Water electrolysis hydrogen production intelligent control system and method based on artificial intelligence

Abstract

The invention discloses a water electrolysis hydrogen production intelligent control system and method based on artificial intelligence. The system and method are suitable for a large-scale water electrolysis hydrogen production scene under the power supply condition of fluctuating renewable energy sources such as wind power and photovoltaic. The system comprises a data acquisition layer, an edge calculation layer, an intelligent control layer and an execution layer. Dynamic modeling of the running state of the hydrogen production system is achieved through multi-parameter real-time monitoring and feature extraction. The intelligent control layer fuses an LSTM prediction module and a reinforcement learning controller, the LSTM prediction module is used for predicting future renewable energy input and hydrogen demand trends, and the reinforcement learning controller calculates an optimal current density set value based on a prediction result so as to maximize hydrogen production efficiency per unit energy consumption. And meanwhile, a digital twinning technology and a safety protection mechanism are combined, thermoelectric dual regulation and control of the electrolysis process are achieved, and the control response speed and the system stability are improved. According to the invention, energy consumption can be effectively reduced, the service life of the stack is prolonged, and the operation efficiency and reliability of the hydrogen production system under complex load are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to water electrolysis hydrogen production technology, in particular to an intelligent control system and method for water electrolysis hydrogen production based on artificial intelligence, and belongs to the technical field of cross between new energy hydrogen production equipment automation control and artificial intelligence application. BACKGROUND

[0002] With the continuous advancement of carbon neutralization goals, hydrogen energy, as an important part of clean energy, has gradually shown great development potential in the fields of industry, transportation and energy storage. Among many hydrogen production technologies, water electrolysis hydrogen production has become a key path for coupling renewable energy to achieve green hydrogen production due to its green process, no carbon emissions, and strong adaptability. Especially in the context of widespread intermittent power sources such as wind power and photovoltaic power, water electrolysis hydrogen production as the central bridge of "electricity-gas" conversion, its intelligent control technology has attracted increasing attention.

[0003] However, the current mainstream water electrolysis system still mainly uses static parameter setting in terms of control strategy, and generally relies on fixed current density operation, supplemented by traditional PID or PI control strategy to adjust the operating state. This type of control method often has problems such as response lag and poor adaptability when facing intermittent and volatile wind power and photovoltaic power supply conditions. On the one hand, the uncertainty of renewable energy output can lead to frequent start and stop of electrolytic cells, which in turn induces the frequent alternation of thermal expansion and cold contraction of the electrolytic cell, causing fatigue aging, deactivation and even rupture of the membrane electrode structure, significantly shortening its service life. On the other hand, in the case of lack of intelligent response capability to external disturbances, the fixed load operation strategy is easy to cause high energy consumption, low hydrogen production efficiency, and even invalid power consumption, thereby increasing the overall operation and maintenance cost and weakening the economic competitiveness of the hydrogen production system.

[0004] At the same time, since the water electrolysis reaction process itself involves multiple physical fields (electricity, heat, mass, and chemical), its operating state is influenced by multiple factors such as temperature, pressure, voltage, current, electrolyte concentration, and gas distribution. In actual operation, these parameters exhibit obvious time-varying, nonlinear, and coupled characteristics, and traditional control systems usually lack multi-parameter coordinated control and dynamic prediction mechanisms, making it difficult to achieve accurate identification and feedforward control of complex working conditions. In addition, most existing systems lack early warning and intelligent response mechanisms for key safety indicators in the electrolysis process, such as local overheating, pressure surge, and abnormal electrolyte flow rate, which may cause serious safety accidents such as thermal runaway, leakage, and explosion in extreme cases. SUMMARY

[0005] To address the aforementioned technical issues, this invention provides an intelligent control system and method for hydrogen production via water electrolysis based on artificial intelligence. This system is applicable to large-scale hydrogen production scenarios using water electrolysis coupled with fluctuating renewable energy sources (wind power / photovoltaics, etc.) and aims to achieve synergistic control objectives of optimal energy efficiency, system safety, and extended lifespan.

[0006] Specifically, the technical solution provided by this invention is as follows: An intelligent control system for hydrogen production via water electrolysis based on artificial intelligence includes: Data acquisition layer: used to collect in real time the operating status parameters of the water electrolysis hydrogen production system and the power supply status parameters of the renewable energy system. The operating status parameters include temperature, pressure, voltage, current, electrolyte concentration, hydrogen production and coolant status. The power supply status parameters include real-time power output information. Edge computing layer: used to preprocess the data collected by the data acquisition layer and transmit it to the intelligent control layer; The intelligent control layer includes an LSTM prediction module and a reinforcement learning controller. The LSTM prediction module is used to predict the renewable energy power fluctuation trend and hydrogen load demand trend within a preset time period based on historical operating status data. The reinforcement learning controller calculates the target current density setpoint based on the current system status and the predicted trend. Execution layer: used to adjust the power output and the working state of the thermal management unit in real time according to the target current density set value. The power output adjusts the electrolysis voltage waveform through high-frequency pulse modulation, and the thermal management unit adjusts the coolant flow rate through a fuzzy PID algorithm to achieve temperature control of the electrolytic cell.

[0007] Furthermore, the system also includes a digital twin module, which combines physical mechanism equations and neural network models to construct a hybrid model for simulating hydrogen production efficiency under different operating conditions. This module supports offline training and simulation optimization of control strategies, and fine-tunes the reinforcement learning controller to adapt to electrolyzer electrode aging and gradual changes in system parameters.

[0008] Furthermore, the edge computing layer includes an anomaly detection module, which is built based on a fusion perception and residual threshold determination mechanism. This module can perform anomaly analysis on multiple operating parameters and trigger an alert when an abnormal state occurs.

[0009] Furthermore, the anomaly detection module employs a multidimensional dynamic sliding analysis method with a time window, and calculates the anomaly score based on the following formula:

[0010] in, Indicates time t The parameter feature vector, is a parameter mean vector within a history window, is a parameter standard deviation vector within a history window, ||·|| is a vector norm, which is used to quantify the deviation degree of the current state from the stable state in history; when the abnormal score value exceeds the set threshold, the system triggers the corresponding early warning response.

[0011] Further, the system has a security protection mechanism triggered automatically by the anomaly detection module, which includes a three-level response control strategy: Primary response: automatically adjust the target current density to the safety threshold when the anomaly appears; Secondary response: automatically switch to the backup cooling system and reduce the electrolysis power when local overheating occurs; Tertiary response: when there is a serious abnormal risk, perform power-off protection and disconnect the main loop of the electrolytic cell.

[0012] An intelligent control method for water electrolysis hydrogen production based on the above system, comprising the steps of: Data acquisition step: real-time acquisition of the operating state parameters of the water electrolysis hydrogen production system and the energy supply state parameters of the renewable energy system; Data preprocessing and feature extraction step: pre-process the collected data and input it into the prediction model; Trend prediction step: based on the LSTM prediction model, short-term prediction of renewable energy power supply trend and hydrogen demand in the future period; Strategy optimization step: input the prediction results and current operating state into the reinforcement learning controller, and output the target current density set value through the strategy network; Instruction execution step: according to the target current density set value, control the pulse power supply to adjust the output waveform parameters.

[0013] Further, the method further includes an anomaly detection and safety control step: continuously monitor the system operating state, calculate the residual score through the edge-deployed anomaly detection module, and if the score exceeds the preset threshold, trigger a multi-level safety protection mechanism, including current limiting protection, cooling switching and emergency power-off operation.

[0014] Preferably, the LSTM prediction model includes a three-layer stacked LSTM network, supplemented by an attention mechanism and a Dropout regularization layer, and outputs new energy power generation prediction values and hydrogen demand prediction values for the next 15 minutes, 30 minutes and 60 minutes through a fully connected layer.

[0015] Further, the reinforcement learning controller selects the current optimal action from a set of action consisting of multiple current density set values in each control cycle, which are arranged at equal intervals from low to high, and adjusted in a step-by-step manner each time. The system will select the optimal action as the set instruction for the next cycle.

[0016] An electronic device comprising a memory, a processor and a computer program stored on the memory, the processor implementing the steps of the above method when executing the program.

[0017] The present application breaks the limitation of traditional water electrolysis system relying on fixed current density and static control parameters, first introduces reinforcement learning and LSTM prediction model to construct dynamic closed-loop control mechanism, so that the system has forward-looking load response ability and real-time strategy optimization ability. Under the condition of coupling fluctuating energy sources such as wind power and photovoltaic, the present application can realize the early perception and strategy pre-adjustment of load disturbance by predicting the future energy input and hydrogen demand trend, thereby significantly improving the system operation stability and hydrogen output stability.

[0018] In addition, the reinforcement learning controller of the present application can dynamically optimize between unit energy consumption optimization and system safety, and train a control strategy that can adapt to various operating conditions through environmental interaction, not only improving hydrogen production rate, but also prolonging the service life of membrane electrode and reducing maintenance frequency. Combined with the virtual simulation platform constructed by the digital twin model, the system can also realize offline training and online optimization update of the control strategy, has the ability of continuous learning and strategy evolution, and has strong robustness in dealing with slow variable interference such as environmental change, electrolyte aging and equipment degradation.

[0019] At the same time, the present application constructs a multi-level safety protection mechanism based on edge computing, which can automatically perform multi-level response operations such as current limiting, switching cooling and emergency power-off at the early stage of potential risks such as local overheating and voltage abnormalities, thereby improving the operation safety and fault prevention and control ability of the system. The execution layer adopts high-frequency pulse power regulation and fuzzy PID thermal management system, which can accurately adjust the voltage waveform and cooling liquid flow, effectively control the thermal stress fluctuation while ensuring the hydrogen production efficiency, and prevent the fatigue damage of the stack.

[0020] In summary, the present application not only realizes intelligent prediction, strategy optimization and closed-loop control of the hydrogen production process, but also constructs a full-link intelligent control system from perception to decision-making to execution, which is superior to the prior art in energy efficiency optimization, system safety, adaptive ability and intelligent level, and has good engineering application value and promotion prospect. BRIEF DESCRIPTION OF DRAWINGS

[0021] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, which together with the embodiments of the present application, serve to explain the present application, and do not constitute a limitation of the present application.

[0022] Figure 1 is a water electrolysis hydrogen production intelligent control system framework schematic diagram provided by an embodiment of the present application. DETAILED DESCRIPTION

[0023] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some embodiments but not all of the present application. Based on the embodiments in the present application, other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0024] Embodiment one The embodiment provides an intelligent control system for hydrogen production by water electrolysis based on artificial intelligence, which mainly comprises a data acquisition layer, an edge computing layer, an intelligent control layer and an execution layer.

[0025] (1) Data acquisition layer As the basic perception layer of the whole intelligent control system, the data acquisition layer undertakes the real-time monitoring task of the whole process, multiple parameters and multiple time domains of the hydrogen production process. By integrating multiple high-precision sensors and monitoring modules, the data acquisition layer constructs a high-time-efficiency and high-reliability physical perception network, which can continuously and stably obtain the state information of the hydrogen production system and the renewable energy system that supplies power for it.

[0026] For the hydrogen production system, the data acquisition layer includes a sensor network deployed inside the electrolytic cell and at key process nodes. The temperature sensors in the network are distributed in the electrolytic cell liquid and wall surface, cooling system inlet and outlet pipeline and other areas to obtain the local and overall thermal field changes in real time; the pressure sensor is used to detect the system pressure fluctuation caused by gas generation during the electrolysis process; at the same time, the current and voltage sensors continuously monitor the input electrical parameters. In addition, chemical sensors are configured to monitor the concentration change of electrolyte, which reflects the evolution trend of catalyst activity and electrolysis efficiency by detecting the concentration change of KOH, NaOH and other alkaline electrolytes; the gas product part is monitored in real time by high-precision hydrogen and oxygen flow meters; the cooling system related sensors record the temperature, flow rate and flow information of the cooling liquid in real time.

[0027] For the renewable energy system, an energy input monitoring module is specially set up to monitor the real-time power supply information on the renewable energy side. By connecting with the direct current or alternating current output port of the photovoltaic module or wind power system, the module can collect power, voltage, current and frequency and other input electrical characteristics in real time, and can also record the fluctuation frequency, power disturbance amplitude and other key indicators of the energy side synchronously.

[0028] All sensors and monitoring modules in the data acquisition layer communicate with the central controller via edge devices, employing standard industrial protocols (such as Modbus TCP, CAN, OPC UA, etc.) to achieve efficient data transmission and redundancy protection. The acquired data stream enters the intelligent control layer through edge computing nodes, forming a bridge for transmitting information from the underlying physical world to the upper-level intelligent control layer.

[0029] (2) Edge computing layer As the intermediary link connecting physical sensing and intelligent decision-making, the edge computing layer undertakes the key functions of data preprocessing and real-time computing. Deployed in edge devices at the hydrogen production site, such as industrial gateways, embedded edge servers, and industrial-grade GPU modules, this computing layer features high reliability, high real-time performance, and low latency, enabling it to process sensing data rapidly within millisecond-level response cycles.

[0030] In actual operation, the edge computing layer first receives a large amount of multi-source heterogeneous information from the data acquisition layer. This information includes dynamic variables such as temperature, pressure, voltage, current, gas production, cooling status, and renewable energy input. To ensure the accuracy and robustness of subsequent control algorithms, a complete data preprocessing mechanism is integrated on the edge device. This mechanism mainly includes three core steps: data filtering, normalization, and feature extraction. Data filtering employs methods such as sliding window mean, median filtering, and FIR filtering to filter out short-term pulse interference caused by sensor noise or operating condition fluctuations. Normalization maps each parameter to a unified dimension range, such as [0,1] or [-1,1], based on its historical distribution range, to facilitate efficient convergence and generalization of the neural network model. In feature extraction, statistical features such as local mean, rate of change, and frequency distribution are calculated based on the temporal characteristics and frequency domain response of different variables to construct a high-dimensional state space for the controller input.

[0031] In some embodiments, the edge computing layer further undertakes the pre-emptive identification task for system security protection, namely, anomaly detection. This function, based on a lightweight algorithm, performs real-time modeling and state classification of the processed data stream. When the model detects anomalies that significantly deviate from the normal state distribution, such as local temperature increases, voltage spikes, or drastic changes in electrolyte concentration, it immediately marks them as potential faults and automatically triggers multi-level security protection mechanisms. In this embodiment, a multi-dimensional dynamic sliding analysis method with a time window is used, and the anomaly score is calculated based on the following formula:

[0032] in, Indicates time t The parameter feature vector, This is a vector of the mean values ​​of parameters within the historical window. Let ||·|| be the standard deviation vector of parameters within the historical window, and ||·|| be the vector norm, used to quantify the deviation of the current state from the historical stable state. When the anomaly score exceeds a set threshold, the system triggers a corresponding early warning response. This process is completed entirely in a closed loop within the local edge device, without relying on cloud computing, ensuring the real-time performance and reliability of the response. This is particularly suitable for the safe operation requirements of hydrogen production systems in highly dynamic environments such as those with drastic fluctuations in wind and solar power.

[0033] (3) Intelligent control layer As the most decision-making and scheduling-capable key component of the system, the intelligent control layer undertakes multiple tasks, including real-time optimization control, predictive regulation strategies, system modeling, and simulation. It not only achieves intelligent perception and response to the dynamic environment of complex industrial sites but also integrates advanced artificial intelligence technologies such as reinforcement learning, sequence modeling, and digital twins. While ensuring system stability and safety, it maximizes hydrogen production efficiency and equipment lifespan.

[0034] The first key component of the intelligent control layer is the reinforcement learning controller. This module is based on mainstream reinforcement learning algorithms such as DQN (Deep Q Network) or PPO (Proximal Policy Optimization). Through continuous observation and trial-and-error mechanisms of the current system state, it learns the optimal electrolyzer current density setting strategy. The state space of the reinforcement learning controller includes multiple real-time parameter dimensions, including but not limited to current temperature, pressure, voltage, current, electrolyte concentration, hydrogen demand, and renewable energy supply status. Through continuous interaction with the environment, the controller adjusts the policy network under the guidance of the reward function to maximize the amount of hydrogen produced per unit of energy consumption, thereby achieving dynamic optimization control under different load conditions. This controller can be fully trained offline using a digital twin module and adapts to real-time field operation data after going online, possessing strong generalization and policy update capabilities.

[0035] Working in conjunction with the reinforcement learning controller is an LSTM (Long Short-Term Memory) prediction module. This module is primarily used for short-term predictions of future renewable energy fluctuations and changes in downstream hydrogen demand. Its prediction time window can be dynamically set according to the operating strategy, typically ranging from 15 minutes to 1 hour. By inputting historical power curves, weather forecast data, hydrogen storage tank status, and user-side hydrogen consumption data, the LSTM prediction module can output predicted supply and demand trends for future time periods. This predictive information serves as a crucial input to the reinforcement learning controller, enabling the control strategy to be forward-looking. It not only optimizes decisions under current conditions but also prepares in advance for upcoming load changes, thereby reducing system start-up and shutdown frequency, mitigating membrane electrode thermal stress fluctuations, and extending equipment lifespan.

[0036] The digital twin module provides virtual modeling and simulation support for the entire intelligent control layer. Its core idea is to establish a virtual model outside the physical world that mirrors the behavior of a real electrolyzer. This model is constructed using both physical mechanisms and data-driven methods. On one hand, based on electrochemical mechanism equations, such as the electrode reaction kinetics described by the Butler-Volmer equation, the potential change patterns revealed by the Nernst equation, and the Tafel slope, the module accurately reproduces the multi-field coupling behavior of heat, electricity, and chemistry during electrolysis. On the other hand, a neural network model trained using historical operating data and experimental data compensates for unknown disturbances and system nonlinearities, thus constructing a high-precision, low-data-dependency composite modeling system. The digital twin module can not only be used for online real-time simulation of electrolyzer operation but also supports offline simulation for control strategy testing, new parameter optimization, and equipment pre-maintenance prediction. Furthermore, this module also serves as an offline training platform, enabling virtual testing and update verification of control strategies without interfering with on-site operation.

[0037] (4) Execution layer The execution layer, as the final carrier of control decisions, directly impacts the power management modules of the electrolyzer, cooling system, and renewable energy system. This execution layer not only enables rapid response transmission from the intelligent control layer to field devices but also, through the introduction of advanced power regulation technology and adaptive thermal management mechanisms, endows the entire system with multiple capabilities, including real-time control, electrical efficiency optimization, and thermal stress regulation. It serves as a crucial interface connecting algorithmic logic and engineering entities.

[0038] In the electrolytic hydrogen production process, the regulation of current density is a key factor affecting hydrogen yield and energy consumption. Traditional constant current power supply methods cannot adapt to the drastic power fluctuations of renewable energy sources such as wind and solar power, often resulting in severe polarization of the electrolyzer, high energy consumption, and even electrode damage. Therefore, this embodiment incorporates an adaptive power management module in the execution layer, which uses high-frequency pulse modulation technology to reshape and regulate the electrolysis voltage waveform in real time. This module can precisely adjust the pulse width, frequency, and duty cycle according to the current density setpoint given by the intelligent control layer, thereby dynamically changing the current input method and ensuring that the electrolyzer remains in the optimal reaction region under different load conditions. This pulsed power supply method not only significantly reduces polarization resistance and energy loss but also activates the electrochemical reaction activity on the electrode surface to a certain extent, improving catalytic efficiency. More importantly, this module has millisecond-level response capability, enabling it to make electrical adjustments in advance to anticipate energy fluctuation trends predicted by LSTM, ensuring the system maintains high stability under dynamic input conditions.

[0039] Furthermore, electrolyzers often face problems such as thermal stress accumulation, localized overheating, and cooling imbalance during long-term operation. Improper control can lead to issues ranging from electrode performance degradation to serious safety hazards like membrane electrode erosion and even explosions. Therefore, this system incorporates an intelligent thermal management system at the execution layer. By continuously monitoring the temperature distribution at key nodes and combining it with fuzzy control logic, it dynamically adjusts the coolant flow rate, circulation frequency, and switching strategy. Compared to traditional constant flow or start-stop control methods, fuzzy PID control can make more precise adjustments based on real-time temperature changes and cooling response delays, maintaining the core temperature of the electrolyzer within the efficient operating range of 88±2℃. This not only helps improve reaction kinetics and hydrogen generation rates but also effectively suppresses the damage to the interface between the membrane electrode and the electrode frame caused by thermal expansion stress, thereby extending the equipment's service life.

[0040] The execution layer units communicate at high frequency via an industrial fieldbus, enabling rapid command issuance and real-time feedback. When the intelligent control layer issues a new current density adjustment command or triggers a safety contingency plan, the execution layer can complete the response action in a very short time and feed the result back to the upper-level system for status confirmation and policy updates.

[0041] Example 2 Based on the above system, this embodiment provides an artificial intelligence-based water electrolysis hydrogen production control method, referencing... Figure 1 As shown, the method mainly includes the following steps: The first stage is dynamic optimization control. In this stage, the on-site deployed sensor network periodically collects comprehensive data on the electrolyzer's operating status and environmental conditions, covering multiple key parameters such as temperature, pressure, current, voltage, hydrogen production, and renewable energy input. This data is sent to an edge computing device in real time, preprocessed, and then input into the LSTM prediction module. Based on historical data sequences, this module makes short-term predictions of the system dynamics for the next 15 minutes to 1 hour. The prediction results mainly include the fluctuation range of renewable energy power and the changing trend of hydrogen consumption.

[0042] In this embodiment, multi-source data highly correlated with system operation over a past period (e.g., the last 2 hours) is continuously collected via an edge device and energy management interface. This data includes, but is not limited to: wind / solar input power time series, grid voltage, current and frequency, external environmental parameters (light intensity, wind speed, temperature and humidity, etc.), actual hydrogen production and consumption, and system operating status (coolant temperature, electrolyte concentration, fuel cell stack voltage, etc.). This multimodal data is aligned with a unified timestamp to form a tensor structure and then fed into the input interface of the LSTM prediction module.

[0043] The LSTM prediction module deploys an LSTM prediction model, with a core structure of a three-layer stacked LSTM network, supplemented by an attention mechanism and a Dropout regularization layer. The model structure design is as follows: Input layer: Accepts tensors of the structure [batch_size, time_step, feature_dim]; The first layer of LSTM: captures short-cycle fluctuations (such as a rapid decrease in wind power within 5 minutes); Second-layer LSTM: Identifies medium-term trends (such as a gradual decline caused by weather changes); The third layer of LSTM: remembers long-term dependence (such as day-night cycle, weekend load cycle); Attention layer: Enhances responsiveness to critical time segments (such as power abrupt changes); Fully connected output layer: outputs the predicted values ​​of new energy power generation and hydrogen demand (load) for the next 15 minutes, 30 minutes and 60 minutes (both are regression results); Loss function: The weighted multi-objective mean square error (MSE) loss function is adopted to balance the energy input forecast error and the load forecast error. Training strategy: Offline training using historical operational data. Considering the changes in wind and light characteristics and system status under different seasons and operational cycles, a periodic online optimization and cloud training mechanism is adopted: past operational data samples are automatically uploaded to the cloud every 24 hours; existing model weights are fine-tuned based on transfer learning or incremental LSTM.

[0044] In actual deployment, the LSTM prediction model is compressed into a lightweight model using methods such as TensorRT or ONNX, supporting high-frequency prediction calls at the second level. During system operation, the prediction process is executed once every 5 seconds.

[0045] Then, based on the above prediction results and the current system state, the reinforcement learning controller of the intelligent control layer calculates the optimal electrolysis current density setpoint. In this embodiment, the control problem is formally modeled as a Markov Decision Process (MDP), which is essentially a reinforcement learning structure that can learn the optimal control strategy through environmental interaction.

[0046] In this modeling framework, the controller exists in a multi-dimensional state space at every moment. This state vector consists of multiple operating parameters closely related to the electrolysis process, denoted as . ,in, Instantaneous power of renewable energy (kW). To predict future H-step power trends using LSTM For downstream hydrogen demand (Nm3 / h), The current voltage (V). This represents the current temperature of the electrolytic cell (°C). This represents the current electrolyte concentration (mol / L). The current hydrogen production rate (Nm 3 / h), The cumulative operating time (h) of the membrane electrode. This refers to the pressure (bar) or capacity percentage (%) of the hydrogen storage tank.

[0047] The corresponding action space design focuses on the control of current density. To facilitate the execution of control commands, this embodiment discretizes the current density space, constructing a finite set of actions composed of multiple setpoints. These current density setpoints are typically arranged at equal intervals from low to high, for example, from 0.1 A / cm². 2 Up to 1.2 A / cm 2 Each adjustment can be made in steps, and the system will select the optimal action as the setting instruction for the next cycle.

[0048] Within each control cycle, the reinforcement learning controller calculates the optimal action selection based on the current state and historical interaction experience. This process is based on an optimization objective function:

[0049] in, For electrical energy input (kWh), The optimal operating temperature is 88℃. , , This is the weighting factor.

[0050] To achieve policy learning and convergence, this embodiment employs the mainstream Proximal Policy Optimization (PPO) algorithm. This algorithm boasts advantages such as high sample efficiency, stable training process, and fast convergence speed. Its training objective function is constrained by the policy probability ratio, improving policy performance while avoiding the risk of policy collapse caused by large step-size updates. The policy network uses a feedforward neural network structure, taking the current state as input and outputting the probability distribution of actions at various current densities. A parallel value function network is used to estimate the expected total reward of the current state and calculate the advantage function, thereby guiding the direction of policy gradient updates.

[0051] During training, the reinforcement learning controller continuously interacts with the digital twin model, collecting trajectory samples of state-action-reward-next state, and optimizing the policy based on these samples. Through continuous iteration, the controller gradually learns how to select the most suitable current density setpoint under different energy input fluctuations and hydrogen demand scenarios, thereby realizing the transformation of the control policy from random exploration to policy convergence.

[0052] In actual operation, whenever the reinforcement learning controller senses a new state, it can complete strategy invocation within tens of milliseconds, output optimal control commands, and send them to the power management and thermal control module of the execution layer via the industrial communication network. The execution layer then dynamically adjusts the pulse power supply parameters and coolant flow rate accordingly, achieving a fast, flexible, and efficient physical control response.

[0053] Meanwhile, the system's safety protection mechanism is constantly under real-time monitoring. Edge computing devices continuously run anomaly detection algorithms, and once signs such as a rapid rise in temperature, sudden pressure changes, or abnormal voltage waveforms are detected in a certain area, a response mechanism is immediately triggered. This mechanism first automatically reduces the electrolysis current to a set safety threshold to slow down the system's response rate; if overheating persists, a backup cooling channel is immediately activated for rapid thermal regulation; in extreme risk situations, the system will forcibly shut down power to prevent fires or explosions. This entire safety response process requires no manual intervention, possessing extremely high autonomy and speed, adapting to the frequent fluctuations in operational risks under renewable energy power supply conditions such as wind and solar power.

[0054] Throughout the continuous operation of the system, the implementation of energy efficiency optimization strategies is ongoing. In the offline phase, researchers can utilize a digital twin module to import different temperature, pressure, and electrolyte ratio conditions into a virtual environment. Combined with a genetic algorithm, they can perform a global search and modeling of the current-hydrogen production response surface, thereby establishing a database of optimal operating conditions under multiple environmental variables. During actual operation, the system collects daily operating data at set intervals and uses an incremental training mechanism to fine-tune the reinforcement learning strategy network, enabling it to continuously adapt to slowly evolving system parameters such as electrolyzer electrode aging, electrolyte aging, or changes in energy efficiency.

[0055] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented using software plus a general-purpose hardware platform, or of course, using hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0056] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; under the concept of the present invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the present invention as described above, which are not provided in detail for the sake of brevity; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. An artificial intelligence-based intelligent control system for hydrogen production by water electrolysis, characterized by, The system comprises a data acquisition layer, an edge computing layer, an intelligent control layer, and an execution layer. The data acquisition layer is configured to collect real-time operation state parameters of a water electrolysis hydrogen production system and energy supply state parameters of a renewable energy system, wherein the operation state parameters include temperature, pressure, voltage, current, electrolyte concentration, hydrogen production, and cooling liquid state, and the energy supply state parameters include real-time power output information. The edge computing layer is configured to pre-process the data collected by the data acquisition layer and transmit the pre-processed data to the intelligent control layer. The intelligent control layer comprises an LSTM prediction module and a reinforcement learning controller, wherein the LSTM prediction module is configured to predict renewable energy power fluctuation trends and hydrogen load demand trends in a future preset time period based on historical operation state data, and the reinforcement learning controller is configured to calculate a target current density set value based on the current system state and the predicted trends.

2. The intelligent control system for hydrogen production by water electrolysis as claimed in claim 1, wherein, The execution layer is configured to adjust the power output and the working state of a thermal management unit in real time based on the target current density set value, wherein the power output is adjusted by a high-frequency pulse modulation method to adjust the electrolysis voltage waveform, and the thermal management unit adjusts the cooling liquid flow rate by a fuzzy PID algorithm to control the temperature of the electrolytic cell.

3. The intelligent control system for hydrogen production by water electrolysis as claimed in claim 1, wherein, The system further comprises a digital twin module, which combines physical mechanism equations and neural network models to construct a hybrid model for simulating hydrogen production efficiency under different working conditions and supporting offline training and simulation optimization of the control strategy, and performs strategy fine-tuning on the reinforcement learning controller to adapt to electrolytic cell electrode aging and slow changes in system parameters.

4. The intelligent control system for hydrogen production by water electrolysis as claimed in claim 3, wherein, The edge computing layer comprises an anomaly detection module, which is constructed based on a fusion perception and residual threshold judgment mechanism and can perform anomaly analysis on multiple operation parameters and trigger a warning when an abnormal state occurs. wherein, represents the time t parameter feature vector at time is the parameter mean vector within the history window, is the parameter standard deviation vector within the history window, and ||·|| is the vector norm, which is used to quantify the deviation degree of the current state from the historical stable state; when the abnormal score value exceeds the set threshold, the system triggers the corresponding early warning response.

5. The intelligent control system for hydrogen production by water electrolysis as claimed in claim 3, wherein, The anomaly detection module uses a multi-dimensional dynamic sliding analysis method with a time window and calculates an anomaly score value based on the following formula: The system has a safety protection mechanism triggered automatically by the anomaly detection module, which comprises a three-level response control strategy: First-level response: automatically adjust the target current density to a safety threshold when an anomaly initially occurs; Second-level response: automatically switch to a backup cooling system and reduce the electrolysis power when local overheating occurs; 6. An intelligent control method for hydrogen production by water electrolysis based on the system according to any one of claims 1 to 5, characterized in that, Third-level response: perform power-off protection and disconnect the main loop of the electrolytic cell when there is a serious abnormal risk. The system comprises the following steps: Data acquisition step: collect real-time operation state parameters of a water electrolysis hydrogen production system and energy supply state parameters of a renewable energy system; Data pre-processing and feature extraction step: pre-process the collected data and input the pre-processed data into a prediction model; Trend prediction step: use an LSTM prediction model to perform short-term prediction of renewable energy power supply trends and hydrogen demand in a future period of time; Strategy optimization step: input the prediction results and the current operation state into a reinforcement learning controller to output a target current density set value through a strategy network; Instruction execution step: control a pulse power supply to adjust output waveform parameters based on the target current density set value.

7. The intelligent control method for hydrogen production by water electrolysis as claimed in claim 6, wherein, The method further comprises an anomaly detection and security control step: continuously monitoring the system operating state, calculating residual scores by the edge-deployed anomaly detection module, and triggering a multi-level security protection mechanism including current limiting protection, cooling switching and emergency power-off operation if the scores exceed the preset threshold.

8. The intelligent control method for hydrogen production by water electrolysis as claimed in claim 6, wherein, The LSTM prediction model comprises a three-layer stacked LSTM network, supplemented by an attention mechanism and a Dropout regularization layer, and outputs new energy power generation prediction values and hydrogen demand prediction values for the next 15 minutes, 30 minutes and 60 minutes through a fully connected layer.

9. The intelligent control method for hydrogen production by water electrolysis as claimed in claim 6, wherein, The reinforcement learning controller selects the current optimal action from a set of action consisting of multiple current density set values in each control cycle, the current density set values are arranged at equal intervals from low to high, and the system selects the optimal action as the set instruction of the next cycle in a step-by-step manner each time.

10. An electronic device comprising a memory, a processor, and a computer program stored on the memory to run thereon, characterized in that, The processor implements the steps of the method of claim 6 when executing the program.

Images