A hydrogen energy storage power station control system

Abstract

A hydrogen energy storage power station control system adopts a hierarchical structure. It uses a reinforcement learning model to comprehensively assess factors such as electricity price information, renewable energy output, ambient temperature, and hydrogen storage tank status data to select hydrogen production, power generation, or standby modes and set power targets for the electrolyzer and fuel cell. The middle layer refines power allocation using optimization algorithms, while the lower layer executes power commands through industrial control algorithms. An independent safety layer monitors indicators such as hydrogen concentration, pressure, and temperature. During optimization, the system simultaneously considers the energy consumption of the electrolyzer and auxiliary equipment, as well as equipment lifespan degradation, achieving full lifecycle economic efficiency, improving the economics of the hydrogen energy storage power station, and extending the lifespan of key equipment.

Description

Technical Field

[0001] This invention relates to the fields of energy management and energy storage control technology, and in particular to a multi-timescale collaborative optimization control system based on reinforcement learning for hydrogen energy storage power plants and its implementation method, which is used to improve the economy of hydrogen energy storage systems, extend equipment life and ensure system operation safety. Background Technology

[0002] In recent years, the large-scale integration of renewable energy and the promotion of dual-carbon policies have led to randomness, volatility, and intermittency in the power system. The grid connection of a high proportion of renewable energy urgently requires large-capacity, long-cycle energy storage technologies to balance supply and demand and smooth fluctuations. Hydrogen energy storage technology adopts an "electricity-hydrogen-electricity" end-to-end energy storage model, which integrates water electrolysis for hydrogen production, high-pressure hydrogen storage, and fuel cell power generation. It has advantages such as large storage capacity, high energy density, and long storage cycle, and can provide ancillary services such as peak shaving and frequency regulation for the power grid. It has become an important support for building integrated energy systems and new power systems.

[0003] However, existing hydrogen energy storage power plant control systems typically employ rule-based, model predictive control (MPC), or heuristic algorithms. These methods rely on precise modeling of equipment such as electrolyzers and fuel cells, as well as the external environment, making it difficult to address the nonlinear, time-varying, and multi-coupling characteristics of hydrogen energy storage systems. When operating conditions or system parameters change, the preset control rules or models quickly become ineffective, making it difficult to maintain system robustness. While traditional heuristic algorithms can solve complex optimization problems, they are computationally intensive and time-consuming, failing to meet the grid's requirements for real-time response to energy storage systems and struggling to achieve global optimization. Simultaneously, hydrogen energy storage systems themselves suffer from low electro-hydrogen-electricity conversion efficiency, limited equipment lifespan, and high operating costs; auxiliary equipment energy consumption and safety risks cannot be ignored. How to improve operational economy and extend the lifespan of critical equipment while ensuring safety has become an urgent technical challenge. Summary of the Invention

[0004] To address the problems of existing hydrogen energy storage power station control systems relying on manual rules or precise models, which are difficult to adapt to changing operating environments and balance economic efficiency and equipment lifespan, this invention provides a multi-timescale collaborative optimization control system and its control method for hydrogen energy storage power stations based on reinforcement learning. Through a hierarchical decoupled control architecture, it achieves improved economic efficiency and extended equipment lifespan for hydrogen energy storage power stations while ensuring operational safety.

[0005] The main objective of this invention is to provide a control system for hydrogen energy storage power plants. By introducing reinforcement learning to achieve macro-level energy dispatch decisions, the system's adaptability under complex electricity price and renewable energy fluctuations is improved. Through a multi-timescale hierarchical control architecture, intelligent decision-making and industrial-grade execution control are decoupled, avoiding the safety and reliability risks caused by reinforcement learning directly controlling the underlying equipment. The operating status and collaborative strategies of auxiliary equipment are incorporated into a unified optimization framework, reducing the energy consumption costs caused by the operation of high-energy-consuming auxiliary equipment during periods of high electricity prices. Equipment lifespan-related factors are introduced into the optimization objective to suppress the adverse effects of frequent start-stop, large power fluctuations, and overheating on equipment lifespan, thereby reducing the total life-cycle operating cost.

[0006] To achieve the above objectives, the present invention adopts the following technical solution.

[0007] This invention provides a hydrogen energy storage power station control system, characterized in that the control system is divided into four independent control layers from top to bottom according to the time scale, including: The reinforcement learning decision layer is used to output the operating mode, target power and auxiliary equipment coordination strategy of the hydrogen energy storage power station based on the state input reflecting the operating status of the hydrogen energy storage power station in the first time scale using reinforcement learning algorithms. The operating modes include at least hydrogen production mode, power generation mode, standby mode, and hybrid mode.

[0008] An optimization coordination layer is used to decompose the target power and auxiliary device coordination strategy output by the reinforcement learning decision layer into specific power allocation and start / stop sequence of each subsystem within a second time scale, and to perform constraint checks on the target power and auxiliary device coordination strategy; when the check fails, a preset conservative strategy is used to replace the output of the reinforcement learning decision layer.

[0009] The equipment control layer is used to execute power commands and start / stop commands issued by the optimization and coordination layer within a third time scale, employing industrial control algorithms such as PID control, model predictive control, or feedforward compensation control, to drive the electrolyzer, fuel cell, and auxiliary equipment to operate stably.

[0010] An independent safety layer is used to monitor safety indicators such as hydrogen concentration, hydrogen storage pressure, and equipment temperature in real time through sensors and actuators that are decoupled from the above layers. When the safety indicators are detected to exceed the preset threshold, emergency shutdown, pressure relief, or power limiting measures are directly triggered.

[0011] The first time scale is 15 minutes to 1 hour, the second time scale is 1–5 minutes, and the third time scale is 1–10 seconds.

[0012] In a preferred embodiment, the state input of the reinforcement learning decision layer includes the external environment state, core energy storage state, auxiliary system readiness, equipment constraint state, and historical operation mode characteristics.

[0013] in: The external environmental conditions include real-time electricity prices and their forecasts, renewable energy forecasts, and ambient temperature. The core energy storage status includes the hydrogen charge state and pressure level of the hydrogen storage tank; The readiness of the auxiliary system includes the readiness of the pure water supply, the readiness of the compression channel, and the power system load rate; The equipment constraints include the current available power limits of the electrolyzer and fuel cell; The historical operation mode features include the operation mode at the previous moment and the duration of the current mode.

[0014] The action space of the reinforcement learning decision layer is a hybrid discrete-continuous space, including operation mode actions, continuous power target actions, and auxiliary device cooperative strategy actions, wherein: The continuous power target action is used to set the target power of the electrolyzer and fuel cell, and its value ranges from zero to the upper limit of the available power of the corresponding equipment. The auxiliary equipment coordination strategy includes at least a compressor operation strategy and a pure water preparation strategy.

[0015] In a preferred embodiment, the reinforcement learning decision layer optimizes the control strategy based on a reward function, which is used to characterize the overall operational economy of the hydrogen energy storage power station.

[0016] The reward function includes at least the following: Revenue items reflect the operating income of the power plant; operating cost items reflect the energy and resource consumption during the operation of the power plant; and equipment life loss items reflect the impact of equipment operating status on its service life.

[0017] The equipment life loss item is determined based on at least one or more of the following: equipment power variation, number of equipment start-ups and shutdowns, and duration of equipment overheating, in order to suppress operating behaviors that are detrimental to equipment life.

[0018] The present invention also provides a control method for a hydrogen energy storage power station based on the above-mentioned control system, which includes the following steps: Within the first time scale, acquire status information characterizing the operating status of the hydrogen energy storage power station; Based on the aforementioned state information, a reinforcement learning decision model is used to output the operating mode, target power, and auxiliary equipment coordination strategy of the hydrogen energy storage power station. Within the second time scale, the operating mode, target power, and auxiliary equipment coordination strategy are decomposed and constrained. Within the third time scale, industrial control algorithms are used to execute power allocation and start / stop commands after constraint checks. Independently monitor safety indicators and trigger safety protection measures when safety indicators exceed limits.

[0019] Compared with the prior art, the present invention has at least the following beneficial effects: By using a multi-time-scale hierarchical control structure, the system effectively decouples intelligent decision-making from industrial-grade execution control, thereby improving the reliability and security of system operation. By introducing a reinforcement learning decision layer, the adaptability of hydrogen energy storage power stations to electricity price fluctuations and the uncertainty of renewable energy can be improved. By incorporating auxiliary equipment coordination strategies into a unified optimization framework, the operating costs of high-energy-consuming auxiliary equipment can be reduced. By incorporating equipment lifespan-related factors into the optimization objectives, frequent start-stop cycles and drastic power fluctuations can be reduced, extending the lifespan of critical equipment and lowering the total lifecycle operating cost. Attached Figure Description

[0020] To better understand this invention, its structure and working principle are described below with reference to the accompanying drawings. The drawings do not limit the scope of protection of this invention.

[0021] Figure 1 It is a control system for a hydrogen energy storage power station. Figure 1 System process flow diagram; Figure 2 This is a schematic diagram of a four-layer control architecture for a hydrogen energy storage power station control system. Detailed Implementation

[0022] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the following embodiments are used to explain the technical solutions of the present invention, and not to limit the scope of protection of the present invention. Equivalent substitutions or modifications made by those skilled in the art to the specific embodiments without departing from the technical concept of the present invention should all fall within the scope of protection of the present invention.

[0023] like Figure 1As shown, this invention is applied to a typical hydrogen energy storage power station physical system. The power grid provides DC power to the electrolyzer via a rectifier and transformer. The electrolyzer converts electrical energy into hydrogen gas through a KOH electrolyte circulation tank. The generated hydrogen gas passes through a gas-liquid separator, a gas cooler, and a purification and drying device to obtain high-purity hydrogen gas, which is then compressed by a hydrogen compressor and stored in a high-pressure hydrogen storage tank. Pure water, obtained from industrial water through a water treatment system including chemical dosing, reverse osmosis, and filtration, is supplied to the electrolyzer. The hot water produced by the electrolyzer is recovered by the cooling system, while the cold water is reused to cool the electrolyzer and compressor. When electrical energy is required, hydrogen gas is sent from the hydrogen storage tank to the fuel cell stack via a hydrogen supply pipeline, where it is converted into AC power by an inverter / transformer and connected to the grid. In addition, the power station is equipped with auxiliary equipment such as cooling pumps, control systems, valves, and sensors.

[0024] like Figure 2 As shown, based on covering all the above-mentioned physical devices, this invention proposes the following four-layer decoupled control architecture: The reinforcement learning decision-making layer, with a decision cycle of 15 minutes to 1 hour, is responsible for macroscopic energy management and collaborative decision-making with auxiliary equipment. The input is a 12-dimensional state vector S processed by feature engineering. t This includes real-time electricity prices and forecasts, predicted renewable energy output of the power plant, ambient temperature, SOC and pressure levels of the hydrogen storage tank, pure water supply readiness, compression channel readiness, power system load rate, available power limits for the electrolyzer and fuel cell, and the operating mode and duration of the previous time step. This layer utilizes deep reinforcement learning algorithms (such as soft actor-commentator SAC, proximal policy optimization PPO, or dual-delay deep deterministic policy gradient TD3) to output a hybrid discrete-continuous action A. t The operation consists of operating modes, target power, and auxiliary strategies. Operating modes include hydrogen production, power generation, standby, and hybrid modes; target power refers to the target power value of the electrolyzer and fuel cell; auxiliary strategies include compressor operating strategies (immediate compression, delayed compression for 30 minutes, or low-power pre-compression) and pure water preparation strategies (immediate water production, delayed water production for 1 hour, or use of stored water). The reinforcement learning decision layer does not directly output execution instructions such as valve opening degree or current value, but provides interpretable and easily manually auditable macroscopic strategies.

[0025] The optimization and coordination layer, with a timescale of 1–5 minutes, decomposes the decision-making layer's target power and strategy into specific execution instructions for each subsystem and performs constraint checks. For example, when the decision-making layer provides the target power for the electrolyzer... After the target power of the fuel cell The coordination layer determines the power allocation between the electrolyzer and compressor based on the characteristic curves of the electrolyzer stack and auxiliary systems, the current hydrogen reserves, and the buffer tank pressure. When the decision-making layer requests a 30-minute delay in compression, the coordination layer records the compressor's start and stop times and checks whether the buffer tank pressure is permissible. The coordination layer also performs a feasibility check on the target power, determining whether it exceeds the current available power range of the equipment, whether the mode switching meets the minimum interval time, and whether the liquid level and pump status of the pure water preparation meet the conditions. If the check fails, the coordination layer automatically uses a conservative default strategy to replace the decision-making layer's output.

[0026] The equipment control layer, with a time scale of 1–10 seconds, primarily uses mature industrial control algorithms (such as PID control, model predictive control, and feedforward compensation) to control equipment such as the electrolyzer BMS, compressor inverter, valves, and fuel cell controller, ensuring stable operation of voltage, current, pressure, and temperature within safe ranges. For example, given the large temperature fluctuations and strong nonlinearity of the electrolyzer, this layer can employ feedforward compensation combined with PID control to reduce overshoot; regarding compressor surge issues, this layer can use model predictive control combined with constraint controllers to ensure the compressor stays away from the surge zone.

[0027] An independent safety layer, completely independent of the three layers mentioned above, implements millisecond-level safety interlocks based on a hardware PLC. This safety layer uses independent sensors to monitor key safety indicators in real time, such as hydrogen concentration, hydrogen storage tank pressure, and electrolyzer temperature. Fixed threshold trigger conditions are set; for example, if the hydrogen concentration exceeds 1000 ppm, the emergency shut-off valve is immediately closed and the system emergency shutdown is initiated; if the hydrogen storage tank pressure exceeds 95% of the design pressure, the pressure relief valve is opened and the compressor stops; if the electrolyzer temperature exceeds the limit, power is forcibly reduced. When any safety trigger occurs, the safety layer can directly control the actuators, without relying on instructions from the decision-making or coordination layers, ensuring system safety.

[0028] To avoid the curse of dimensionality and improve training stability, this invention simplifies the state space design, keeping the dimensionality within twelve dimensions. External environmental states include real-time electricity prices and four-hour forecasts, renewable energy output forecasts, and ambient temperature. Core energy storage states include the SOC and pressure level of the hydrogen storage tank. Auxiliary system readiness is composed of three-dimensional features: pure water supply readiness, compression channel readiness, and power system load rate, calculated from the pure water tank level and reverse osmosis operating status, buffer tank pressure and compressor status, and inverter load rate, respectively. Equipment constraint states include the current available power limits of the electrolyzer and fuel cell. Historical operating modes include the operating mode at the previous moment and the duration of the current mode. Sensor data related to safety, such as hydrogen concentration and flame arrester pressure difference, are not included in the reinforcement learning state space but are processed by a separate safety layer.

[0029] The reinforcement learning action space adopts a hybrid discrete-continuous form. Discrete actions represent operating modes, with a value set of {hydrogen production mode, power generation mode, standby mode, and hybrid mode}. Continuous actions represent the target power of the electrolyzer and fuel cell, with values ​​ranging from zero to the upper limit of the equipment's available power. Discrete actions also include auxiliary equipment coordination strategies: compressor strategies (immediate compression, delayed compression for 30 minutes, or low-power pre-compression) and pure water preparation strategies (immediate water production, delayed water production for 1 hour, or use of stored water). This high-level action design does not directly control current or valves, but outputs strategies with clear business implications, facilitating engineer understanding and debugging.

[0030] The reward function of this invention takes the economic efficiency throughout the entire life cycle as the sole optimization objective and consists of three parts: net revenue, explicit costs, and lifetime depreciation costs. Net revenue includes electricity sales revenue, subsidies for participating in ancillary services such as peak shaving and frequency regulation, and carbon trading revenue. Explicit costs include the electricity cost of hydrogen production via electrolyzers, the electricity cost of auxiliary equipment such as compressors and water purifiers, and water consumption costs. Lifetime depreciation costs are calculated based on the power variation amplitude ΔP and the number of start-stop cycles N. start and overheating duration T overheat Calculation, i.e., C 寿命 =k1·ΔP + k2·N start + k3·T overheat This reward design eliminates the trade-offs of security factors, with security handled by an independent hardware protection layer. As a result, reinforcement learning is not affected by security constraints, allowing it to focus on economic optimization.

[0031] Since industrial hydrogen energy storage systems cannot be directly used as reinforcement learning training environments, this invention proposes a training deployment scheme combining offline training, shadow mode verification, and online fine-tuning. First, 1–2 years of historical operating data are acquired through the SCADA system of the hydrogen energy storage power station. Offline training is then performed using batch reinforcement learning algorithms (such as BCQ and CQL) to construct a data-driven environment model, learning system dynamics and initial strategies. Subsequently, the system runs in shadow mode for 3–6 months, where reinforcement learning outputs decision suggestions but does not execute them; only the differences from manual scheduling and the execution effect of strategy predictions are recorded. In the initial stage of formal deployment, a "reinforcement learning suggestion + manual review" model is adopted, where the dispatcher can accept, modify, or reject algorithm suggestions, and the manual correction information is used for online fine-tuning of strategy parameters. Once sufficient empirical data (e.g., 1000 empirical tuples) is accumulated or predetermined performance indicators are reached, continuous online fine-tuning is performed, eventually gradually relaxing manual review to achieve autonomous decision-making.

[0032] The decision-making process of this invention includes the following steps: State awareness: The SCADA system collects raw sensor data such as electricity price, SOC, temperature, pressure, and liquid level every 15 minutes, and obtains a 12-dimensional state vector S through feature fusion calculation. tAnd input it into the reinforcement learning decision layer.

[0033] Strategy Reasoning: The Actor network, based on the current state S t Forward propagation outputs actions, including operating modes, power targets, and auxiliary equipment strategies. Attention mechanisms, or SHAP values, are used to provide rationale for decisions and enhance interpretability.

[0034] Constraint check: The optimization and coordination layer performs a rationality check on the action. If the power target exceeds the available range of the equipment, the mode switching interval is too short, or the start-stop conditions of the auxiliary machine are not met, a conservative strategy is adopted. If the check is passed, proceed to the next step.

[0035] Command issuance: The coordination layer translates macroscopic objectives into specific equipment power allocation and start / stop timing, and sends them to the equipment control layer. The equipment control layer uses algorithms such as PID / MPC to drive each actuator, and the physical system runs one decision cycle Δt according to the command.

[0036] Feedback and Rewards: Record the system's operating benefits, explicit costs, and lifespan depreciation costs, and calculate the reward value R. t And store experience tuples (S) t A t ,R t ,S t+1 ).

[0037] Online learning (optional): When enough experience data has been accumulated, the strategy model is fine-tuned online, and the new strategy is verified in a simulation environment. The updates are gradually deployed through A / B testing.

[0038] To highlight the inventiveness of this invention, a comparative study was conducted using rule-based control, model predictive control (MPC), and the method of this invention. Experiments simulated the operation of a hydrogen energy storage power station under scenarios with stochastic electricity price fluctuations and renewable energy output, primarily comparing indicators such as economic benefits, auxiliary equipment energy consumption, and equipment lifespan. The results show: Compared to rule-based control methods, this invention does not require manual pre-setting of complex IF-THEN rules. Instead, it automatically learns the optimal strategy, demonstrating strong capabilities to handle complex scenarios and eliminating the need to maintain a large rule base. RL can discover optimization opportunities that human experts may miss, such as starting the compressor in advance for pre-compression during periods of low electricity prices and stopping compression and using stored hydrogen during periods of high electricity prices, thereby achieving off-peak electricity consumption for auxiliary equipment. Experiments show that BOP energy consumption costs are reduced by approximately 20–30%, and annual economic benefits are increased by approximately 15–25%.

[0039] Compared to model predictive control (MPC) methods, this invention eliminates the need to construct high-precision mathematical models, thus avoiding the effects of model mismatch and equipment aging. Comparative analysis under typical operating conditions shows that MPC performance declines by 15–20% after equipment aging, while this solution maintains performance fluctuations within ±5% through online learning; simultaneously, the policy network forward propagation time is less than 1 second, and the calculation speed is 100–1000 times faster, meeting real-time control requirements.

[0040] Compared to heuristic optimization algorithms, the policy network of this invention directly infers, eliminating repeated iterative optimization at each decision point, thus improving decision-making speed by two orders of magnitude. Reinforcement learning, through long-term reward optimization, makes the policy more stable, significantly reducing the number of electrolyzer start-ups and shutdowns and power fluctuations. Experimental results show that heuristic algorithms, pursuing short-term profit maximization, result in electrolyzer start-ups and shutdowns exceeding 200 times per year, with a lifespan of approximately 60,000 hours. This invention, by guiding the use of lifespan depreciation costs, reduces the number of start-ups and shutdowns to less than 50 times per year, extending the electrolyzer lifespan to approximately 80,000 hours and extending the equipment overhaul cycle by approximately 30%.

[0041] This invention incorporates auxiliary equipment into the optimization process, enabling high-energy-consuming equipment such as compressors and water purifiers to operate during off-peak electricity prices, avoiding operation during peak electricity price periods, thus significantly reducing BOP (Balance of Plant) energy consumption. In experiments, the off-peak electricity strategy, which allows compressors to operate during off-peak electricity prices (0.3 yuan / kWh), saves approximately 75% of energy costs compared to operating during peak electricity prices (1.2 yuan / kWh), resulting in annual operating cost savings of 500,000–1,000,000 yuan / MW of installed capacity. Electrolyzer utilization is increased from approximately 70% using traditional methods to 85%, and the annual utilization hours of the hydrogen energy storage power station increase by 15–20%, leading to a corresponding increase in return on investment.

[0042] Although this invention employs reinforcement learning decision-making to achieve multi-timescale collaborative optimization of hydrogen energy storage power plants, other alternative technical solutions exist, such as rolling optimization based on model predictive control (MPC), fuzzy logic control based on expert rule bases, decision-making methods based on supervised learning or heuristic algorithms, and implementation methods using different reinforcement learning algorithms. These solutions can achieve similar optimization objectives, but all suffer from strong model dependence, poor real-time performance, or unsatisfactory long-term results. The core of this invention lies in its four-layer decoupled control architecture and full lifecycle economic reward function. Any equivalent substitution method falling within the scope of protection of the claims of this invention is also within the scope of protection of this invention.

[0043] This invention proposes a reinforcement learning-based multi-timescale collaborative optimization control system for hydrogen energy storage power plants. By constructing a four-layer decoupled control architecture, introducing auxiliary equipment readiness features, designing a reward function that includes explicit costs and lifetime attrition costs, and formulating an engineering-oriented training and deployment strategy, the system achieves a comprehensive improvement in the economy, reliability, and safety of hydrogen energy storage systems. Experimental results show that this system can significantly reduce the energy consumption of auxiliary equipment, increase the economic benefits of hydrogen energy storage power plants, and extend the lifespan of critical equipment while ensuring safety. The technical solution of this invention is applicable to various electrolyzer-fuel cell type hydrogen energy storage devices and can be extended to other energy storage systems that include high-energy-consuming auxiliary equipment.

Claims

1. A control system for a hydrogen energy storage power station, characterized in that, The hydrogen energy storage power station includes an electrolyzer, a fuel cell, a compressor, a pure water machine, and a cooling system. The control system is divided into four independent control layers from top to bottom according to the time scale. The four independent control layers include a reinforcement learning decision layer, an optimization coordination layer, an equipment control layer, and an independent safety assurance layer. The reinforcement learning decision layer operates on a first timescale of 15 minutes to 1 hour, and is used to output the operating mode, target power, and auxiliary equipment coordination strategy of the hydrogen energy storage power station based on the state input of the control system using a deep reinforcement learning algorithm. The state inputs of the reinforcement learning decision layer include external environment state, core energy storage state, auxiliary system readiness, equipment constraint state, and historical operating mode characteristics. The external environment state is 3-dimensional, including real-time electricity price and forecast for the next four hours, renewable energy predicted output, and ambient temperature. The core energy storage state is 2-dimensional, including the hydrogen charge state and pressure level of the hydrogen storage tank. The auxiliary system readiness is 3-dimensional, including pure water supply readiness, compression channel readiness, and power system load rate. The equipment constraint state is 2-dimensional, including the current available power limit of the electrolyzer and fuel cell. The historical operation mode features are two-dimensional, including the operation mode at the previous moment and the duration of the current mode; The action space of the reinforcement learning decision layer is a hybrid discrete-continuous space, including: operation mode actions, whose values ​​include hydrogen production mode, power generation mode, standby mode, and hybrid mode; continuous power target actions, used to set the target power of the electrolyzer and the target power of the fuel cell, whose values ​​range from zero to the upper limit of the available power of the corresponding equipment; and auxiliary equipment coordination strategy actions, including compressor operation strategy and pure water preparation strategy; the compressor operation strategy includes immediate compression, delayed compression for 30 minutes, or low-power pre-compression, and the pure water preparation strategy includes immediate water production, delayed water production for 1 hour, or use of stored water; The reward function of the deep reinforcement learning is used to characterize the comprehensive operational economy of the hydrogen energy storage power station. The reward function includes at least a revenue item reflecting the operating benefits of the power station, an operating cost item reflecting the energy and resource consumption during the operation of the power station, and an equipment life loss item reflecting the impact of equipment operating status on the service life. The equipment life loss item is determined based on at least one or more of the following: power change, number of start-stop cycles, and overheating duration. The optimization and coordination layer operates on a second timescale of 1–5 minutes. It decomposes the target power and coordination strategy output by the reinforcement learning decision layer into specific power allocations and start / stop sequences for each subsystem, and performs constraint checks on the target power and auxiliary equipment coordination strategy. If the power target exceeds the available range of the equipment, the mode switching interval is less than the minimum time, or the start / stop conditions of the auxiliary equipment are not met, a preset strategy is used as a substitute. The optimization and coordination layer uses one or more methods from linear programming, quadratic programming, dynamic programming, or model predictive control to solve for the equipment power allocation. The device control layer operates on a third time scale of 1–10 seconds, and is used to execute power commands and start / stop commands issued by the optimization coordination layer using PID control, model predictive control, or feedforward compensation control. The independent safety assurance layer is decoupled from the reinforcement learning decision layer, optimization coordination layer, and equipment control layer. The independent safety assurance layer includes independent sensors and actuators, and is triggered by physical thresholds. The thresholds include at least hydrogen concentration exceeding a predetermined ppm range, hydrogen storage tank pressure exceeding the design pressure, and electrolyzer temperature exceeding the safety limit. When the independent safety assurance layer detects the over-limit conditions, it triggers emergency shutdown, pressure relief, or power limiting measures. The training method for the reinforcement learning decision layer includes: collecting historical operating data containing equipment status, control actions, and operating results from the SCADA system of the hydrogen energy storage power station; using a batch reinforcement learning algorithm to train the historical operating data offline to obtain an initial policy network; running the initial policy network in shadow mode, recording the differences between policy suggestions and manual scheduling results, and evaluating policy performance; and fine-tuning the policy online to achieve preset performance indicators and realize autonomous decision-making.

2. The hydrogen energy storage power station control system according to claim 1, characterized in that, The hardware thresholds for triggering emergency shutdown, pressure relief, or power limiting measures in the independent safety protection layer are: hydrogen concentration exceeding 1000ppm, hydrogen storage tank pressure exceeding 95% of design pressure, and electrolyzer temperature exceeding the safety limit.

3. The hydrogen energy storage power station control system according to claim 1, characterized in that, The deep reinforcement learning uses SAC, PPO, or TD3 algorithms to construct the policy network.

4. A control method for a hydrogen energy storage power station according to any one of claims 1-3, characterized in that, Includes the following steps: S1. Within a first timescale of 15 minutes to 1 hour, acquire state information characterizing the operating status of the hydrogen energy storage power station. The state information includes external environment status, core energy storage status, auxiliary system readiness, equipment constraint status, and historical operating mode characteristics, and construct a twelve-dimensional state vector. S2. Based on the deep reinforcement learning decision model, the operation mode, target power and auxiliary equipment coordination strategy of the hydrogen energy storage power station are output. The reinforcement learning is only based on the comprehensive operation economic reward function optimization. The reward function eliminates the trade-off of safety factors. The reinforcement learning is not affected by safety constraints. S3. Within a second timescale of 1–5 minutes, the power allocation and start-up / shutdown sequence of the electrolyzer, fuel cell and auxiliary equipment are decomposed according to the operating mode, target power and auxiliary equipment coordination strategy, and the target power and auxiliary equipment coordination strategy are constrained and checked. If the constraints are not met, a preset conservative strategy is used instead. S4. When the constraint check passes, within a third time scale of 1–10 seconds, the power distribution and start / stop commands are executed using PID control, model predictive control, or feedforward compensation control. S5. Independent of steps S1–S4, an independent safety layer monitors hydrogen concentration, hydrogen storage pressure, and equipment temperature safety indicators in real time. When the safety indicators exceed the preset threshold, emergency shutdown, pressure relief, or power limiting measures are triggered. S6. Reinforcement learning network training steps: Collect SCADA historical operation data and conduct offline batch training to obtain the initial policy network; evaluate the policy through shadow mode trial operation; Manual review and online fine-tuning until autonomous decision-making is achieved.

Images