Hydrogen liquefaction control method and system based on AI decision

Abstract

The invention discloses a hydrogen liquefaction control method and system based on AI decision making. The method comprises the steps that real-time operation data and parameters of a hydrogen liquefaction device are obtained; based on the real-time operation data and the parameters, a fuzzy clustering algorithm is utilized to identify the current working condition so as to determine the working condition type to which the current working condition belongs; based on the working condition type, a reinforcement learning algorithm is utilized to generate a target control strategy, and the target control strategy comprises at least one control strategy of the yield, the energy consumption and the equipment service life early warning of the hydrogen liquefaction device; and according to the target control strategy, utilizing a particle swarm optimization algorithm to adjust PID parameters of the hydrogen liquefaction device so as to control at least one of yield, energy consumption and equipment life early warning of the hydrogen liquefaction device. The response speed, stability and anti-disturbance capability of the control loop of the hydrogen liquefaction device can be improved, dependence on experience of operators is reduced, performance degradation of the hydrogen liquefaction device is avoided, and efficient operation of the hydrogen liquefaction device is guaranteed.

Description

Technical Field

[0001] This application relates to the field of hydrogen liquefaction control technology, specifically to a hydrogen liquefaction control method and system based on AI decision-making. Background Technology

[0002] A hydrogen liquefaction unit is an industrial device that converts hydrogen gas into a liquid state. It typically consists of key components such as a multi-stage hydrogen compressor, a liquid nitrogen precooling tank, a multi-stage heat exchanger, a multi-stage hydrogen turbine, and throttling valves. In actual operation, the control strategies for each stage need to comprehensively consider multiple parameters, including temperature, pressure, flow rate, vibration, and energy consumption.

[0003] In related technologies, the control strategy of hydrogen liquefaction equipment is generally set in advance by relevant technical personnel, which relies on their experience and judgment. This can easily lead to performance degradation of the hydrogen liquefaction equipment and make it impossible to guarantee its efficient operation. Summary of the Invention

[0004] This application provides a hydrogen liquefaction control method and system based on AI decision-making, which aims to avoid performance degradation of hydrogen liquefaction equipment and ensure its efficient operation.

[0005] Firstly, this application provides a hydrogen liquefaction control method based on AI decision-making, the method comprising: Obtain real-time operating data and parameters of the hydrogen liquefaction unit; Based on the real-time operating data and parameters, a fuzzy clustering algorithm is used to identify the current operating condition in order to determine the type of operating condition to which the current operating condition belongs. Based on the operating condition type, a target control strategy is generated using a reinforcement learning algorithm. The target control strategy includes a control strategy for at least one of the following: the output of the hydrogen liquefaction unit, energy consumption, and equipment life warning. According to the target control strategy, the PID parameters of the hydrogen liquefaction unit are adjusted using a particle swarm optimization algorithm to control at least one of the following: the output of the hydrogen liquefaction unit, energy consumption, and equipment life warning.

[0006] In the above embodiments, a hierarchical AI decision-making and control architecture was constructed. The upper layer utilizes fuzzy clustering algorithm for macroscopic operating condition perception and reinforcement learning algorithm to generate globally optimal target setpoints; the lower layer uses particle swarm optimization algorithm to dynamically optimize the parameters of the PID controller in the execution layer, ensuring accurate and stable tracking of the upper-layer target. This method combines global optimization with dynamic control, significantly improving the automation level, operating efficiency, and operating condition adaptability of the hydrogen liquefaction unit, and realizing intelligent closed-loop control of the entire process.

[0007] In conjunction with some embodiments of the first aspect, in some embodiments, the reinforcement learning algorithm includes an action value update formula, and the generation of a target control strategy based on the working condition type using the reinforcement learning algorithm includes: Generate a current state vector that includes the real-time operating data and parameters, and the membership degree of the operating condition type; Based on the current state vector and the preset action vector space, the action value of the state vector-action vector pair is iteratively updated using the action value update formula. The action value update formula is updated based on the current state vector, the action vector of the currently executed action, the immediate reward calculated based on the reward function, and the maximum expected action value of the next state vector. The reward function includes a reward term and a penalty term. The reward term is positively correlated with the output and operating efficiency of the hydrogen liquefaction unit, and the penalty term is negatively correlated with the degree to which the hydrogen liquefaction unit deviates from the output target, increased energy consumption, and reduced equipment life. The target control strategy is determined based on the converged action value.

[0008] In the above embodiments, the environmental perception capability is enhanced by constructing a composite state vector that integrates physical data and fuzzy semantics, and the learning direction is guided by a composite reward function that includes performance rewards and action penalties, ultimately outputting the optimal control objective. This scheme deeply integrates reinforcement learning theory with hydrogen liquefaction process control, ensuring the robustness and optimization capability of the control strategy.

[0009] In conjunction with some embodiments of the first aspect, in some embodiments, after acquiring the real-time operating data and parameters of the hydrogen liquefaction device, the method further includes: Based on the real-time operating data and parameters, the equipment-level health status of the hydrogen liquefaction unit is determined; If a downward trend in the equipment-level health is detected, the remaining life of the corresponding equipment in the hydrogen liquefaction unit is predicted based on the equipment-level health using the Weibull lifetime distribution model, and the predicted remaining life of the equipment is obtained. The scale parameter and shape parameter in the Weibull lifetime distribution model are determined based on the trend index and volatility index of the equipment-level health. Based on the predicted remaining lifespan of the device, an early warning is issued.

[0010] In the above embodiments, the dynamic characteristics (trends and fluctuations) of equipment-level health are dynamically correlated with the key parameters (scale and shape) of the Weibull life distribution model, thereby realizing real-time and adaptive prediction of the remaining life of the equipment. Finally, through an early warning mechanism, the prediction results are transformed into maintenance decision support with practical guidance, improving the convenience of equipment management.

[0011] In conjunction with some embodiments of the first aspect, in some embodiments, determining the equipment-level health of the hydrogen liquefaction unit based on the real-time operating data and parameters includes: Extract the time-domain statistical features of the real-time running data and parameters; The parametric health level of each time-domain statistical feature is determined based on the deviation between each time-domain statistical feature and the corresponding health baseline value. The equipment-level health of the hydrogen liquefaction device is obtained by weighted summation of the parameter-level health of multiple time-domain statistical features.

[0012] In the above embodiments, multi-dimensional time-domain statistical features are extracted from the original time-series signal to condense information. Then, a non-linear exponential decay function is used to map the deviation of each physical feature into a standardized parameter-level health score, accurately depicting the degradation trajectory of a single dimension. Finally, multiple parameter-level health scores are integrated into a comprehensive device-level health score, realizing a refined and multi-dimensional assessment of device-level health.

[0013] In conjunction with some embodiments of the first aspect, in some embodiments, after predicting the remaining lifetime of the corresponding equipment in the hydrogen liquefaction unit using the Weibull lifetime distribution model, the method further includes: Based on the scale parameters and the shape parameters, determine the instantaneous failure rate function of the corresponding equipment in the hydrogen liquefaction unit; Based on the instantaneous failure rate function, the instantaneous failure rate of the corresponding equipment in the hydrogen liquefaction unit is determined, wherein the warning information used for early warning includes the instantaneous failure rate.

[0014] In the above embodiments, by constructing and solving the instantaneous failure rate function, not only are the dynamic failure modes of the equipment revealed, but also its failure risk at any given moment is quantified. Then, the core risk indicator of instantaneous failure rate is directly integrated into the early warning information, complementing the predicted remaining lifespan, and providing maintenance personnel with a dual basis for decision-making: "how long can it still be used?" and "how dangerous is it now?"

[0015] In conjunction with some embodiments of the first aspect, in some embodiments, after acquiring the real-time operating data and parameters of the hydrogen liquefaction device, the method further includes: Based on the real-time operating data and parameters, the linear correlation parameters between the inlet pressure and output power of the turbine expander in the hydrogen liquefaction unit are determined, as well as the nonlinear correlation parameters between the rotational speed and efficiency of the turbine expander are determined. Determine the associated feature vectors corresponding to the linear correlation parameter and the nonlinear correlation parameter; Based on the associated feature vectors, the operating status of the turbine expander is assessed using a support vector machine model to obtain a health score.

[0016] In the above embodiments, by calculating linear and nonlinear correlation parameters, deep coupling features are extracted from the physical mechanism level. Then, a composite correlation feature vector containing dynamic information is constructed. Finally, by utilizing the nonlinear modeling capability of the support vector machine model, these deep features are accurately mapped into a quantified health score, thus achieving an accurate assessment that is highly sensitive to and robust to equipment performance degradation.

[0017] In conjunction with some embodiments of the first aspect, in some embodiments, after assessing the health status of the turbine expander using a support vector machine model based on the associated feature vector to obtain a health score, the method further includes: Using the real-time operating data and parameters, the associated feature vector, and the health score as input, a long short-term memory network model is used to predict the trend of the outlet pressure and outlet temperature of the turbine expander, and the trend prediction results are obtained. Based on the trend prediction results, feedforward compensation processing is performed on the PID parameters of the hydrogen liquefaction device.

[0018] In the above embodiments, by inputting multi-source information into the LSTM model, high-precision, long-term prediction of key output parameters is achieved, and the predicted future trend is used as a feedforward signal to dynamically compensate and optimize the PID control loop, fundamentally solving the problem of lag in feedback control when dealing with slow equipment degradation.

[0019] In conjunction with some embodiments of the first aspect, in some embodiments, before performing feedforward compensation processing on the controller of the hydrogen liquefaction device based on the trend prediction results, the method further includes: Generate energy consumption models for each device in the hydrogen liquefaction unit; Based on the energy consumption model, a multi-objective optimization function is determined with the goal of minimizing net energy consumption. The constraints of the multi-objective optimization function include at least one of the following: safety constraints, process constraints, equipment constraints, and production capacity constraints. Solve the multi-objective optimization function to determine the target operating parameter combination, which is applied to the controller of the hydrogen liquefaction unit.

[0020] In the above embodiments, by constructing a high-precision hybrid energy consumption model, the system energy efficiency problem is transformed into a solvable mathematical optimization problem, and the optimal operating setpoint is obtained by using a global optimization algorithm. This model works in conjunction with subsequent feedforward compensation, with the former responsible for determining the optimal target and the latter responsible for accurately resisting disturbances, together ensuring the long-term, safe, and economical operation of the hydrogen liquefaction device.

[0021] In conjunction with some embodiments of the first aspect, in some embodiments, solving the multi-objective optimization function to determine the combination of objective operating parameters includes: Multiple global optimization algorithms are employed in parallel to calculate a set of candidate combinations of running parameters. The multiple global optimization algorithms include at least one of the following: improved genetic algorithm, particle swarm optimization algorithm, and simulated annealing algorithm. The target combination of operating parameters is determined based on multiple sets of candidate operating parameter combinations.

[0022] In the above embodiments, multiple global optimization algorithms with different mechanisms are run in parallel to form complementary advantages, and a robust optimization fusion mechanism is designed to determine the final solution. This ensures that the hydrogen liquefaction device can continuously and stably operate at the global optimum jointly confirmed by "collective wisdom", thereby maximizing energy saving and consumption reduction.

[0023] Secondly, embodiments of this application provide an AI-based hydrogen liquefaction control system, which is used to execute the method described in any possible implementation of the first aspect. The beneficial effects achievable by the AI-based hydrogen liquefaction control system can be found in the beneficial effects described in the corresponding method, and will not be repeated here.

[0024] One or more technical solutions provided in this application have at least the following technical effects or advantages: A hierarchical AI decision-making and control architecture is constructed. The upper layer utilizes fuzzy clustering algorithms for macroscopic operating condition perception and reinforcement learning algorithms to generate globally optimal target setpoints; the lower layer uses particle swarm optimization algorithms to dynamically optimize the parameters of the PID controller in the execution layer, ensuring accurate and stable tracking of the upper-layer target. This method combines global optimization with dynamic control, significantly improving the automation level, operating efficiency, and operating condition adaptability of the hydrogen liquefaction unit, and realizing intelligent closed-loop control throughout the entire process. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating the hydrogen liquefaction control method based on AI decision-making in an embodiment of this application. Figure 2 This is a system architecture diagram of the hydrogen liquefaction control system based on AI decision-making in the embodiments of this application. Detailed Implementation

[0026] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification of this application, the singular expressions “a,” “an,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to any or all possible combinations including one or more of the listed items.

[0027] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0028] The purpose of this application is to provide a hydrogen liquefaction control method and system based on AI (Artificial Intelligence) decision-making, particularly suitable for the operation control of large-scale hydrogen liquefaction production plants. A hydrogen liquefaction plant is an industrial device that converts hydrogen gas into a liquid state, typically consisting of key components such as a multi-stage hydrogen compressor, a liquid nitrogen precooling tank, a multi-stage heat exchanger, a multi-stage hydrogen turbine, and throttling valves. In actual operation, each stage needs to comprehensively consider multi-dimensional parameters such as temperature, pressure, flow rate, vibration, and energy consumption.

[0029] The hydrogen liquefaction control method and system based on AI decision-making provided in this application have intelligent decision-making capabilities, enabling adaptive control strategy optimization, autonomous start-up and shutdown process management, real-time status monitoring and early warning, and dynamic energy consumption optimization and adjustment, thereby improving the safety, stability, and economy of large-scale hydrogen liquefaction plant operation.

[0030] Specifically, the AI-based decision-making method and system for hydrogen liquefaction control has the following advantages: 1. Achieve adaptive control strategy optimization: It can autonomously adjust control parameters according to the real-time operating conditions of the hydrogen liquefaction unit. Through AI algorithms, it learns historical operating data and process rules, automatically optimizes PID (Proportional-Integral-Derivative) parameters, interlocking logic and control strategies, get rid of fixed control modes, reduce dependence on operator experience, and realize the autonomous and intelligent operation of related units.

[0031] 2. Enable autonomous start-up and shutdown process management: It can automatically formulate start-up and shutdown sequences, predict and respond to abnormal operating conditions during the start-up and shutdown process, and autonomously adjust key process parameters. It can safely and stably complete the start-up and shutdown operations of related equipment without manual intervention, thereby improving the stability of the start-up and shutdown process.

[0032] 3. Real-time status monitoring and early warning: By monitoring the operating parameters of the device in real time, AI algorithms analyze the health status of the equipment, predict potential failure risks, and issue early warning information in advance, replacing the manual inspection mode, realizing predictive maintenance, and reducing equipment failure rate and maintenance costs.

[0033] 4. Achieve dynamic energy consumption optimization and adjustment: It can optimize the operating parameters of each piece of equipment in real time according to the production capacity target, process conditions and energy efficiency characteristics, automatically find the optimal energy consumption configuration point, without the need for manual adjustment by process personnel, and achieve continuous optimization of the energy consumption of the unit operation and energy saving and consumption reduction goals.

[0034] Reference Figure 1 This application provides a hydrogen liquefaction control method based on AI (Artificial Intelligence) decision-making.

[0035] S101. Obtain real-time operating data and parameters of the hydrogen liquefaction unit.

[0036] In the embodiments of this application, a hydrogen liquefaction device refers to a complete process system for cooling gaseous hydrogen to below its boiling point, thereby converting it into liquid hydrogen. It typically includes a compressor, a precooling unit, a purification unit, a main heat exchanger, and a turboexpander as the core refrigeration component. Real-time operating data and parameters refer to a series of high-frequency acquired dynamic data that characterize the physical state and performance of the hydrogen liquefaction device and its core components (such as the turboexpander) at the current moment.

[0037] In some embodiments of this application, real-time operating data and parameters include at least one or more of the following: inlet pressure, inlet temperature, outlet pressure, outlet temperature, rotational speed, braking current, output power, and real-time operating efficiency of the turbine expander. High-frequency acquisition means that the sampling period is short enough (e.g., on the order of seconds or sub-seconds) to accurately capture transient changes in the process, thereby providing timely and accurate environmental awareness for the control algorithm.

[0038] S102. Based on real-time operating data and parameters, use fuzzy clustering algorithm to identify the current operating condition in order to determine the type of operating condition to which the current operating condition belongs.

[0039] In the embodiments of this application, fuzzy clustering algorithm is an unsupervised machine learning method that classifies data samples into different categories based on their similarity. Its "fuzzy" characteristic is that it allows a sample to belong to multiple categories simultaneously with different membership degrees. The current operating condition is a comprehensive snapshot of the state formed by real-time operating data and parameters at a specific moment. The operating condition type is a predefined discrete category that summarizes the typical operating states of a hydrogen liquefaction unit, such as "stable operating condition," "high load operating condition," and "low efficiency operating condition."

[0040] In some embodiments of this application, the Fuzzy C-Means (FCM) algorithm is used as the specific fuzzy clustering algorithm. This scheme first constructs a feature vector from the acquired multi-dimensional real-time operating data and parameters, and then calculates the membership degree between this feature vector and the cluster centers of each preset operating condition type. Finally, the operating condition type with the highest membership degree value is determined as the current operating condition type. This membership-based "soft" classification method can more precisely describe the transitional states between operating conditions. For example, when the system simultaneously exhibits 70% "stable operating condition" characteristics and 30% "fluctuating operating condition" characteristics, this information can be fully transmitted, making subsequent control decisions more robust and forward-looking, effectively avoiding control abrupt changes caused by hard switching of operating conditions.

[0041] S103. Based on the operating condition type, a target control strategy is generated using a reinforcement learning algorithm. The target control strategy includes a control strategy for at least one of the following: the output of the hydrogen liquefaction unit, energy consumption, and equipment life warning.

[0042] In the embodiments of this application, reinforcement learning algorithms are machine learning paradigms in which an agent learns by interacting with the environment and through trial and error in order to maximize cumulative rewards, such as Q-learning algorithms.

[0043] The target control strategy is a high-level, guiding set of optimization objectives output by a reinforcement learning algorithm. It does not directly specify concrete actuator actions, but rather sets macroscopic operational goals for the lower-level control system to pursue under the current operating conditions. The agent can learn which macroscopic strategy (action) will yield the greatest long-term cumulative reward under different operating conditions (states). After learning convergence, for any given current operating condition, it can output an optimal macroscopic strategy. For example, when the operating condition is identified as stable and the equipment is healthy, the algorithm might output "Strategy A"; when the equipment performance is identified as deteriorating, it might output "Strategy D". This output macroscopic strategy is the "target control strategy," which points the way for lower-level optimization.

[0044] In the control strategy for at least one of the following early warning mechanisms in a hydrogen liquefaction unit—production, energy consumption, and equipment lifespan—a target inlet valve opening and a target braking current may be included. These are key control variable setpoints that can be directly executed. The inlet valve opening primarily regulates the hydrogen flow rate into the turbine expander, thereby affecting its power and cooling capacity; the braking current regulates the expander's load and speed through the energy recovery system. Through the self-learning and optimization capabilities of reinforcement learning, the target control strategy can break free from dependence on precise mathematical models, dynamically adapt to slowly changing factors such as equipment aging, continuously seek optimization online, and achieve dynamic optimization of global performance.

[0045] S104. According to the target control strategy, the PID parameters of the hydrogen liquefaction unit are adjusted using the particle swarm optimization algorithm to control at least one of the following: the output of the hydrogen liquefaction unit, energy consumption, and equipment life warning.

[0046] In the embodiments of this application, the Particle Swarm Optimization (PSO) algorithm is a swarm intelligence optimization algorithm that simulates the foraging behavior of bird flocks and excels at searching for optimal solutions in complex multidimensional spaces. The Proportional-Integral-Derivative (PID) parameters refer to the three core tuning parameters (Kp, Ki, Kd) in the proportional-integral-derivative controller, whose values ​​directly determine the dynamic response performance and steady-state control accuracy of the control loop. Essentially, after the reinforcement learning algorithm determines the control objective, the PSO algorithm is used to optimize the performance of the PID controller, ensuring that the execution layer can quickly, smoothly, and accurately track the objective issued by the upper layer.

[0047] In some embodiments of this application, in the control strategy for at least one of the following—production, energy consumption, and equipment lifespan warning of a hydrogen liquefaction unit—the target inlet valve opening and the target braking current can be used as setpoints for two independent PID control loops. Taking the inlet valve control loop as an example, the objective function of the particle swarm optimization algorithm is constructed to minimize the integral error between the actual valve opening and the target opening, such as the integral of absolute error (IAE) or the integral of time-weighted absolute error (ITAE), while applying penalties for overshoot and settling time. Each particle in the particle swarm represents a set of proportional-integral-derivative parameters (Kp, Ki, Kd), and the parameter combination that minimizes the objective function is found through iterative optimization. This optimal parameter combination is then applied to the PID controller. This scheme achieves online self-tuning of the proportional-integral-derivative parameters, enabling dynamic adjustment of the controller's performance according to different operating conditions and control objectives, thereby achieving excellent dynamic tracking performance and steady-state control accuracy throughout the entire operating range.

[0048] In summary, this application's embodiments construct a hierarchical AI decision-making and control architecture. The upper layer utilizes fuzzy clustering algorithms for macroscopic operating condition perception and reinforcement learning algorithms to generate globally optimal target setpoints. The lower layer uses particle swarm optimization algorithms to dynamically optimize the parameters of the PID controller in the execution layer, ensuring accurate and stable tracking of the upper-layer target. This method combines global optimization with dynamic control, significantly improving the automation level, operating efficiency, and adaptability of the hydrogen liquefaction unit, achieving intelligent closed-loop control throughout the entire process.

[0049] In some embodiments of this application, the reinforcement learning algorithm includes the Q-value (Quality value, i.e., action value) update formula from the Q-learning algorithm. Based on the work condition type, generating a target control strategy using the reinforcement learning algorithm may include: S201. Generate a current state vector that includes real-time operating data and parameters, as well as the membership degree of the current operating condition type.

[0050] In the embodiments of this application, the current state vector serves as the basis for the reinforcement learning agent's decision-making, representing a digital representation of the environmental information it can observe at the decision-making moment. The membership degree, calculated using a fuzzy clustering algorithm, is a numerical value characterizing the similarity between the current system state and various preset working condition types, with a value range of 0 to 1. Incorporating membership degrees, rather than single working condition category labels, into the state vector provides the agent with richer working condition information, enabling it to perceive smooth transitions between working conditions and thus make more refined and adaptive decisions.

[0051] In some embodiments of this application, the current state vector is a concatenated vector, constructed by concatenating multiple normalized real-time operating data and parameters (such as inlet pressure and rotational speed) with membership values ​​representing the current operating condition (e.g., a membership of 0.8 for "stable operating condition" and 0.2 for "fluctuating operating condition"). This vectorized representation method preserves the details of the underlying physical parameters while incorporating a higher-level understanding of the operating condition, thus providing a comprehensive and multi-dimensional description of the current system state. This enables reinforcement learning algorithms to learn differentiated optimal control strategies under different operating conditions and their transition regions.

[0052] S202. Based on the current state vector and the preset action vector space, the action value of the state vector-action vector pair is iteratively updated using the action value update formula.

[0053] In the embodiments of this application, the action vector space is the set of all possible actions that the agent can take. For example, it can be a two-dimensional space consisting of the opening degree of the inlet valve and the adjustment amount of the braking current. The action value can be represented as Q(s, a), which is an estimate of the expected value of the future cumulative reward that the agent can obtain after performing action a in state s. The action value update formula is the core of the reinforcement learning algorithm. It defines how to optimize the estimation of action value based on the agent's experience in interacting with the environment. The action value update formula is updated based on the current state vector, the action vector of the currently executed action, the immediate reward calculated based on the reward function, and the maximum expected action value of the next state vector.

[0054] The reward function is used to calculate immediate rewards, which directly guide the agent's learning direction. It includes reward terms and penalty terms. The reward terms are positively correlated with the output and operating efficiency of the hydrogen liquefaction unit; the penalty terms are negatively correlated with the degree to which the hydrogen liquefaction unit deviates from the output target, increased energy consumption, and reduced equipment lifespan.

[0055] In some embodiments of this application, an algorithmic framework based on Deep Q-Network (DQN) is employed, using a deep neural network to fit the action value. The action value update is reflected in the update of the network weights, which is achieved by minimizing the mean squared error loss function between the current predicted Q value and the target Q value. The calculation of the target Q value follows the Bellman equation: Target Q value = Immediate reward + Discount factor × Maximum expected action value of the next state vector. The reward function can be specifically designed as: R = w1 × P_out + w2 × η - w3 × |ΔV_open| - w4 × |ΔI_b|. Here, R is the immediate reward, which may include P_out and η, where P_out is the output power and η is the operating efficiency. The penalty term may include ΔV_open and ΔI_b, where ΔV_open and ΔI_b are the changes in the inlet valve opening and braking current, respectively. w1, w2, w3, and w4 are weighting coefficients. This scheme introduces a deep neural network, enabling the reward function to flexibly balance control objectives (high power, high efficiency) and process stability (small adjustment), guiding the agent to learn a control strategy that is both efficient and stable.

[0056] S203. Determine the target control strategy based on the converged action value.

[0057] In the embodiments of this application, the converged action value refers to the estimated value of the action that tends to stabilize after sufficient iterative updates, indicating that the agent has learned a stable perception of the environment. The target control strategy is the final mapping rule from state to action formed based on this converged action value.

[0058] In some embodiments of this application, the target control strategy employs a greedy policy. Specifically, for any given current state vector, the agent traverses all possible actions in a preset action vector space, calculates the action value corresponding to each action using a converged Q-network, and selects the action vector with the maximum action value as the output. The specific values ​​contained in this output action vector are determined as the target inlet valve opening and the target braking current. This scheme ensures that during the decision-making phase, the agent always selects the optimal action that it believes will bring the greatest long-term reward, thereby directly transforming the learned knowledge into the optimal control command, achieving determinism and optimality in control decision-making.

[0059] In summary, the embodiments of this application enhance environmental perception capabilities by constructing a composite state vector that integrates physical data and fuzzy semantics, and guide the learning direction using a composite reward function that includes performance rewards and action penalties, ultimately outputting the optimal control objective. This scheme deeply integrates reinforcement learning theory with hydrogen liquefaction process control, ensuring the robustness and optimization capability of the control strategy.

[0060] In some embodiments of this application, an alternative method for generating target control strategies is provided. This method aims to solve the complex coupled control problem of multi-stage series systems in hydrogen liquefaction plants by introducing a hierarchical reinforcement learning architecture, thereby generating target control strategies with stronger logical coherence, synergy, and global optimality. Specifically, generating target control strategies based on operating conditions using reinforcement learning algorithms may include: I. Constructing a Hierarchical Reinforcement Learning (HRL) algorithm framework.

[0061] In this embodiment of the application, the reinforcement learning algorithm is a hierarchical reinforcement learning (HRL) algorithm, which includes a high-level policy network and one or more low-level policy networks.

[0062] Hierarchical reinforcement learning is a reinforcement learning paradigm that decomposes complex decision-making tasks into multiple hierarchical sub-tasks for learning. In the hydrogen liquefaction control scenario, this algorithm mimics the decision-making logic of human experts who "first formulate an overall strategy, then execute specific steps," decomposing the coordinated control problem of the entire device into two levels: high-level macroscopic goal setting and low-level specific action execution. This decomposition effectively reduces the dimensionality of the decision space, accelerates the convergence of the learning process, and improves the performance and interpretability of the final control strategy.

[0063] The high-level strategy network, acting as the strategic decision-maker in the hierarchical architecture, functions to set a series of abstract, guiding macro-level goals over a longer timescale based on the system's global state, rather than directly outputting specific control commands to the actuators. The high-level strategy network is a deep neural network. Its input is the system's global state information, including the current operating condition type, the real-time net energy consumption per unit of liquid hydrogen in the hydrogen liquefaction unit, liquid hydrogen production, and the equipment-level health of key equipment. Its output is a macro-level target vector. For example, for the core process of "compression-precooling-expansion" in the hydrogen liquefaction system, this macro-level target vector can be defined as [target pressure difference_ij, target temperature drop_k, target flow distribution coefficient_l]. These three terms represent "the target pressure difference range between the outlet of the i-th stage compressor and the inlet of the j-th stage heat exchanger," "the target isentropic temperature drop range of the k-th stage turbine expander," and "the target flow distribution ratio of hydrogen flowing through the l-th heat exchanger," respectively.

[0064] As a tactical executor in the hierarchical architecture, the low-level strategy network receives macro-level objectives from the higher-level strategy network and, in conjunction with the real-time status of the local subsystems it is responsible for, generates and outputs specific, executable control actions on a relatively short timescale. In this embodiment, multiple low-level strategy networks exist, each corresponding to a key regulation loop or subsystem in the hydrogen liquefaction unit. Taking the "low-level strategy network for the turbine expander cooling loop" as an example, the input of this network includes two parts: one part is the macro-level objective issued by the higher-level strategy network, such as "target temperature drop_k"; the other part is the local real-time operating data and parameters of this loop, such as the real-time inlet pressure, inlet temperature, and rotational speed of the turbine expander. The output of this network is the target inlet valve opening and target braking current in the control strategy for at least one of the following: hydrogen liquefaction unit output, energy consumption, and equipment life warning.

[0065] II. Generating target control strategies based on hierarchical reinforcement learning algorithms.

[0066] In the embodiments of this application, the process of generating target control policies using hierarchical reinforcement learning algorithms is specifically defined as a collaborative process in which a high-level policy network generates macro-level targets and a low-level policy network executes specific actions, and is trained using a specific intrinsic reward mechanism.

[0067] The high-level policy network makes decisions based on the current operating condition type (e.g., "high load condition" identified by fuzzy clustering) and the global system state, using its action value update formula. Its reward function is designed to evaluate the long-term overall performance of the system after executing a macro-level objective. For example, the reward term is positively correlated with the average liquid hydrogen production and average operating efficiency within a macro-level decision-making cycle (e.g., 15 minutes), while the penalty term is negatively correlated with the volatility of system energy consumption within that cycle. Through training, the high-level policy network can learn to plan optimal, long-term macro-level objective sequences for different operating conditions.

[0068] After receiving a macro-level objective from a higher-level network, the task of the lower-level policy network is to achieve that objective within a short decision-making cycle (e.g., 1 minute) by adjusting specific actuators (such as inlet valves). To effectively guide the learning of the lower-level network, its reward function introduces intrinsic rewards in addition to system-level external rewards. Intrinsic rewards measure the degree to which the lower-level policy network completes the sub-tasks assigned by the higher-level network. For example, if the objective from the higher level is "to maintain the temperature drop within a certain range," then when the lower-level policy network adjusts to bring the actual temperature drop into that range, it will receive a significant positive intrinsic reward. This intrinsic reward allows the lower-level network to focus on efficiently and stably achieving local objectives without directly relating to complex global final energy efficiency, thus greatly simplifying the learning task.

[0069] Through the aforementioned hierarchical collaboration and intrinsic reward mechanism, the higher-level policy network learns "what to do," while the lower-level policy network learns "how to do it." The combination of these two approaches results in an optimal target control strategy under the current operating conditions, ultimately formed by the action sequences output by multiple lower-level policy networks. By learning deep collaborative patterns among devices, more refined and efficient dynamic control of the hydrogen liquefaction unit is achieved, maximizing the overall operating efficiency and energy-saving level of the system.

[0070] In some embodiments of this application, after obtaining the real-time operating data and parameters of the hydrogen liquefaction device, the following may also be included: S301. Determine the equipment-level health status of the hydrogen liquefaction unit based on real-time operating data and parameters.

[0071] Equipment-level health is a comprehensive quantitative indicator, typically ranging from 0 to 1. It characterizes the degree to which the current operating state of core equipment (such as turbine expanders) in a hydrogen liquefaction unit deviates from its ideal health state. A value of 1 represents that the equipment is in perfect health, while a value closer to 0 indicates more severe performance degradation or potential malfunctions.

[0072] S302. If a downward trend in equipment-level health is detected, the remaining life of the corresponding equipment in the hydrogen liquefaction unit is predicted based on the equipment-level health using the Weibull lifetime distribution model, and the predicted remaining life of the equipment is obtained.

[0073] A downward trend refers to a statistically consistent monotonic or fluctuating decline in equipment health over multiple consecutive monitoring periods. The Weibull lifetime distribution model is a probability distribution model used in reliability engineering that can flexibly describe the failure patterns of products at different stages. Predicted remaining lifespan (PWW) refers to the expected time that equipment, given its current health condition, can continue to operate normally until a functional failure occurs.

[0074] In some embodiments of this application, the scale parameter and shape parameter in the Weibull lifetime distribution model are determined based on the trend index and volatility index of equipment-level health. The trend index is calculated by linear regression or moving average of historical equipment-level health time series data, and its negative slope characterizes the average rate of health decline. The volatility index is quantified by calculating the standard deviation or kurtosis of the time series, reflecting the stability of the equipment's operating status. The scale parameter η (i.e., characteristic lifetime) of the Weibull distribution model is inversely proportional to the average rate of health decline, that is, the faster the decline, the shorter the characteristic lifetime; the shape parameter β (i.e., failure mode) is positively correlated with the volatility index, that is, the more drastic the fluctuation, the higher the risk of early failure, and the larger the β value. By establishing the mapping relationship between these indicators and model parameters, the Weibull lifetime distribution model can be updated in real time. Subsequently, based on the updated model and the current operating time, the current cumulative failure probability of the equipment is calculated, and the remaining time required to reach the preset failure threshold is deduced, which is the predicted remaining lifetime of the equipment. This solution constructs a data-driven, adaptive life prediction model whose parameters can be dynamically adjusted according to the actual degradation process of the equipment, making the prediction results closer to the real working conditions and more accurate.

[0075] S303. Provide early warnings based on the predicted remaining lifespan of the equipment.

[0076] Early warning alerts are a proactive operation and maintenance information feedback mechanism designed to transform predictive analysis results into clear warnings that can be used by operators or maintenance engineers to make decisions.

[0077] In some embodiments of this application, the early warning system employs a multi-level threshold strategy. For example, when the predicted remaining lifespan of a device first falls below a relatively long preset threshold (e.g., 180 days), the system triggers a level 3 warning, highlighting the corresponding device in yellow via the human-machine interface and prompting, "It is recommended to include this device in the maintenance schedule for the next maintenance window." When the predicted remaining lifespan of the device further decreases and falls below a more urgent threshold (e.g., 30 days), the system triggers a level 4 warning, alerting the system through audible and visual alarms and sending messages to designated maintenance personnel, prompting, "Equipment degradation is accelerating, with a risk of short-term failure; please arrange maintenance immediately." This tiered early warning system provides differentiated prompts based on the urgency of the risk, helping the maintenance team to rationally allocate resources, achieving a shift from reactive maintenance to predictive maintenance, and effectively avoiding significant economic losses caused by unplanned downtime.

[0078] In summary, the embodiments of this application dynamically correlate the dynamic characteristics (trends and fluctuations) of equipment-level health with the key parameters (scale and shape) of the Weibull life distribution model, thereby achieving real-time and adaptive prediction of the remaining life of the equipment. Finally, through an early warning mechanism, the prediction results are transformed into maintenance decision support with practical guidance, improving the convenience of equipment management.

[0079] In some embodiments of this application, determining the equipment-level health of a hydrogen liquefaction unit based on real-time operating data and parameters may include: S401. Extract the time-domain statistical features of real-time running data and parameters.

[0080] Time-domain statistical features are a series of numerical values ​​that reflect the inherent laws and state information of a signal, calculated using mathematical statistical methods from the real-time running data and parameters of the original time series. By extracting features, high-dimensional, dynamic time series data can be transformed into low-dimensional, relatively static feature vectors, achieving dimensionality reduction and information condensation.

[0081] In some embodiments of this application, the extracted time-domain statistical features for the vibration acceleration signal of a turbine expander may include one or more of the following: mean, root mean square (RMS), variance, standard deviation, peak-to-peak value, kurtosis, and margin factor. For example, the RMS reflects the energy intensity of the vibration signal, and its increase is usually associated with rotor imbalance or accelerated bearing wear; kurtosis is sensitive to the impact components in the signal, and its abnormal increase often indicates early failures such as localized pitting in rolling bearings. By comprehensively analyzing these features, subtle early signs of equipment performance degradation can be captured from different dimensions, thereby greatly improving the sensitivity and accuracy of fault diagnosis.

[0082] S402. Determine the parametric level of health of each time-domain statistical feature based on the deviation between each time-domain statistical feature and the corresponding health baseline value.

[0083] The health baseline value is a reference value for various time-domain statistical characteristics extracted from the device under ideal health conditions and during long-term stable operation. The parameter-level health score is a normalized health index between 0 and 1 calculated for a single time-domain statistical characteristic. It maps the characteristic deviation of the physical domain to a standardized health score, facilitating the unified evaluation of characteristics with different dimensions.

[0084] In some embodiments of this application, taking the parametric health of a turbine expander vibration signal as an example, the calculation formula for the parametric health may include, for example: H_i=1-|(V_current-V_baseline) / V_max_deviation| Where H_i is the parameter-level health status, V_current is the statistical characteristics of the current vibration signal (such as the root mean square value), V_baseline is the baseline value in the healthy state, and V_max_deviation is the maximum allowable deviation threshold.

[0085] In some embodiments of this application, the parameter-level health level is determined using a method based on an exponential decay function, specifically calculated as follows: H_i=exp(-α×|(F_i_current-F_i_baseline) / (F_i_threshold-F_i_baseline)|) Where H_i is the parametric health status of the i-th time-domain statistical feature, F_i_current is its current value, F_i_baseline is its health baseline value, F_i_threshold is the preset alarm threshold that marks the device entering a failure state, and α is the decay coefficient. This exponential function can non-linearly reflect the health status decay process: when the feature deviation is small, the health status decreases slowly; once the deviation approaches the alarm threshold, the health status will drop sharply. This is highly consistent with the actual degradation pattern of most devices, enabling parametric health status to more realistically and accurately depict the device performance degradation trajectory under a single dimension.

[0086] S403. The parameter-level health of multiple time-domain statistical features is weighted and summed to obtain the equipment-level health of the hydrogen liquefaction unit.

[0087] Weighted summation is a multi-indicator comprehensive evaluation method. It assigns a weight coefficient to each parameter-level health score, and then adds the results of multiplying all parameter-level health scores by their corresponding weights to obtain the equipment-level health score that comprehensively reflects the overall health status of the equipment.

[0088] In some embodiments of this application, the determination of weight coefficients employs a dynamic weighting strategy combining information entropy weighting and the Analytic Hierarchy Process (AHP). First, using information entropy weighting, objective weights are assigned based on the dispersion of each time-domain statistical feature over a period of time; features with lower information entropy provide more information and thus have higher weights. Second, expert knowledge is introduced, and the AHP is used to compare the importance of each feature pairwise, constructing a judgment matrix and calculating subjective weights. Finally, the objective and subjective weights are linearly combined to obtain the final comprehensive weight. This approach utilizes the statistical regularities of the data itself, avoiding the arbitrariness of purely subjective weighting, while incorporating the experience and knowledge of domain experts to compensate for potential cognitive biases in purely data-driven methods. This combined subjective and objective weighting method makes the final device-level health indicators more scientific, reasonable, and reliable.

[0089] In summary, the embodiments of this application extract multi-dimensional time-domain statistical features from the original time-series signal to condense information. Then, a nonlinear exponential decay function is used to map the deviations of each physical feature into standardized parameter-level health, accurately depicting the degradation trajectory of a single dimension. Finally, multiple parameter-level health values ​​are integrated into a comprehensive device-level health value, realizing a refined and multi-dimensional assessment of device-level health.

[0090] In some embodiments of this application, after predicting the remaining lifetime of the corresponding equipment in the hydrogen liquefaction unit using the Weibull lifetime distribution model, the method may further include: S501. Determine the instantaneous failure rate function of the corresponding equipment in the hydrogen liquefaction unit based on the dimensional and shape parameters.

[0091] The instantaneous failure rate function, mathematically represented as λ(t), describes the instantaneous probability that a device will fail within the next smallest time unit after operating normally for time t. This function not only focuses on when the device will fail (provided by its remaining lifespan), but also reveals the magnitude of the current failure risk and the trend of future risk changes, thus providing a deeper understanding of the dynamic failure characteristics of the device.

[0092] In some embodiments of this application, the instantaneous failure rate function is derived based on the probability density function and reliability function of the Weibull distribution, and its mathematical expression is: λ(t) = (β / η) × (t / η)^(β-1). Here, β is the shape parameter, η is the scale parameter, and t is the equipment's operating time. The shape of the instantaneous failure rate function is dominated by the shape parameter β: when β > 1, the instantaneous failure rate increases with time, indicating that the equipment is in a wear-out period; when β = 1, the instantaneous failure rate is constant, indicating that the equipment is in an occasional failure period; when β < 1, the instantaneous failure rate decreases with time, indicating that the equipment is in an early failure period. By determining the instantaneous failure rate function, maintenance personnel can not only obtain a quantitative risk model but also determine the current failure stage of the equipment, providing a theoretical basis for developing more targeted maintenance strategies.

[0093] S502. Based on the instantaneous failure rate function, determine the instantaneous failure rate of the corresponding equipment in the hydrogen liquefaction unit, and apply the instantaneous failure rate to the early warning prompt.

[0094] The instantaneous failure rate is a specific value calculated by substituting the total operating time t_current of the equipment into the instantaneous failure rate function λ(t), i.e., λ(t_current). This value intuitively represents the failure risk level of the equipment at the current moment. Including this information in early warning prompts can provide operators with another key decision-making dimension besides remaining lifetime: the urgency of the risk. Since remaining lifetime is a long-term predictive indicator, while the instantaneous failure rate is an immediate risk indicator, combining the two can effectively avoid the decision-making errors caused by relying on a single indicator.

[0095] In summary, by constructing and solving the instantaneous failure rate function, this application not only reveals the dynamic failure modes of the equipment but also quantifies its failure risk at any given moment. Then, the core risk indicator of instantaneous failure rate is directly integrated into the early warning information, complementing the predicted remaining lifespan and providing maintenance personnel with a dual basis for decision-making: "how long can it still be used?" and "how dangerous is it now?"

[0096] In some embodiments of this application, after obtaining the real-time operating data and parameters of the hydrogen liquefaction device, the following may also be included: S601. Based on real-time operating data and parameters, determine the linear correlation parameters between the inlet pressure and output power of the turbine expander in the hydrogen liquefaction unit, as well as the nonlinear correlation parameters between the rotational speed and efficiency.

[0097] Linear correlation parameters are statistical indicators used to quantify the strength and direction of the linear relationship between two variables. Nonlinear correlation parameters, on the other hand, are indicators used to measure the nonlinear dependence between two variables. Because there are inherent physical relationships between the input, output, and state variables of a turbine expander, changes in equipment health will first be reflected in deviations from these physical relationships. Therefore, quantifying these relationships allows for the construction of features that are highly sensitive to equipment performance degradation.

[0098] In some embodiments of this application, the linear correlation parameter is calculated using the Pearson correlation coefficient. Multiple sets of synchronous data pairs of inlet pressure and output power are collected within a sliding time window, and their Pearson correlation coefficients are calculated. Under healthy conditions, inlet pressure and output power should exhibit a strong positive correlation, and the Pearson correlation coefficient should remain stable at a relatively high level close to 1. The nonlinear correlation parameter is calculated using the Maximum Information Coefficient (MIC). Rotational speed and efficiency typically have a nonlinear relationship, and the MIC can effectively measure the strength of this relationship.

[0099] S602. Determine the correlation feature vectors corresponding to the linear correlation parameters and nonlinear correlation parameters.

[0100] The correlation feature vector is an ordered numerical array composed of multiple correlation parameters calculated in the previous step, which serves as the standardized input to the machine learning model. This step integrates multiple indicators describing different physical correlation dimensions of the device into a unified data structure, facilitating pattern recognition by the machine learning algorithm.

[0101] S603. Based on the associated feature vectors, the health status of the turbine expander is assessed using a support vector machine model to obtain a health score.

[0102] Support Vector Machine (SVM) is a supervised learning model, particularly adept at handling classification and regression problems with small sample sizes, nonlinearity, and high dimensionality. In this application, it is used to establish a nonlinear mapping from associated feature vectors to health scores.

[0103] In some embodiments of this application, a Support Vector Regression (SVR) model is employed. This model maps the original correlated feature vectors to a higher-dimensional feature space using a nonlinear kernel function, such as a Radial Basis Function (RBF) kernel, and searches for an optimal hyperplane within this space. The model's training data comes from historical data, where the input consists of correlated feature vectors under different operating conditions and wear stages, and the output is a pre-labeled health score. During actual operation, the real-time generated correlated feature vectors are input into the trained SVR model, which then outputs a real-time health score. SVR effectively captures the complex nonlinear relationship between features and health, and has strong generalization ability, providing stable and reliable health assessment results even under fluctuating operating conditions.

[0104] In summary, the embodiments of this application extract deeply coupled features from the physical mechanism level by calculating linear and nonlinear correlation parameters, then construct a composite correlation feature vector containing dynamic information, and finally utilize the nonlinear modeling capability of the support vector machine model to accurately map these deep features into a quantified health score, thereby achieving a highly sensitive and robust accurate assessment of equipment performance degradation.

[0105] In some embodiments of this application, after assessing the health status of the turbine expander using a support vector machine model based on the associated feature vectors and obtaining a health score, the process may further include: S701. Using real-time operating data and parameters, associated feature vectors, and health scores as input, a long short-term memory network model is used to predict the trend of the outlet pressure and outlet temperature of the turbine expander, and the trend prediction results are obtained.

[0106] Long Short-Term Memory (LSTM) networks are a type of recurrent neural network (RNN) whose internal gating structure effectively learns long-term dependencies in time-series data. Trend prediction results refer to the model's output of predicted values ​​for turbine expander outlet pressure and outlet temperature over one or more future time steps. The LSTM network model can be obtained through the preceding model training process, which will not be elaborated upon here.

[0107] S702. Based on the trend prediction results, feedforward compensation processing is performed on the parameters of the proportional-integral-derivative controller of the hydrogen liquefaction unit.

[0108] The Proportional-Integral-Derivative (PID) controller is the core component of the control loop in a hydrogen liquefaction plant. Feedforward compensation is a control strategy that, based on the prediction of future disturbances, generates a compensating control quantity in advance to proactively counteract the effects of the disturbances. In this application, the slow drift in outlet pressure and temperature caused by equipment performance degradation is considered an endogenous disturbance.

[0109] In some embodiments of this application, the feedforward compensation processing is specifically manifested as an adaptive PID controller parameter dynamic optimization module. This module calculates the future predicted deviation sequence, i.e., the difference between the predicted value and the current setpoint, based on the trend prediction results. Subsequently, based on the amplitude and rate of change of the predicted deviation sequence, a pre-trained fuzzy logic controller or a small neural network is used to calculate online the fine-tuning amount of the current PID controller parameters. For example, if the LSTM predicts that the outlet temperature will continue to rise due to equipment degradation, the feedforward compensation module will increase the integral term weight slightly and smoothly in advance to enhance the controller's ability to eliminate steady-state errors, thereby maintaining the temperature near the setpoint before it actually deviates. This scheme combines the "predictive" capability of predictive maintenance with the "adjustment" capability of process control, thereby anticipating and proactively adapting to equipment performance degradation, effectively suppressing the decline in control quality caused by equipment aging, and ensuring the long-term stability and efficiency of the hydrogen liquefaction process.

[0110] In summary, the embodiments of this application achieve high-precision, long-term prediction of key output parameters by inputting multi-source information into the LSTM model, and use the predicted future trend as a feedforward signal to dynamically compensate and optimize the PID control loop, fundamentally solving the problem of lag in feedback control when dealing with slow equipment degradation.

[0111] In some embodiments of this application, before performing feedforward compensation processing on the controller of the hydrogen liquefaction unit based on trend prediction results, the following may also be included: S801, Energy consumption model of each device in the hydrogen liquefaction unit.

[0112] An energy consumption model is a model that describes the mathematical relationship between a device's energy consumption or production capacity and its key operating parameters. For example, for energy-consuming equipment, the model input is the operating parameters, and the output is the electrical power consumed; for production-producing equipment, the input is the operating parameters, and the output is the cooling capacity generated. The purpose of generating energy consumption models is to provide a calculable and predictable description of the underlying component behavior for subsequent system-level energy efficiency optimization.

[0113] S802. Based on the energy consumption model, determine an optimization function with the objective of minimizing net energy consumption. The constraints of the optimization function include at least one of the following: safety constraints, process constraints, equipment constraints, and production capacity constraints.

[0114] An optimization function is a mathematical expression used to find the optimal operating point of a system, with the goal of minimizing the net energy consumption of the entire hydrogen liquefaction system. Net energy consumption is defined as the total energy consumption of all energy-consuming equipment minus the equivalent energy output of all production-producing equipment. Constraints are a series of boundaries set to ensure the feasibility and safety of the optimization results.

[0115] In some embodiments of this application, the optimization function is specifically defined as: Minimize F(X) = ΣP_comp(X) - k × P_exp(X), where P_comp is the total power of all compressors, P_exp is the output power of the turbine expander, k is the energy conversion factor, and X is the set of operating parameter variables to be optimized. Constraints can be finely set as follows: 1) Safety constraints: the pressure and temperature of critical nodes must not exceed design limits; 2) Process constraints: the outlet temperature of the interstage cooler must be within a specific range; 3) Equipment constraints: the speed of the turbine expander must not exceed the maximum allowable speed; 4) Production capacity constraints: the liquid hydrogen yield must not be lower than the requirements of the current production task. This scheme abstracts the complex plant operation problem into a mathematical optimization problem with a clear optimization objective and strict boundaries.

[0116] S803. Solve the optimization function to determine the target combination of operating parameters, and apply the target combination of operating parameters to the controller of the hydrogen liquefaction unit.

[0117] Solving for the optimization function involves using optimization algorithms to find a set of operating parameter values ​​that minimize the objective function, while satisfying all constraints. This set represents the "optimal setpoint" of the system under the current state and will be updated to the new setpoint of the controller (such as a PID controller).

[0118] In summary, the embodiments of this application construct a high-precision hybrid energy consumption model, transforming the system energy efficiency problem into a solvable mathematical optimization problem, and using a global optimization algorithm to obtain the optimal operating setpoint. This model works in conjunction with subsequent feedforward compensation, with the former responsible for determining the optimal target and the latter responsible for accurately resisting disturbances, together ensuring the long-term, safe, and economical operation of the hydrogen liquefaction device.

[0119] In some embodiments of this application, solving a multi-objective optimization function to determine the combination of objective operating parameters may include: S901. Multiple global optimization algorithms are used in parallel to calculate a set of candidate running parameter combinations. The multiple global optimization algorithms include at least one of the improved genetic algorithm, particle swarm optimization algorithm, and simulated annealing algorithm.

[0120] Global optimization algorithms are a class of heuristic algorithms designed to find the globally optimal solution to a complex function within a given domain. Parallel execution refers to simultaneously launching and running multiple different optimization algorithms, each independently solving the same optimization function. Candidate combinations of running parameters represent the optimal solution output by each global optimization algorithm after completing its search. The core idea of ​​this step is to examine the complex solution space from the perspectives of different algorithms, thereby overcoming the inherent limitations of a single algorithm.

[0121] In some embodiments of this application, the parallel framework is specifically deployed as follows: the system simultaneously enables the Improved Genetic Algorithm (IGA), Particle Swarm Optimization (PSO), and Simulated Annealing (SA) algorithms. These three algorithms represent different optimization philosophies based on biological evolution, swarm intelligence, and physical processes, respectively. Running them in parallel maximizes the exploration of the solution space, forming a complementary advantage.

[0122] Improved genetic algorithm: Based on the standard genetic algorithm, an adaptive crossover and mutation probability adjustment mechanism and niche technology are introduced to avoid premature convergence and ensure population diversity.

[0123] Particle Swarm Optimization (PSO) algorithm: Each "particle" represents a potential solution and updates itself by tracking the individual's historical best and the group's historical best, which has the advantage of fast convergence speed.

[0124] Simulated annealing algorithm: derived from the physical process of solid annealing, it accepts solutions that are worse than the current solution with a certain probability, thus having the ability to escape local optima.

[0125] S902. Based on multiple candidate combinations of operating parameters, determine the target combination of operating parameters.

[0126] This step involves finalizing and fusing multiple candidate solutions computed in parallel to determine a final target combination of operating parameters, ensuring that the final output solution is globally optimal or very close to globally optimal.

[0127] In some embodiments of this application, a two-stage optimal fusion strategy is adopted to determine the final target combination of operating parameters: Phase 1: Consistency Check and Optimal Selection. First, calculate the objective function value corresponding to the candidate combinations of running parameters obtained by each algorithm. If the relative difference between the objective function values ​​of each algorithm is less than a preset threshold, it indicates that different algorithms have converged to the same optimal region. At this point, the optimal set can be selected as the final combination of objective running parameters.

[0128] The second stage: Elite secondary optimization. If the relative difference between the objective function values ​​of each algorithm is greater than or equal to a preset threshold, it indicates that the algorithm may have converged to different local optima. At this point, these candidate combinations of running parameters are used as the initial population, and a smaller but more accurate local search algorithm, such as Sequential Quadratic Programming (SQP), is started for secondary fine optimization. The final solution obtained by the SQP algorithm is determined as the final combination of objective running parameters.

[0129] In a specific implementation, the logic of the optimal fusion strategy is as follows: Consistency check: After running the IGA, PSO, and SA algorithms in parallel, three sets of candidate solutions and their corresponding objective function values ​​are obtained. The coefficient of variation (CV) of these objective function values ​​is calculated. If the CV is less than a preset threshold ε, the results are considered to have high consistency, and the set with the smallest objective function value is selected as the final result.

[0130] Elite Quadratic Optimization: If CV is greater than or equal to ε, then quadratic optimization is initiated. The three candidate solutions are considered elite individuals, and slightly disturbed dynamic individuals are generated, collectively forming the initial search point set for the SQP algorithm. The SQP algorithm is executed, and the final converged solution is the final target combination of running parameters.

[0131] This scheme utilizes the wide-area search capability of the global optimization algorithm and greatly improves the reliability and optimality of the final solution through consistency checks and elite quadratic optimization mechanisms, thus ensuring the quality of the optimization results.

[0132] In summary, the embodiments of this application achieve complementary advantages by running multiple global optimization algorithms with different mechanisms in parallel, and design a robust optimization fusion mechanism to determine the final solution. This ensures that the hydrogen liquefaction device can continuously and stably operate at the global optimum jointly confirmed by "collective wisdom", thereby maximizing energy saving and consumption reduction.

[0133] Those skilled in the art should understand that the steps of solving multi-objective optimization functions to determine the combination of objective operating parameters, and the steps of generating objective control strategies using reinforcement learning algorithms, together constitute a hierarchical intelligent control architecture, which is mutually synergistic in terms of functional positioning and time scale.

[0134] Specifically, the step of determining the target combination of operating parameters describes a higher-level steady-state real-time optimization (RTO) layer. This layer is based on a steady-state energy consumption model and performs a global optimization calculation on a longer time scale (e.g., every 15 minutes or every hour). The aim is to determine a set of static setpoints that are theoretically the most energy-efficient under the current operating conditions for the entire hydrogen liquefaction unit, i.e., the target combination of operating parameters.

[0135] The step of determining the target control strategy is a lower-level Advanced Process Control (APC) layer. The core task of the reinforcement learning algorithm in this layer is to receive and dynamically track the combination of target operating parameters issued by the upper optimization layer. On shorter timescales of milliseconds or seconds, it senses the dynamic changes and disturbances of the process in real time, generating specific and continuous control actions, including the target inlet valve opening and the target braking current—the target control strategy. This ensures that the hydrogen liquefaction unit can smoothly, quickly, and accurately reach and maintain near the optimal operating point, while handling various transient disturbances.

[0136] It can be seen that the two constitute a complete closed loop from static global optimization to dynamic precise control, jointly ensuring that the hydrogen liquefaction unit can achieve long-term, stable and energy-efficient operation under all constraints.

[0137] In some embodiments of this application, an AI-based hydrogen liquefaction control system is also provided, which is used in any of the AI-based hydrogen liquefaction control methods described in the above embodiments.

[0138] In some embodiments of this application, reference is made to Figure 2 AI-based hydrogen liquefaction control systems may include: Traditional control layers consist of PLC (Programmable Logic Controller) or DCS (Distributed Control System) control systems, retaining safety control functions such as interlocking protection and basic regulation; The communication interface layer enables data interaction between the traditional control layer and the AI ​​decision-making layer through standard industrial protocols. The AI ​​decision-making layer includes a data acquisition module, an intelligent analysis module, a decision optimization module, and an instruction issuance module. The data acquisition module acquires real-time operating parameters of the large-scale hydrogen liquefaction unit, such as temperature, pressure, flow rate, vibration, and energy consumption. The intelligent analysis module analyzes the equipment's operating status and process trends based on machine learning algorithms. The decision optimization module generates optimal control strategies and parameter adjustment schemes based on the analysis results. The instruction issuance module transmits the AI ​​decision results to the traditional control layer for execution through the communication interface layer. When the system detects an operational anomaly or a failure in the AI ​​decision-making layer, it automatically switches to the traditional control mode to ensure the safe operation of the unit.

[0139] Specifically, the communication interface layer includes: a protocol conversion unit that supports multiple standard industrial communication protocols; a data caching unit that caches and preprocesses the collected real-time data; an instruction forwarding unit that converts the control instructions generated by the AI ​​decision layer into a format recognizable by the traditional control system and issues them for execution; and a status monitoring unit that monitors the status of the communication link in real time to ensure the reliability of data transmission.

[0140] Specifically, the intelligent analysis module includes: a parameter correlation analysis unit, which learns the correlation patterns between process parameters such as temperature, pressure, and flow rate based on historical data; an equipment status assessment unit, which uses machine learning algorithms to establish an equipment health evaluation model and evaluate the operating status of key equipment such as multi-stage compressors, heat exchangers, and turbines in real time; a process trend prediction unit, which predicts the changing trends of process parameters based on time series analysis; and an anomaly detection unit, which identifies equipment operating anomalies and process deviations through pattern recognition technology.

[0141] Specifically, the decision optimization module includes: a control strategy optimization unit, which adaptively adjusts PID parameters and control strategies according to the current operating conditions, replacing manual experience-based adjustments; a start-stop sequence planning unit, which formulates the optimal start-stop operation sequence based on AI algorithms to achieve unattended start-stop; an energy consumption optimization unit, which finds the optimal operating point in real time and automatically adjusts the parameters of each device to minimize energy consumption; and an early warning decision unit, which generates early warning information and maintenance suggestions based on the equipment status assessment results.

[0142] Specifically, the safety switching module is used to monitor the operating status of the AI ​​decision-making layer; when an AI decision-making layer failure, communication interruption, or abnormal decision is detected, it automatically switches to the traditional PLC or DCS control mode to ensure the safe operation of the device; when the AI ​​decision-making layer returns to normal, it can be manually or automatically switched back to the AI ​​decision control mode.

[0143] Specifically, the instruction issuance module includes: a security verification unit, which performs security checks on the control instructions generated by AI to ensure that the instructions are within a safe range; a priority management unit, which sets priorities for different types of control instructions to ensure that critical safety instructions are executed first; and an execution feedback unit, which monitors the execution of control instructions and provides feedback to the AI ​​decision-making layer.

[0144] This application achieves intelligent replacement of manual operation by superimposing AI decision-making functions on the traditional control system, reducing the workload of operators by more than 90%; through AI equipment status assessment and early warning, the fault prediction accuracy reaches more than 95%, and the unplanned equipment downtime is reduced by 70%; through intelligent optimization of process energy consumption, the overall energy consumption of the unit is reduced by 10%, which significantly improves the safety, reliability and economic benefits of large-scale hydrogen liquefaction unit operation.

[0145] In some embodiments of this application, the hydrogen liquefaction control method based on AI decision-making in any of the above embodiments is described by way of example.

[0146] The various modules in the AI ​​decision-making layer interact with each other via a high-speed memory bus. A microservice architecture is adopted, with each module deployed independently and capable of service calls. Specifically, this includes: Data Acquisition Module: Real-time acquisition of operating parameters of large-scale hydrogen liquefaction units, including multi-dimensional data such as temperature, pressure, flow rate, vibration, and energy consumption. The data acquisition module includes a data receiving unit, a data preprocessing unit, a data storage unit, and a data quality assessment unit.

[0147] The data receiving unit acquires real-time data from the traditional control layer through the communication interface layer, including: multi-stage compressor parameters (inlet temperature T1, inlet pressure P1, exhaust temperature T2, exhaust pressure P2, speed N, power W, vibration value V, etc.), precooling system parameters (liquid nitrogen temperature TN2, liquid level LN2, hydrogen inlet temperature TH2_in, outlet temperature TH2_out, etc.), heat exchanger parameters (inlet and outlet temperatures of hot and cold fluids, flow rate, pressure difference, etc.), and turbine parameters (inlet pressure, outlet pressure, speed, power, etc.).

[0148] The data preprocessing unit filters, denoises, detects outliers, and performs unit conversion on the raw data. A Kalman filter algorithm is used to smooth data such as temperature and pressure.

[0149] Intelligent Analysis Module: Based on machine learning algorithms, this module analyzes equipment operating status and process trends, including a parameter correlation analysis unit, an equipment status assessment unit, a process trend prediction unit, and an anomaly detection unit.

[0150] The parameter correlation analysis unit performs correlation analysis on the filtered data to identify linear and nonlinear dependencies between key variables, providing multi-parameter input support for process control. Taking turbine control as an example in the control process: Based on the key parameters of turbine operation (inlet pressure Pin, inlet temperature Tin, speed N, braking current Ib, outlet pressure Pout, outlet temperature Tout, target power Ptarget, and real-time efficiency η), a correlation analysis is performed on the filtered data.

[0151] Linear correlation calculation:

[0152] Used to determine the direct impact of changes in inlet pressure on output power.

[0153] Nonlinear correlation calculation:

[0154] Used to identify the nonlinear effect of speed variation on efficiency.

[0155] Form a correlation feature vector F that includes power and efficiency. assoc This provides input for subsequent health assessments, trend predictions, and control optimization.

[0156] The equipment status assessment unit classifies the turbine's operating status and scores its health based on the support vector machine (SVM) model, and introduces power and efficiency indicators.

[0157] Health model:

[0158] in, It is the ratio of actual efficiency to design value. This represents the ratio of actual power to target power. H in the first and second terms... therm Different thermal health parameters are considered, such as bearing temperature, winding temperature, or inlet / outlet temperature difference. mech For mechanical health, such as vibration and shaft displacement, W1, W2, W3, W4, and W5 are the corresponding weighting coefficients. When the health deteriorates, the controller reduces the aggressiveness of power pursuit; when efficiency is high and the health is good, it allows for a greater pursuit of power maximization.

[0159] The process trend prediction unit uses a long short-term memory network (LSTM) to make short-term and medium-term predictions for the process, such as for key turbine outputs (e.g., Pout, Tout).

[0160] Core structure:

[0161] Among them, the Gate of Oblivion Input gate Output gate From historical state The predictions are calculated based on the current input, including associated features and health status. The prediction results provide feedforward compensation to the controller, enabling it to adjust the inlet opening and braking current in advance before load fluctuations or changes in operating conditions, thus avoiding over-temperature or over-pressure.

[0162] The anomaly detection unit comprehensively utilizes observation residuals, prediction biases, statistical distances, and physical constraints for verification to identify abnormal states during turbine operation.

[0163] Residual statistics: Calculate the difference between the measured value and the predicted value, and determine whether it exceeds the normal range by using Mahalanobis distance.

[0164] Physical consistency: Check whether the temperature drop and pressure drop are within the design operating condition window.

[0165] Fusion judgment: The alarm score is generated by weighting multiple detection results. When the score exceeds the threshold, an early warning or protection action is triggered.

[0166] The adaptive control strategy is optimized as follows: Adaptive control strategies comprise four main steps: condition identification, strategy generation, parameter adjustment, and performance evaluation. By collecting operational data in real time and combining it with AI algorithms to learn from historical operational data and process patterns, automatic control and adaptive adjustment of PID parameters are achieved for each production stage. This breaks away from traditional fixed control modes, reduces reliance on operator experience, and enables the plant to operate autonomously and intelligently. An example is the adaptive control strategy in turbine control.

[0167] The operating condition identification system uses the fuzzy C-means clustering (FCM) algorithm to identify the current operating condition. It automatically identifies the current operating condition category (such as stable, high load, low efficiency, etc.) based on multi-dimensional operating parameters. This enables automatic classification of complex operating states, providing accurate operating condition labels for strategy selection, allowing control strategies to be tailored to different operating conditions.

[0168] The formula is:

[0169] in, The sample feature vector contains turbine operation-related parameters (inlet pressure Pin, inlet temperature Tin, speed N, braking current Ib, outlet pressure Pout, outlet temperature Tout, target power Ptarget, real-time efficiency η, etc.). The j-th operating condition cluster center represents a typical operating state, such as normal operating condition; For the sample Cluster The membership degree is (0–1); m is the fuzzy index.

[0170] The smaller j is, the closer the sample is to its cluster center, and the better the clustering effect. By minimizing j, the cluster centers that best distinguish different operating conditions (such as stable, high load, low efficiency, etc.) are automatically found, providing accurate operating condition labels for subsequent strategy selection.

[0171] The control strategy is generated based on the Q-learning reinforcement learning algorithm, which combines power maximization and efficiency reward to generate the optimal inlet valve opening and braking current control strategy, thereby achieving a dynamic balance between power and efficiency.

[0172] Q-value update formula:

[0173] in, The state vector consists of (operating parameters + operating condition membership). Let [Vopen, Ib] be the action vector; α is the learning rate (0–1). The discount factor is (0–1).

[0174] Based on the current state Actions to be performed Rewards received The maximum expected value of the next state is then used to update the Q value of the current state-action pair. Through repeated trial and error, the strategy learns which actions yield higher long-term returns under different operating conditions, gradually forming a control strategy that can automatically select the optimal inlet valve opening and braking current under different operating states.

[0175] Reward function:

[0176] in, The primary optimization objective is to optimize the output power. Efficiency (0–1), reward item; The amount of change in movement serves as a penalty for adjusting excessive actions. These are the weighting coefficients.

[0177] Power output, efficiency, and motion smoothness are combined into a numerical reward; higher power and efficiency result in a larger reward, while more drastic motion changes result in a smaller reward. This guides AI strategies to pursue high power without sacrificing efficiency and stability, achieving a balance between economy and safety.

[0178] Adaptive PID parameter adjustment utilizes a particle swarm optimization (PSO) algorithm to adjust PID parameters online, ensuring the control loop maintains optimal dynamic and steady-state performance under different operating conditions. This eliminates the performance degradation problem of fixed PID parameters under changing operating conditions, ensuring the turboexpander has a fast response, low overshoot, and low oscillation.

[0179] Equipment status monitoring and early warning are as follows: It includes five steps: data acquisition, feature extraction, status assessment, fault prediction, and early warning generation.

[0180] First, multi-dimensional data acquisition is performed. This involves acquiring the time-domain characteristic parameters of the equipment's operation. Taking turbine expander control as an example: (inlet pressure Pin, inlet temperature Tin, rotational speed N, braking current Ib, outlet pressure Pout, outlet temperature Tout, target power Ptarget, real-time efficiency η, vibration V).

[0181] Feature extraction, specifically time-domain statistical features: The calculation of time-domain statistical features follows an analytical sequence from the overall to the detailed, and from stationarity to anomalies. First, the operating baseline for each parameter is established using the mean, providing a reference for long-term drift monitoring. Then, the standard deviation is calculated to measure the fluctuation range of these parameters around the mean, thereby assessing operational stability and control quality. Based on this, skewness analysis is introduced to address the asymmetry of the data distribution, identifying whether the operating state is consistently biased towards high or low values, revealing potential regulation anomalies or structural deviations. Finally, kurtosis characterizes the sharpness and heavy-tailed nature of the distribution, thus identifying subtle anomalies during equipment operation. Time-domain statistical features can monitor both chronic degradation (such as seal leakage and efficiency degradation) and sudden mechanical failures (such as bearing pitting and rotor imbalance), providing a data foundation for subsequent frequency domain analysis, health assessment, and life prediction.

[0182] Equipment health status assessment. A multi-level health assessment model is established, consisting of equipment-level health and system-level health. Parameter-level health reflects the status of individual indicators, equipment-level health integrates multiple indicators to form an equipment profile, and system-level health merges the status of multiple devices into a global health level. This hierarchical and quantitative link of "parameter → equipment → system" is used to prioritize the identification of equipment health changes, and then assess their impact on the operational safety and performance of the entire hydrogen liquefaction system, providing a quantitative basis for maintenance decisions and operational optimization.

[0183] The formula for parametric-level health assessment is:

[0184] in, The health status of the i-th monitored parameter (0~1, 1 is completely healthy); These are current measured values ​​(such as average vibration value, average outlet pressure value, etc.). This is the baseline value for health (the average value under normal operating conditions). This represents the maximum allowable deviation for this parameter.

[0185] The formula for assessing equipment-level health is:

[0186] in, Device-level health status (0~1, 1 is fully healthy); Let be the weight of the i-th parameter. The health parameters of the equipment, such as turbine vibration V, are summed in order of importance and weighted. The health score is calculated using expert experience, FMEA (Failure Mode and Effects Analysis), or optimization based on historical downtime losses. It outputs a health score from 0 to 1.

[0187] The system-level health assessment formula is:

[0188] in, System-level health status (0~1); The device-level health status of the qth critical device; Let q be the importance coefficient of the q-th device. The health of multiple key devices within the system (such as the turbine expander, compressor, and main cooler) is productized and aggregated. This product structure exhibits a "low-value amplification effect": a very low health of one device will significantly lower the overall health of the system. For example, when the health of the turbine expander... When the health status of other devices declines, the system-level health status also declines. It will also drop rapidly, thus triggering a system-level device health alarm.

[0189] Equipment remaining life prediction is necessary when equipment health shows a downward trend, allowing for the prediction of its remaining usable life to facilitate advance maintenance planning. This embodiment employs the Weibull life distribution model to probabilistically model lifespan and updates parameters based on real-time features.

[0190] The Weibull distribution model is as follows:

[0191] in, The cumulative failure probability is 0 ~ 1; t is the running time; η is determined by the health status. The long-term trend (mean decay rate) and historical failure data are fitted together; β is a volatility index estimated by the volatility of health (standard deviation, kurtosis, etc.) and the changing trend of vibration characteristics; the lifetime distribution curve is characterized by the scale parameter η and the shape parameter β, and the lifetime curve is plotted by combining the historical failure data of the turbine expander with the current degradation trajectory (such as the continuous increase of kurtosis of vibration V and the enhancement of characteristic frequency amplitude), and the remaining lifetime interval corresponding to the current moment is deduced.

[0192] The failure rate function is:

[0193] in, Instantaneous failure rate (1 / Day); Based on the forecast results, a tiered early warning system is generated, and the warning level is determined as follows:

[0194] Where rul is the estimated remaining operating time of the equipment under the current operating conditions, calculated by the life prediction model.

[0195] Dynamic energy consumption optimization adjustment is as follows: After inputting the target liquid hydrogen production capacity, the system dynamically coordinates the operating parameters of key equipment throughout the entire process, including multi-stage hydrogen compressors, liquid nitrogen precooling tanks, multi-stage heat exchangers, multi-stage hydrogen turbine expanders, and throttling valves. Under strict adherence to safety and process constraints, it dynamically balances compression work, precooling work, turbine cooling, and throttling to achieve coordinated regulation and optimal energy consumption across the entire process. The system continuously monitors and adaptively corrects its operating strategies, ensuring stable production capacity while minimizing energy consumption per unit of liquid produced, significantly improving the overall energy efficiency, operational stability, and economy of the hydrogen liquefaction process.

[0196] The process includes five steps: energy consumption modeling, objective function construction, constraint setting, optimization solution, and result verification.

[0197] Energy consumption modeling involves creating individual energy consumption models for all equipment in the hydrogen liquefaction plant, using these models as the basic input. For example, for a multi-stage hydrogen compressor, the total mass flow rate is used. Total pressure difference Process average density Overall electrical efficiency The formula for directly assessing the total electrical power of the compression section is: It provides the energy consumption principal term for the compression segment of the objective function, and supports significantly influencing the total energy consumption by adjusting the flow rate and pressure difference distribution.

[0198] The multi-objective optimization function and constraints aim to minimize net energy consumption while meeting production capacity requirements, encouraging turbine priority, minimum energy saving, and high heat exchange efficiency. The underlying principle and formula are as follows:

[0199] Constraints include safety constraints, process constraints, equipment constraints, and capacity constraints. For example, equipment constraints are... .

[0200] Traffic and efficiency boundary constraints were implemented to ensure that the equipment operates within a healthy and controllable range.

[0201] The algorithm selection and optimization command issuance process involves using intelligent optimization algorithms to search for the optimal combination of operating parameters, while satisfying the objective function and constraints. The results are then safely sent to the control system for execution. Multiple algorithms are used simultaneously to calculate and evaluate the optimal solution for the command parameters after each algorithm is applied. Global optimization methods such as improved genetic algorithm (GA), particle swarm optimization (PSO), and simulated annealing (SA) can be selected. The specific process involves initializing the population, generating a set of candidate solutions (each solution being a combination of equipment operating parameters) based on the equipment's operating range and constraints, and calculating the objective function for each candidate solution using the modeling formula. , Then substitute the parameters of each process device into the input. The objective function is calculated to obtain the F-value; constraint checks are performed to eliminate or penalize solutions that violate safety, process, equipment, or capacity constraints; selection is made to retain solutions with better objective function values ​​for the next generation; convergence is determined when the change in the optimal solution over multiple consecutive generations is less than a threshold, or when the maximum number of iterations is reached, the search is stopped.

[0202] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application 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. Such 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. A hydrogen liquefaction control method based on AI decision-making, characterized in that, The AI-based decision-making hydrogen liquefaction control method includes: Obtain real-time operating data and parameters of the hydrogen liquefaction unit; Based on the real-time operating data and parameters, a fuzzy clustering algorithm is used to identify the current operating condition in order to determine the type of operating condition to which the current operating condition belongs. Based on the operating condition type, a target control strategy is generated using a reinforcement learning algorithm. The target control strategy includes a control strategy for at least one of the following: the output of the hydrogen liquefaction unit, energy consumption, and equipment life warning. According to the target control strategy, the PID parameters of the hydrogen liquefaction unit are adjusted using a particle swarm optimization algorithm to control at least one of the following: the output of the hydrogen liquefaction unit, energy consumption, and equipment life warning.

2. The hydrogen liquefaction control method based on AI decision-making as described in claim 1, characterized in that, The reinforcement learning algorithm includes an action value update formula. The step of generating a target control strategy based on the work condition type using the reinforcement learning algorithm includes: Generate a current state vector that includes the real-time operating data and parameters, and the membership degree of the operating condition type; Based on the current state vector and the preset action vector space, the action value of the state vector-action vector pair is iteratively updated using the action value update formula. The action value update formula is updated based on the current state vector, the action vector of the currently executed action, the immediate reward calculated based on the reward function, and the maximum expected action value of the next state vector. The reward function includes a reward term and a penalty term. The reward term is positively correlated with the output and operating efficiency of the hydrogen liquefaction unit, and the penalty term is negatively correlated with the degree to which the hydrogen liquefaction unit deviates from the output target, increased energy consumption, and reduced equipment life. The target control strategy is determined based on the converged action value.

3. The hydrogen liquefaction control method based on AI decision-making as described in claim 1, characterized in that, After acquiring the real-time operating data and parameters of the hydrogen liquefaction unit, the method further includes: Based on the real-time operating data and parameters, the equipment-level health status of the hydrogen liquefaction unit is determined; If a downward trend in the equipment-level health is detected, the remaining life of the corresponding equipment in the hydrogen liquefaction unit is predicted based on the equipment-level health using the Weibull lifetime distribution model, and the predicted remaining life of the equipment is obtained. The scale parameter and shape parameter in the Weibull lifetime distribution model are determined based on the trend index and volatility index of the equipment-level health. Based on the predicted remaining lifespan of the device, an early warning is issued.

4. The hydrogen liquefaction control method based on AI decision-making as described in claim 3, characterized in that, The determination of the equipment-level health status of the hydrogen liquefaction unit based on the real-time operating data and parameters includes: Extract the time-domain statistical features of the real-time running data and parameters; The parametric health level of each time-domain statistical feature is determined based on the deviation between each time-domain statistical feature and the corresponding health baseline value. The equipment-level health of the hydrogen liquefaction device is obtained by weighted summation of the parameter-level health of multiple time-domain statistical features.

5. The hydrogen liquefaction control method based on AI decision-making as described in claim 3, characterized in that, After predicting the remaining lifespan of the corresponding equipment in the hydrogen liquefaction unit using the Weibull lifetime distribution model, the method further includes: Based on the scale parameters and the shape parameters, determine the instantaneous failure rate function of the corresponding equipment in the hydrogen liquefaction unit; Based on the instantaneous failure rate function, the instantaneous failure rate of the corresponding equipment in the hydrogen liquefaction unit is determined, wherein the warning information used for early warning includes the instantaneous failure rate.

6. The hydrogen liquefaction control method based on AI decision-making as described in claim 1, characterized in that, After acquiring the real-time operating data and parameters of the hydrogen liquefaction unit, the method further includes: Based on the real-time operating data and parameters, the linear correlation parameters between the inlet pressure and output power of the turbine expander in the hydrogen liquefaction unit are determined, as well as the nonlinear correlation parameters between the rotational speed and efficiency of the turbine expander are determined. Determine the associated feature vectors corresponding to the linear correlation parameter and the nonlinear correlation parameter; Based on the associated feature vectors, the operating status of the turbine expander is assessed using a support vector machine model to obtain a health score.

7. The hydrogen liquefaction control method based on AI decision-making as described in claim 6, characterized in that, After assessing the health status of the turbine expander using a support vector machine model based on the associated feature vectors to obtain a health score, the process further includes: Using the real-time operating data and parameters, the associated feature vector, and the health score as input, a long short-term memory network model is used to predict the trend of the outlet pressure and outlet temperature of the turbine expander, and the trend prediction results are obtained. Based on the trend prediction results, feedforward compensation processing is performed on the PID parameters of the hydrogen liquefaction device.

8. The hydrogen liquefaction control method based on AI decision-making as described in claim 7, characterized in that, Before performing feedforward compensation processing on the controller of the hydrogen liquefaction unit based on the trend prediction results, the process further includes: Generate energy consumption models for each device in the hydrogen liquefaction unit; Based on the energy consumption model, a multi-objective optimization function is determined with the goal of minimizing net energy consumption. The constraints of the multi-objective optimization function include at least one of the following: safety constraints, process constraints, equipment constraints, and production capacity constraints. Solve the multi-objective optimization function to determine the target operating parameter combination, which is applied to the controller of the hydrogen liquefaction unit.

9. The hydrogen liquefaction control method based on AI decision-making as described in claim 8, characterized in that, Solving the multi-objective optimization function to determine the combination of objective operating parameters includes: Multiple global optimization algorithms are employed in parallel to calculate a set of candidate combinations of running parameters. The multiple global optimization algorithms include at least one of the following: improved genetic algorithm, particle swarm optimization algorithm, and simulated annealing algorithm. The target combination of operating parameters is determined based on multiple sets of candidate operating parameter combinations.

10. A hydrogen liquefaction control system based on AI decision-making, characterized in that, The AI-based decision-making hydrogen liquefaction control system is used to execute the AI-based decision-making hydrogen liquefaction control method according to any one of claims 1 to 9.

Images