Adaptive hydrogen charging management method based on cloud reinforcement learning and online self-tuning

Abstract

The present application relates to the field of hydrogen energy storage, and particularly to an adaptive hydrogen filling management method based on cloud reinforcement learning and online self-tuning. The method formalizes the hydrogen filling process of a hydrogen storage container as a Markov decision process, taking the container pressure, target pressure difference, bottle mouth temperature, bottle surface temperature, filling unit pressure and temperature, environmental parameters, cumulative cycle number, and control action at the previous time as the state space, taking the filling unit switching, proportional valve opening, and pre-cooling temperature as the mixed action space, and constructing a reward function with the filling efficiency, energy consumption, temperature rise, pressure change rate, and target pressure achievement. Then, a simulation model is used to train a policy network for generating control instructions using a reinforcement learning algorithm in the cloud. The policy network is deployed in an edge intelligent controller and continuously trained and updated with dynamic updated operation data. The present application can overcome the shortcomings of existing solutions in terms of adaptability, real-time performance, and end-to-end decision-making ability.

Description

Technical Field

[0001] This invention relates to the field of hydrogen energy storage, and in particular to an adaptive hydrogen charging management method based on cloud-based reinforcement learning and online self-tuning, as well as the corresponding computer program product and hydrogen charging system. Background Technology

[0002] As the global energy structure shifts towards cleaner and lower-carbon energy, hydrogen energy, with its high energy density, zero carbon emissions, and renewable characteristics, is considered a crucial component of the future energy system. The hydrogen energy industry has also been listed as a key development direction for strategic emerging industries and future industries. In the field of hydrogen energy applications, fuel cell vehicles, especially fuel cell heavy-duty trucks (referred to as "fuel cell heavy-duty trucks"), have become the priority breakthrough point for hydrogen energy transportation applications due to their advantages in medium- and long-distance transportation, and are widely considered the main body for large-scale hydrogen energy application in the medium to long term. In the technological system of fuel cell heavy-duty trucks, the on-board hydrogen storage system is one of the key core components. To meet the requirements of vehicle range and space layout, large-capacity, high-pressure Type IV hydrogen storage cylinders (i.e., plastic-lined carbon fiber fully wound cylinders) have become the mainstream technological route.

[0003] According to mandatory safety standards for relevant products, hydrogen storage cylinders must undergo high-pressure hydrogen pressure cycling durability tests under both ambient and extreme temperature conditions to verify their structural integrity and safety under long-term alternating loads. In actual testing, hydrogen cycling tests typically involve frequent hydrogen charging and discharging operations, with strict regulations on the time, pressure rate, and cycle waveform for each charging operation. However, existing hydrogen cycling testing systems suffer from high energy consumption, long charging times, and low testing efficiency when performing tests on 70MPa high-pressure, large-capacity Type IV hydrogen storage cylinders. Therefore, optimizing the hydrogen charging control strategy of the hydrogen cycling testing system to reduce system energy consumption and improve pressure control stability has become a pressing technical challenge in this field.

[0004] In existing technologies, to improve hydrogen charging efficiency and control temperature rise, engineers typically use a combination of offline simulation and orthogonal experiments to generate static hydrogen charging strategy maps for typical container specifications and several discrete operating points. This involves pre-designing several stages of hydrogen charging units, fixed switching pressure points, and fixed valve openings, and then executing the strategy according to the map lookup table. However, long-term engineering practice has verified that this type of static strategy has the following limitations: 1. The static strategy switching logic is fixed, making it impossible to perform online global collaborative optimization by combining real-time residual pressure, temperature, and availability information of each hydrogen charging unit. 2. Existing control decisions are separated from safety thresholds, lacking an end-to-end decision-making mechanism that incorporates both "goal achievement" and "safety boundary" into a unified optimization objective. Summary of the Invention

[0005] To address the shortcomings of existing hydrogen refueling solutions in terms of adaptability, real-time performance, and end-to-end decision-making capabilities, this invention provides an adaptive hydrogen refueling management method based on cloud-based reinforcement learning and online self-tuning, along with its corresponding computer program product and hydrogen refueling system.

[0006] The technical solution provided by this invention is as follows: An adaptive hydrogen charging management method based on cloud-based reinforcement learning and online self-tuning includes: A simulation model was created to simulate the gas circulation process of filling a hydrogen storage container with hydrogen from a high-pressure filling tank.

[0007] The hydrogen charging process is modeled using MDP, defining states, actions, state transition probabilities, rewards, and discount factors.

[0008] The simulation model is encapsulated into an interface that can be called by reinforcement learning algorithms. A reinforcement learning algorithm based on the Actor-Critic architecture is used in the cloud to solve the Markov decision of the hydrogen charging process in order to complete the training of the policy network.

[0009] The trained policy network is deployed to the edge intelligent controller, and a safety fence mechanism is designed for the edge intelligent controller to take over control of the policy network in the event of an over-limit state.

[0010] The system collects real operational data from each successful hydrogen refueling task performed by the edge intelligent controller. Based on this data, it performs online self-tuning of the safety barrier mechanism to dynamically adjust the preset constraints. The operational data is then anonymized and uploaded to the cloud for continuous training of the policy network. After offline verification, simulation playback verification, and safety rule verification, the trained policy network generates new policy parameters and distributes them to the edge intelligent controller during off-peak hours.

[0011] In offline mode, the hydrogen charging process is managed using an updated edge intelligent controller.

[0012] As a further improvement of the present invention, the simulation model is used to simulate the hydrogen charging response under different ambient temperatures, different initial pressures, different target pressures, different cylinder cyclic aging degrees, and different hydrogen charging unit states.

[0013] The simulation model uses the Peng-Robinson real gas law to calculate the compressibility factor, density, and enthalpy of high-pressure hydrogen at different pressures and temperatures; a lumped thermal resistance model is used to describe the heat transfer process between the gas inside the cylinder, the cylinder wall, and the external environment; a first-order inertial model is used to describe the precooler outlet temperature response; and a valve flow characteristic table or equivalent flow model is used to describe the relationship between the proportional valve opening and the mass flow rate.

[0014] As a further improvement of the present invention, in the MDP, the state space S includes measurable and derived quantities for reflecting the high-pressure filling tank, hydrogen storage container and gas circulation process.

[0015] Action space A includes a mixed action space for hydrogen charging unit switching commands, valve opening degree, and precooling temperature.

[0016] The state transition probability P is obtained through interactive sampling with the physical system or equivalent simulation environment.

[0017] The reward function R is used to comprehensively evaluate hydrogen charging efficiency, energy consumption, temperature safety, pressure change rate safety, and target pressure achievement. The expression for the reward function is as follows: ; in, Let t be the reward value at time t; The reward for hydrogen charging progress can be expressed as the increase in the mass of hydrogen charged per unit time or the degree to which the pressure approaches the target pressure; This is an energy consumption penalty item; This is a penalty item for exceeding temperature limits; Penalty item for exceeding the pressure change rate limit; Rewards for achieving targets under pressure; These are non-negative weighting coefficients.

[0018] As a further improvement of the present invention, the training strategy for training the policy network based on the reinforcement learning algorithm includes: The strategy network Actor includes discrete action output heads and continuous action output heads. The discrete action output head outputs the hydrogen charging unit selection results using the Gumbel-Softmax method, while the continuous action output head outputs the proportional valve opening degree and precooling temperature using the tanh function normalization.

[0019] The Critic valuation network simultaneously receives state vectors and mixed action codes, and outputs the corresponding Q-values.

[0020] The training process employs an experience replay pool, soft updates to the target network, exploration of noise, and a hierarchical sampling mechanism based on operating conditions.

[0021] Continuous actions are explored using Ornstein-Uhlenbeck noise or Gaussian noise, while discrete actions are explored using an ε-greedy strategy.

[0022] The training process first obtains an initial policy in a simulation environment, and then fine-tunes the policy based on actual test system running data.

[0023] As a further improvement of this invention, the simulation model provides interfaces during the training phase of the policy network that include at least reset, step, observation_space, action_space, and render. Specifically, reset is used to initialize the simulation conditions; step is used to advance a control cycle based on the current action and return the next state, reward value, and termination flag; observation_space is used to define the state space; action_space is used to define the action space; and render is used to output the current simulation state or process curve.

[0024] As a further improvement of the present invention, the operating parameters are randomized during each reset; the operating parameters include, but are not limited to, ambient temperature, initial pressure of hydrogen storage container, target pressure, and cumulative number of cycles.

[0025] As a further improvement of the present invention, in the hydrogen filling task of each hydrogen storage container, the edge intelligent controller executes a real-time control loop with a fixed sampling period, including: (i) Collect data on hydrogen storage container pressure and temperature, hydrogen filling unit pressure and temperature, environmental parameters, valve status, and safety status; (ii) Filter, normalize, remove outliers, and construct features from the collected data; (iii) Input the current state into the strategy network to obtain the hydrogen charging unit selection, proportional valve opening degree and precooling temperature setpoint; (iv) Perform secondary verification on the policy output action according to the hard threshold protection rule; (v) Convert actions that pass safety verification into control commands for valves, proportional valves, precoolers, and related actuators; (vi) Record the current state, actions, rewards, next state and termination flag to form operational experience data.

[0026] As a further improvement of the present invention, the triggering condition of the safety barrier mechanism includes any one of the following: (1) The temperature at the bottle opening or the surface temperature of the bottle body reaches the set safe temperature threshold; (2) The pressure change rate reaches the set safe pressure change rate threshold; (3) The pressure of the hydrogen storage container exceeds the target pressure by a set percentage; (4) Communication with key sensors timed out; (5) The inference time of the policy network continuously exceeds the set threshold; (6) Abnormal hydrogen concentration, valve feedback or emergency stop status.

[0027] As a further improvement of the present invention, when the safety barrier mechanism takes over, it executes a preset safety action sequence, including limiting the opening of the proportional valve, closing the main hydrogen charging valve, stopping the output of the high-pressure hydrogen charging unit, opening the pressure relief or return valve, triggering an audible and visual alarm, and recording abnormal events.

[0028] As a further improvement of the present invention, the online self-tuning method of the safety barrier mechanism includes: The peak temperature, peak pressure change rate, maximum valve opening, charging time, and target pressure error of each successful hydrogen charging process are statistically analyzed. Quantile indices are calculated within a rolling window, and the following decisions are made: (1) If the peak temperature and the peak pressure change rate of the hydrogen charging process for a preset number of consecutive times are both lower than the current safety threshold, the control threshold will be tightened within the hard threshold range according to a preset ratio.

[0029] (2) If the peak temperature and the peak pressure change rate are close to the safety boundary, the hydrogen charging efficiency reward weight shall be temporarily reduced or the upper limit of valve opening shall be tightened.

[0030] As a further improvement of the present invention, the edge intelligent controller retains the historically available policy network; when the new version of the policy network malfunctions or fails to verify, it automatically reverts to the previous stable version.

[0031] The present invention also includes a computer program product, which includes a computer program that, when executed by a processor, implements the aforementioned adaptive hydrogen charging management method based on cloud reinforcement learning and online self-tuning; thereby updating the parameters of the edge intelligent controller and using it to generate operating instructions for each device in the hydrogen charging process.

[0032] The present invention also includes a hydrogen charging system, which includes an edge intelligent controller for managing the hydrogen charging process; the edge intelligent controller uses the aforementioned adaptive hydrogen charging management method based on cloud reinforcement learning and online self-tuning to update parameters and generate operating instructions for each device in the hydrogen charging process.

[0033] The present invention has the following beneficial effects: This invention formalizes the hydrogen filling process of a hydrogen storage container as a Markov decision process, defining a state space, a hybrid action space, and a reward function. Combined with a created simulation model, a deep reinforcement learning algorithm based on an Actor-Critic architecture is used to train a policy network for generating the hydrogen filling control strategy. The trained policy network is deployed in an edge intelligent controller, and a safety barrier mechanism is designed for the edge intelligent controller. The edge intelligent controller performs state acquisition, policy reasoning, safety barrier judgment, command issuance, and experience recording in real time according to a fixed sampling period. Using the edge intelligent controller of this invention, the hydrogen filling strategy can be adaptively adjusted based on the initial pressure of the hydrogen storage container, ambient temperature, cylinder aging degree, and the state of each level of the hydrogen filling unit. This shortens the hydrogen filling time while meeting safety constraints on temperature, pressure, and pressure change rate, improving the automation level and operational efficiency of the hydrogen storage container calibration and filling process. Furthermore, this invention can continuously train the policy network based on operational data during the operation of the hydrogen filling system and perform online self-tuning of the safety barrier mechanism, thereby improving the robustness and generalization of the system. Attached Figure Description

[0034] Figure 1 This is a flowchart of the adaptive hydrogen charging management method based on cloud reinforcement learning and online self-tuning provided in Embodiment 1 of the present invention.

[0035] Figure 2 This is a diagram of the hydrogen charging system provided in Embodiment 2 of the present invention.

[0036] Figure 3 This is an architecture diagram of the automated testing system for hydrogen storage container verification provided in Embodiment 3 of the present invention.

[0037] Figure 4 This is a graph showing the convergence of the reinforcement learning training reward in an example of the present invention.

[0038] Figure 5 This is a comparison chart of pressure curves between the strategy of this invention and the traditional static staged hydrogen charging strategy in an example of this invention.

[0039] Figure 6 This is a comparison chart of temperature curves between the strategy of this invention and the traditional static staged hydrogen charging strategy in an embodiment of this invention. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the embodiments described below are only some embodiments of the present invention, and not all embodiments. Other implementation methods obtained by those skilled in the art based on the disclosed technical content of the present invention without creative effort should all fall within the protection scope of the present invention.

[0041] Unless otherwise defined, the technical and scientific terms used herein have the meanings commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein is for describing specific embodiments only and is not intended to limit the scope of protection of this invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0042] Example 1 To address the shortcomings of existing static hydrogen charging solutions in adapting to different working scenarios, and the common deficiencies in real-time performance and end-to-end decision-making capabilities of dynamic optimization control strategies, this embodiment provides an adaptive hydrogen charging management method based on cloud-based reinforcement learning and online self-tuning. This method formalizes the hydrogen charging process of a hydrogen storage container as a Markov decision process. The state space is defined by container pressure, target pressure difference, bottle opening temperature, bottle surface temperature, hydrogen charging unit pressure and temperature, environmental parameters, cumulative cycle count, and the control action at the previous moment. The mixed action space is defined by hydrogen charging unit switching, proportional valve opening, and pre-cooling temperature. The reward function is constructed using hydrogen charging efficiency, energy consumption, temperature rise, pressure change rate, and target pressure achievement. Then, a policy network for generating hydrogen charging control strategies is trained in the cloud using a deep reinforcement learning algorithm based on the Actor-Critic architecture.

[0043] The trained policy network is deployed in an edge intelligent controller, for which a safety barrier mechanism is designed. Locally, state acquisition, policy inference, safety barrier judgment, command issuance, and experience recording are performed in real-time according to a fixed sampling period to achieve automatic control of the hydrogen charging process. Furthermore, the edge intelligent controller uses operational data to perform online self-tuning of the safety barrier mechanism and uploads it to a cloud optimization platform for continuous training and version updates of the policy network.

[0044] Specifically, such as Figure 1 As shown, the adaptive hydrogen charging management method based on cloud-based reinforcement learning and online self-tuning provided in this embodiment includes the following steps: S1: Create a simulation model to simulate the gas circulation process of filling hydrogen from a high-pressure gas tank to a hydrogen storage container.

[0045] To support training the policy network in the cloud using reinforcement learning algorithms, this embodiment creates a simulation model to model the hydrogen filling response of a hydrogen filling system under different ambient temperatures, initial pressures, target pressures, cylinder aging levels, and hydrogen filling unit states. This simulation model can simulate the dynamic changes in the state of each device within the hydrogen filling system as input control commands occur during gas circulation; thus assisting the reinforcement learning algorithm in generating corresponding states based on new actions in each round.

[0046] In practical applications, the simulation model in this embodiment is based on the real gas law of high-pressure hydrogen, the heat transfer model of the hydrogen storage container, the dynamic model of the precooler, the flow characteristic model of the proportional valve, and the resistance model of the hydrogen filling pipeline; and constitutes a twin network corresponding to the actual hydrogen filling system. Specifically, the simulation model in this embodiment uses the Peng-Robinson real gas law to calculate the compressibility factor, density, and enthalpy of high-pressure hydrogen at different pressures and temperatures; uses a lumped thermal resistance model to describe the heat transfer process between the gas inside the cylinder, the cylinder wall, and the external environment; uses a first-order inertial model to describe the precooler outlet temperature response; and uses a valve flow characteristic table or equivalent flow model to describe the relationship between the proportional valve opening and the mass flow rate.

[0047] S2: Perform MDP modeling on the hydrogen charging process, defining states, actions, state transition probabilities, rewards, and discount factors.

[0048] This embodiment formalizes the hydrogen charging process as a Markov decision process (MDP) and uses a quintuple...<S, A, P,R, γ> Description. In the MDP, the state space S includes measurable and derived quantities reflecting the high-pressure filling tank, hydrogen storage container, and gas circulation process. The action space A includes a mixed action space of hydrogen filling unit switching commands, valve opening degrees, and precooling temperatures. The state transition probability P is obtained through interactive sampling with the physical system or equivalent simulation environment. The reward function R is used to comprehensively evaluate hydrogen filling efficiency, energy consumption, temperature safety, pressure change rate safety, and target pressure achievement; its expression is as follows: ; in, Let t be the reward value at time t; The reward for hydrogen charging progress can be expressed as the increase in the mass of hydrogen charged per unit time or the degree to which the pressure approaches the target pressure; This is an energy consumption penalty item; This is a penalty item for exceeding temperature limits; Penalty item for exceeding the pressure change rate limit; Rewards for achieving targets under pressure; These are non-negative weighting coefficients.

[0049] S3: The simulation model is encapsulated into an interface that can be called by reinforcement learning algorithms. A reinforcement learning algorithm based on the Actor-Critic architecture is used in the cloud to solve the Markov decision of the hydrogen charging process in order to complete the training of the policy network.

[0050] This embodiment employs a reinforcement learning algorithm to train the hydrogen charging control strategy. To achieve this, the simulation training environment is encapsulated as a reinforcement learning interface. Specifically, in practical applications, the interface provided by the simulation model during the training phase of the policy network includes at least reset, step, observation_space, action_space, and render. Reset initializes the simulation conditions; step advances the control cycle based on the current action and returns the next state, reward value, and termination flag; observation_space defines the state space; action_space defines the action space; and render outputs the current simulation state or process curve. Furthermore, this embodiment randomizes the operating parameters during each reset; these parameters include, but are not limited to, ambient temperature, initial pressure of the hydrogen storage container, target pressure, and cumulative cycle count.

[0051] In this embodiment, the reinforcement learning algorithm based on the Actor-Critic architecture includes a policy network (Actor) and a valuation network (Critic). The Actor network outputs hydrogen charging control actions based on the current state, while the Critic network evaluates the value function of the actions corresponding to the current state and the actions. Since the hydrogen charging action includes both discrete actions such as switching hydrogen charging units and continuous actions such as proportional valve opening and precooling temperature, this embodiment modifies the conventional DDPG algorithm by performing a hybrid action space modification.

[0052] Specifically, the policy network (Actor) includes discrete action output heads and continuous action output heads. The discrete action output head outputs the hydrogen charging unit selection results using the Gumbel-Softmax method, while the continuous action output head outputs the proportional valve opening and precooling temperature using tanh function normalization. The evaluation network (Critic) simultaneously receives the state vector and the mixed action encoding and outputs the corresponding Q value. During training, an empirical replay pool, soft updates of the target network, exploration noise, and hierarchical sampling of operating conditions are employed. Continuous actions are explored using Ornstein-Uhlenbeck noise or Gaussian noise, while discrete actions are explored using an ε-greedy strategy. The training process can adopt a two-stage approach of "simulation pre-training + real-world fine-tuning," that is, first obtaining the initial policy in a simulation environment, and then fine-tuning the policy based on actual test system operating data.

[0053] S4: Deploy the trained policy network to the edge intelligent controller and design a safety fence mechanism for the edge intelligent controller to take over control of the policy network in over-limit states.

[0054] This embodiment deploys the trained policy network into an edge intelligent controller, thereby enabling adaptive control of the hydrogen charging system using the edge intelligent controller. Specifically, when the edge intelligent controller in this embodiment is working, it executes a real-time control loop with a fixed sampling period. This real-time control loop includes: collecting data on hydrogen storage container pressure and temperature, hydrogen charging unit pressure and temperature, environmental parameters, valve status, and safety status. The collected data is then filtered, normalized, outlier removed, and feature constructed. The current state is input into the Actor network to obtain the hydrogen charging unit selection, proportional valve opening, and precooling temperature setpoint. The policy output actions are then subjected to secondary verification according to hard threshold protection rules. Actions that pass the safety verification are converted into control commands for valves, proportional valves, precoolers, and related actuators. Finally, the current state, actions, rewards, next state, and termination flag are recorded to form operational experience data.

[0055] Furthermore, to prevent the policy network obtained based on reinforcement learning from outputting unsafe actions under abnormal conditions, this embodiment also sets up a safety barrier mechanism between the policy network and the execution mechanism. This safety barrier mechanism has a higher priority than the policy network and is used to take over control of the policy network when triggering conditions are met. In practical applications, the triggering conditions for the safety barrier mechanism include any one of the following: (1) The temperature at the bottle opening or the surface temperature of the bottle reaches the set safe temperature threshold.

[0056] (2) The pressure change rate reaches the set safe pressure change rate threshold.

[0057] (3) The pressure of the hydrogen storage container exceeds the target pressure by a set percentage.

[0058] (4) Key sensor communication timeout.

[0059] (5) The inference time of the strategy network exceeds the set threshold continuously.

[0060] (6) Abnormal hydrogen concentration, valve feedback or emergency stop status.

[0061] Once the safety barrier mechanism takes over, the system will execute a preset safety action sequence, including limiting the opening of the proportional valve, closing the main hydrogen charging valve, stopping the output of the high-pressure hydrogen charging unit, opening the pressure relief or reflux valve, triggering audible and visual alarms, and recording abnormal events.

[0062] S5: Statistically analyze the actual operating data of the edge intelligent controller each time it successfully executes the hydrogen charging task, and perform online self-tuning of the safety barrier mechanism and continuous training of the policy network based on the operating data.

[0063] In this embodiment, after each sequential hydrogen charging task, the edge intelligent controller generates operational experience data based on real process data, recording the current state, action, reward, next state, and termination flag. This data can be used to retrain the policy network to improve its robustness and generalization in real-world application scenarios. Furthermore, in the initial state, if the constraints of the safety barrier mechanism are set too loosely, it may reduce the system's security; if the constraints are set too tightly, they may be triggered frequently, affecting the control efficiency of the policy network and making it difficult to achieve control targets such as hydrogen charging efficiency and energy consumption. Therefore, this embodiment also utilizes real operational experience data to perform online self-tuning of the safety barrier mechanism.

[0064] Specifically, the online self-tuning method of the safety barrier mechanism in this embodiment includes: statistically analyzing the peak temperature, peak pressure change rate, maximum valve opening, hydrogen charging time, and target pressure error of each successful hydrogen charging process; calculating the quantile index within a rolling window; and making the following decisions: (1) If the peak temperature and peak pressure change rate of the hydrogen charging process for a consecutive preset number of times are both lower than the current safety threshold, then the control threshold is tightened by a preset ratio within the hard threshold range; making the strategy more stable. (2) If the peak temperature and peak pressure change rate are close to the safety boundary, then the hydrogen charging efficiency reward weight is temporarily reduced or the upper limit of the valve opening is tightened; making the strategy more conservative. It should be emphasized that all online self-tuning results in this embodiment must not exceed the hard safety boundary of the relevant technical standards.

[0065] In addition, to continuously optimize the policy network in the edge intelligent controller, the operational experience data generated by the edge intelligent controller is anonymized and periodically uploaded to the cloud optimization platform. The cloud optimization platform deduplicates, classifies, cleans, and labels the experience data from different test nodes and different hydrogen storage containers, and continuously trains the policy network based on the newly added experience data. After offline verification, simulation playback verification, and safety rule verification, the trained policy network generates new version policy parameters and is distributed to the edge intelligent controller during off-peak hours. Furthermore, the edge intelligent controller can retain all historically available policy networks. When the new version policy network malfunctions or fails verification, the system can automatically revert to the previous stable version.

[0066] S6: In offline mode, the hydrogen charging process is managed using an updated edge intelligent controller.

[0067] In this embodiment, the edge intelligent controller relies on its internal policy network and safety barrier mechanism to generate control commands, thereby improving hydrogen charging efficiency and reducing energy consumption while ensuring basic safety. After each hydrogen charging task is completed, the edge intelligent controller can also automatically summarize the data from the entire process and generate a standardized report. The report mainly includes the hydrogen storage container number, container specifications, environmental conditions, target pressure, initial pressure, maximum pressure, minimum pressure, maximum temperature at the container opening, maximum temperature inside the container, charging time, peak pressure change rate, valve action records, alarm records, and policy version number.

[0068] Example 2 Based on the adaptive hydrogen charging management method based on cloud reinforcement learning and online self-tuning provided in Embodiment 1, this embodiment further provides a computer program product, which includes a computer program. When the computer program is executed by a processor, it implements the adaptive hydrogen charging management method based on cloud reinforcement learning and online self-tuning as in Embodiment 1; thereby updating the parameters of the edge intelligent controller and using it to generate operation instructions for each device in the hydrogen charging process.

[0069] This embodiment also provides a hydrogen charging system, such as Figure 2 As shown, it includes a high-pressure gas filling assembly, a gas source, a precooler, a circulation assembly, a valve assembly, a pressure gauge, and an edge intelligent controller.

[0070] The high-pressure filling assembly is used for temporary hydrogen storage. Within this assembly, the hydrogen is under high pressure, enabling spontaneous filling of the gas cylinder. This embodiment's high-pressure filling assembly includes multiple filling units with preset pressures and volumes; each filling unit consists of at least one high-pressure storage tank. Each high-pressure storage tank in this embodiment includes an inlet end and an outlet end. The inlet end is used to replenish the high-pressure storage tank, while the outlet end is used to fill the hydrogen storage container with the high-pressure gas from the high-pressure storage tank.

[0071] A gas source is used to safely and stably store hydrogen under low pressure. In practical applications, the gas source includes at least one low-pressure storage tank for storing hydrogen, each low-pressure storage tank including an inlet and an outlet. Compared to high-pressure storage tanks, hydrogen is typically stored at a lower pressure in low-pressure storage tanks, thus making it safer. The gas source in this embodiment is used to replenish hydrogen in any high-pressure storage tank when it is consumed during the refueling of the hydrogen storage container. The outlet of the low-pressure storage tank is connected to the inlet of the high-pressure storage tank. In practical applications, unlike high-pressure storage tanks, the volume of the low-pressure storage tank in this embodiment is much larger than that of a high-pressure storage tank; therefore, only one low-pressure storage tank for storing hydrogen can be installed in a system.

[0072] A precooler is a key device that cools a working medium before it enters a specific equipment or container through heat exchange. In this embodiment, the precooler is used to precool hydrogen gas being charged from a high-pressure storage tank into a hydrogen storage container. In practical applications, the precooler's inlet can be connected to the exhaust end of each high-pressure storage tank via pipes and manifold blocks, while the precooler's exhaust end is connected to the inlet of the hydrogen storage container to form a charging channel. A pressure gauge is installed on the hydrogen storage container to be calibrated and used to measure its real-time pressure.

[0073] In the hydrogen filling system of this embodiment, the filling of the hydrogen storage container from the high-pressure storage tank can be completed spontaneously by utilizing the pressure difference between the containers. However, the replenishment of the high-pressure storage tank from the low-pressure storage tank cannot be completed spontaneously. Therefore, the circulation component in this embodiment includes pipelines and a compressor. The pipelines connect the various high-pressure and low-pressure storage tanks to form a replenishment channel. The compressor pumps gas from any designated low-pressure storage tank along the filling channel to any designated high-pressure storage tank to complete the replenishment. The valve assembly includes multiple solenoid valves. The solenoid valves manage the opening and closing of the gas transmission channels between the low-pressure and high-pressure storage tanks, and between the high-pressure storage tank and the storage container.

[0074] The edge intelligent controller is electrically connected to the compressor, pressure gauge, and valve assembly. As the control center of the entire hydrogen charging system, the edge intelligent controller can issue commands to various actuators based on detection signals, control the system's operation, and complete the hydrogen charging task. Specifically, the edge intelligent controller in this embodiment employs an adaptive hydrogen charging management method based on cloud-based reinforcement learning and online self-tuning, as described in Embodiment 1, to update parameters and generate operating instructions for each device during the hydrogen charging process.

[0075] Example 3 Based on the scheme in Example 1, such as Figure 3 As shown in the figure, this embodiment also provides an automated testing system for the verification of hydrogen storage containers, which includes: an environmental chamber, a high-pressure gas filling component, a gas source, a circulation component, a precooler, a valve component, a pressure gauge, and an edge intelligent controller.

[0076] The environmental chamber includes a temperature and humidity control subsystem for providing the necessary environmental conditions for the hydrogen storage container to be calibrated. The high-pressure charging assembly includes multiple charging units with preset pressures and volumes; each charging unit consists of at least one high-pressure storage tank; each high-pressure storage tank includes an inlet and an outlet. The gas source includes at least one low-pressure storage tank for storing the target gas, each storage tank including an inlet and an outlet. Furthermore, the gas source may also include at least one low-pressure storage tank for storing nitrogen; and the high-pressure charging tank also includes a redundant charging unit for receiving nitrogen. This allows for purging of the hydrogen storage container by charging it with nitrogen through the redundant charging unit after each charging / discharging cycle.

[0077] The circulation assembly includes piping and a compressor. Piping connects the various high-pressure and low-pressure storage tanks to form a replenishment channel, and connects the hydrogen storage container and the low-pressure storage tank to form a recovery channel. The compressor pumps gas from any designated low-pressure storage tank along the replenishment channel to any designated high-pressure storage tank to complete the replenishment. The precooler inlet is connected to the exhaust end of each high-pressure storage tank via piping and a manifold block, and its exhaust end is connected to the inlet end of the hydrogen storage container to form a replenishment channel. A pressure gauge is installed on the hydrogen storage container to be calibrated and is used to measure its real-time pressure.

[0078] The valve assembly includes multiple solenoid valves and one two-way control valve. The solenoid valves independently control the on / off state of the inlet and outlet ends of each charging unit. The two-way control valve is installed at the external port of the hydrogen storage container and is used to control the on / off state and / or opening degree of the charging and recovery channels. The two-way control valve includes a one-way inlet end, a one-way outlet end, and a two-way port. The two-way port is used to connect to the external port of the hydrogen storage container to be calibrated; the one-way inlet end is connected to the outlet end of the precooler via a pipeline; and the one-way outlet end is connected to the inlet end of the low-pressure storage tank via a pipeline.

[0079] The edge intelligent controller is electrically connected to the temperature and humidity control subsystem, compressor, pressure gauge, and valve assembly. As the control center of the entire automated testing system for hydrogen storage container calibration, the edge intelligent controller can issue commands to various actuators based on detection signals, controlling the system to operate automatically and thus completing the calibration task of the hydrogen storage container. Specifically, the edge intelligent controller in this embodiment employs the cloud-based reinforcement learning and online self-tuning adaptive hydrogen filling management method as described in Embodiment 1 to update parameters and generate operating instructions for each device during the hydrogen filling process.

[0080] Simulation and Testing To verify the performance of the adaptive hydrogen charging management method based on cloud-based reinforcement learning and online self-tuning provided by this invention, the technicians formulated an experimental plan, implemented and tested the relevant schemes, and the specific experimental process is as follows: This experiment uses a vehicle-mounted Type IV hydrogen storage cylinder with a rated operating pressure of 70 MPa and a water volume of 393 L as the hydrogen to be filled. This hydrogen storage cylinder is used to simulate the large-volume, high-pressure hydrogen storage container in the on-board hydrogen storage system of a fuel cell vehicle or fuel cell heavy-duty truck. The hydrogen filling system adopts a three-stage hydrogen filling unit structure. The three stages of hydrogen filling units are 35 MPa / 300 L, 70 MPa / 300 L, and 105 MPa / 300 L, respectively. Each stage of hydrogen filling unit is connected to the main pipeline via an independent proportional valve, and the main pipeline is connected to the hydrogen storage cylinder to be filled after passing through a precooler. By controlling the switching state of the three-stage hydrogen filling units, the opening degree of the proportional valve, and the precooling temperature, the hydrogen filling pressure rise process, temperature rise process, and hydrogen filling time are comprehensively optimized.

[0081] First, a hydrogen charging process simulator was established. The simulator was built using SimulationX and MATLAB / Simulink. The simulator includes a real gas equation of state module, a thermal model of the hydrogen storage container, a dynamic model of the precooler, a proportional valve flow model, a pipeline resistance model, and a safety constraint module. The real gas equation of state module uses the Peng-Robinson equation of state to calculate the density, enthalpy, and compressibility factor of high-pressure hydrogen at different pressures and temperatures. The thermal model of the hydrogen storage container uses a lumped parameter model, equating the heat transfer process between the gas inside the container, the inner liner, the carbon fiber winding layer, and the external environment to a multi-node thermal resistance-capacity network. The dynamic model of the precooler uses a first-order inertial element to describe the response of the precooler outlet temperature to the set temperature. The proportional valve flow model calculates the instantaneous mass flow rate based on the inlet and outlet pressures, hydrogen temperature, and valve opening.

[0082] Secondly, the hydrogen filling simulator is encapsulated as a reinforcement learning environment. During each reset, the operating parameters are randomized. Specifically, the ambient temperature T_amb is uniformly sampled within the range of -20℃ to 50℃; the initial cylinder pressure P_cyl,0 is uniformly sampled within the range of 0 MPa to 10 MPa; the target pressure P_target is uniformly sampled within the range of 60 MPa to 75 MPa; the cumulative number of cylinder cycles n_cyc is uniformly sampled within the range of 0 to 1500 cycles; and the ambient wind speed v_wind is uniformly sampled within the range of 0 m / s to 10 m / s. Through this randomization, the training process covers typical scenarios such as low temperature, high temperature, low initial pressure, high initial pressure, new cylinders, aged cylinders, and different heat dissipation conditions.

[0083] In the created MDP, the state vector is set to 20 dimensions, including the difference between the current pressure and the target pressure ΔP, the bottle neck temperature T_neck, the bottle surface temperature T_surf, the hydrogen filling time t_elapsed, the cumulative hydrogen mass filled m_fill, the pressure and temperature of each of the three hydrogen filling units, the ambient temperature, the ambient wind speed, the initial pressure, the target pressure, the cumulative number of cycles of the gas cylinder n_cyc, and the action at the previous moment a_t-1.

[0084] The action vector consists of two parts: discrete actions and continuous actions. The discrete actions include four components, corresponding to the first-stage hydrogen charging unit, the second-stage hydrogen charging unit, the third-stage hydrogen charging unit, and the hydrogen charging stop action. The continuous actions include the proportional valve opening θ and the precooling temperature T_precool, where the proportional valve opening θ ranges from 0 to 1, and the precooling temperature T_precool ranges from -20 ℃ to 20 ℃.

[0085] This experiment uses an improved DDPG algorithm for policy training. The Actor network consists of three hidden layers with 256, 256, and 128 neurons respectively, using ReLU as the activation function. The Actor network includes discrete and continuous action output heads. The discrete action output head uses Gumbel-Softmax to output the action selected by the hydrogen charging unit, while the continuous action output head uses the tanh function for normalization. The Critic network consists of four hidden layers with 256, 256, 128, and 64 neurons respectively, used to estimate the action value function corresponding to the current state and action.

[0086] The training hyperparameters are as follows: the learning rate of the Actor network is 1×10^-4, the learning rate of the Critic network is 1×10^-3, and the target network's soft update coefficient is... The value is 0.005, the experience replay pool capacity is 1×10^6, the batch size is 128, the total training steps are 2×10^6, and the discount factor is... It is 0.99.

[0087] The weights of each term in the reward function are set as follows: ω1=1.0, ω2=0.2, ω3=2.0, ω4=1.5, ω5=10.0. The safe temperature threshold T_safe is set to 80 ℃, the safe pressure change rate dp / dt_safe is set to 8 MPa / s, and the target pressure tolerance is set to... Set to 0.5 MPa. When the hydrogen charging process rapidly approaches the target pressure, the temperature does not exceed the safety threshold, the pressure change rate does not exceed the safety threshold, and the energy consumption is low, the system provides a higher reward; when the temperature exceeds the limit, the pressure change rate exceeds the limit, the hydrogen charging unit is selected inappropriately, or the hydrogen charging stops before the target pressure is reached, the system imposes a penalty.

[0088] like Figure 4 As shown, the reward curve gradually converges after training. When the number of training steps reaches approximately 1.8 × 10^6, the reward value stabilizes around 440, indicating that the policy network has learned the coordination relationship between the switching of the three-stage hydrogen charging unit, the adjustment of the proportional valve opening, and the adjustment of the precooling temperature under different operating conditions.

[0089] The trained Actor policy network is deployed to the edge intelligent controller. The edge intelligent controller is connected to pressure sensors, temperature sensors, environmental sensors, hydrogen concentration sensors, proportional valves, switching valves, a precooler, and a safety interlock module. During actual hydrogen charging, the edge intelligent controller executes a control loop with a control period of 100ms. Each control loop includes the following steps: The first step is to collect data on the current pressure of the hydrogen storage container, the temperature at the bottle opening, the surface temperature of the bottle body, the pressure and temperature of the three-stage hydrogen filling unit, the ambient temperature, the ambient wind speed, the valve status, and the safety status.

[0090] The second step is to normalize and determine the validity of the collected data. When the data from a key sensor exceeds the physically reasonable range or a communication timeout occurs, the system enters the safety barrier process.

[0091] The third step is to input the current state vector into the Actor strategy network to obtain the hydrogen charging unit selection, proportional valve opening degree, and precooling temperature setpoint for the current control cycle.

[0092] Fourth, the safety barrier layer performs a second judgment on the output action of the strategy network. If the current bottle temperature is close to the safety threshold, or the pressure change rate is close to the safety threshold, the safety barrier layer reduces the upper limit of the proportional valve opening; if the pressure, temperature, hydrogen concentration, or communication status reaches the hard interlock condition, the safety barrier layer directly takes over control, closes the main hydrogen charging valve, and performs pressure relief or backflow actions.

[0093] The fifth step is to send the control actions verified by the safety barrier to the corresponding actuators.

[0094] The sixth step is to record the current state, control action, reward value, next state, and termination flag, and write them to the local experience pool.

[0095] In this experiment, if any of the following conditions are met, the edge intelligent controller will immediately enter the hard threshold protection state: (1) the surface temperature of the bottle is not lower than 85 ℃; (2) the pressure change rate is not lower than 10 MPa / s; (3) the pressure of the hydrogen storage container is not lower than 1.05 times the target pressure; (4) the communication timeout of any key sensor exceeds 500 ms; (5) the single inference time of the Actor network exceeds 30 ms and occurs 3 times consecutively; (6) the hydrogen concentration alarm or emergency stop signal is triggered. After the hard threshold protection is triggered, the system closes the main hydrogen charging valve, stops the output of the current hydrogen charging unit, opens the pressure relief return valve or executes the preset safety action sequence, and triggers the audible and visual alarm and records the alarm event.

[0096] To verify the effectiveness of this invention, its adaptive hydrogen charging strategy was compared with a traditional static staged hydrogen charging strategy. The traditional static staged hydrogen charging strategy charges hydrogen according to preset pressure switching points. For example, it uses the first-stage hydrogen charging unit in the low-pressure stage, switches to the second-stage hydrogen charging unit in the medium-pressure stage, and switches to the third-stage hydrogen charging unit in the high-pressure stage, with the proportional valve opening and pre-cooling temperature executed according to preset values. The adaptive hydrogen charging strategy of this invention, however, adjusts the hydrogen charging unit and proportional valve opening in real time based on the current hydrogen storage container pressure, temperature, target pressure difference, the status of each stage of the hydrogen charging unit, and environmental parameters. In the initial stage of hydrogen charging, the strategy can select a larger allowable flow rate based on the pressure difference and temperature rise; when approaching the target pressure, the strategy automatically reduces the valve opening to suppress pressure overshoot; when the cylinder temperature rises rapidly, the strategy automatically reduces the hydrogen charging rate or adjusts the pre-cooling temperature.

[0097] The pressure and temperature change curves for the two schemes during hydrogen charging were finally obtained as follows: Figure 5 and Figure 6 As shown in the figure, analysis of the data reveals that under the same target pressure conditions, the traditional static staged hydrogen charging strategy requires approximately 384 seconds to complete hydrogen charging, while the adaptive hydrogen charging strategy of this invention requires approximately 318 seconds, significantly shortening the charging time. Simultaneously, the peak temperature of the hydrogen storage container under the traditional static staged hydrogen charging strategy can reach approximately 99 °C, exceeding the safe temperature threshold; while the peak temperature of the hydrogen storage container under the adaptive hydrogen charging strategy of this invention is approximately 70 °C, below the safe temperature threshold of 80 °C, indicating that this invention can effectively suppress temperature rise while shortening the hydrogen charging time.

[0098] Therefore, this invention can not only achieve a higher hydrogen charging rate, but also achieve comprehensive optimal control while meeting safety constraints such as temperature, pressure change rate and target pressure accuracy.

[0099] The above-described embodiments are merely one implementation of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. An adaptive hydrogen charging management method based on cloud-based reinforcement learning and online self-tuning, characterized in that, It includes: Create a simulation model to simulate the gas circulation process of filling hydrogen from a high-pressure filling tank to a hydrogen storage container. The hydrogen charging process is modeled using MDP, defining states, actions, state transition probabilities, rewards, and discount factors. The simulation model is encapsulated into an interface that can be called by reinforcement learning algorithms. A reinforcement learning algorithm based on the Actor-Critic architecture is used in the cloud to solve the Markov decision of the hydrogen charging process in order to complete the training of the policy network. The trained policy network is deployed to the edge intelligent controller, and a safety fence mechanism is designed for the edge intelligent controller to take over control of the policy network in the event of an over-limit state. The actual operating data of the edge intelligent controller each time it successfully executes a hydrogen charging task is statistically analyzed, and the safety barrier mechanism is self-tuned online based on the operating data to dynamically adjust the preset constraints. The running data is anonymized and uploaded to the cloud for continuous training of the policy network. After offline verification, simulation playback verification and security rule verification, the trained policy network generates new version policy parameters and is distributed to the edge intelligent controller during off-peak hours. In offline mode, the hydrogen charging process is managed using an updated edge intelligent controller.

2. The adaptive hydrogen charging management method based on cloud-based reinforcement learning and online self-tuning as described in claim 1, characterized in that: The simulation model is used to simulate the hydrogen charging response under different ambient temperatures, different initial pressures, different target pressures, different cylinder aging degrees, and different hydrogen charging unit states. The simulation model uses the Peng-Robinson real gas equation of state to calculate the compressibility factor, density, and enthalpy of high-pressure hydrogen at different pressures and temperatures; it uses a lumped thermal resistance model to describe the heat transfer process between the gas inside the cylinder, the cylinder wall, and the external environment; it uses a first-order inertial model to describe the precooler outlet temperature response; and it uses a valve flow characteristic table or equivalent flow model to describe the relationship between the proportional valve opening and the mass flow rate.

3. The adaptive hydrogen charging management method based on cloud-based reinforcement learning and online self-tuning as described in claim 1, characterized in that: In MDP, the state space S includes measurable and derived quantities that reflect the high-pressure filling tank, hydrogen storage container, and gas circulation process. Action space A includes a mixed action space for hydrogen charging unit switching commands, valve opening degree, and precooling temperature; The state transition probability P is obtained through interactive sampling with the physical system or equivalent simulation environment; The reward function R is used to comprehensively evaluate hydrogen charging efficiency, energy consumption, temperature safety, pressure change rate safety, and target pressure achievement; its expression is as follows: ； in, Let t be the reward value at time t; The reward for hydrogen charging progress can be expressed as the increase in the mass of hydrogen charged per unit time or the degree to which the pressure approaches the target pressure; This is an energy consumption penalty item; This is a penalty item for exceeding temperature limits; Penalty item for exceeding the pressure change rate limit; Rewards for achieving targets under pressure; These are non-negative weighting coefficients.

4. The adaptive hydrogen charging management method based on cloud-based reinforcement learning and online self-tuning as described in claim 3, characterized in that, Training strategies for policy networks based on reinforcement learning algorithms include: The strategy network Actor includes discrete action output heads and continuous action output heads. The discrete action output head outputs the hydrogen charging unit selection results using the Gumbel-Softmax method, while the continuous action output head outputs the proportional valve opening and precooling temperature using the tanh function normalization. The Critic valuation network simultaneously receives state vectors and mixed action codes, and outputs the corresponding Q-values. The training process employs an experience replay pool, soft updates to the target network, exploration of noise, and a hierarchical sampling mechanism based on operating conditions. Continuous actions are explored using Ornstein-Uhlenbeck noise or Gaussian noise, while discrete actions are explored using an ε-greedy strategy. The training process first obtains an initial policy in a simulation environment, and then fine-tunes the policy based on actual test system running data.

5. The adaptive hydrogen charging management method based on cloud-based reinforcement learning and online self-tuning as described in claim 1, characterized in that: The simulation model provides interfaces during the training phase of the policy network that include at least reset, step, observation_space, action_space, and render. Specifically, reset is used to initialize the simulation conditions; step is used to advance a control cycle based on the current action and return the next state, reward value, and termination flag; observation_space is used to define the state space; action_space is used to define the action space; and render is used to output the current simulation state or process curve. Alternatively, the operating parameters may be randomized during each reset; these operating parameters include, but are not limited to, ambient temperature, initial pressure of the hydrogen storage container, target pressure, and cumulative number of cycles.

6. The adaptive hydrogen charging management method based on cloud-based reinforcement learning and online self-tuning as described in claim 1, characterized in that: During the hydrogen filling task of each hydrogen storage container, the edge intelligent controller executes a real-time control loop with a fixed sampling period, including: (i) Collect data on hydrogen storage container pressure and temperature, hydrogen filling unit pressure and temperature, environmental parameters, valve status, and safety status; (ii) Filter, normalize, remove outliers, and construct features from the collected data; (iii) Input the current state into the strategy network to obtain the hydrogen charging unit selection, proportional valve opening degree and precooling temperature setpoint; (iv) Perform secondary verification on the policy output action according to the hard threshold protection rule; (v) Convert actions that pass safety verification into control commands for valves, proportional valves, precoolers, and related actuators; (vi) Record the current state, actions, rewards, next state and termination flag to form operational experience data.

7. The adaptive hydrogen charging management method based on cloud-based reinforcement learning and online self-tuning as described in claim 1, characterized in that: The triggering conditions for the safety barrier mechanism include any one of the following: (1) The temperature at the bottle opening or the surface temperature of the bottle body reaches the set safe temperature threshold; (2) The pressure change rate reaches the set safe pressure change rate threshold; (3) The pressure of the hydrogen storage container exceeds the target pressure by a set percentage; (4) Communication with key sensors timed out; (5) The inference time of the policy network continuously exceeds the set threshold; (6) Abnormal hydrogen concentration, valve feedback, or emergency stop status; And / or, when the safety barrier mechanism takes over, it executes a preset safety action sequence, including limiting the opening of the proportional valve, closing the main hydrogen charging valve, stopping the output of the high-pressure hydrogen charging unit, opening the pressure relief or return valve, triggering audible and visual alarms, and recording abnormal events.

8. The adaptive hydrogen charging management method based on cloud-based reinforcement learning and online self-tuning as described in claim 1, characterized in that: The online self-tuning method for the safety barrier mechanism includes: The peak temperature, peak pressure change rate, maximum valve opening, charging time, and target pressure error of each successful hydrogen charging process are statistically analyzed. Quantile indices are calculated within a rolling window, and the following decisions are made: (1) If the peak temperature and the peak pressure change rate of the hydrogen charging process for a consecutive preset number of times are both lower than the current safety threshold, the control threshold shall be tightened within the hard threshold range according to the preset ratio. (2) If the peak temperature and the peak pressure change rate are close to the safety boundary, the hydrogen charging efficiency reward weight shall be temporarily reduced or the upper limit of valve opening shall be tightened. And / or, the edge intelligent controller retains historically available policy networks; when a new version of the policy network malfunctions or fails verification, it automatically reverts to the previous stable version.

9. A computer program product comprising a computer program, characterized in that: When the computer program is executed by the processor, it implements the adaptive hydrogen charging management method based on cloud reinforcement learning and online self-tuning as described in any one of claims 1-8; thereby updating the parameters of the edge intelligent controller and using it to generate operation instructions for each device in the hydrogen charging process.

10. A hydrogen charging system, characterized in that: It includes an edge intelligent controller for managing the hydrogen charging process; the edge intelligent controller uses the adaptive hydrogen charging management method based on cloud reinforcement learning and online self-tuning as described in any one of claims 1-8 to update parameters and generate operation instructions for each device in the hydrogen charging process.

Images