Chemical industrial park energy scheduling method based on heterogeneous graph reinforcement learning

Abstract

This invention relates to the field of energy management technology in chemical industrial parks, and discloses an energy scheduling method for chemical industrial parks based on heterogeneous graph reinforcement learning. The method includes: acquiring the operating status of physical equipment and external environmental data of the park system to generate the current state; constructing a node feature matrix and a heterogeneous adjacency matrix based on the fixed physical connections of the park system, and inputting them into a GNN encoder to generate a current state embedding vector, which is then input into an action decision network to generate a current composite action vector; applying the action vector to a physical simulation environment to acquire the interaction variables and carbon emissions of physical equipment within the park system, and calculating a comprehensive reward; inputting the state embedding vector and action vector into a value network, and combining the comprehensive reward, backpropagating to update the parameters of the GNN encoder, action policy network, and value network, and obtaining the optimal comprehensive reward through iterative optimization to achieve optimal energy scheduling in the park; this invention realizes intelligent scheduling and optimization of energy in chemical industrial parks.

Description

Technical Field

[0001] This invention relates to the field of energy management technology in chemical industrial parks, specifically to an energy dispatching method for chemical industrial parks based on heterogeneous graph reinforcement learning. Background Technology

[0002] High-energy-consuming chemical industrial parks, as a key and challenging area for low-carbon transformation, face core challenges that are difficult to solve with existing technologies. These challenges manifest in several ways: Taking the Shanghai Jinshan Chemical Industrial Park as an example, there is a spatiotemporal mismatch between stable industrial by-product hydrogen supply and drastically fluctuating electricity load. The "self-generation and self-consumption, grid-connected but not connected to the grid" operation mode places extremely high demands on scheduling accuracy. Furthermore, the heterogeneous system coupled with the four media of "electricity-hydrogen-heat-carbon" includes various heterogeneous devices such as chemical units, energy conversion equipment, and energy storage equipment, making it difficult for traditional centralized or decentralized control to achieve coordinated optimization. Existing model-driven methods are limited by system nonlinearity and prediction uncertainty, and conventional reinforcement learning cannot adapt to heterogeneous systems, making it difficult to meet actual scheduling needs. Therefore, existing energy scheduling methods for chemical industrial parks suffer from model fragmentation, inability to handle complex constraints, and inability to achieve cross-domain coordination, thus preventing energy scheduling from reaching global optimum. Summary of the Invention

[0003] To address the aforementioned shortcomings in existing technologies, this invention provides an energy scheduling method for chemical industrial parks based on heterogeneous graph reinforcement learning. This method solves the problems of existing energy scheduling methods for chemical industrial parks, such as model fragmentation, inability to handle complex constraints, and inability to achieve cross-domain collaboration, which prevent energy scheduling from reaching the global optimum.

[0004] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: An energy dispatching method for chemical industrial parks based on heterogeneous graph reinforcement learning includes the following steps: S1. Obtain operational status data and external environment data to generate the current status of the chemical industrial park system. ; S2. Construct the node feature matrix and heterogeneous adjacency matrix of the physical device, and input them into the GNN encoder to generate the current state embedding vector. ; S3. Embed the current state into the vector. Input the action policy network to generate the current composite action vector. ; S4. Based on the current composite action vector Obtain the interaction variables of physical devices and carbon emissions to generate the next-time state of the chemical industrial park system. ; S5. Calculate the current composite vector. Comprehensive rewards ; S6. Embed the current state into the vector. With the current composite action vector Input value network, combined with comprehensive rewards To maximize the cumulative comprehensive rewards and achieve optimal energy scheduling for the chemical industrial park system.

[0005] Furthermore, the physical equipment includes chemical process units, hydrogen compressors, hydrogen storage tanks, fuel cells, gas boilers, air separation systems, and thermal storage units.

[0006] Furthermore, the operational status data includes the hydrogen storage quality in the hydrogen storage tank and the health status of the fuel cell.

[0007] Furthermore, external environmental data includes electricity prices, carbon intensity, and ambient temperature.

[0008] Furthermore, step S1 specifically includes: First, real-time operating data of physical equipment within the chemical industrial park is collected and used as node feature vectors. Simultaneously, external environmental data is collected and used as a global state vector or extended feature vector, which is then concatenated into the node features of each physical device to construct a feature matrix containing information about the entire system, serving as the current state. .

[0009] Furthermore, step S2 specifically includes: S21. Change the current status of the chemical industrial park system. As a node characteristic, physical devices are constructed 3D node feature matrix ,in, This represents the total number of nodes, i.e., the total number of physical devices. This represents the feature dimension of each node, which is the sum of the dimensions of the current state data of each physical device and the external environment data. S22. Obtain the fixed physical connection relationships of the chemical industrial park system and construct a heterogeneous adjacency matrix. ; S23, Transform the node feature matrix Heterogeneous adjacency matrix Inputting a GNN encoder, the system generates a current-moment state embedding vector that integrates spatiotemporal information of the entire chemical industrial park system through message passing and clustering. .

[0010] Furthermore, heterogeneous adjacency matrix The calculation formula is:

[0011] in, This represents the adjacency matrix of a power network, used to describe the wire or bus connection relationships between physical devices. This represents the hydrogen network adjacency matrix, used to describe the hydrogen pipeline connection relationships between physical devices. This represents the thermal network connection matrix, used to describe the connection relationships of heating pipe networks between physical devices.

[0012] Furthermore, the current composite action vector Control commands for different physical devices, namely:

[0013] in, This indicates a fine-tuning command for a chemical process unit. Indicates fuel cell scheduling instructions, This indicates a gas-fired boiler dispatching instruction.

[0014] Furthermore, based on the current composite action vector Obtaining the interaction variables of physical devices and carbon emissions specifically includes: The current composite action vector When applying a physical simulation environment to a chemical industrial park system, if the chemical process unit within the system is a chlor-alkali electrolyzer, the formulas for calculating the interaction variables and carbon emissions between the chlor-alkali electrolyzer and other physical equipment, including hydrogen compressors, hydrogen storage tanks, fuel cells, gas-fired boilers, air separation systems, and thermal storage units, are as follows:

[0015]

[0016]

[0017]

[0018]

[0019]

[0020]

[0021]

[0022]

[0023]

[0024]

[0025]

[0026]

[0027]

[0028]

[0029] in, Indicates the current time Hydrogen production rate of chlor-alkali electrolyzer, This indicates the molar mass of hydrogen. Indicates the number of chlor-alkali electrolytic cells. Indicates Faraday efficiency. Indicates the current time The total current flowing through the chlor-alkali electrolyzer, Represents electron transfer measurement. Denotes Faraday's constant. Indicates the current time Total power consumption of the chlor-alkali electrolytic cell Indicates the voltage of a single cell. Indicates the current time current density, Indicates the current time temperature, Indicates the current time The heat power required for the chlor-alkali electrolytic cell This represents the function of current and ambient temperature. Indicates the current time Ambient temperature, Indicates the current time Direct carbon dioxide emission rate of chlor-alkali electrolyzer. Carbon emission factor representing production load, Indicates the current time The electrical power consumed by the hydrogen compressor, A physical model representing the power consumption of a hydrogen compressor. Indicates the current time The pressure inside the hydrogen storage tank, Indicates the next moment The hydrogen storage capacity of the hydrogen storage tank Indicates the current time The hydrogen storage capacity of the hydrogen storage tank Indicates the number of fuel cells (FC). Indicates the current time No. The hydrogen flow rate consumed by Taiwan's fuel cell FC Indicates the time step. Indicates the current time No. The power generation capacity of Taiwan's fuel cell FC This indicates the actual efficiency of the fuel cell (FC). Indicates the current time No. The health status of Taiwan's fuel cell FC This indicates the lower calorific value of hydrogen. Indicates the current time No. Taiwan's fuel cell FC can recover waste heat power. This indicates the proportion of non-recoverable heat loss in the chemical industrial park system. Indicates the next moment No. The health status of Taiwan's fuel cell FC , , These represent the steady-state, dynamic, and cyclic decay coefficients, respectively. Indicates the previous moment No. The power generation capacity of Taiwan's fuel cell FC This represents the total number of fuel cells. Indicates the current time No. Start-stop status of Taiwan's fuel cell FC Indicates the current time No. Indication functions for start-stop events of Taiwan's fuel cell FC. Indicates the current time The heat output of a gas-fired boiler Indicates the current time Input the natural gas power of the gas boiler. This indicates the thermal efficiency of a gas-fired boiler. Indicates the current time The direct carbon emission rate of gas-fired boilers, The carbon emission factor of natural gas. Indicates the current time The indirect carbon emission rate of the power grid, Indicates the current time Real-time carbon intensity factor of the power grid Indicates the current time Start-up and shutdown status of the air separation system ASU. Indicates the current time The power purchase capacity of the power grid This indicates the rated power of the air separation system ASU. Indicates the current time The normal electrical load, , They represent the current time. The heat release and charging power of the thermal storage unit Indicates the current time The normal heat load, Indicates the current time Total carbon emission rate of the chemical industrial park system.

[0030] Further, calculate the current composite vector. Comprehensive rewards The formula is:

[0031]

[0032]

[0033]

[0034]

[0035] in, , These represent the weighting coefficients for the economy and the environment, respectively. Indicates the current time Economic benefits items Indicates the current time Environmental cost item Indicates the current time The profits from chemical products, Indicates the current time The cost of purchasing electricity, Indicates the current time The cost of purchasing gas, Indicates the current time The equipment degradation cost of fuel cell arrays Indicates the current time The cost of energy waste Indicates the first The initial investment cost of Taiwan's fuel cell FC, The weighting coefficient representing the penalty for waste. This indicates taking the maximum value. Indicates the current time Total power generation Indicates the current time Total electrical load Indicates the current time The carbon price.

[0036] The present invention has the following beneficial effects: The proposed energy dispatching method for chemical industrial parks based on heterogeneous graph reinforcement learning, through heterogeneous graph characterization of the correlation between stable industrial by-product hydrogen and fluctuating power load in chemical industrial parks, combined with a dynamic dispatching strategy based on reinforcement learning, not only achieves precise matching between stable energy supply and fluctuating energy demand, but also achieves global optimization of energy dispatching. Simultaneously, based on the global state perception of the GNN encoder and the millisecond-level response of the action policy network, a dynamic balance between power generation and consumption is achieved, avoiding energy waste in the grid-connected but not grid-connected mode and reducing economic losses. Furthermore, the physical connections of multi-media devices involving electricity, hydrogen, heat, and carbon are integrated through a heterogeneous adjacency matrix. By combining the message passing mechanism of GNN, the collaborative bottleneck of traditional distributed control is broken through, thereby improving the overall operating efficiency of the system. In addition, the constructed comprehensive reward integrates indicators such as equipment interaction economy and carbon emission intensity. Through the collaborative iteration of action policy network and value network, a multi-objective balance of minimum energy consumption, minimum carbon emission, and optimal economy is achieved simultaneously, effectively increasing the carbon emission reduction of chemical industrial parks. Finally, this method does not rely on precise system models or fixed rules. Through real-time interaction between reinforcement learning and simulation environment, it can adapt to dynamic scenarios such as load fluctuations and equipment parameter changes, significantly improving the anti-interference capability of chemical industrial park systems. Attached Figure Description

[0037] Figure 1 This is a flowchart illustrating the energy scheduling method for chemical industrial parks based on heterogeneous graph reinforcement learning proposed in this invention. Figure 2 This is a schematic diagram of the energy dispatching method for chemical industrial parks based on heterogeneous graph reinforcement learning in the embodiment. Detailed Implementation

[0038] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0039] like Figures 1-2 As shown, the energy scheduling method for chemical industrial parks based on heterogeneous graph reinforcement learning includes the following steps S1-S6: S1. Obtain the operating status data of all physical equipment within the chemical industrial park system, as well as external environmental data, and generate the current status of the chemical industrial park system. .

[0040] Specifically, the physical equipment includes chemical process units, hydrogen compressors, hydrogen storage tanks, fuel cells, gas boilers, air separation systems, and thermal storage units.

[0041] Specifically, external environmental data includes electricity prices, carbon intensity, and ambient temperature.

[0042] In this embodiment, real-time operating status data of all physical equipment within the chemical industrial park system is collected, such as the current time. Hydrogen storage capacity of hydrogen storage tank Current moment Total power consumption of chlor-alkali electrolyzer Current moment No. Health status of Taiwan fuel cell FC Simultaneously collect external environmental data, such as electricity price, carbon intensity, and ambient temperature, to generate the current state of the chemical industrial park system. Specifically: First, real-time operating data (such as power, pressure, and temperature) of physical equipment in the chemical industrial park, including chlor-alkali electrolyzers, fuel cells, and hydrogen storage tanks, are collected and used as node feature vectors. Simultaneously, external environmental data (such as electricity price, carbon intensity, and ambient temperature) are collected and used as global state vectors or extended features, which are then concatenated into the node features of each physical device. This constructs a feature matrix containing information about the entire system, serving as the current state. Therefore, by mapping the physical state of heterogeneous equipment with external environmental data in a unified dimension, the problem of aligning multi-source heterogeneous data is solved, providing a standardized data foundation for the subsequent comprehensive perception of the spatiotemporal coupling characteristics of the chemical industrial park by graph neural networks.

[0043] S2, Current state based on the chemical industrial park system By combining the fixed physical connections of the chemical industrial park system, a node feature matrix and a heterogeneous adjacency matrix of physical equipment are constructed. The node feature matrix and the heterogeneous adjacency matrix are then input into the GNN encoder to generate the current state embedding vector. .

[0044] In this embodiment, the message passing mechanism of heterogeneous graph neural networks is used to effectively extract the topological coupling features between different energy networks such as electricity, hydrogen, and heat, thus solving the problem that traditional methods have difficulty in handling strong coupling relationships between multiple energy flows.

[0045] Specifically, step S2 includes S21-S23: S21. Change the current status of the chemical industrial park system. As a node characteristic, physical devices are constructed 3D node feature matrix ,in, This represents the total number of nodes, i.e., the total number of physical devices. This represents the feature dimension of each node, which is the sum of the dimensions of the current state data of each physical device and the external environment data.

[0046] In this embodiment, each physical device within the chemical industrial park system is abstracted as a node. Since each physical device carries a series of data describing its current state, such as pressure, power, and temperature, arranging this data in a column yields a feature vector for that physical device, i.e., the node. Therefore, based on this, a feature vector can be constructed. 3D node feature matrix The node feature matrix describes all physical devices and the feature dimensions of each physical device.

[0047] S22. Obtain the fixed physical connection relationships of the chemical industrial park system and construct a heterogeneous adjacency matrix. ,Right now:

[0048] in, This represents the adjacency matrix of a power network, used to describe the wire or bus connection relationships between physical devices. This represents the hydrogen network adjacency matrix, used to describe the hydrogen pipeline connection relationships between physical devices. This represents the thermal network connection matrix, used to describe the connection relationships of heating pipe networks between physical devices.

[0049] In this embodiment, the power network adjacency matrix describes which physical devices within the chemical industrial park system are connected by wires or buses. For example, if a fuel cell is connected to a power bus by a wire, then the corresponding position in the power network adjacency matrix is ​​marked with 1, indicating the flow path of electrons or electrical energy. The hydrogen network adjacency matrix describes which physical devices within the chemical industrial park system are connected by hydrogen pipelines. For example, if there is a pipeline connection between a chemical process unit and a hydrogen storage tank, then the corresponding position in the hydrogen network adjacency matrix is ​​marked with 1, indicating the path of hydrogen physical flow. The thermal network connection matrix describes which physical devices within the chemical industrial park system are connected by a heating network. For example, if the waste heat generated by the fuel cell can be transported to the chemical process unit for heating, then the corresponding position in the thermal network connection matrix is ​​marked with 1, indicating the path of heat energy flow.

[0050] S23, Transform the node feature matrix Heterogeneous adjacency matrix Inputting a GNN encoder, the system generates a current-moment state embedding vector that integrates spatiotemporal information of the entire chemical industrial park system through message passing and clustering. .

[0051] In this embodiment, the GNN encoder exchanges and aggregates information between nodes in the graph through a message passing mechanism, ultimately generating the current state embedding vector of the chemical industrial park system that integrates the spatiotemporal information of the entire graph. For example, hydrogen storage tank nodes not only know their own pressure, but can also sense the hydrogen production rate of the chemical process units connected to them and the hydrogen consumption rate of the fuel cells.

[0052] S3. Embed the current state into the vector input action policy network to generate the current composite action vector. .

[0053] Specifically, the current composite action vector Control commands for different physical devices, namely:

[0054] in, This indicates a fine-tuning command for a chemical process unit. Indicates fuel cell scheduling instructions, This indicates a gas-fired boiler dispatching instruction.

[0055] In this embodiment, the current state is embedded into the vector. As input to the action policy network, the action policy network outputs the current composite action vector. Furthermore, the current composite action vector consists of control commands for different physical devices, such as fine-tuning commands for chemical process units and scheduling commands for fuel cells. The action policy network is a multilayer perceptron (MLP) that embeds the current state into the vector. After inputting into the multilayer perceptron, multiple independent output heads are designed, each responsible for generating the current composite action vector. This is a specific type of instruction used to adapt to the control needs of different physical devices. Ultimately, the specific values ​​output by all the heads are concatenated together to form the final current composite action vector. ,for: To make the current composite action vector This serves as input to subsequent modeling formulas, thereby driving the evolution of the physical simulation environment of the chemical industrial park system and generating the next-time state of the chemical industrial park system. .

[0056] S4. Change the current composite action vector A physical simulation environment is applied to the chemical industrial park system to obtain the interaction variables and carbon emissions of all physical devices within the system, and to generate the next-time state of the chemical industrial park system. .

[0057] In this embodiment, the construction of a high-fidelity physical simulation environment can simulate the nonlinear dynamic changes of the actual chemical industrial park system, enabling the intelligent agent to accumulate experience in a low-cost trial-and-error manner through interaction with the environment, thus avoiding the safety risks associated with training directly in the real system.

[0058] Interaction variables describe the flow of matter or energy between physical devices (such as hydrogen flow rate, exchange power, heat flow, etc.); while operational status data focuses on describing the internal state attributes of the physical devices (such as pressure in the hydrogen storage tank, health status of the fuel cell, ambient temperature), serving as the basis for decision-making; therefore, when the intelligent agent issues an action command... Then, the system first calculates the "interaction variables" (flow rates) between physical devices based on physical equations. These flow rates then lead to changes in the "operating status data" inside the devices (e.g., airflow causes pressure to rise), thereby generating the next-moment state of the chemical industrial park system. Therefore, in order to obtain the interaction variables and carbon emissions of all physical devices within the chemical industrial park system, this invention models the energy transfer between all physical devices and constructs a physical simulation environment, including electrochemical equations, hydrogen compression equations, energy storage dynamic equations, fuel cell equations, and carbon emission equations. Figure 2 The electrochemical reaction equation is the same as the equation for the subsequent chlor-alkali electrolysis cell. Figure 2 The hydrogen compression equation is the same as the equation for the hydrogen compressor in the subsequent steps. Figure 2 The energy storage dynamic equation is the corresponding equation for the hydrogen storage tank in the subsequent steps. Figure 2 The fuel cell FC efficiency equation is the corresponding equation for the subsequent fuel cell FC steps. Figure 2 The carbon emission equation is the corresponding equation for the carbon emissions of subsequent chemical processes. For different chemical process units within a chemical industrial park system, only the modeling equations for those chemical process units differ. This embodiment uses a chlor-alkali electrolyzer as an example to construct the model, thereby illustrating the modeling principle of the present invention, but this does not constitute a limitation of the present invention. Those skilled in the art should understand that the core idea of ​​the present invention is to model chemical process units as producer-consumer entities, which is also applicable to other chemical process units or other industries.

[0059] Therefore, the current composite action vector When applying a physical simulation environment to a chemical industrial park system, if the chemical process unit within the system is a chlor-alkali electrolyzer, the calculation formulas for the interaction variables and carbon emissions between the chlor-alkali electrolyzer and other physical equipment, including hydrogen compressors, hydrogen storage tanks, fuel cells, gas boilers, air separation systems, and thermal storage units, are as follows: 1. Modeling of chemical process units (taking chlor-alkali electrolyzer as an example): 1) Hydrogen production rate of chlor-alkali electrolyzer:

[0060] in, Indicates the current time Hydrogen production rate of chlor-alkali electrolyzer, in kg / s. This indicates the molar mass of hydrogen gas, expressed in kg / mol. Indicates the number of chlor-alkali electrolytic cells. The Faraday efficiency is expressed as the current. The function, Indicates the current time The total current flowing through the chlor-alkali electrolyzer, expressed in amperes (A). Represents the electron transfer stoichiometry, with units of mol e. - / mol, and the value in this invention is [value missing]. , This represents the Faraday constant, with units of C / mol.

[0061] 2) Power consumption in chemical industry:

[0062] in, Indicates the current time The total power consumption of the chlor-alkali electrolyzer, in watts (W). This represents the voltage of a single cell, in volts (V), and the current density. With temperature The function, Indicates the current time Current density, in A / cm² 2 , Indicates the current time Temperature, in °C.

[0063] 3) Direct carbon emissions from chemical process units:

[0064] in, Indicates the current time The direct carbon dioxide emission rate of a chlor-alkali electrolyzer, expressed in kgCO2 / s. The carbon emission factor representing production load is expressed in kg CO2 / (A·s).

[0065] 4) Heat load of chemical process unit:

[0066] in, Indicates the current time The thermal power required for a chlor-alkali electrolytic cell is expressed in W. This represents the relationship between current and ambient temperature; that is, the heat power required by the chlor-alkali electrolytic cell is a function of current and ambient temperature. Indicates the current time The ambient temperature, expressed in Kelvin (K).

[0067] 2. Modeling of a hydrogen compressor Power consumption:

[0068] in, Indicates the current time The electrical power consumed by the hydrogen compressor, measured in watts (W). Physical models representing the power consumption of a hydrogen compressor, such as polytropic compression models. Indicates the current time The pressure inside the hydrogen storage tank, measured in Pa.

[0069] 3. Modeling of hydrogen storage tanks Mass balance:

[0070] in, Indicates the next moment The hydrogen storage capacity of the hydrogen storage tank is expressed in kg. Indicates the current time The hydrogen storage capacity of the hydrogen storage tank is expressed in kg. Indicates the number of fuel cells, Indicates the current time No. The hydrogen flow rate consumed by the fuel cell FC in Taiwan is expressed in kg / s. Indicates the time step, in seconds (s).

[0071] 4. Fuel Cell Array Modeling 1) Hydrogen consumption rate:

[0072] in, Indicates the current time No. The power generation capacity of a fuel cell FC in Taiwan is measured in watts (W). The actual efficiency of a fuel cell (FC) is a function of power and state of health. Indicates the current time No. The health status of Taiwan's fuel cell FC This indicates the lower heating value of hydrogen, expressed in J / kg.

[0073] 2) Waste heat production power:

[0074] in, Indicates the current time No. The recoverable waste heat power of a fuel cell (FC), measured in watts (W). This indicates the proportion of non-recoverable heat loss in the chemical industrial park system.

[0075] 3) Decline in State of Health (SOH)

[0076] in, Indicates the next moment No. The health status of Taiwan's fuel cell FC , , These represent the steady-state, dynamic, and cyclic decay coefficients, respectively, and are empirical parameters. Indicates the previous moment No. The power generation capacity of a fuel cell FC in Taiwan is measured in watts (W). Indicates the current time No. Start-stop status of Taiwan's fuel cell FC Indicates the current time No. The indicator function for the start-stop event of the fuel cell FC is set to 1 when a start-up occurs and 0 otherwise.

[0077] 5. Modeling of auxiliary heat source (gas boiler) and external power grid 1) Heat generation from gas-fired boilers:

[0078] in, Indicates the current time The heat output power of a gas-fired boiler is measured in W. Indicates the current time Input the natural gas power of the gas-fired boiler, in watts (W). This indicates the thermal efficiency of a gas-fired boiler.

[0079] 2) Carbon removal from gas-fired boilers:

[0080] in, Indicates the current time The direct carbon emission rate of a gas-fired boiler, expressed in kg CO2 / s. This indicates the carbon emission factor of natural gas, expressed in kg CO2 / J.

[0081] 3) Indirect carbon emissions from the power grid:

[0082] in, Indicates the current time The indirect carbon emission rate of the power grid, expressed in kg CO2 / s. Indicates the current time Real-time carbon intensity factor of the power grid, in kg CO2 / J. Indicates the current time The power purchased by the power grid is measured in watts (W).

[0083] 6. System power balance modeling

[0084] in, This represents the total number of fuel cells. Indicates the current time The start / stop status of the air separation system ASU is 0 or 1. This indicates the rated power of the air separation system ASU, in watts (W). Indicates the current time The standard electrical load is expressed in watts (W).

[0085] The purpose of performing power balance modeling is to: for the current moment The total power generation and total power consumption of the internal chemical industrial park system are calculated to determine the power that the chemical industrial park system must purchase from the power grid. (Used to calculate electricity purchase costs) or power wasted due to non-grid restrictions. (Used to calculate penalty costs).

[0086] 7. System thermodynamic balance modeling

[0087] in, , They represent the current time. The heat release and charge power of the thermal storage unit, measured in W. Indicates the current time The conventional heat load is expressed in W; and this conventional heat load is different from the aforementioned chemical process and electrolytic cell heat loads, mainly including the baseline heat load required to maintain the daily operation of the chemical industrial park system (such as building heating, domestic hot water, etc.).

[0088] The purpose of performing thermal equilibrium modeling is to: for the current moment The total heat production and total heat consumption of the internal chemical industrial park system are calculated to determine the net charging / discharging power of the thermal storage unit. , The result is used to update the state of the thermal storage unit at the next moment.

[0089] 8. Total Carbon Emission Modeling

[0090] in, Indicates the current time The total carbon emission rate of the chemical industrial park system, expressed in kg CO2 / s.

[0091] The purpose of total carbon emissions modeling is to aggregate all direct carbon emissions (such as chemical process units and gas-fired boilers) and indirect carbon emissions (such as electricity purchases) to obtain the total carbon emission rate of the chemical industrial park. The result will be used to calculate the environmental cost item. .

[0092] S5. Calculate the current composite vector based on the interaction variables of all physical equipment within the chemical industrial park system and carbon emissions. Comprehensive rewards .

[0093] In this embodiment, a comprehensive reward function that incorporates both economic and environmental considerations is constructed. This transforms the multi-objective optimization problem into a reward maximization problem using reinforcement learning, guiding the agent to automatically find the optimal balance point (Nash equilibrium) between economic costs and carbon emissions in a complex scheduling space, thereby achieving optimal energy scheduling in the chemical industrial park. Therefore, the process of constructing the comprehensive reward is as follows: 1. Integrated Reward (Dual Objective) Modeling

[0094] in, Indicates the current time The total reward is in yuan. , These represent the weighting coefficients for the economy and the environment, respectively. Indicates the current time The economic benefits are expressed in yuan. Indicates the current time The environmental cost item is in yuan.

[0095] 2. Modeling of Downgrade Benefits

[0096]

[0097]

[0098] in, Indicates the current time Revenue from chemical products, in yuan. Indicates the current time The cost of electricity purchase, in yuan. Indicates the current time The cost of purchasing gas, in yuan. Indicates the current time The equipment degradation cost of a fuel cell array, in yuan. Indicates the current time Energy waste penalty cost, in yuan. Indicates the first The initial investment cost of a fuel cell FC in Taiwan, in yuan. This represents the weighting coefficient for penalties for waste, expressed in yuan / J. This indicates taking the maximum value. Indicates the current time Total power generation, in watts (W). Indicates the current time Total electrical load, in watts (W).

[0099] 3. Environmental Costs

[0100] in, Indicates the current time The carbon price is expressed in yuan / kg CO2.

[0101] S6. Embed the current state into the vector. With the current composite action vector Input the value network to generate a long-term value assessment result at the current moment, and combine it with comprehensive rewards. By updating the parameters of the GNN encoder, action policy network, and value network through backpropagation, the cumulative comprehensive reward is maximized, thereby achieving optimal energy scheduling for the chemical industrial park system.

[0102] In this embodiment, a comprehensive reward is obtained. Then, the value network is used to evaluate the current state embedding vector. With the current composite action vector long-term value And construct experience tuples ( , , , ),in, The next-time state embedding vector is used to embed the next-time state of the chemical industrial park system. The input is calculated by a GNN encoder; then, based on long-term value, algorithms such as Temporal Difference (TD-Error) are used. Comprehensive rewards Backpropagation updates the network parameters of the GNN encoder, action policy network, and value network, enabling the agent to obtain higher cumulative rewards in future decisions, thereby achieving optimal energy scheduling in the chemical industrial park system.

[0103] In summary, the energy scheduling method for chemical industrial parks based on heterogeneous graph reinforcement learning proposed in this invention achieves global optimization of energy scheduling. By breaking down the decision-making barriers between production and energy, it achieves the lowest global cost that traditional decoupling optimization cannot reach. Simultaneously, it taps into the potential flexibility of the chemical industrial park system, transforming chemical process units and large adjustable loads (air separation systems) into virtual batteries that can be scheduled, greatly enhancing the system's ability to cope with electricity price fluctuations and grid disconnection constraints. Specifically, through the action policy network, it flexibly adjusts the current density of chlor-alkali electrolyzers and the start-up and shutdown of the air separation system, increasing energy consumption during periods of low electricity prices or renewable energy surplus (equivalent to battery charging) and reducing energy consumption during periods of high electricity prices or insufficient supply (equivalent to battery discharging). This exhibits "peak shaving and valley filling" characteristics similar to energy storage batteries through source-load interaction. Furthermore, it considers long-term interests by using SOH and carbon emission models, ensuring that the agent's decisions take into account both the long-term health of physical equipment and the park's environmental goals, rather than just short-term operating costs.

[0104] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

[0105] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. An energy dispatching method for chemical industrial parks based on heterogeneous graph reinforcement learning, characterized in that, Includes the following steps: S1. Obtain operational status data and external environment data to generate the current status of the chemical industrial park system. ; S2. Construct the node feature matrix and heterogeneous adjacency matrix of the physical device, and input them into the GNN encoder to generate the current state embedding vector. ; S3. Embed the current state into the vector. Input the action policy network to generate the current composite action vector. ; S4. Based on the current composite action vector Obtain the interaction variables of physical devices and carbon emissions to generate the next-time state of the chemical industrial park system. ; S5. Calculate the current composite vector. Comprehensive rewards ; S6. Embed the current state into the vector. With the current composite action vector Input value network, combined with comprehensive rewards To maximize the cumulative comprehensive rewards and achieve optimal energy scheduling for the chemical industrial park system.

2. The energy dispatching method for chemical industrial parks based on heterogeneous graph reinforcement learning according to claim 1, characterized in that, The physical equipment includes chemical process units, hydrogen compressors, hydrogen storage tanks, fuel cells, gas boilers, air separation systems, and thermal storage units.

3. The energy dispatching method for chemical industrial parks based on heterogeneous graph reinforcement learning according to claim 1, characterized in that, Operational status data includes the hydrogen storage quality in the hydrogen storage tank and the health status of the fuel cell.

4. The energy dispatching method for chemical industrial parks based on heterogeneous graph reinforcement learning according to claim 1, characterized in that, External environmental data include electricity price, carbon intensity, and ambient temperature.

5. The energy dispatching method for chemical industrial parks based on heterogeneous graph reinforcement learning according to claim 1, characterized in that, Step S1 specifically includes: First, real-time operational data of physical equipment within the chemical industrial park is collected and used as node feature vectors. Simultaneously, external environmental data is collected and used as a global state vector or extended feature vector, which is then concatenated into the node features of each physical device to construct a feature matrix containing information about the entire system, serving as the current state. .

6. The energy dispatching method for chemical industrial parks based on heterogeneous graph reinforcement learning according to claim 1, characterized in that, Step S2 specifically includes: S21. Change the current status of the chemical industrial park system. As a node characteristic, physical devices are constructed. 1D node feature matrix ,in, This represents the total number of nodes, i.e., the total number of physical devices. This represents the feature dimension of each node, which is the sum of the dimensions of the current state data of each physical device and the external environment data. S22. Obtain the fixed physical connection relationships of the chemical industrial park system and construct a heterogeneous adjacency matrix. ; S23, Transform the node feature matrix Heterogeneous adjacency matrix Inputting a GNN encoder, the system generates a current-moment state embedding vector that integrates spatiotemporal information of the entire chemical industrial park system through message passing and clustering. .

7. The energy dispatching method for chemical industrial parks based on heterogeneous graph reinforcement learning according to claim 6, characterized in that, Heterogeneous adjacency matrix The calculation formula is: in, This represents the adjacency matrix of a power network, used to describe the wire or bus connection relationships between physical devices. This represents the hydrogen network adjacency matrix, used to describe the hydrogen pipeline connection relationships between physical devices. This represents the thermal network connection matrix, used to describe the connection relationships of heating pipe networks between physical devices.

8. The energy dispatching method for chemical industrial parks based on heterogeneous graph reinforcement learning according to claim 1, characterized in that, Current composite action vector These are control commands for different physical devices, namely: in, This indicates a fine-tuning command for a chemical process unit. Indicates fuel cell scheduling instructions, This indicates a gas-fired boiler dispatching instruction.

9. The energy dispatching method for chemical industrial parks based on heterogeneous graph reinforcement learning according to claim 1, characterized in that, Based on the current composite action vector Obtaining the interaction variables of physical devices and carbon emissions specifically includes: The current composite action vector When applying a physical simulation environment to a chemical industrial park system, if the chemical process unit within the system is a chlor-alkali electrolyzer, the formulas for calculating the interaction variables and carbon emissions between the chlor-alkali electrolyzer and other physical equipment, including hydrogen compressors, hydrogen storage tanks, fuel cells, gas-fired boilers, air separation systems, and thermal storage units, are as follows: in, Indicates the current time Hydrogen production rate of chlor-alkali electrolyzer, This indicates the molar mass of hydrogen. Indicates the number of chlor-alkali electrolytic cells. Indicates Faraday efficiency. Indicates the current time The total current flowing through the chlor-alkali electrolyzer, Represents electron transfer measurement. Denotes Faraday's constant. Indicates the current time Total power consumption of the chlor-alkali electrolytic cell Indicates the voltage of a single cell. Indicates the current time current density, Indicates the current time temperature, Indicates the current time The heat power required for the chlor-alkali electrolytic cell This represents the function of current and ambient temperature. Indicates the current time Ambient temperature, Indicates the current time Direct carbon dioxide emission rate of chlor-alkali electrolyzer. Carbon emission factor representing production load, Indicates the current time The electrical power consumed by the hydrogen compressor, A physical model representing the power consumption of a hydrogen compressor. Indicates the current time The pressure inside the hydrogen storage tank, Indicates the next moment The hydrogen storage capacity of the hydrogen storage tank Indicates the current time The hydrogen storage capacity of the hydrogen storage tank Indicates the number of fuel cells (FC). Indicates the current time No. The hydrogen flow rate consumed by Taiwan's fuel cell FC Indicates the time step. Indicates the current time No. The power generation capacity of Taiwan's fuel cell FC This indicates the actual efficiency of the fuel cell (FC). Indicates the current time No. The health status of Taiwan's fuel cell FC This indicates the lower calorific value of hydrogen. Indicates the current time No. Taiwan's fuel cell FC can recover waste heat power. This indicates the proportion of non-recoverable heat loss in the chemical industrial park system. Indicates the next moment No. The health status of Taiwan's fuel cell FC , , These represent the steady-state, dynamic, and cyclic decay coefficients, respectively. Indicates the previous moment No. The power generation capacity of Taiwan's fuel cell FC This represents the total number of fuel cells. Indicates the current time No. Start-stop status of Taiwan's fuel cell FC Indicates the current time No. Indication functions for start-stop events of Taiwan's fuel cell FC. Indicates the current time The heat output of a gas-fired boiler Indicates the current time Input the natural gas power of the gas boiler. This indicates the thermal efficiency of a gas-fired boiler. Indicates the current time The direct carbon emission rate of gas-fired boilers, The carbon emission factor of natural gas. Indicates the current time The indirect carbon emission rate of the power grid, Indicates the current time Real-time carbon intensity factor of the power grid Indicates the current time Start-up and shutdown status of the air separation system ASU. Indicates the current time The power purchase capacity of the power grid This indicates the rated power of the air separation system ASU. Indicates the current time The normal electrical load, , They represent the current time. The heat release and charging power of the thermal storage unit Indicates the current time The normal heat load, Indicates the current time Total carbon emission rate of the chemical industrial park system.

10. The energy dispatching method for chemical industrial parks based on heterogeneous graph reinforcement learning according to claim 9, characterized in that, Calculate the current composite vector Comprehensive rewards The formula is: in, , These represent the weighting coefficients for the economy and the environment, respectively. Indicates the current time Economic benefits items Indicates the current time Environmental cost item Indicates the current time The profits from chemical products, Indicates the current time The cost of purchasing electricity, Indicates the current time The cost of purchasing gas, Indicates the current time The equipment degradation cost of fuel cell arrays Indicates the current time The cost of energy waste Indicates the first The initial investment cost of Taiwan's fuel cell FC, The weighting coefficient representing the penalty for waste. This indicates taking the maximum value. Indicates the current time Total power generation Indicates the current time Total electrical load Indicates the current time The carbon price.

Images