Multi-stage coordinated scheduling method and system of hydrogen-electricity coupling comprehensive energy system

Abstract

The invention discloses a multi-stage coordinated scheduling method and system for a hydrogen-electricity coupling comprehensive energy system, and the method comprises the steps: solving an initial scheduling scheme according to a day-ahead scheduling model of the hydrogen-electricity coupling comprehensive energy system, the initial scheduling scheme comprises the unit output of a thermal power generation unit in the hydrogen-electricity coupling comprehensive energy system, the renewable energy grid-connected proportion of a renewable energy power station and the electricity and hydrogen purchasing amount of a hydrogen production center; monitoring the safety confidence of the hydrogen-electricity coupled comprehensive energy system within specified time after the hydrogen-electricity coupled comprehensive energy system executes the initial scheduling scheme, if the safety confidence is greater than a preset threshold value, ending and exiting, and otherwise, skipping to the next step; and an intra-day correction network based on an MAPPO algorithm is adopted to correct the day-ahead scheduling model, and a corrected scheduling scheme is issued to the hydrogen-electricity coupling integrated energy system for execution. According to the method, the economical efficiency, the safety and the calculation efficiency performance of the hydrogen-electricity coupling system in a complex and uncertain environment can be guaranteed.

Description

Technical Field

[0001] This invention relates to the field of energy system optimization and scheduling technology, specifically to a multi-stage coordinated scheduling method and system for a hydrogen-electric coupled integrated energy system. Background Technology

[0002] As the penetration rate of renewable energy in power systems continues to increase, its uncertainties pose significant challenges to the safe and stable operation of these systems. Hydrogen-electric coupled integrated energy systems (EH-IES) offer a promising solution, leveraging the flexibility of hydrogen energy to address the uncertainties of renewable energy and promote its integration. However, EH-IES dispatch faces numerous challenges, such as coordinating conflicting interests between power and hydrogen energy systems, and quantifying the impact of uncertainties on both systems.

[0003] Currently, the main solutions for uncertain scheduling in EH-IES systems include single-stage and two-stage uncertainty scheduling methods. Single-stage methods describe uncertainties using mathematical models and incorporate them into the scheduling model, but suffer from high computational costs, conservative decision-making schemes, and difficulties in balancing computational efficiency and system security. While two-stage methods improve system operating economy to some extent, day-ahead uncertainty scheduling remains relatively conservative, and using specific distributions to describe uncertainties may lead to omissions and redundancies of critical scenarios, failing to ensure safe system operation under extreme conditions. Furthermore, existing solutions fail to effectively balance system operating economy, security, and computational burden, easily leading to problems such as power flow imbalances, overload, high operating costs, and scheduling delays. Therefore, a new scheduling scheme is urgently needed that can effectively handle the complex uncertainties in EH-IES, balance system operating economy, security, and computational burden, and improve energy utilization efficiency and system stability. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a multi-stage coordinated scheduling method and system for a hydrogen-electric coupled integrated energy system, addressing the aforementioned problems in the prior art.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A multi-stage coordinated scheduling method for a hydrogen-electric coupled integrated energy system includes the following steps: S101. An initial scheduling scheme is obtained by solving the day-ahead scheduling model of the hydrogen-electricity coupled integrated energy system. The day-ahead scheduling model aims to minimize the total scheduling cost of the hydrogen-electricity coupled integrated energy system. The initial scheduling scheme includes the unit output of the thermal power generation unit in the hydrogen-electricity coupled integrated energy system, the grid connection ratio of renewable energy power plants, and the amount of electricity and hydrogen purchased by the hydrogen production center. S102, monitor the safety confidence level of the hydrogen-electric coupling integrated energy system within a specified time after the initial scheduling scheme is executed. If the safety confidence level is greater than the preset threshold, the process ends and exits; otherwise, proceed to the next step. S103, an intraday correction network based on the MAPPO algorithm is used to correct the intraday scheduling model to ensure that the generated corrected scheduling scheme meets deterministic and chance constraints, and the corrected scheduling scheme is then issued to the hydrogen-electric coupling integrated energy system for execution; wherein the intraday scheduling model aims to minimize the total scheduling cost of the hydrogen-electric coupling integrated energy system at the time point of intraday scheduling.

[0006] Furthermore, in step S101, when obtaining the initial scheduling scheme based on the day-ahead scheduling model of the hydrogen-electric coupled integrated energy system, the day-ahead scheduling model includes an objective function. The objective function aims to minimize the total scheduling cost of the hydrogen-electric coupled integrated energy system, and its expression is: , , , , in, For the time of day, For thermal power generation unit number The cost of electricity generation at a given point in time. For renewable energy power plants Cost of power curtailment at a given point in time For the hydrogen production center Operating costs at each point in time. This refers to the number of thermal power generation units. , and For the first Preset fuel cost parameters for each thermal power generation unit For the first The first thermal power generation unit Power output at each time point For time resolution at a given point in time, and These refer to the number of wind farms and solar power plants in renewable energy power plants, respectively. and The unit penalty costs for wind power and solar power curtailment are respectively. and The first The wind farm in the first The absorption rate and predicted power generation at each point in time. and The first The solar power station was in the first The absorption rate and predicted power generation at each point in time. The number of hydrogen production centers; The levelized cost of hydrogen production by electrolyzer. Hydrogen production center In the The amount of hydrogen produced at each point in time. To levelize the cost of hydrogen storage systems, and Hydrogen production centers In the The amount of hydrogen injected and released at each time point. It is the price of electricity. Hydrogen production center In the Electricity purchased from the grid at a specific point in time. It is the price of purchasing hydrogen. Hydrogen production center In the The amount of hydrogen purchased at each point in time.

[0007] Furthermore, the day-ahead scheduling model also includes constraints, which include power system safety operation constraints and hydrogen production center safety operation constraints. The power system safety operation constraints include power balance constraints, upper limit constraints on renewable energy injection into the grid, and safety operation constraints of generating units and transmission lines. The hydrogen production center safety operation constraints include hydrogen energy balance equations, electrolyzer operation model constraints, and hydrogen storage system operation constraints. The functional expression for the power balance constraint is:

[0008] in, , , and These refer to the number of thermal power generation units, wind farms, solar power plants, and hydrogen production centers, respectively. For the first The thermal power generation unit in the first Power output at each time point and The i-th wind farm and the j-th solar power station are respectively located at the i-th wind farm and the j-th solar power station. The power injected into the grid at each point in time. For the first Electricity load at a given time point For the first The hydrogen production center is in the first Electricity purchased from the grid at a specific point in time; The functional expression for the upper limit constraint on the power injected into the grid by renewable energy is:

[0009]

[0010] The functional expression of the hydrogen energy balance equation is:

[0011] in, , and The k-th hydrogen production center is located at the k-th hydrogen production center in ... The amount of hydrogen produced at a given point in time using on-site wind power, solar power, and purchased electricity. For the k-th hydrogen production center in the th... Hydrogen load at each time point and The k-th hydrogen production center is located at the k-th hydrogen production center in ... The amount of hydrogen released and injected at each time point, For the k-th hydrogen production center in the th... The amount of hydrogen purchased at each point in time; The functional expression for the constraints of the electrolytic cell operation model is:

[0012] in, Let be the input power of the electrolyzer in the k-th hydrogen production center. This is the maximum input power of the electrolytic cell. , and They are respectively...; The functional expression for the operational constraints of the hydrogen storage system is:

[0013]

[0014]

[0015]

[0016]

[0017] in, and The hydrogen storage system of the k-th hydrogen production center is located in the time period. and time period Hydrogen storage capacity, and These refer to the injection efficiency and release efficiency of the hydrogen storage system, respectively. This is the maximum capacity of the hydrogen storage tank. and These represent the maximum injection and release amounts in a single time period, respectively. and These represent the hydrogen storage levels at the beginning and end of the scheduling cycle, respectively. The amount of hydrogen injected at time t. Let be the amount of hydrogen released at time t. Let be the amount of hydrogen stored at time t.

[0018] Furthermore, in step S102, when monitoring the safety confidence level of the hydrogen-electric coupled integrated energy system within a specified time after the execution of the initial scheduling scheme, seven types of variables are used as inputs: the power output of the thermal power generation unit at the current moment, the power injected into the grid by the renewable energy power station, the renewable energy power and purchased power in the hydrogen production center for hydrogen production, the power load, the hydrogen load, and the latest predicted value of the renewable energy power station. A machine learning model is used to obtain the safety confidence level of the hydrogen-electric coupled integrated energy system within a specified time. The machine learning model is pre-trained to establish a mapping relationship between the seven types of input variables and the safety confidence level of the hydrogen-electric coupled integrated energy system within a specified time. The safety confidence level of the hydrogen-electric coupled integrated energy system within a specified time includes the safe operation probability of the thermal power generation unit, transmission lines, and hydrogen production center.

[0019] Furthermore, the machine learning model includes an input layer, an attention layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer connected in sequence, and the loss function used by the machine learning model during the training phase is mean squared error.

[0020] Further, step S103 includes: An intraday correction network based on the MAPPO algorithm is used to transform the intraday scheduling model into a multi-agent Markov game for solution. Each agent represents a decision-making unit in the hydrogen-electricity coupled integrated energy system. Solving the problems of each decision-making unit generates a corrected scheduling scheme that satisfies deterministic and chance constraints. The objective function expression of the intraday correction network based on the MAPPO algorithm is as follows: , In the above formula, The objective function used in the MAPPO algorithm. For thermal power generation unit number The cost of electricity generation at a given point in time. For renewable energy power plants Cost of power curtailment at a given point in time For the hydrogen production center Operating costs at each point in time. The safety confidence penalty coefficient. , , The safety confidence level of the hydrogen-electric coupled integrated energy system within a specified time period is determined by the safe operation probability of the thermal power generation unit, transmission lines, and hydrogen production center.

[0021] Furthermore, the decision-making units involved include: the power output increment of thermal power generation units. The power injected into the grid by renewable energy power plants Renewable energy power used for hydrogen production The electricity purchased by the hydrogen production center Hydrogen purchased by the hydrogen production center Operational plan for hydrogen storage in hydrogen production center ,in and These represent the amounts of hydrogen injected and released, respectively.

[0022] Furthermore, in step S103, the intraday correction network based on the MAPPO algorithm updates the network parameters of the agent during training using a dynamic learning rate, and the calculation function expression of the dynamic learning rate is: , in, For dynamic learning rate, As the baseline learning rate, and These are the sampling probabilities for the latest policy and the old policy, respectively. For policy network parameter symbols, Let be the sampling probability of the i-th agent based on the folding strategy. Let be the sampling probability of the i-th agent based on the old policy. Let t be the sampled action and system state of the i-th agent at time t. The system state is the action sampled by the i-th agent at time t. Let be the system state perceived by the i-th agent at time t. For the clipping function, This is the cutting factor.

[0023] A multi-stage coordinated scheduling system for a hydrogen-electric coupled integrated energy system includes an interconnected microprocessor and a memory, wherein the microprocessor is programmed or configured to execute a multi-stage coordinated scheduling method for the hydrogen-electric coupled integrated energy system.

[0024] A computer-readable storage medium storing a computer program / instructions programmed or configured to execute a multi-stage coordinated scheduling method for a hydrogen-electric coupled integrated energy system via a processor.

[0025] Compared with the prior art, the advantages of the present invention are as follows: This invention derives an initial scheduling scheme from the day-ahead scheduling model of a hydrogen-electric coupled integrated energy system. The day-ahead scheduling model allows for rapid solution generation, yielding the most economically optimal solution within a finite timeframe and providing a reliable benchmark for subsequent scheduling. Furthermore, this invention monitors the safety confidence level of the hydrogen-electric coupled integrated energy system within a specified timeframe after executing the initial scheduling scheme. This enables real-time assessment of the system status, and subsequent steps are initiated only when necessary for correction (when the safety confidence level is less than a preset threshold), reducing redundant calculations and computational overhead from frequent adjustments. This invention employs an intraday correction network based on the MAPPO algorithm to correct the day-ahead scheduling model. This allows for flexible adjustments to the initial scheduling scheme based on the real-time changes and operating status of each subunit in the hydrogen-electric coupled integrated energy system, enhancing the system's adaptability and anti-interference capabilities. Through the economic benchmark of the day-ahead scheduling model, real-time detection of the safety confidence level, and dynamic adjustment of the intraday correction network, the multi-stage coordinated scheduling of these three elements ensures the economy, safety, and high computational efficiency of the hydrogen-electric coupled system under complex and uncertain environments. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating the multi-stage coordinated scheduling method for a hydrogen-electric coupled integrated energy system according to an embodiment of the present invention.

[0027] Figure 2 This is a schematic diagram of the structure of an online security monitoring network in a specific application embodiment.

[0028] Figure 3 This is a schematic diagram of the training mechanism of the intraday correction network in a specific application embodiment. Detailed Implementation

[0029] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0030] like Figure 1 As shown, the multi-stage coordinated scheduling method for a hydrogen-electric coupled integrated energy system according to an embodiment of the present invention includes the following steps: S101. An initial scheduling scheme is obtained by solving the day-ahead scheduling model of the hydrogen-electricity coupled integrated energy system. The day-ahead scheduling model aims to minimize the total scheduling cost of the hydrogen-electricity coupled integrated energy system. The initial scheduling scheme includes the unit output of the thermal power generation unit in the hydrogen-electricity coupled integrated energy system, the grid connection ratio of renewable energy power plants, and the amount of electricity and hydrogen purchased by the hydrogen production center. S102, monitor the safety confidence level of the hydrogen-electric coupling integrated energy system within a specified time after the initial scheduling scheme is executed. If the safety confidence level is greater than the preset threshold, the process ends and exits; otherwise, proceed to the next step. S103 uses an intraday correction network based on the MAPPO algorithm to correct the day-ahead scheduling model to ensure that the generated corrected scheduling scheme meets deterministic and chance constraints, and then distributes the corrected scheduling scheme to the hydrogen-electricity coupled integrated energy system for execution.

[0031] In specific application embodiments, the hydrogen-electricity coupled integrated energy system model is generally constructed first. In the hydrogen-electricity coupled integrated energy system, the power system and the hydrogen energy system are coupled through a hydrogen production center (HPC). The HPC is the key hub connecting the two systems, used for the production, storage, and distribution of hydrogen energy. The HPC consists of transformers, rectifiers, electrolyzers, and compressors. The transformers are responsible for converting the high voltage from the grid into a low voltage suitable for the electrolyzers; the rectifiers convert alternating current (AC) to direct current (DC) to meet the power requirements of the electrolyzers; the electrolyzers use DC to decompose water into hydrogen and oxygen, thereby realizing hydrogen production; and the compressors compress the produced hydrogen for storage and transportation.

[0032] In a hydrogen-electricity coupled integrated energy system, the hydrogen produced by the HPC (High-Potential Hydrogen Processing Unit) is used to meet local hydrogen demand, such as for hydrogen refueling stations for fuel cell vehicles and industrial hydrogen users. When the produced hydrogen exceeds local demand, the excess hydrogen is injected into a hydrogen storage system (HSS) for storage. Conversely, when local hydrogen demand is high and the HPC's production capacity cannot meet it, the hydrogen stored in the HSS is released, and hydrogen can also be purchased from the external hydrogen market to supplement the supply. Based on the energy source, HPCs can be divided into three types: PV-HPC: Deployed near solar power plants, these systems primarily utilize electricity generated by solar photovoltaic panels to produce hydrogen. Solar energy is a renewable energy source with advantages such as being clean and pollution-free. When sunlight is abundant, PV-HPC can fully leverage solar energy for efficient hydrogen production. It can also serve as a flexible power load, responding to peak-shaving demands on the grid based on time-of-use pricing. For example, during periods of low electricity prices, PV-HPC can increase hydrogen production to consume excess electricity; during periods of high electricity prices, it can reduce hydrogen production, feeding more electricity into the grid.

[0033] Wind-HPC: Built around wind farms, it uses electricity generated by wind turbines to produce hydrogen. Wind energy is a rich renewable energy source, and wind-HPC can effectively absorb the electricity from wind farms, reducing wind curtailment. Similarly, it can flexibly adjust its hydrogen production capacity according to the grid's needs, participating in peak shaving for the grid.

[0034] Grid electricity - HPC: For regions that are far from renewable energy power plants but have a high demand for hydrogen, grid electricity - HPC is a suitable option. It mainly relies on purchasing electricity from the grid to produce hydrogen, which can meet the local stable hydrogen demand.

[0035] In step S101 of this embodiment, when obtaining the initial scheduling scheme based on the day-ahead scheduling model of the hydrogen-electric coupled integrated energy system, the day-ahead scheduling model includes an objective function. The objective function aims to minimize the total scheduling cost of the hydrogen-electric coupled integrated energy system, and its function expression is: , , , , in, This represents the number of time points within a day in the current day's scheduling model. For thermal power generation unit number The cost of electricity generation at a given point in time. For renewable energy power plants Cost of power curtailment at a given point in time For the hydrogen production center Operating costs at each point in time. This refers to the number of thermal power generation units (thermal power generation units are traditional power generation equipment that generate electricity by burning fossil fuels such as coal and natural gas). , and For the first The preset fuel cost parameters for each thermal power generation unit (these parameters are related to factors such as the type and efficiency of the thermal power generation unit). For the first The first thermal power generation unit Power output at each time point (in kilowatts). The time resolution for a given point in time (can be set to 15 minutes, meaning that a 24-hour day is divided into multiple 15-minute time intervals for scheduling). and These refer to the number of wind farms and solar power plants in renewable energy power plants, respectively. and The unit penalty cost for wind power and solar power curtailment (in yuan / kW; when renewable energy generation cannot be fully absorbed and is curtailed, a certain economic loss will occur, and this unit penalty cost is used to measure this loss). and The first The wind farm in the first The absorption rate and predicted power generation at each time point (the absorption rate represents the proportion of actual power generated by the wind farm that is utilized, and the predicted power generation is the wind farm's power generation capacity predicted in advance based on meteorological data, etc.) and The first The solar power station was in the first The absorption rate and predicted power generation at each point in time. The number of hydrogen production centers; The levelized cost of hydrogen production by electrolyzer (unit: yuan / kg, representing the average cost of producing one kilogram of hydrogen). Hydrogen production center In the The amount of hydrogen produced at each time point (in kilograms). The levelized cost of hydrogen storage systems (in yuan / kg, used to measure the cost of hydrogen storage). and Hydrogen production centers In the The amount of hydrogen injected and released at each time point (in kilograms). This is the price of electricity (in yuan / kilowatt). Hydrogen production center In the Electricity purchased from the grid at a given time point (in kilowatts). This is the price of hydrogen (in yuan / kilogram). Hydrogen production center In the The amount of hydrogen purchased at each point in time (in kilograms).

[0036] In this embodiment, the day-ahead scheduling model further includes constraints, which include power system safety operation constraints and hydrogen production center safety operation constraints. The power system safety operation constraints include power balance constraints, upper limit constraints on renewable energy injection into the grid, and safety operation constraints of generating units and transmission lines. The hydrogen production center safety operation constraints include hydrogen energy balance equations, electrolyzer operation model constraints, and hydrogen storage system operation constraints. The functional expression for the power balance constraint is:

[0037] The power balance constraint states that in each time period, the sum of the generating power of thermal power units and the power injected into the grid by renewable energy power plants equals the sum of the power load and the power purchased from the grid by HPC, ensuring the power balance of the power system; where, , , and These refer to the number of thermal power generation units, wind farms, solar power plants, and hydrogen production centers, respectively. For the first The thermal power generation unit in the first Power output (MW) at each time point. and The i-th wind farm and the j-th solar power station are respectively located at the i-th wind farm and the j-th solar power station. Power injected into the grid at each time point (MW). For the first Electricity load (MW) at each time point. For the first The hydrogen production center is in the first Electricity (MW) purchased from the grid at each point in time; The functional expression for the upper limit constraint on the power injected into the grid by renewable energy is:

[0038]

[0039] The upper limit constraint on the power of renewable energy injected into the grid limits the power injected into the grid by wind farms and solar power plants to not exceed their predicted power generation, so as to avoid over-generation. Safety constraints for generating units and transmission lines: The power output of thermal power units must be within their allowable range, and the transmission power of transmission lines must not exceed their capacity limits in order to ensure the safe and stable operation of the power system. The functional expression of the hydrogen energy balance equation is:

[0040] The hydrogen balance equation states that in each time period, the amount of hydrogen produced by HPC using on-site wind and solar power, purchased electricity, and the amount of hydrogen purchased, plus the amount of hydrogen released from the hydrogen storage system, equals the sum of the local hydrogen load demand and the amount of hydrogen injected into the hydrogen storage system, ensuring a balance between hydrogen supply and demand; where, , and The k-th hydrogen production center is located at the k-th hydrogen production center in ... The amount of hydrogen (kg) produced at each point in time using on-site wind power, solar power, and purchased electricity. For the k-th hydrogen production center in the th... Hydrogen load (kg) at each time point. and The k-th hydrogen production center is located at the k-th hydrogen production center in ... The amount of hydrogen released and injected at each time point (kg). For the k-th hydrogen production center in the th... The amount of hydrogen purchased at each point in time (kg); The functional expression for the constraints of the electrolytic cell operation model is:

[0041] There is a certain relationship between the operating power and hydrogen production of an electrolyzer, and specific operating curves and efficiency requirements need to be met to ensure the safe and efficient operation of the electrolyzer; among them, Let be the input power (MW) of the electrolyzer in the k-th hydrogen production center. The maximum input power (MW) of the electrolyzer. The functional expression for the operating constraints of the hydrogen storage system is:

[0042]

[0043]

[0044]

[0045]

[0046] Hydrogen storage systems have upper and lower limits on their storage capacity, and the rates of hydrogen injection and release must also be within a reasonable range to prevent overcharging or over-discharging; among these, and The hydrogen storage system of the k-th hydrogen production center is located in the time period. and time period Hydrogen storage capacity (kg). and The values ​​represent the injection efficiency and release efficiency (%) of the hydrogen storage system, respectively. This refers to the maximum capacity (kg) of the hydrogen storage tank. and These represent the maximum injection and release rates per single time period (kg / h), respectively. and These represent the hydrogen storage levels at the beginning and end of the scheduling cycle (which need to be balanced). The amount of hydrogen injected at time t. Let be the amount of hydrogen released at time t. Let be the amount of hydrogen stored at time t.

[0047] It should be noted that the models for thermal power generation units, renewable energy power plants, and hydrogen production centers are specifically distributed within the objective functions and constraints of the day-ahead dispatch model. Thermal power generation units are the system's basic power source. The core of the model revolves around "power generation cost calculation" and "safe operation constraints." Its objective function (power generation cost) is the fuel cost, represented by a quadratic function (consistent with the energy consumption characteristics of thermal power units). Safe operation constraints include power output range constraints (the actual output of thermal power units must be between the minimum and maximum technical output), ramp rate constraints (the rate of change in thermal power unit output cannot exceed the maximum ramp / ramp capacity to avoid frequent start-ups, shutdowns, or overloads), and power balance constraints (the total output of thermal power units must participate in the power system's power balance, along with renewable energy output and loads). Renewable energy power plants include wind farms and solar power plants. The core of the model is "curtailment loss cost" and "grid-connected power constraints." Its objective function is the curtailment loss cost (due to the volatility of renewable energy output such as wind and solar power, unconsumed electricity will result in curtailment losses). Safety operation constraints include grid-connected power limit constraints (the actual power injected into the grid by renewable energy cannot exceed the predicted power generation to avoid exceeding the grid's capacity); the hydrogen production center is the core hub of "hydrogen-electricity coupling", responsible for hydrogen production, storage and consumption. The core of the model is "operating cost calculation" and "hydrogen energy / equipment constraints". Its objective function (operating cost) includes the cost of hydrogen production by electrolyzer, the cost of hydrogen storage system, the cost of electricity purchase, and the cost of hydrogen purchase. Safety operation constraints include hydrogen energy balance constraints (the supply and demand of hydrogen at the hydrogen production center must be balanced, i.e., production + hydrogen purchase + hydrogen storage release = load + hydrogen storage injection), electrolyzer operation constraints (the input power of the electrolyzer must be within the rated range, and the hydrogen production is linearly related to the input power (determined by the efficiency of the electrolyzer)), and hydrogen storage system constraints (the capacity of the hydrogen storage tank and the injection / release rate must meet safety limits, and the hydrogen storage volume must be balanced at the beginning and end of the scheduling cycle to avoid long-term deficits or overflows).

[0048] In step S102 of this embodiment, when monitoring the safety confidence level of the hydrogen-electric coupled integrated energy system within a specified time after the execution of the initial scheduling scheme, seven types of variables are used as inputs: the power output of the thermal power generation unit at the current moment, the power injected into the grid by the renewable energy power station, the renewable energy power and purchased power in the hydrogen production center for hydrogen production, the power load, the hydrogen load, and the latest predicted value of the renewable energy power station. A machine learning model is used to obtain the safety confidence level of the hydrogen-electric coupled integrated energy system within a specified time. The machine learning model is pre-trained to establish a mapping relationship between the seven types of input variables and the safety confidence level of the hydrogen-electric coupled integrated energy system within a specified time. The safety confidence level of the hydrogen-electric coupled integrated energy system within a specified time includes the safe operation probability of the thermal power generation unit, transmission line, and hydrogen production center, providing a probabilistic basis for the subsequent generation of correction schemes.

[0049] In this embodiment, the machine learning model includes an input layer, an attention layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer connected in sequence, and the loss function used by the machine learning model during the training phase is mean squared error.

[0050] In a specific application embodiment, the machine learning model is specifically an online safety monitoring network. It employs a deep learning approach combined with MCS (Multi-Channel System) and takes seven types of variables as input: the current power output of the thermal power generation unit, the power injected into the grid from the renewable energy power plant, the renewable energy power used for hydrogen production and the purchased power in the hydrogen production center, the electricity load, the hydrogen load, and the latest predicted value of the renewable energy power plant. The output represents the safety confidence level of the system, using the safe operation probability of the thermal power generation unit, transmission lines, and electrolyzers in the HPC (Hyper-Channel System). This allows for real-time monitoring of the system's safety confidence level (or safety margin) over a future period (e.g., 4 hours). The network structure satisfies the following:

[0051] in, For pooling layer output, , These are the weights and biases for the fully connected layer.

[0052] like Figure 2 As shown, the structure of the online security monitoring network includes an attention layer, a convolutional layer, a pooling layer, and a fully connected layer, wherein: Attention layer: Input data first enters the attention layer, which assigns different weights to different input channels. Because different input variables have varying degrees of impact on the system's safety margin, the attention mechanism can highlight key input features, improving the network's learning efficiency and accuracy. The vector processed by the attention layer is denoted as... .

[0053] Convolutional layer: The input is fed into a convolutional layer, which extracts local features through a sliding convolution operation between the kernel and the input vector. The convolutional kernel can be viewed as a feature extractor; it slides across the input vector, performing convolution operations on each local region to obtain a series of feature maps. These feature maps contain local feature information from the input data, helping the network learn details of the system's operational state.

[0054] Pooling Layers: Following the convolutional layers are pooling layers. The main function of pooling layers is to reduce the size of features while preserving the main features, thereby reducing computational burden and the number of features. Common pooling operations include max pooling and average pooling. This invention can select the appropriate pooling method based on the actual situation.

[0055] Fully connected layer: Key feature vectors processed by convolution and pooling are input into the fully connected layer. The fully connected layer integrates these feature vectors to learn the complex relationship between the system's operating state and safety margin. The final output is the safe operating probability of the thermal power generation unit, transmission line, and electrolyzer in HPC.

[0056] The main steps for training an online security monitoring network are as follows: A large number of training samples are generated by MCS, specifically based on ultra-short-term (e.g., 4-hour) historical prediction error datasets. Renewable energy fluctuation scenarios are randomly sampled to simulate the energy flow of the system under a given operating state, so that the generated training data can truly reflect the impact of renewable energy uncertainty on safety confidence. Count the number of security constraint violations in each scenario and calculate the confidence level label; Gradient descent is used to train and optimize network parameters, with the mean squared error as the loss function.

[0057] in, This is the network output (i.e., the predicted probability of safe operation). This represents the actual confidence level (i.e., the actual probability of safe operation) under the opportunity constraint given the corresponding input vector. By continuously adjusting the network parameters, the value of the loss function is minimized, thereby improving the network's prediction accuracy.

[0058] In this embodiment, step S103 includes: An intraday correction network based on the MAPPO algorithm is used to transform the intraday scheduling model into a multi-agent Markov game for solution. Each agent represents a decision-making unit in the hydrogen-electricity coupled integrated energy system. Solving the problems of each decision-making unit generates a corrected scheduling scheme that satisfies deterministic and chance constraints. The objective function expression of the intraday correction network based on the MAPPO algorithm is as follows: , In the above formula, The objective function used in the MAPPO algorithm. For thermal power generation unit number The cost of electricity generation at a given point in time. For renewable energy power plants Cost of power curtailment at a given point in time For the hydrogen production center Operating costs at each point in time. The safety confidence penalty coefficient. , , The safety confidence level of the hydrogen-electric coupled integrated energy system within a specified time period is determined by the safe operation probability of the thermal power generation unit, transmission lines, and hydrogen production center.

[0059] In this embodiment, the decision-making unit involved includes: the power output increment of the thermal power generation unit. The power injected into the grid by renewable energy power plants Renewable energy power used for hydrogen production The electricity purchased by the hydrogen production center Hydrogen purchased by the hydrogen production center Operational plan for hydrogen storage in hydrogen production center ,in and These represent the amounts of hydrogen injected and released, respectively.

[0060] In this embodiment, in step S103, the intraday correction network based on the MAPPO algorithm updates the network parameters of the agent using a dynamic learning rate during training, and the calculation function expression of the dynamic learning rate is: , in, For dynamic learning rate, In step S103 of the basic embodiment, the intraday correction network based on the MAPPO algorithm updates the network parameters of the agent during training using a dynamic learning rate, and the calculation function expression of the dynamic learning rate is the quasi-learning rate. and These are the sampling probabilities for the latest policy and the old policy, respectively. For policy network parameter symbols, Let be the sampling probability of the i-th agent based on the "Fold Policy". Let be the sampling probability of the i-th agent based on the "Old Policy". Let t be the sampled action and system state of the i-th agent at time t. The system state is the action sampled by the i-th agent at time t. Let be the system state perceived by the i-th agent at time t. The symbol for a clip function is used to ensure that the function output is within the specified range. and between, This is the cutting factor.

[0061] In a specific application embodiment, an intraday correction network is designed based on the Multi-Agent Proximity Policy Optimization (MAPPO) algorithm. The objective of the intraday scheduling model is also to minimize the total scheduling cost of the system. The objective function is similar to that of the day-ahead scheduling model, but the time points are the same as those for intraday scheduling. Due to the uncertainty of renewable energy sources, the actual output of thermal power generation units, grid power flow, and the actual input power of electrolyzers will all change during actual operation. To ensure the safe operation of the system under uncertain conditions, deterministic constraints and chance constraints with acceptable probability of violation are formulated.

[0062] It should be noted that both the intraday scheduling model and the intraday correction network serve intraday scheduling optimization, and their ultimate goal is to achieve "economic efficiency + safety" in the intraday operation of the hydrogen-electricity coupling system: the intraday scheduling model defines the objective of "minimizing total scheduling cost + satisfying safety constraints," and the intraday correction network generates specific correction schemes by solving this model to ensure that the objective is achieved. The training and decision inputs of the intraday correction network must be based on the objective function (such as cost calculation logic) and constraints (such as equipment output limits) of the intraday scheduling model to ensure that the schemes generated by the network meet the model requirements; the output of the intraday correction network (the corrected scheduling scheme) is the "optimal solution" of the intraday scheduling model, which is directly used for system execution, transforming the model from a "theoretical framework" into "actual operational instructions." For example, if the intraday scheduling model requires "minimizing the cost of power curtailment losses," the intraday correction network achieves this objective by adjusting the proportion of renewable energy grid connection and the allocation of hydrogen production and consumption; regarding the "hydrogen storage system capacity constraint" of the intraday scheduling model, the intraday correction network will adjust the hydrogen storage injection / release volume to avoid exceeding the limit.

[0063] The intraday scheduling model, intraday correction network, and day-ahead scheduling model together constitute a multi-stage scheduling system. The day-ahead scheduling model generates an initial plan. If online monitoring finds that the safety confidence level is not up to standard, the intraday scheduling model will update the objectives and constraints based on real-time data. The intraday correction network then solves the updated model to generate a correction plan, ensuring that the system smoothly transitions from "day-ahead planning" to "intraday actual operation".

[0064] The intraday correction network, based on the MAPPO algorithm, transforms the intraday scheduling model into a multi-agent Markov game for solution. Through constraint projection and feasible sample learning mechanisms, it ensures that the correction scheme satisfies deterministic and chance constraints to achieve safe and economical system operation. The objective function is:

[0065] in, The safety confidence penalty coefficient. , , The confidence levels (%) for power flow, unit output, and electrolytic cell operation safety are respectively.

[0066] Optionally, the intraday scheduling model can also adopt a "specialized and refined version" of the intraday opportunity-constrained scheduling model. This model is only required when there are significant uncertainties in the system (such as large fluctuations in wind / solar power output and high load forecasting errors). If the system operating environment is stable (uncertainty can be ignored), the basic intraday scheduling model can meet the requirements.

[0067] In a multi-agent system, each agent represents a decision-making unit, such as a thermal power generation unit, a renewable energy power plant, or an HPC (High-Performance Computing) system. The actions of the agents include: 1) Output increment of thermal power generation unit This indicates the power output that a thermal power generation unit needs to adjust within a day, based on the day-ahead dispatch schedule.

[0068] 2) Power injected into the grid by renewable energy sources The power injected into the grid is adjusted according to the actual renewable energy generation situation and system needs.

[0069] 3) Renewable energy power used for hydrogen production , and rationally allocate renewable energy power for hydrogen production.

[0070] 4) Purchased electricity HPC determines the amount of electricity it purchases from the grid based on its hydrogen demand and the grid electricity price.

[0071] 5) Purchased hydrogen gas When local hydrogen production is insufficient, HPC purchases hydrogen from external markets.

[0072] 6) Operational plan for hydrogen storage in HPC This includes the amount of hydrogen injected and released to balance the supply and demand of hydrogen.

[0073] The following describes the "system state" and "actions performed" of each agent, as well as the system state of all agents.

[0074] Each agent in the multi-agent system represents a decision-making unit in the hydrogen-electric coupling system. The decision-making unit includes core components such as thermal power generation units, renewable energy power plants, and hydrogen production centers. The system state and execution actions of each agent specifically refer to the following: 1. Intelligent Agent for Thermal Power Generation Unit: The system state it relies on refers to key real-time / predictive information affecting the output adjustment of the thermal power unit, specifically including: real-time power load (total power demand of the system in the current period). This requires matching the balance between thermal power output and load, and real-time grid-connected power of renewable energy (the actual power injected into the grid from wind farms and solar power plants). To avoid overcapacity caused by the superposition of thermal power output and renewable energy output, and to ensure the proper functioning of thermal power units (current output). Maximum / Minimum Technical Output Climbing / Descending Rate (Ensure actions comply with equipment safety constraints) and safety confidence level (current probability of safe operation of the thermal power unit). (The actions need to be adjusted through safety confidence penalty items to avoid safety risks); the actions executed refer to the adjustable decision variables of the thermal power unit, specifically the "power output increment of the thermal power generation unit". Specifically, this refers to the initial output during the day-ahead scheduling. Based on this, the power deviation value is adjusted in real time during the day, for example = +5MW means an increase of 5MW in output compared to the initial plan. = -3MW means a reduction of 3MW in output); the action must meet the "safety constraints of thermal power units": the adjusted total output. + It must be within the range, and the absolute value must not exceed the product of the climbing / descending rate and the time resolution. ).

[0075] 2. Renewable Energy Power Plant Intelligent Agent (including wind farms and solar power plants): The system status it is based on refers to key information affecting renewable energy power distribution (grid connection / hydrogen production), specifically including: real-time predicted output of renewable energy (ultra-short-term predicted values ​​for wind farms and solar power plants). (e.g., rolling forecasts within 4 hours), serving as the basis for power allocation, grid acceptance capacity (the maximum allowable grid connection power of renewable energy in the current period, to avoid exceeding the transmission line capacity), and energy demand of the hydrogen production center (the renewable energy hydrogen production power that the hydrogen production center can currently accept). This involves coordinating the allocation of "grid-connected power" and "hydrogen production power," as well as the cost of curtailment losses (the unit cost of curtailment penalties, which must be minimized through actions). The specific actions to be performed refer to the power injected into the grid from renewable energy sources. (Real-time power output of wind farms / solar power plants directly connected to the grid must meet the upper limit constraint on grid-connected power output), renewable energy power used for hydrogen production. (The real-time power supplied directly to the hydrogen production center by the wind farm / solar power station must meet the requirement that the total allocated power does not exceed the predicted output and must coordinate with the actions of the intelligent agent of the hydrogen production center.)

[0076] 3. Hydrogen Production Center Intelligent Agent: The system status it relies on refers to the key information affecting the hydrogen production center's energy purchase, hydrogen storage, and hydrogen procurement decisions, specifically including: hydrogen supply and demand status (current hydrogen load). Current hydrogen storage capacity of the hydrogen storage system Maximum capacity of hydrogen storage tank (ensuring hydrogen energy balance) and electricity market prices (current grid purchase price) This requires optimizing power purchase strategies through pricing, and improving direct renewable energy supply capacity (the direct hydrogen production capacity output by the intelligent body of the renewable energy power plant). (As the preferred source of electricity for hydrogen production), and safety confidence (the probability of safe operation of the electrolyzers in the hydrogen production center needs to be adjusted through penalty adjustments to avoid over-power operation of the electrolyzers). The specific actions to be performed refer to: the electricity purchased by the hydrogen production center. (The real-time power purchased from the power grid for hydrogen production needs to be coordinated with...) Together, they meet the power constraints of the electrolytic cell. )≤ Hydrogen purchased from the hydrogen production center (When local hydrogen production) +Hydrogen storage release When the hydrogen load cannot be met, the amount of hydrogen purchased from the external hydrogen market must meet the hydrogen energy balance requirements, and the operation plan for hydrogen storage (real-time injection and release rates of the hydrogen storage system) must be in place. and iterative constraints on hydrogen storage capacity It needs to meet the following conditions: Within the range).

[0077] The system state for all agents is: "Shared core state + Independent dedicated state". The intraday correction network and multi-agent game logic of the MAPPO algorithm show that all agents do not use completely independent system states, nor do they completely share the same system state. Instead, they "share the core global state + retain independent dedicated states", as detailed below: 1. The shared "core global state" refers to common system information that affects the decisions of all agents. The "online safety monitoring network input variables" and "intraday scheduling model objective function" must be shared globally, including: real-time power load and real-time hydrogen load: the actions of all agents must be coordinated around the "electricity-hydrogen load balance", such as adjusting thermal power output, grid connection of renewable energy, and matching power load for hydrogen production and power purchase; ultra-short-term forecast values ​​of renewable energy: these are the common basis for renewable energy power plant agents to allocate power and for hydrogen production center agents to formulate power purchase strategies; system-level safety confidence: the probability of safe operation of thermal power units, transmission lines, and hydrogen production centers. The "safety confidence penalty" must be borne by all agents, and action adjustments must avoid global safety risks.

[0078] 2. The "independent and exclusive state" of each intelligent agent refers to personalized information that only affects the decision-making of a single intelligent agent, including: equipment operating parameters: such as the ramp rate of the thermal power unit, the maximum input power of the electrolyzer, and the capacity of the hydrogen storage tank. Only the corresponding intelligent agent needs to adjust its actions based on these parameters (e.g., the intelligent agent of the hydrogen production center does not need to pay attention to the ramp rate of the thermal power unit); local supply and demand relationship: such as the real-time hydrogen storage of the hydrogen production center, which only affects its own actions and is unrelated to the thermal power unit and renewable energy power station; the real-time output fluctuation of the wind farm only directly affects the power allocation of the renewable energy power station intelligent agent and does not require other intelligent agents to be fully aware of it.

[0079] 3. The intraday correction network "transforms the intraday scheduling model into a multi-agent Markov game," and the core characteristic of Markov games is that "agents optimize collaboratively through interaction"—if all agents use completely independent states, global constraints such as power balance and hydrogen energy balance cannot be achieved; if the same state is used, it will lead to redundant calculations (e.g., thermal power units do not need to be aware of hydrogen storage tank parameters). Therefore, "shared core state + independent dedicated state" is the implicit design logic to ensure a balance between collaboration and efficiency.

[0080] like Figure 3 As shown, the training mechanism of the intraday correction network includes the following steps: 1) Generate training samples. Multiple agents continuously interact with the environment, generate correction schemes based on the policy network, and transmit the joint actions to EH-IES to obtain rewards and the next state, forming transition samples stored in the sample pool. Rewards can be designed based on metrics such as the system's total scheduling cost and safety margin. For example, if the correction scheme reduces system costs and ensures safe system operation, a positive reward is given; otherwise, a negative reward is given. The reward function is as follows:

[0081] The calculation basis and sources of each parameter are as follows: thermal power cost Based on the actual output of thermal power in the next moment ,according to Calculation, parameters Preset cost coefficients for thermal power units; Cost of power curtailment Based on the projected renewable energy output at the next moment ( ) and actual consumption ( ),according to calculate, This refers to the absorption rate; Hydrogen production center costs Based on the hydrogen production, electricity / hydrogen purchase, and hydrogen storage actions at the next moment, according to calculate; Safety penalty compensation items : The security confidence penalty coefficient is preset for the file. (Probability of safe operation of thermal power plants) (Probability of safe operation of transmission lines) The probability of safe operation of the electrolytic cell is output by the "online safety monitoring network" (based on real-time status prediction). The higher the safety level, the greater the compensation and the higher the reward.

[0082] 2) Train the value network by extracting samples from the sample pool. ,in: For the current / next state of the intelligent agent of the thermal power unit, For the current / next state of the intelligent agent of the renewable energy power plant, This represents the current / next state of the hydrogen production center's intelligent agent. These are the individual actions of three types of intelligent agents; The total reward for joint operations.

[0083] Calculate the state value function and the objective value function:

[0084] Formula meaning: Target value = Cumulative reward of discounts over the next K time steps + Discount value of the state at step K ( This refers to a rolling time window for intraday scheduling, such as 16 15-minute time slots, corresponding to 4 hours, consistent with Item 4 of the document, "Designated Time for Online Security Monitoring"); : No. Instant rewards for time steps, ( The negative sign transforms "minimum cost" into "maximum reward"; Discount factor (implicit hyperparameter, value) The weight used to decay long-term rewards is consistent with the EH-IES intraday scheduling characteristic that "the recent state has a greater impact"; : No. The value prediction of the time step state is output by the "value network with frozen parameters" (the current parameters are fixed during training to avoid fluctuations in the target value); : respectively the first The costs of thermal power generation, waste power loss, and hydrogen production centers.

[0085] A loss function is constructed to update the parameters of the value network. The state value function represents the expected long-term cumulative reward that the system can obtain after taking a series of actions in a certain state; the target value function is the target value calculated based on the current policy and environmental feedback. The loss function is constructed and optimized to update the parameters of the value network, enabling the value network to more accurately estimate state values.

[0086] The loss function is constructed with the objective of minimizing the sum of squared errors between the predicted state value and the target value, as shown in the following formula:

[0087] In the above formula, : The number of samples extracted from the sample pool (e.g., batch size is 64); : No. Predicted state value of each sample; : No. The target value (true label) of each sample; Design rationale: MSE can effectively penalize large error samples (such as samples in EH-IES where "low security confidence leads to a sharp drop in reward"), ensuring that the value network's evaluation of critical states is more accurate.

[0088] Optimize the loss function using gradient descent. Update value network parameters The steps are as follows: Calculate the gradient: for the loss function Please provide information about the parameters. partial derivatives This reflects "the direction and magnitude of the impact of parameter changes on loss"; Dynamic learning rate adaptation: using the dynamic learning rate formula Adjust the learning rate based on the stability of the policy network update (e.g., reduce it when the policy fluctuates greatly). (Avoid value network parameter oscillations) Parameter iterative update: Update the parameters according to the following formula until the loss function converges (e.g., the MSE decreases by less than 1 / 2 epochs over 5 consecutive epochs). ):

[0089] Key constraint: The updated value network must meet the requirement of "higher value assessment of the safety status of EH-IES" (e.g., hydrogen storage within a reasonable range, and a safety confidence level of ≥95% for thermal power plants). (It should be significantly higher than the unsafe state), ensuring that the value assessment is consistent with the document's scheduling objectives of "economy + safety".

[0090] 3) Train the policy network, introduce a dominance function to evaluate the economics of the policy network output, and update the policy network parameters by combining dynamic learning rate and constraints. The dominance function represents the additional reward that can be obtained by taking a certain action relative to the average action in a certain state.

[0091] The core formula for the advantage function (adapted to EH-IES multi-timestep returns) is as follows:

[0092] Formula meaning: Advantage function = (Cumulative reward with discounts over the next K time steps + Discounted value of the state at the Kth step) - Value of the current state, i.e., "Taking an action" The difference between the "actual cumulative profit" and the "average profit in the current state"; The current system status (including power dimension: thermal power output, renewable energy grid-connected power; hydrogen energy dimension: hydrogen storage, hydrogen load; coupling dimension: hydrogen production and power purchase, renewable energy hydrogen production power) is completely consistent with the "online safety monitoring network input variables". Policy network in state Downward output actions (such as thermal power output increment) Hydrogen production capacity from renewable energy sources Hydrogen production and electricity purchase ), corresponding to "action variables of the decision-making unit"; : No. The immediate reward at each time step is calculated using a modified MAPPO objective function. The negative sign transforms "minimum cost" into "maximum reward," where: : No. Step-by-step thermal power cost ( ); : No. Step-by-step power curtailment loss cost ( ); : No. Cost of hydrogen production center ; : No. Step-by-step security confidence level penalty compensation item ( Preset coefficients for the file. The probability of safe operation of thermal power plants / transmission lines / hydrogen production centers is output by the "online safety monitoring network". Discount factor (implied hyperparameter in the file, value) ), adapted to the EH-IES intraday scheduling characteristic that "short-term economic benefits have a greater impact"; The daily rolling time window (e.g., 16 15-minute intervals, corresponding to 4 hours) is consistent with the "specified time for online security monitoring"; , : Current state and the first The value function of each step state, output by the "trained value network", reflects the average economic return of the state.

[0093] When optimizing the policy network using the advantage function, the parameters are updated by "maximizing the logarithm of the action probability weighted by the advantage function through gradient ascent". The core formula is as follows:

[0094] Formula meaning: Policy network parameters The direction of the update is determined by the expected value of the product of the "logarithm of the action probability" and the "advantage function"; like (Motion Economy), then If the value is positive, the parameter update will "increase the output probability of the action" (the policy network is more inclined to generate economic actions); like If the action is uneconomical, the product will be negative, and the parameter update will "reduce the output probability of the action" (the policy network avoids generating uneconomical actions). Combining the "dynamic learning rate", the update process is carried out through... Control the step size to ensure that the policy network is balanced between "economic optimization" and "security constraint satisfaction" (e.g., when the action is economical but the security confidence is low, reduce the advantage function through the security penalty term to avoid excessive pursuit of economy and neglect of security).

[0095] 4) Evaluate the state value based on the trained value network and policy network.

[0096] After the value network and policy network are trained, the evaluation of state value needs to focus on the "electric-hydrogen coupling state", "multi-agent collaborative decision-making" and "dual objectives of economy and safety" of EH-IES (hydrogen-electric coupled integrated energy system). The specific steps are as follows: 1. Determine the range of the "system status" to be evaluated. The status to be evaluated must cover the "seven types of input variables of the online safety monitoring network" and key information on the electro-hydrogen coupling to ensure a complete reflection of the EH-IES operating status, specifically including: Electricity Dimension Status: Real-time output of thermal power generation units ( ), power injected into the grid from wind farms / solar power plants ( ), power load ( ), probability of safe operation of thermal power units / transmission lines ( ); Hydrogen Energy Status: Real-time Hydrogen Storage Capacity at Hydrogen Production Center ( ), hydrogen load ( ), Electrolytic cell input power ( ), probability of safe operation of electrolytic cell ( ); Electricity-hydrogen coupling dimension status: Hydrogen production capacity directly supplied by renewable energy ( ), the power purchased from the grid by the hydrogen production center ( ).

[0097] All state data must be standardized according to the format used in the training phase (e.g., time resolution should be uniformly set to a specific value). (Units can be converted to MW, kg, etc.)

[0098] 2. Freeze the dual network parameters and determine the calling logic. After training, the value network (parameters) ) and policy network (parameters) With the parameters fixed, the logic of "dual network collaborative invocation" needs to be clearly defined when evaluating the value of the state: The value network directly outputs the "state-based value" (reflecting the average cumulative revenue of that state). The policy network outputs the "optimal action probability distribution" to help determine the "value gain corresponding to the optimal action in this state", and finally forms a complete state value assessment of "basic value + optimal action gain" (an extension of the MAPPO multi-agent game logic).

[0099] 3. After training, the fundamental value of the state to be evaluated is directly calculated through forward propagation of the value network. The specific process is as follows: State input and feature extraction (adapting to EH-IES coupling characteristics) The standardized system state is input into the value network, using the same network structure as the training phase (including attention layers, convolutional layers, and fully connected layers). Key processing steps: Attention layer weighting: for critical states of the electro-hydrogen coupling (e.g.) , (This is an implicit adaptation improvement in the document, distinct from the equal characteristics treatment of a single energy system) to give it higher weight, ensuring that the impact of coupling relationships on value assessment is fully captured; Feature extraction from convolutional and pooling layers: Extracting local correlation features in the state (such as "hydrogen storage capacity-hydrogen load matching degree", "thermal power output-electricity load balance degree", "renewable energy power-hydrogen production and consumption demand matching degree"), which are directly related to "system security constraints" and "economic cost objectives"; Fully connected layer mapping: Maps the extracted high-dimensional features to scalar values. This refers to the "basic value of the state," with the same unit as the reward (yuan; a negative sign indicates cumulative cost, and a positive sign indicates cumulative revenue, consistent with the reward calculation logic).

[0100] 4. After training, the "optimal action" for the current state is obtained through the policy network, and the value gain of this action compared to the average action is calculated to help improve the state value evaluation: Status to be evaluated Input the policy network and output the action probability distribution of each agent (thermal power unit, renewable energy power plant, hydrogen production center). Select the optimal joint action according to the "maximum probability" principle: Intelligent Agent for Thermal Power Units: Optimal Output Increment (like (probability 55%) Intelligent agents for renewable energy power plants: optimal grid-connected power Optimal hydrogen production power (like , (60% probability of joint action) Intelligent Agent of Hydrogen Production Center: Optimal Power Purchase Optimal hydrogen storage injection / release rate (like , (The probability of a joint event is 58%).

[0101] All optimal actions must satisfy "system safety constraints" (such as...). It meets the requirements for the ramp-up rate of thermal power plants. (Renewable energy forecast output).

[0102] The value gain of the optimal action (actual action gain / average state gain) is calculated based on the "advantage function" logic. After training, the value gain of the optimal action can be calculated using either a "pre-stored sample library" or "rapid simulation". : Method 1: Pre-stored sample pool query: Filter samples from the training sample pool that match the current state. Similar samples, extract the "optimal action" from these samples. The mean of the dominance function, as (such as in similar samples) The average advantage is -20 yuan, meaning that this action reduces costs by 20 yuan compared to the average action, and the gain is +20 yuan. Method 2: EH-IES Fast Simulation: Optimal Actions Input the EH-IES simulation model and calculate the reward at a single time step. (According to the reward formula in Article 6 of the document), and combined with the next state value output by the value network. Gain is calculated using timing differential: ( The implicit discount factor in the document, with a value of [value missing]. ).

[0103] 5. Integrate basic value and optimal action gain to output final state value. The final state value needs to be combined with the "basic value (average reward)" and the "optimal action gain (additional reward for the optimal decision)" to form a comprehensive evaluation result, as shown in the following formula:

[0104] The formula means: Final state value = basic state value + additional value gain from the optimal action, which fully reflects the "average benefit of the current state" and the "upper limit of the benefit that the optimal decision can achieve in this state".

[0105] Compared with the prior art, this embodiment has the following main advantages: 1) Effectively balance economy, security and computational burden: The multi-stage collaborative scheduling method in this embodiment formulates an initial scheduling plan through day-ahead deterministic scheduling, monitors the system's security margin in real time through online security monitoring, and flexibly corrects the scheduling plan according to the actual situation within the day. This can reduce the total scheduling cost of the system while ensuring the safe operation of the system, and avoid the huge computational burden caused by complex uncertainty modeling.

[0106] 2) Improve the capacity for renewable energy absorption: Utilizing HPC to convert renewable energy into hydrogen energy for storage and utilization reduces the curtailment of renewable energy, increases the absorption rate of renewable energy, and promotes the large-scale application of renewable energy.

[0107] 3) Enhance system flexibility and adaptability: By adopting the multi-agent Markov game method for intraday correction, the actions of each decision-making unit can be flexibly adjusted according to the real-time changes of renewable energy and the operating status of the system, making the system more adaptable and resistant to interference.

[0108] 4) Ensure the safe and stable operation of the system: The online security monitoring network can monitor the system's security margin in real time, promptly identify potential security risks, and make adjustments through the intraday calibration network to ensure that the system can operate safely and stably under various uncertainties.

[0109] The following are scheduling optimization methods in two specific application scenarios: 1: Scheduling optimization of a hydrogen-electricity coupled integrated energy system in a certain region The region's integrated hydrogen-electricity energy system comprises two 300MW thermal power units (G1 and G2), three 100MW wind farms (WT1-WT3), two 50MW solar power plants (PV1-PV2), and a 20MW hydrogen production center (HPC) equipped with 2000kg hydrogen storage tanks. The region's electricity and hydrogen loads exhibit some fluctuation, and the generation of renewable energy is uncertain. Typical daily load data is as follows:

[0110] 1) Day-ahead scheduling phase Objective function solution: Establish a day-ahead scheduling model with the goal of minimizing the total system scheduling cost. Substitute the following parameters: thermal power cost (G1: a=1000 yuan / h, b=200 yuan / MWh, c=0.05 yuan / (MW²h); G2: a=1200 yuan / h, b=180 yuan / MWh, c=0.04 yuan / (MW²h)), curtailment penalty cost (wind power 0.6 yuan / kWh, photovoltaic 0.5 yuan / kWh), and hydrogen production cost (electrolyzer levelized cost 3 yuan / kg, electricity purchase price 300 yuan / kg). (Based on a price of 550 yuan / MWh and a hydrogen purchase price of 40 yuan / kg), considering the safety operation constraints of the power system and HPC, an optimization algorithm is used to solve the model and obtain the optimal scheduling scheme for thermal power units, renewable energy power plants, and HPC. For example, during periods of low electricity prices, HPC is arranged to purchase more electricity from the grid for hydrogen production, while the power generation capacity of thermal power units is increased; during periods of high electricity prices, the power generation capacity of thermal power units is reduced to increase the consumption of renewable energy.

[0111] Dispatch results: From 00:00 to 06:00, G1 outputs 200MW, G2 outputs 50MW, WT1-WT3 operate at full capacity (total 100MW), PV outputs nothing, HPC purchases 50MW of electricity to produce 150kg / h of hydrogen (hydrogen stored in hydrogen storage tanks); From 12:00 to 18:00, G1 and G2 operate at full capacity (total 600MW), WT1-WT3 outputs 30MW, PV1-PV2 operate at full capacity (total 50MW), HPC stops purchasing electricity, and utilizes wind and solar power to produce 80kg / h of hydrogen, with an additional 50kg / h of hydrogen purchased for any shortfall.

[0112] 2) Online security monitoring phase Data Acquisition: Real-time acquisition of data such as unit power output, renewable energy power injected into the grid, renewable energy power used for hydrogen production and power purchase in HPC, power load, hydrogen load, and the latest forecast values ​​of renewable energy, with a sampling frequency of 15 minutes / time.

[0113] Safety confidence calculation: The safety confidence is output through a machine learning model (7-dimensional input layer, 8-head attention layer, 3×3 convolutional kernels, 2×2 max pooling layer, and 3 fully connected layers). When the actual output of WT1 suddenly increases to 90MW at 14:00 (40MW over the prediction), the safe operation probability of transmission line L1 drops to 82% (preset threshold 85%), triggering the correction mechanism.

[0114] 3) Intraday Correction Phase MAPPO algorithm correction: Six agents (G1, G2, WT1 - WT3, PV1 - PV2, HPC power purchase, HPC hydrogen purchase, and hydrogen storage system) make collaborative decisions, with a dynamic learning rate baseline of 0.001 and δ=0.2.

[0115] Correction results: G1 output decreased to 280MW, G2 output decreased to 270MW; WT1 increased hydrogen production capacity by 10MW (hydrogen production increased by 3.3kg / h); HPC reduced hydrogen purchases by 10kg / h; hydrogen storage tanks released 5kg / h of hydrogen. After correction, the safe operation probability of transmission line L1 increased to 92%, and the time-period cost increased by 500 yuan (original cost 85,000 yuan), but losses from safety accidents were avoided.

[0116] Effect comparison: 1) Cost Comparison (Typical Day)

[0117] 2) Efficiency Comparison

[0118] 2: Application of a hydrogen-electricity coupled integrated energy system in an industrial park The park includes three 50MW PV-HPC systems, two 80MW wind-HPC systems, and one 100MW grid electricity-HPC system, with a total hydrogen storage capacity of 5000kg. It serves 10 chemical companies (with an average daily electricity demand of 800MW and a hydrogen demand of 500kg / h). Key parameters are as follows:

[0119] 1) Day-ahead scheduling phase Production plan matching: According to the enterprise's production plan, the chemical enterprise operates at full capacity from 8:00 to 18:00 (electricity demand 900MW, hydrogen demand 600kg / h), and at low capacity during other periods (electricity demand 500MW, hydrogen demand 300kg / h).

[0120] Dispatch plan: 8:00 - 18:00, PV-HPC operates at full capacity (150MW, producing 4500kg of hydrogen), wind-HPC operates at full capacity (160MW, producing 5120kg of hydrogen), grid electricity-HPC purchases 80MW of electricity (producing 2240kg of hydrogen), and the hydrogen storage tank releases 1140kg of hydrogen; during other periods, PV-HPC and wind-HPC operate at 50% capacity, grid electricity-HPC purchases 20MW of electricity, and the hydrogen storage tank stores hydrogen.

[0121] 2) Online security monitoring phase Anomaly Detection: At 10:00, a chemical company suddenly increased its hydrogen demand by 100 kg / h. The input power of the HPC electrolyzer exceeded the rated value by 10%, and the probability of safe operation dropped to 78% (threshold 85%), triggering correction.

[0122] 3) Intraday Correction Phase Coordinated adjustments: PV - HPC output increased by 10MW (hydrogen production 300kg / h), wind - HPC output increased by 5MW (hydrogen production 160kg / h), grid electricity - HPC electricity purchase increased by 15MW (hydrogen production 420kg / h), hydrogen storage tank released 220kg of hydrogen, safe operation was restored within 15 minutes, correction cost increased by 1200 yuan, avoiding production stoppage losses (average daily production stoppage loss of 500,000 yuan).

[0123] Effect comparison: 1) Cost Comparison (Monthly Average)

[0124] 2) Stability Comparison

[0125] The present invention further provides a multi-stage coordinated scheduling system for a hydrogen-electric coupled integrated energy system, including a microprocessor and a memory interconnected thereto, wherein the microprocessor is programmed or configured to execute a multi-stage coordinated scheduling method for the hydrogen-electric coupled integrated energy system.

[0126] The present invention further provides a computer-readable storage medium storing a computer program / instruction, the computer program / instruction being programmed or configured to execute a multi-stage coordinated scheduling method for a hydrogen-electric coupled integrated energy system via a processor.

[0127] The system and medium of the present invention, corresponding to the methods described above, also have the advantages described above.

[0128] The present invention can implement all or part of the processes in the methods of the above embodiments, or it can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, it can implement the steps of the above method embodiments. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. Computer-readable media include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. The memory is used to store computer programs and / or modules. The processor implements various functions by running or executing the computer programs and / or modules stored in the memory, and by calling data stored in the memory. The memory may include high-speed random access memory, as well as non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart media cards (SMC), secure digital (SD) cards, flash cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0129] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A multi-stage coordinated scheduling method for a hydrogen-electric coupled integrated energy system, characterized in that, Includes the following steps: S101. An initial scheduling scheme is obtained by solving the day-ahead scheduling model of the hydrogen-electricity coupled integrated energy system. The day-ahead scheduling model aims to minimize the total scheduling cost of the hydrogen-electricity coupled integrated energy system. The initial scheduling scheme includes the unit output of the thermal power generation unit in the hydrogen-electricity coupled integrated energy system, the grid connection ratio of renewable energy power plants, and the amount of electricity and hydrogen purchased by the hydrogen production center. S102, monitor the safety confidence level of the hydrogen-electric coupling integrated energy system within a specified time after the initial scheduling scheme is executed. If the safety confidence level is greater than the preset threshold, the process ends and exits; otherwise, proceed to the next step. S103, an intraday correction network based on the MAPPO algorithm is used to correct the intraday scheduling model to ensure that the generated corrected scheduling scheme meets deterministic and chance constraints, and the corrected scheduling scheme is then issued to the hydrogen-electric coupling integrated energy system for execution; wherein the intraday scheduling model aims to minimize the total scheduling cost of the hydrogen-electric coupling integrated energy system at the time point of intraday scheduling.

2. The multi-stage coordinated scheduling method for a hydrogen-electric coupled integrated energy system according to claim 1, characterized in that, In step S101, when obtaining the initial scheduling scheme based on the day-ahead scheduling model of the hydrogen-electric coupled integrated energy system, the day-ahead scheduling model includes an objective function. The objective function aims to minimize the total scheduling cost of the hydrogen-electric coupled integrated energy system, and its expression is: ， ， ， ， in, For the time of day, For thermal power generation unit number The cost of electricity generation at a given point in time. For renewable energy power plants Cost of power curtailment at a given point in time For the hydrogen production center Operating costs at each point in time. This refers to the number of thermal power generation units. , and For the first Preset fuel cost parameters for each thermal power generation unit For the first The first thermal power generation unit Power output at each time point For time resolution at a given point in time, and These refer to the number of wind farms and solar power plants in renewable energy power plants, respectively. and The unit penalty costs for wind power and solar power curtailment are respectively. and The first The wind farm in the first The absorption rate and predicted power generation at each point in time. and The first The solar power station was in the first The absorption rate and predicted power generation at each point in time. The number of hydrogen production centers; The levelized cost of hydrogen production by electrolyzer. Hydrogen production center In the The amount of hydrogen produced at each point in time. To levelize the cost of hydrogen storage systems, and Hydrogen production centers In the The amount of hydrogen injected and released at each time point. It is the price of electricity. Hydrogen production center In the Electricity purchased from the grid at a specific point in time. It is the price of purchasing hydrogen. Hydrogen production center In the The amount of hydrogen purchased at each point in time.

3. The multi-stage coordinated scheduling method for the hydrogen-electric coupled integrated energy system according to claim 2, characterized in that, The day-ahead scheduling model also includes constraints, which include power system safety operation constraints and hydrogen production center safety operation constraints. The power system safety operation constraints include power balance constraints, upper limit constraints on renewable energy power injected into the grid, and safety operation constraints of generating units and transmission lines. The hydrogen production center safety operation constraints include hydrogen energy balance equations, electrolyzer operation model constraints, and hydrogen storage system operation constraints. The functional expression for the power balance constraint is: in, , , and These refer to the number of thermal power generation units, wind farms, solar power plants, and hydrogen production centers, respectively. For the first The thermal power generation unit in the first Power output at each time point and The i-th wind farm and the j-th solar power station are respectively located at the i-th wind farm and the j-th solar power station. The power injected into the grid at each point in time. For the first Electricity load at a given time point For the first The hydrogen production center is in the first Electricity purchased from the grid at a specific point in time; The functional expression for the upper limit constraint on the power injected into the grid by renewable energy is: The functional expression of the hydrogen energy balance equation is: in, , and The k-th hydrogen production center is located at the k-th hydrogen production center in ... The amount of hydrogen produced at a given point in time using on-site wind power, solar power, and purchased electricity. For the k-th hydrogen production center in the th... Hydrogen load at each time point and The k-th hydrogen production center is located at the k-th hydrogen production center in ... The amount of hydrogen released and injected at each time point, For the k-th hydrogen production center in the th... The amount of hydrogen purchased at each point in time; The functional expression for the constraints of the electrolytic cell operation model is: in, Let be the input power of the electrolyzer in the k-th hydrogen production center. This is the maximum input power of the electrolytic cell; The functional expression for the operational constraints of the hydrogen storage system is: in, and The hydrogen storage system of the k-th hydrogen production center is located in the time period. and time period Hydrogen storage capacity, and These refer to the injection efficiency and release efficiency of the hydrogen storage system, respectively. This is the maximum capacity of the hydrogen storage tank. and These represent the maximum injection and release amounts per single time period, respectively. and These represent the hydrogen storage levels at the beginning and end of the scheduling cycle, respectively. Let t be the amount of hydrogen injected at time t. Let be the amount of hydrogen released at time t. Let t be the amount of hydrogen stored at time t.

4. The multi-stage coordinated scheduling method for a hydrogen-electric coupled integrated energy system according to claim 1, characterized in that, In step S102, when monitoring the safety confidence level of the hydrogen-electric coupled integrated energy system within a specified time after the execution of the initial scheduling scheme, seven types of variables are used as inputs: the power output of the thermal power generation unit at the current moment, the power injected into the grid by the renewable energy power station, the renewable energy power and purchased power in the hydrogen production center for hydrogen production, the power load, the hydrogen load, and the latest predicted value of the renewable energy power station. A machine learning model is used to obtain the safety confidence level of the hydrogen-electric coupled integrated energy system within a specified time. The machine learning model is pre-trained to establish a mapping relationship between the seven types of input variables and the safety confidence level of the hydrogen-electric coupled integrated energy system within a specified time. The safety confidence level of the hydrogen-electric coupled integrated energy system within a specified time includes the safe operation probability of the thermal power generation unit, transmission line, and hydrogen production center.

5. The multi-stage coordinated scheduling method for the hydrogen-electric coupled integrated energy system according to claim 4, characterized in that, The machine learning model comprises an input layer, an attention layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer connected in sequence, and the loss function used by the machine learning model during the training phase is mean squared error.

6. The multi-stage coordinated scheduling method for a hydrogen-electric coupled integrated energy system according to claim 1, characterized in that, Step S103 includes: An intraday correction network based on the MAPPO algorithm is used to transform the intraday scheduling model into a multi-agent Markov game for solution. Each agent represents a decision-making unit in the hydrogen-electricity coupled integrated energy system. Solving the problems of each decision-making unit generates a corrected scheduling scheme that satisfies deterministic and chance constraints. The objective function expression of the intraday correction network based on the MAPPO algorithm is as follows: ， In the above formula, The objective function used in the MAPPO algorithm. For thermal power generation unit number The cost of electricity generation at a given point in time. For renewable energy power plants Cost of power curtailment at a given point in time For the hydrogen production center Operating costs at each point in time. The safety confidence penalty coefficient. , , The safety confidence level of the hydrogen-electric coupled integrated energy system within a specified time period is determined by the safe operation probability of the thermal power generation unit, transmission line, and hydrogen production center.

7. The multi-stage coordinated scheduling method for a hydrogen-electric coupled integrated energy system according to claim 6, characterized in that, The decision-making units involved include: the power output increment of thermal power generation units. The power injected into the grid by renewable energy power plants Renewable energy power used for hydrogen production The electricity purchased by the hydrogen production center Hydrogen purchased by the hydrogen production center Operational plan for hydrogen storage in hydrogen production center ,in and These represent the amounts of hydrogen injected and released, respectively.

8. The multi-stage coordinated scheduling method for a hydrogen-electric coupled integrated energy system according to claim 7, characterized in that, In step S103, the intraday correction network based on the MAPPO algorithm updates the network parameters of the agent using a dynamic learning rate during training, and the calculation function expression of the dynamic learning rate is: in, For dynamic learning rate, As the baseline learning rate, and These are the sampling probabilities for the latest policy and the old policy, respectively. For policy network parameter symbols, Let be the sampling probability of the i-th agent based on the folding strategy. Let be the sampling probability of the i-th agent based on the old policy. Let t be the sampled action and system state of the i-th agent at time t. The system state is the action sampled by the i-th agent at time t. Let be the system state perceived by the i-th agent at time t. For the clipping function, This is the cutting factor.

9. A multi-stage coordinated scheduling system for a hydrogen-electric coupled integrated energy system, comprising interconnected microprocessors and a memory, characterized in that, The microprocessor is programmed or configured to execute the multi-stage coordinated scheduling method for the hydrogen-electric coupled integrated energy system according to any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program / instructions, characterized in that, The computer program / instructions are programmed or configured to execute, via a processor, the multi-stage coordinated scheduling method for the hydrogen-electric coupled integrated energy system according to any one of claims 1 to 8.

Images