluctuations; Step two, generating optimized control parameters through a look-up table-mixed constraint optimizer, combining look-up table method and constrained gradient descent to optimize hydrogen yield, energy efficiency and life; Step three, generating scheduling instructions through a life integrator-dynamic thermal standby scheduler, balancing yield and life based on life integration and thermal standby strategy; Step four, generating final control instructions through an arbitration bus, based on priority arbitration and multi-unit coordination, driving the AWE system to run.

[0007] Preferably, in step one, the adaptive power predictor comprises the following steps: Step S1-1, based on irradiance , wind speed and real-time power , generating baseline power through multi-scale wavelet transform: ; Wherein, is the physical baseline power, unit: kW; is the real-time irradiance, unit: W / m 2 ; is the real-time wind speed, unit: m / s; is the discrete wavelet transform; is the wavelet basis function; is the jth layer scale coefficient; J is the number of decomposition layers; , is the environmental coefficient; is the wavelet weight; is the real-time power, unit: kW; Step S1-2, calculate the error and generate the correction value through adaptive kernel regression; Step S1-3: Dynamically optimize weights based on operating conditions; Steps S1-4: Integrate baselines and correction values ​​to generate predicted power .

[0008] Preferably, in step S1-2, the error between the real-time power and the baseline power is calculated: ; Corrected values ​​are generated using adaptive kernel regression: ; in, The Gaussian kernel function; This is the physical baseline power, measured in kW. Real-time power, unit: kW; For adaptive weights, N The length of the window; This represents the prediction error; To predict power The variance.

[0009] Preferably, in steps S1-3, the weights are dynamically adjusted based on the operating mode and power fluctuation characteristics. Kernel regression bandwidth h and wavelet weights ; The weight update combines the mean square error and power fluctuation, and the objective function is: ; in, For operating condition sensing factors; Indicates the rate of change of power; This represents the prediction error; To predict power The variance; The weight update formula is: ; The kernel regression bandwidth update formula is: ; Where J is the weighted sum of the mean square error and the power volatility; The formula for updating wavelet weights is: ; in, This is the learning rate.

[0010] Preferably, in steps S1-4, the output is updated and predicted online in real time; Integrating baseline and correction values ​​to generate predicted power: ; Online model parameter updates This is achieved through incremental optimization, and the formula is: ; Output prediction curve It supports a 5–30 minute prediction window to meet real-time control requirements.

[0011] Preferably, step two includes the following steps: Step S2-1: Obtain initial control parameters through a pre-built operating condition table. ;in: I For current; F For traffic; Pr For pressure; Step S2-2: Construct the objective function, taking into account hydrogen yield, energy efficiency, and lifetime degradation; Step S2-3: Optimize control parameters using constrained gradient descent; Step S2-4: Update the table lookup content and constraint weights online.

[0012] Preferably, in step S2-1: ; in, This is a lookup function that outputs the initial current, flow rate, and pressure. It is a predicted power; Indicates the load mode.

[0013] Preferably, in step S2-2, the objective function is defined as: ; in, Hydrogen yield, in mol / s; Energy efficiency, unit: ; This refers to the lifespan degradation rate of the electrolytic cell. To constrain weights; The constraints include power constraints, current density range, flow rate range, and pressure range: ; In the initial setting: ; ; ; ; ; .

[0014] Preferably, steps S2-3 include the following steps: Perform real-time optimization by adjusting the initial parameters through an iterative algorithm to minimize the objective function. ; Using the constrained gradient descent method, the update formula is as follows: ; in, To optimize the step size and adapt to the rapidly fluctuating AWE load, L2 dynamically adjusts the constraint weights based on the operating mode. ; The weight update formula is: ; in, is the learning rate for weights.

[0015] Preferably, in steps S2-4, the optimized control parameter set is output. And update the lookup table content and constraint weights online; The table lookup and update are based on real-time feedback, and the formula is: ; The output control parameters are executed by the L1 scheduler, supporting control cycles of 5–30 minutes, with a calculation time of 2 ms.

[0016] Preferably, step three includes the following steps: Step S3-1: Estimate the cumulative damage of the electrolytic cell by lifetime integration, using the following formula: ; in, , The degradation coefficient; For reference temperature; For temperature scale; N This is the length of the integration window; Step S3-2: Assess thermal standby requirements based on temperature and lifetime degradation rate; Step S3-3: Generate scheduling instructions through constrained optimization; Step S3-4: Update weights and degradation coefficients online.

[0017] Preferably, in step S3-2, the thermal standby strategy is triggered when the temperature is below a threshold or the lifetime degradation rate exceeds the limit, and low power consumption parameters are set: ; in, ; ; .

[0018] Preferably, in step S3-3, the objective function is defined as: ; in, , , As a weighting factor; Hydrogen yield, in mol / s; Energy efficiency, unit: ; This refers to the lifespan degradation rate of the electrolytic cell. Scheduling constraints include power constraints, parameter ranges, and state transition frequencies: ; in, Maximum switching frequency; Using the constrained gradient descent method, the update formula is as follows: ; in, The scheduling step size is used; the weighting factor is dynamically adjusted, and the formula is: ; in, The learning rate is the weight. Output scheduling instruction set The lifetime integral model and weighting factors are updated online; the update formula is: ; The output scheduling instructions are executed by L0, supporting scheduling cycles of 5–30 minutes, with a computation time of <3ms.

[0019] Preferably, step four includes the following steps: Step S4-1: Prioritize based on lifespan degradation rate and operating mode, using the following formula: ; in, , As a weighting factor; Let be the lifetime degradation rate of the k-th electrolytic cell; Operating condition mode factor; Step S4-2: Detect scheduling instruction conflicts and verify global constraints; Step S4-3: Optimize and adjust instructions through multi-unit coordination; Step S4-4: Generate final control commands based on real-time feedback.

[0020] Preferably, in step S4-2, conflicts in L1 scheduling instructions are detected, and it is verified whether global power constraints and unit operation constraints are met; the global power constraints ensure that the total power does not exceed the available power. ; in, This represents the actual power consumption of the k-th unit; Unit operation constraints include: ; in, ; ; ; ; ; If a conflict is detected, proceed to the coordination step.

[0021] Preferably, in step S4-3, conflicting instructions are adjusted through a coordination algorithm to optimize the global objective function, taking into account hydrogen production rate, energy efficiency, and lifespan. The objective function is defined as: ; in, , , As a weighting factor; Hydrogen yield, in mol / s; Energy efficiency, unit: ; This refers to the lifespan degradation rate of the electrolytic cell. Using the constrained gradient descent method, the update formula is as follows: ; in, The scheduling step size is used; the weighting factor is dynamically adjusted, and the formula is: ; in, is the learning rate for weights.

[0022] Preferably, in step S4-4, the final instruction is adjusted based on real-time feedback, and control is executed; the feedback adjustment formula is: ; in, For feedback on the learning rate; Real-time power, unit: kW; Final instruction set Drive the AWE system and update the priority table: ; The output instructions support an execution cycle of 1–5 seconds, with a calculation time of <1ms.

[0023] Compared with the prior art, the present invention has the following advantages and beneficial effects: I. This invention provides an intelligent hierarchical control method for alkaline water electrolysis hydrogen production, which improves power prediction accuracy and load adaptability, while optimizing hydrogen yield and system energy efficiency; by optimizing start-stop sequences and hot standby scheduling, it reduces the number of electrolyzer start-stop cycles, significantly extending equipment life; by using a dynamic arbitration mechanism to eliminate multi-level command conflicts, it improves system stability under complex operating conditions and multi-constraint scenarios; it supports embedded real-time operation and can flexibly adapt to the operating requirements of fixed or mobile AWE equipment. Attached Figure Description

[0024] Figure 1 This is a diagram of the four-layer control architecture of the AWE system in this invention; Figure 2 This is a flowchart of the L3 adaptive power predictor in this invention; Figure 3 This is a flowchart of the L2 lookup table-hybrid constraint optimizer in this invention; Figure 4 This is a flowchart of the L1 lifetime integrator-dynamic hot standby scheduler in this invention; Figure 5 This is the flowchart for the L0 arbitration bus. Detailed Implementation

[0025] The present invention will be further described in detail below with reference to embodiments, but the implementation of the present invention is not limited thereto.

[0026] Example 1 like Figure 1 As shown, this embodiment provides an intelligent hierarchical control system for alkaline water electrolysis to produce hydrogen, including an L3 adaptive power predictor, an L2 lookup table-hybrid constraint optimizer, an L1 lifetime integrator-dynamic hot standby scheduler, and an L0 dynamic arbitration bus. L3 Adaptive Power Predictor: Generates power prediction curves through an adaptive power predictor, based on multi-scale wavelet transform and adaptive kernel regression, to adapt to rapid load fluctuations. L2 lookup table-hybrid constraint optimizer: By using efficiency and loss lookup tables, hybrid constraint objective functions and dynamic weight optimization algorithms, the current density and electrolyte flow rate are adjusted in real time to achieve a multi-objective balance of hydrogen production, energy efficiency and thermal stability. L1 Lifetime Integrator - Dynamic Hot Standby Scheduler: Based on multi-dimensional lifetime weight calculation, dynamic programming model and adaptive temperature threshold, it optimizes the start-up and shutdown sequence of the electrolyzer; when a potential lifetime loss risk is detected, it triggers the hot standby mode to protect the equipment. L0 Dynamic Arbitration Bus: Employs a fixed priority mechanism and dynamically generated JSON interlocking rules to coordinate control commands at all levels in real time and output conflict-free control.

[0027] Example 2 This embodiment employs a control method based on the intelligent stratified control system for alkaline water electrolysis hydrogen production shown in Example 1, including the following steps: Step 1: Generate a power prediction curve using an adaptive power predictor, based on multi-scale wavelet transform and adaptive kernel regression, to adapt to rapid load fluctuations. Step 2: Generate optimal control parameters using a lookup-hybrid constraint optimizer, and combine the lookup method and constrained gradient descent to optimize hydrogen yield, energy efficiency, and lifetime. Step 3: Generate scheduling instructions through the lifetime integrator-dynamic hot standby scheduler, and balance productivity and lifetime based on lifetime integration and hot standby strategy; Step 4: Generate final control commands through the arbitration bus, and drive the AWE system to run based on priority arbitration and multi-unit coordination.

[0028] Example 3 like Figures 2-5 As shown, this embodiment employs a control method based on an intelligent stratified control system for alkaline water electrolysis to produce hydrogen, as illustrated in Example 1. The method includes the following steps: Step 1: Generate a power prediction curve using an adaptive power predictor, based on multi-scale wavelet transform and adaptive kernel regression, to adapt to rapid load fluctuations. Step 2: Generate optimal control parameters using a lookup-hybrid constraint optimizer, and combine the lookup method and constrained gradient descent to optimize hydrogen yield, energy efficiency, and lifetime. Step 3: Generate scheduling instructions through the lifetime integrator-dynamic hot standby scheduler, and balance productivity and lifetime based on lifetime integration and hot standby strategy; Step 4: Generate final control commands through the arbitration bus, and drive the AWE system to run based on priority arbitration and multi-unit coordination.

[0029] The L3 adaptive power predictor employs multi-scale wavelet transform to decompose the power signal, combines adaptive kernel regression to correct errors, and improves prediction accuracy through condition-aware adaptive weight optimization and multi-parameter dynamic update mechanisms. This method supports embedded real-time computation (exemplary processing time <1 ms), requires no offline pre-training, and meets the real-time control requirements of AWE equipment under dynamic operating conditions. The output of the L3 adaptive power predictor provides reliable power prediction input for subsequent L2 (lookup table-hybrid constraint optimizer) and L1 (lifetime integrator-dynamic thermal standby scheduler), optimizing hydrogen yield, energy efficiency, and electrolyzer lifespan, and is suitable for stationary or mobile AWE systems.

[0030] In step one, the adaptive power predictor includes the following steps: Step S1-1: Generate physical baseline predictions through multi-scale wavelet transform. Based on irradiance. Wind speed and real-time power Baseline power is generated through multi-scale wavelet transform: ; in, This is the physical baseline power, in kW, ranging from 0 to 100kW. This is real-time irradiance, measured in W / m². 2 Range 0-1000W / m 2 ; This is the real-time wind speed, in m / s, ranging from 0 to 20 m / s, with a load mode (exemplary classifications: high fluctuation, medium fluctuation, low fluctuation). It is a discrete wavelet transform (e.g., the Daubechies wavelet). These are wavelet basis functions; is the scaling factor for the j-th layer; J is the number of decomposition layers; , Environmental coefficients (initially set at 0.005 and 0.002, to be updated based on actual parameters). Wavelet weights; Real-time power, unit: kW; Step S1-2: Calculate the error and generate a correction value through adaptive kernel regression; Step S1-3: Dynamically optimize weights based on operating conditions; Steps S1-4: Integrate baselines and correction values ​​to generate predicted power .

[0031] In step S1-2, the error between the real-time power and the baseline power is calculated: ; Corrected values ​​are generated using adaptive kernel regression: ; in, The Gaussian kernel function (wideband h=0.1); This is the physical baseline power, measured in kW. Real-time power, unit: kW; For adaptive weights, N The window length is 128 points. Kernel regression captures the nonlinear load characteristics of AWE, has a small number of parameters (<50k), and is suitable for embedded devices. This represents the prediction error.

[0032] In steps S1-3, the weights are dynamically adjusted based on the operating mode and power fluctuation characteristics. Kernel regression bandwidth h and wavelet weights ; The weight update combines the mean square error and power fluctuation, and the objective function is: ; in, The working condition perception factor is initially set to 0.1. Indicates the rate of change of power; This represents the prediction error; To predict power The variance; The weight update formula is: ; The kernel regression bandwidth update formula is: ; Where J is the weighted sum of the mean square error and the power volatility; The formula for updating wavelet weights is: ; in, Let be the learning rate (initially set to 0.01, 0.005, 0.005).

[0033] In steps S1-4, the output is updated and predicted online in real time. Integrating baseline and correction values ​​to generate predicted power: ; Online model parameter updates This is achieved through incremental optimization, and the formula is: ; Output prediction curve It supports a 5–30 minute prediction window to meet real-time control requirements.

[0034] The L2 lookup-hybrid constraint optimizer uses the power prediction curve output by the L3 adaptive power predictor. This method optimizes the operating parameters of the AWE system (such as current, flow rate, and pressure) to maximize hydrogen production, optimize energy efficiency, and extend the electrolyzer's lifespan. L2 combines lookup table methods with hybrid constraint optimization, using pre-built operating condition tables, dynamic constraint adjustments, and real-time optimization algorithms to adapt to the rapid load fluctuations and nonlinear response characteristics of the AWE system. This method supports embedded real-time computation (exemplary processing time <2 ms), eliminating the need for offline training and ensuring efficient control under dynamic operating conditions. L2 receives the predicted power input from L3 and collaborates with L1 (lifespan integrator-dynamic thermal standby scheduler) and L0 (arbitration bus) to form a four-layer closed-loop control architecture.

[0035] The L2 inputs include real-time parameters, configuration parameters, and the L3 output. Real-time parameters cover the current power... (Unit: kW, exemplary range 0–100 kW), electrolyzer conditions (e.g., current density) (Unit: A / cm², range 0.1–0.5 A / cm²); Temperature T(t) (unit: °C, range 30–80 °C), operating mode (high, medium, low fluctuation). Configuration parameters include lookup table resolution (100 operating points), constraint weight (initially set to 0.5), optimization step size (initially set to 0.01), and prediction window length (128 points, representing 5–30 min). L3 output is a power prediction curve. (Unit: kW), the output of L2 is the optimized set of control parameters, including current ( (Unit: A), Flow Rate ( (Unit: L / min) and "Pressure ( (Unit: bar), for execution by the L1 scheduler.

[0036] Step two includes the following steps: Step S2-1: Obtain initial control parameters through a pre-built operating condition table. ;in: I For current; F For traffic; Pr For pressure; Step S2-2: Construct the objective function, taking into account hydrogen yield, energy efficiency, and lifetime degradation; Step S2-3: Optimize control parameters using constrained gradient descent; Step S2-4: Update the table lookup content and constraint weights online.

[0037] First, initial control parameters are obtained using a lookup table method. A pre-built operating condition table stores recommended parameters (such as current, flow rate, and pressure) for the AWE system under different power inputs and operating conditions, generated based on historical operating data and equipment characteristics. The lookup process is based on the output of L3. In step S2-1, regarding the operating mode: ; in, The function is a lookup table function that outputs the initial current, flow rate, and pressure (with an initial resolution of 100 operating points). It predicts power. This method utilizes empirical mapping of the load characteristics of the AWE system to reduce computational complexity; Indicates the load mode.

[0038] In this process, a hybrid constraint optimization model is constructed by comprehensively considering hydrogen yield, energy efficiency, and electrolyzer lifespan. In step S2-2, the objective function is defined as: ; in, The hydrogen yield is expressed in mol / s, based on Faraday efficiency. Energy efficiency, unit: ; The electrolyzer lifetime degradation rate (exemplary cumulative damage model based on current density and temperature). Set the constraint weights (initially 0.5, 0.3, 0.2); The constraints include power constraints, current density range, flow rate range, and pressure range: ; In the initial setting: ; ; ; ; ; .

[0039] Steps S2-3 include the following steps: Perform real-time optimization by adjusting the initial parameters through an iterative algorithm to minimize the objective function. ; Using the constrained gradient descent method, the update formula is as follows: ; in, To optimize the step size (initially set at 0.01); to adapt to the rapidly fluctuating AWE load, L2 is increased according to the operating mode (e.g., increasing under high fluctuation conditions). The constraint weights are dynamically adjusted based on the priority of hydrogen production rate. ; The weight update formula is: ; in, The weighted learning rate is set to 0.005 initially to ensure that the optimization process adapts to dynamic operating conditions.

[0040] In steps S2-4, the optimized control parameter set is output. And update the lookup table content and constraint weights online; The table lookup update is based on real-time feedback (such as the deviation between actual and predicted power), and the formula is: ; The output control parameters are executed by the L1 scheduler, supporting control cycles of 5–30 minutes, with a calculation time of 2 ms.

[0041] The L1 lifetime integrator-dynamic hot standby scheduler dynamically schedules the operating state (operating, hot standby, shutdown) of the AWE system based on the optimized control parameters (such as current, flow rate, and pressure) output by the L2 lookup table-hybrid constraint optimizer, to balance hydrogen yield, energy efficiency, and electrolyzer lifetime. L1 uses a lifetime integrator to estimate cumulative damage to the electrolyzer, combined with a dynamic hot standby strategy and a real-time scheduling algorithm, adapting to the rapid load fluctuations and high-temperature operating characteristics of the AWE system (exemplary temperature 50–80°C). This method supports embedded real-time computation (exemplary processing time <3ms), eliminating the need for offline training and ensuring efficient scheduling under dynamic operating conditions. L1 receives the optimized parameters from L2 and collaborates with L3 (power predictor) and L0 (arbitration bus).

[0042] The inputs to L1 include real-time parameters, configuration parameters, and the L2 output. The real-time parameters cover the current power... (Unit: kW, range 0–100kW), Electrolyzer status (e.g., current density) (Unit: A / cm², range 0.1–0.5 A / cm²); Temperature T(t) (Unit: °C, range 50–80°C), operating mode (classified as high, medium, and low fluctuation). Configuration parameters include lifespan integration window length (256 points, representing 10–60 min), hot standby temperature threshold (60°C), scheduling cycle (5 min), weighting factor (0.4), and scheduling step size (0.01). L2 output is the optimized control parameter set. , representing current (unit: A / cm²), flow rate (unit: L / min), and pressure (unit: bar), respectively. The output of L1 is a set of scheduling instructions, including the operating state S(t) (operating, hot standby, off) and adjusted control parameters. It is used for L0 arbitration bus execution.

[0043] Step three includes the following steps: Step S3-1: Estimate the cumulative damage of the electrolytic cell by lifetime integration, using the following formula: ; in, , The degradation coefficients are 0.01 and 0.02. Reference temperature (50℃); Temperature scale (10°C); N The integral window length is 256 points. This formula quantifies the electrolyzer's lifespan loss, providing a basis for subsequent scheduling decisions.

[0044] Step S3-2: Assess thermal standby requirements based on temperature and lifetime degradation rate; Step S3-3: Generate scheduling instructions through constrained optimization; Step S3-4: Update weights and degradation coefficients online.

[0045] The assessment of standby requirements determines whether to enter standby mode based on temperature and lifetime degradation rate. If the temperature T(t) is below the standby threshold (60°C) or the lifetime degradation rate... If the safety threshold (0.05) is exceeded, hot standby mode is triggered, and the running status is set. And adjust the control parameters to low power consumption values: ; in, ; ; .

[0046] Among these steps, dynamic scheduling optimization is performed, comprehensively considering hydrogen production rate, energy efficiency, and lifetime degradation. In step S3-3, the objective function is defined as: ; in, , , The weighting factors are (0.4, 0.3, 0.3). Hydrogen yield, in mol / s; Energy efficiency, unit: ; This refers to the lifespan degradation rate of the electrolytic cell. Scheduling constraints include power constraints, parameter ranges, and state transition frequencies: ; in, Maximum switching frequency (1 time / 5min); Using the constrained gradient descent method, the update formula is as follows: ; in, The scheduling step size is 0.01; the weight factor is dynamically adjusted using the following formula: ; in, The weighted learning rate is 0.005. Output scheduling instruction set The lifetime integral model and weighting factors are updated online; the update formula is: ; The output scheduling instructions are executed by L0, supporting scheduling cycles of 5–30 minutes, with a computation time of <3ms.

[0047] The L0 arbitration bus receives the scheduling instruction set output by the L1 lifetime integrator-dynamic hot standby scheduler, coordinates the operating status and control parameters of multiple electrolyzer units within the AWE system, and executes the final control decision to maximize hydrogen yield, optimize energy efficiency, and ensure safe equipment operation. L0 employs a priority arbitration mechanism, conflict detection and coordination algorithms, and real-time feedback adjustment to adapt to the rapid load fluctuations and multi-unit collaborative characteristics of the AWE system (such as non-uniform load distribution). This method supports embedded real-time computation (executive processing time <1ms), eliminating the need for offline training and ensuring efficient control under dynamic operating conditions. As the bottom-level execution module of the four-layer control architecture, L0 collaborates with L3, L2, and L1 to form a closed-loop control system suitable for both fixed and mobile AWE systems.

[0048] The inputs to L0 include real-time parameters, configuration parameters, and the L1 output. The real-time parameters cover the current power... (Unit: kW, range 0–100kW), Electrolyzer status (e.g., current density) (Unit: A / cm², range 0.1–0.5 A / cm²); Temperature T (t) (Unit: °C, range 50–80 °C); k represents the k-th electrolytic cell unit, "k=1,2,…,K", "K=4"), operating mode (classified as high, medium, and low fluctuation). Configuration parameters include arbitration priority table (4 priority levels), conflict detection threshold (power deviation 0.5kW), coordination step size (0.01), and feedback update cycle (1s). L1 output is the scheduling instruction set. , representing the operating status (working, hot standby, off), current (A / cm²), flow rate (L / min), and pressure (bar) of the k-th electrolytic cell unit, respectively. The output of L0 is the final control command set. It directly drives the AWE system to execute.

[0049] Step four includes the following steps: Step S4-1: Prioritize based on lifespan degradation rate and operating mode, using the following formula: ; in, , Weighting factors (0.6, 0.4); Let be the lifetime degradation rate of the k-th electrolytic cell; The operating condition factor is 0.8 for high volatility, 0.5 for medium volatility, and 0.2 for low volatility. High-priority units execute L1 scheduling instructions first, while low-priority units may be adjusted to meet global constraints.

[0050] Step S4-2: Detect scheduling instruction conflicts and verify global constraints; Step S4-3: Optimize and adjust instructions through multi-unit coordination; Step S4-4: Generate final control commands based on real-time feedback.

[0051] In step S4-2, conflicts in L1 scheduling instructions are detected, and it is verified whether global power constraints and unit operation constraints are met. Global power constraints ensure that the total power does not exceed the available power. ; in, This represents the actual power consumption of the k-th unit; Unit operation constraints include: ; in, ; ; ; ; ; If a conflict is detected (such as total power exceeding the limit), proceed to the coordination step.

[0052] In step S4-3, conflicting instructions are adjusted through a coordination algorithm to optimize the global objective function, taking into account hydrogen production rate, energy efficiency and lifespan. By coordinating conflicting instructions using a coordination algorithm, the global objective function is optimized, comprehensively considering hydrogen production rate, energy efficiency, and lifetime. The objective function is defined as follows: ; in, , , The weighting factors are (0.4, 0.3, 0.3). Hydrogen yield, in mol / s; Energy efficiency, unit: ; This refers to the lifespan degradation rate of the electrolytic cell. Using the constrained gradient descent method, the update formula is as follows: ; in, The scheduling step size is 0.01; the weight factor is dynamically adjusted using the following formula: ; in, The weighted learning rate is 0.005.

[0053] In step S4-4, based on real-time feedback (such as actual power) With predicted power The deviation is used to adjust the final instruction and execute control; the feedback adjustment formula is: ; in, The feedback learning rate is 0.005. Real-time power, unit: kW; Final instruction set Drive the AWE system and update the priority table: ; The output instructions support an execution cycle of 1–5 seconds, with a calculation time of <1ms.

[0054] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications or equivalent changes made to the above embodiments based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. An intelligent stratified control system for alkaline water electrolysis to produce hydrogen, characterized in that: It includes an L3 adaptive power predictor, an L2 lookup table-hybrid constraint optimizer, an L1 lifetime integrator-dynamic hot standby scheduler, and an L0 dynamic arbitration bus. L3 Adaptive Power Predictor: Generates power prediction curves through an adaptive power predictor, based on multi-scale wavelet transform and adaptive kernel regression, to adapt to rapid load fluctuations. L2 lookup table-hybrid constraint optimizer: By using efficiency and loss lookup tables, hybrid constraint objective functions and dynamic weight optimization algorithms, the current density and electrolyte flow rate are adjusted in real time to achieve a multi-objective balance of hydrogen production, energy efficiency and thermal stability. L1 Lifetime Integrator - Dynamic Hot Standby Scheduler: Based on multi-dimensional lifetime weight calculation, dynamic programming model and adaptive temperature threshold, it optimizes the start-up and shutdown sequence of the electrolyzer; when a potential lifetime loss risk is detected, it triggers the hot standby mode to protect the equipment. L0 Dynamic Arbitration Bus: Employs a fixed priority mechanism and dynamically generated JSON interlocking rules to coordinate control commands at all levels in real time and output conflict-free control.

2. The control method for the intelligent stratified control system for alkaline water electrolysis to produce hydrogen according to claim 1, characterized in that, Includes the following steps: Step 1: Generate a power prediction curve using an adaptive power predictor, based on multi-scale wavelet transform and adaptive kernel regression, to adapt to rapid load fluctuations. Step 2: Generate optimal control parameters using a lookup-hybrid constraint optimizer, and combine the lookup method and constrained gradient descent to optimize hydrogen yield, energy efficiency, and lifetime. Step 3: Generate scheduling instructions through the lifetime integrator-dynamic hot standby scheduler, and balance productivity and lifetime based on lifetime integration and hot standby strategy; Step 4: Generate final control commands through the arbitration bus, and drive the AWE system to run based on priority arbitration and multi-unit coordination.

3. The intelligent stratified control method for alkaline water electrolysis to produce hydrogen according to claim 2, characterized in that, In step one, the adaptive power predictor includes the following steps: Step S1-1, based on irradiance Wind speed and real-time power Baseline power is generated through multi-scale wavelet transform: ； in, This is the physical baseline power, measured in kW. This is real-time irradiance, measured in W / m². 2 ; This is the real-time wind speed, measured in m / s. It is a discrete wavelet transform; These are wavelet basis functions; is the scaling factor for the j-th layer; J is the number of decomposition layers; , For environmental factors; Wavelet weights; Real-time power, unit: kW; Step S1-2: Calculate the error and generate a correction value through adaptive kernel regression; Step S1-3: Dynamically optimize weights based on operating conditions; Steps S1-4: Integrate baselines and correction values ​​to generate predicted power .

4. The intelligent stratified control method for alkaline water electrolysis to produce hydrogen according to claim 3, characterized in that, In step S1-2, the error between the real-time power and the baseline power is calculated: ； Corrected values ​​are generated using adaptive kernel regression: ； in, The Gaussian kernel function; This is the physical baseline power, measured in kW. Real-time power, unit: kW; For adaptive weights, N The length of the window; This represents the prediction error.

5. The intelligent stratified control method for alkaline water electrolysis hydrogen production according to claim 4, characterized in that, In steps S1-3, the weights are dynamically adjusted based on the operating mode and power fluctuation characteristics. Kernel regression bandwidth h and wavelet weights ; The weight update combines the mean square error and power fluctuation, and the objective function is: ； in, For operating condition sensing factors; Indicates the rate of change of power; This represents the prediction error; To predict power The variance; The weight update formula is: ； The kernel regression bandwidth update formula is: ； Where J is the weighted sum of the mean square error and the power volatility; The formula for updating wavelet weights is: ； in, This is the learning rate.

6. The intelligent stratified control method for alkaline water electrolysis to produce hydrogen according to claim 5, characterized in that, In steps S1-4, the output is updated and predicted online in real time. Integrating baseline and correction values ​​to generate predicted power: ； Online model parameter updates This is achieved through incremental optimization, and the formula is: ； Output prediction curve It supports a 5–30 minute prediction window to meet real-time control requirements.

7. The intelligent stratified control method for alkaline water electrolysis to produce hydrogen according to claim 6, characterized in that, Step two includes the following steps: Step S2-1: Obtain initial control parameters through a pre-built operating condition table. ;in: I For current; F For traffic; Pr For pressure; Step S2-2: Construct the objective function, taking into account hydrogen yield, energy efficiency, and lifetime degradation; Step S2-3: Optimize control parameters using constrained gradient descent; Step S2-4: Update the table lookup content and constraint weights online.

8. The intelligent stratified control method for alkaline water electrolysis to produce hydrogen according to claim 7, characterized in that, In step S2-1: ； in, This is a lookup function that outputs the initial current, flow rate, and pressure. It is a predicted power; Indicates the load mode.

9. The intelligent stratified control method for alkaline water electrolysis to produce hydrogen according to claim 8, characterized in that, In step S2-2, the objective function is defined as: ； in, Hydrogen yield, in mol / s; Energy efficiency, unit: ; This refers to the life degradation rate of the electrolytic cell. To constrain weights; The constraints include power constraints, current density range, flow rate range, and pressure range: ； In the initial setting: ; ; ; ; ; .

10. The intelligent stratified control method for alkaline water electrolysis to produce hydrogen according to claim 9, characterized in that, Steps S2-3 include the following steps: Perform real-time optimization by adjusting the initial parameters through an iterative algorithm to minimize the objective function. ; Using the constrained gradient descent method, the update formula is as follows: ； in, To optimize the step size and adapt to the rapidly fluctuating AWE load, L2 dynamically adjusts the constraint weights based on the operating mode. ; The weight update formula is: ； in, is the learning rate for weights.

11. The intelligent stratified control method for alkaline water electrolysis to produce hydrogen according to claim 10, characterized in that, In steps S2-4, the optimized control parameter set is output. And update the lookup table content and constraint weights online; The table lookup and update are based on real-time feedback, and the formula is: ； The output control parameters are executed by the L1 scheduler, supporting control cycles of 5–30 minutes, with a calculation time of 2 ms.

12. The intelligent stratified control method for alkaline water electrolysis to produce hydrogen according to claim 11, characterized in that, Step three includes the following steps: Step S3-1: Estimate the cumulative damage of the electrolytic cell by lifetime integration, using the following formula: ； in, , The degradation coefficient; For reference temperature; For temperature scale; N This is the length of the integration window; Step S3-2: Assess thermal standby requirements based on temperature and lifetime degradation rate; Step S3-3: Generate scheduling instructions through constrained optimization; Step S3-4: Update weights and degradation coefficients online.

13. The intelligent stratified control method for alkaline water electrolysis to produce hydrogen according to claim 12, characterized in that, In step S3-2, the thermal standby strategy is triggered when the temperature is below a threshold or the lifetime degradation rate exceeds the limit, and low power consumption parameters are set: ； in, ; ; 。 14. The intelligent stratified control method for alkaline water electrolysis to produce hydrogen according to claim 13, characterized in that, In step S3-3, the objective function is defined as: ； in, , , As a weighting factor; Hydrogen yield, in mol / s; Energy efficiency, unit: ; This refers to the life degradation rate of the electrolytic cell. Scheduling constraints include power constraints, parameter ranges, and state transition frequencies: ； in, Maximum switching frequency; Using the constrained gradient descent method, the update formula is as follows: ； in, The scheduling step size is used; the weighting factor is dynamically adjusted, and the formula is: ； in, The learning rate is the weight. Output scheduling instruction set The lifetime integral model and weighting factors are updated online; the update formula is: ； The output scheduling instructions are executed by L0, supporting scheduling cycles of 5–30 minutes, with a computation time of <3ms.

15. The intelligent stratified control method for alkaline water electrolysis to produce hydrogen according to claim 14, characterized in that, Step four includes the following steps: Step S4-1: Prioritize based on lifespan degradation rate and operating mode, using the following formula: ； in, , As a weighting factor; Let be the lifetime degradation rate of the k-th electrolytic cell; Operating condition mode factor; Step S4-2: Detect scheduling instruction conflicts and verify global constraints; Step S4-3: Optimize and adjust instructions through multi-unit coordination; Step S4-4: Generate final control commands based on real-time feedback.

16. The intelligent stratified control method for alkaline water electrolysis to produce hydrogen according to claim 15, characterized in that, In step S4-2, conflicts in L1 scheduling instructions are detected, and it is verified whether global power constraints and unit operation constraints are met; the global power constraints ensure that the total power does not exceed the available power. ； in, This represents the actual power consumption of the k-th unit; Unit operation constraints include: ； in, ; ; ; ; ; If a conflict is detected, proceed to the coordination step.

17. The intelligent stratified control method for alkaline water electrolysis to produce hydrogen according to claim 16, characterized in that, In step S4-3, conflicting instructions are adjusted through a coordination algorithm to optimize the global objective function, taking into account hydrogen production rate, energy efficiency and lifespan. The objective function is defined as: ； in, , , As a weighting factor; Hydrogen yield, in mol / s; Energy efficiency, unit: ; This refers to the life degradation rate of the electrolytic cell. Using the constrained gradient descent method, the update formula is as follows: ； in, The scheduling step size is used; the weighting factor is dynamically adjusted, and the formula is: ； in, is the learning rate for weights.

18. The intelligent stratified control method for alkaline water electrolysis to produce hydrogen according to claim 17, characterized in that, In step S4-4, the final instruction is adjusted based on real-time feedback, and control is executed; the feedback adjustment formula is: ； in, For feedback on the learning rate; Real-time power, unit: kW; Final instruction set Drive the AWE system and update the priority table: ； The output instructions support an execution cycle of 1–5 seconds, with a calculation time of <1ms.

Images