A Power System Dispatch Optimization Method Based on Graph Neural Networks

Abstract

This invention relates to the field of power system regulation technology, specifically to a power system dispatch optimization method based on graph neural networks. The method involves initializing the operating parameters and load power of wind and photovoltaic power plants, hydrogen generators, energy storage, and thermal power plants in the initial power system; constructing a day-ahead dispatch optimization model based on the stepped ramping of thermal power plants; and inputting the initialized operating parameters and load power of these components into the model to obtain the adjustable resource capacity information for the day; then inputting this adjustable resource capacity information into a trained real-time frequency regulation optimization model to obtain the optimized power regulation scheme. This invention addresses the operational characteristics of isolated energy systems, studying system dispatch schemes from a multi-timescale collaborative perspective. By integrating the stepped regulation of thermal power plants, the rapid response of energy storage systems, and the flexible regulation capabilities of hydrogen production, it improves the renewable energy absorption level and system operational stability.

Description

Technical Field

[0001] This invention relates to the field of power system regulation technology, and more specifically, to a power system dispatch optimization method based on graph neural networks. Background Technology

[0002] Integrated energy cyber-physical systems (IEPS) are a crucial technological approach for building a clean and low-carbon society. With the widespread application of digital technologies and demand response equipment, the intelligence level of IPS has significantly improved, promoting efficient interaction between energy sources and enabling interconnection and mutual support among various energy sources. However, this also results in significant traffic uncertainty due to large volumes of communication data transmission and user-driven demand response decisions. These factors impact the planning and secure operation of the system's communication network. Therefore, studying the uncertainty of user demand response and balancing communication network traffic transmission is of great significance for the system's planning, operation, and economic efficiency.

[0003] Existing research on dispatch optimization largely focuses on dispatch optimization or capacity allocation problems at a single time scale, primarily studying day-ahead dispatch optimization and frequency capacity reserves. However, optimization combining day-ahead frequency capacity reserves with real-time frequency regulation is rarely addressed. Furthermore, existing methods still have room for improvement in evaluating the performance of thermal power units' stepped regulation and studying frequency regulation modes for deep peak-shaving loads, especially in integrating flexible resources such as power-to-hydrogen and battery energy storage systems for multi-time-scale dispatch, where a systematic solution has yet to be developed. Current frequency regulation optimization often employs traditional heuristic algorithms to handle nonlinear problems, but their computational efficiency decreases significantly with increasing system scale. Although neural network technology (such as convolutional networks) has made progress in areas such as grid load forecasting and fault location, it has not been applied to multi-time-scale dispatch optimization, particularly research integrating spatiotemporal feature learning with rapid dispatch decision-making. Summary of the Invention

[0004] The purpose of this invention is to provide a power system dispatch optimization method based on graph neural networks. Existing power systems face insufficient multi-timescale coordination capabilities when dealing with high-proportion renewable energy integration, making it difficult to effectively balance intermittent power sources and load demand. Traditional dispatch methods have limitations in the coordinated optimization of electricity, hydrogen, and storage, failing to fully leverage the advantages of rapid response from energy storage systems and flexible adjustment from hydrogen energy systems, resulting in limited system operating economy and renewable energy absorption efficiency. This invention proposes a multi-timescale dispatch method based on the synergy of linear optimization and deep learning. By constructing an electricity-hydrogen-storage joint operation model, coordinated optimization is achieved at both the day-ahead and real-time timescales. This method utilizes deep learning to analyze system operating characteristics, combined with linear optimization algorithms for multi-objective decision-making, optimizing energy storage charging and discharging strategies and hydrogen production load allocation, effectively improving the system's flexible adjustment capabilities, enhancing renewable energy absorption, and simultaneously meeting the ancillary service needs of grid peak shaving and frequency regulation.

[0005] To achieve the above objectives, the embodiments of this application provide the following technical solutions:

[0006] This application provides a power system dispatch optimization method based on graph neural networks, applicable to power systems including wind and photovoltaic generators, hydrogen generators, energy storage, and thermal power units. The method includes: initializing the operating parameters and load power of the wind and photovoltaic generators, hydrogen generators, energy storage, and thermal power units in the power system; constructing a day-ahead dispatch optimization model based on the stepped ramping of thermal power units, and inputting the initialized operating parameters and load power of the wind and photovoltaic generators, hydrogen generators, energy storage, and thermal power units into the model to obtain the adjustable resource capacity information for the day. The adjustable resource capacity information is used to characterize the upper and lower limits of the operating parameters of the wind and photovoltaic generators, hydrogen generators, energy storage, and thermal power units in the power system; inputting the adjustable resource capacity information into a trained real-time frequency regulation optimization model to obtain an optimized power regulation scheme. The real-time frequency regulation optimization model is a neural network model, including an optimal frequency regulation unit and a graph neural network and prediction profile correction unit.

[0007] Optionally, a day-ahead scheduling optimization model based on the stepped ramping of thermal power units is constructed, including constructing a stepped ramping sub-model of thermal power units, a stepped ramping constraint sub-model taking into account the up and down reserve capacity, and a stepped frequency regulation performance sub-model.

[0008] Among them, the thermal power unit stepped ramp sub-model is used to characterize the ramping capability of the unit in different depth peak shaving operation ranges in a stepped ramp form, so as to ensure that the planned power of the day-ahead meets the achievable upward and downward changes between adjacent time periods.

[0009] A step ramping constraint sub-model that takes into account both uplink and downlink reserve capacity is used to provide the system with callable uplink and downlink reserves while ensuring that the unit's daily output meets the operating constraints, and to limit the reserve capacity margin and step ramping capability of the unit.

[0010] A tiered frequency regulation performance sub-model is used to evaluate mileage and comprehensive regulation capabilities based on unit regulation performance, segmented by oil injection depth peak shaving, deep peak shaving, and conventional peak shaving. Combining uplink and downlink frequency capacity and frequency regulation costs, it generates comprehensive performance values ​​for day-ahead dispatching and frequency regulation capacity allocation. The power system allocates frequency regulation capacity to units with uplink or downlink regulation capabilities based on the response speed and adjustable upper and lower limits of different regulation resources.

[0011] Optionally, the method for constructing the stepped ramp sub-model of the thermal power unit is as follows:

[0012] The output of thermal power units is managed in segments according to the deep peak shaving stage. Each different peak shaving stage includes conventional peak shaving, deep peak shaving and oil injection deep peak shaving. Each different peak shaving stage corresponds to different up and down ramp rates and output upper and lower limits. Secondly, the current peak shaving stage and the corresponding feasible ramp range are determined by combining the start-up status of thermal power units and the corresponding deep peak shaving boundary.

[0013] Alternatively, the step ramp constraint sub-model that takes into account both uplink and downlink reserve capacity can be constructed as follows:

[0014] Based on the stepped ramp sub-model of thermal power units, the uplink and downlink frequency capacity reserves that thermal power units can provide in each time period are calculated. At the same time, thermal power units are subject to the joint constraints of the stage ramp rate, unit capacity, start-up and shutdown status and deep peak shaving boundary, thus forming the day-ahead adjustable resource capacity information, which is used as the safety boundary for frequency reserve capacity allocation.

[0015] Optionally, the step-based frequency regulation performance sub-model can be constructed as follows:

[0016] Establish segmented regulation performance evaluation standards for different peak shaving stages, and give a weighted score by combining mileage regulation performance and overall regulation performance; when in the deep peak shaving or deep oil injection peak shaving stage, introduce different regulation performance performance and regulation performance score.

[0017] Based on the weighted scoring or regulation performance score, the corresponding uplink and downlink capacity, response delay and ramp rate characteristics are obtained, and the power regulation scheme is coordinated and optimized based on the uplink and downlink capacity, response delay and ramp rate characteristics corresponding to each peak shaving stage.

[0018] Optionally, the day-ahead scheduling optimization model based on the stepped ramping of thermal power units also includes the construction of a flexible resource operation sub-model for the system, wherein the flexible resource operation sub-model for the system includes a hydrogen production and storage sub-model and an energy storage sub-model.

[0019] Optionally, the hydrogen production and storage sub-model in the flexible resource operation sub-model of the system can be constructed as follows:

[0020] Excess electricity is prioritized for hydrogen production, which is then injected into storage tanks and sold to the market. The operating power of the hydrogen generator is constrained by the equipment's maximum power and minimum stable output ratio, and the corresponding uplink / downlink frequency adjustment capability is generated based on the operating point of the hydrogen generator in the system.

[0021] Secondly, the capacity of the hydrogen storage tank is limited by upper and lower limits and charging and discharging conditions, and the initial and final hydrogen storage amounts are assumed to be equal; thus forming an optimization model for hydrogen production revenue and available backup capacity.

[0022] Optionally, the energy storage sub-model in the flexible resource operation sub-model of the system can be constructed as follows:

[0023] The energy storage state is constrained by preset upper and lower limits; only charging or discharging states are allowed during any optimization period; taking into account charging / discharging efficiency, power limits and equivalent uplink / downlink frequency regulation capabilities, it can perform the function of rapid power regulation while meeting safety boundaries;

[0024] Furthermore, the consistency condition of the initial and final states of charge is applied during the day-to-day rolling process to ensure the sustainability of its compensation plan.

[0025] Optionally, flexible backup and planned output optimization schemes can be constructed based on hydrogen production and storage sub-models and energy storage sub-models to enable the consumption of renewable energy, thereby reducing wind and solar power curtailment, and alleviating the pressure of thermal power units on ramping and deep peak shaving under load fluctuations and frequency regulation requirements.

[0026] Secondly, the adjustable resource capacity information formed by its flexible backup and planned output optimization scheme serves as a security boundary for real-time frequency allocation, thereby supporting the comprehensive optimization of the system's economy, security, and flexibility.

[0027] Optionally, a day-ahead scheduling optimization model based on the stepped ramping of thermal power units is constructed, including constructing the stepped ramping of thermal power units, implemented as follows:

[0028] Ramp speed limit:

[0029] ;

[0030] ;

[0031] in , , and These represent the output state, ramping state, deep peak shaving state, and startup state of the nth thermal power unit at time t, respectively.

[0032] , , and These represent the output state, ramping state, deep peak shaving state, and startup state of the nth thermal power unit at time t-1, respectively.

[0033] The ramp value represents the peak-shaving value under the oil injection depth condition; Represents the ramp value under deep peak shaving conditions; This represents the ramp value under normal peak-shaving conditions.

[0034] This represents the lower limit of power output under peak-shaving conditions at the depth of oil injection. This represents the lower limit of power under normal peak-shaving conditions; This represents the lower limit of power under deep peak shaving conditions.

[0035] This represents the ramp value under peak-shaving conditions. If If the output power of the nth thermal power unit at time t is in a normal deep peak-shaving state, then... The climbing state is ;if Thermal power output is at Its climbing state is ;if The unit power is And, the climbing stage is ; For deep peak-shaving power boundary, , and This indicates that the changes in the unit's peak-shaving state before and after are considered nearest neighbors, i.e. ; This indicates that when the unit's peak-shaving state changes twice, then... ;

[0036] ;

[0037] in and These represent the uplink and downlink frequency regulation capabilities of the nth thermal power unit in the t-th optimal interval, respectively. and These represent the uplink and downlink frequency regulation capabilities of the nth thermal power unit in the (t-1)th optimal interval, respectively.

[0038] The formula for calculating the maintenance cost of thermal power plants is as follows:

[0039] ;

[0040] in , and These represent the maintenance cost, rotor cracking cycle, and fuel consumption of the nth thermal power unit during the tth optimal time interval, respectively. This represents the squared output value of the nth generator unit at time t; , and These are the preset coal consumption characteristic function parameters for the nth thermal power unit; , and These represent the investment price, carbon price, and oil price of the nth thermal power unit, respectively. and These represent the operating loss coefficients of the nth thermal power unit under oil-filled and oil-free depth adjustment, respectively;

[0041] The specific calculation method for constructing the regulation performance model of thermal power units is as follows:

[0042] ;

[0043] ;in Let be the comprehensive value of the frequency regulation performance of the nth thermal power unit at time t. , It refers to the frequency regulation cost and peak-shaving status of the nth thermal power unit with uplink frequency capacity. and This represents the mileage regulation performance and total regulation performance score of the nth thermal power unit in the current interval; and This indicates the mileage regulation performance and overall regulation performance score when the frequency regulation is upward; and These represent the mileage regulation performance of the nth thermal power unit under deep peak shaving and conventional peak shaving when the frequency regulation is upward; and This indicates the mileage regulation performance of the nth thermal power unit when the frequency regulation is upward, encompassing both deep peak shaving and conventional peak shaving. This indicates the boundary of the peak-shaving state.

[0044] Optionally, a day-ahead scheduling optimization model based on the stepped ramping of thermal power units is constructed, including the construction of flexible resource operation constraints for the system, implemented as follows:

[0045] P2H hydrogen production constraint calculation formula:

[0046] ;

[0047] in The amount of hydrogen stored in the hydrogen storage tank during the (t+1)th optimal time interval. , and These represent the hydrogen storage capacity of the hydrogen storage tank, the operating power of the P2H equipment, and the flow rate of hydrogen sold from the hydrogen storage tank to the hydrogen market during the t-th optimal time interval, respectively. and Indicates the efficiency of P2H equipment and hydrogen storage tanks; This indicates the energy conversion rate between electrical energy and hydrogen energy;

[0048] ;

[0049] in This represents the maximum power of the P2H device; This represents the minimum operating power ratio of P2H equipment; and This represents the frequency adjustment capability of the P2H device in the t-th optimal range.

[0050] In the current scheduling process, the capacity of the hydrogen storage tank in the termination optimization interval should be equal to the capacity of the initial optimization interval, as detailed below:

[0051] ;

[0052] in and These represent the capacity status of the hydrogen storage tank at the initial optimization and the end of the optimization, respectively.

[0053] The constraint calculation formula for the energy storage model is as follows:

[0054] ;

[0055] in , and These represent the SOC, charging power, and discharging power of BESS during the t-th optimal time interval, respectively. This represents the SOC of BESS during the (t+1)th optimal time interval; It is BESS's maximum SOC; and This represents the charge and discharge efficiency of BESS;

[0056] Secondly, within an optimization range, the energy storage system can only be in one of two states of charge or discharge, therefore capacity constraints must also be met. The calculation method is as follows:

[0057] ;

[0058] in and It is a binary variable representing the discharge / charge state of the BESS within the t-th optimal interval; and These represent the uplink and downlink frequency regulation capabilities of BESS at the t-th optimal interval, respectively.

[0059] In daytime dispatching, the SOC of the energy storage system in the termination optimization interval should be equal to the SOC of the initial optimization interval. The calculation formula is as follows:

[0060] ;

[0061] in and These represent the State of Charge (SOC) of the BESS at the initial and final intervals, respectively. Optionally, a day-ahead scheduling optimization model based on the stepped ramping of thermal power units is constructed, including the construction of the day-ahead scheduling model, implemented as follows:

[0062] For day-ahead energy balance, the sum of regulating resources should equal the load demand for each time control interval, as shown below:

[0063] ;

[0064] in , and These represent the optimal wind power output, solar power output, and load power for the t-th optimal interval, respectively.

[0065] During the current optimization phase, the resource optimization methods for peak shaving and frequency regulation capabilities are as follows:

[0066] ;

[0067] in , , and These represent the hydrogen production profit, the start-up and operation cost of thermal power units, the renewable energy reduction, and the frequency cost for the t-th optimal time interval, respectively.

[0068] How to calculate the profit from hydrogen production:

[0069] ;

[0070] in That is the selling price of hydrogen, and This refers to hydrogen density; the formula for calculating the penalty for renewable energy reduction is:

[0071] ;

[0072] in This represents the penalty factor for each kilowatt-hour of renewable energy reduction; Represents the minimum reduction rate for wind power. This represents the minimum reduction rate for photovoltaic power. and Predicted power output representing solar and wind energy; and This represents the uplink frequency adjustment threshold for wind power and photovoltaic power at time t.

[0073] The constraint on the adjustment capability is:

[0074] ;

[0075] in This represents the maximum ratio of the frequency capacity of the k-th regulating resource to its power capacity. This represents the mileage regulation performance of the k-th regulation resource. , , and These represent the uplink and downlink frequency modulation capacity and frequency modulation mileage of the k-th frequency modulation resource within the t-th optimal interval, respectively. , , and This represents the uplink / downlink capacity requirements and FM mileage requirements for the day.

[0076] Optionally, the real-time frequency adjustment optimization model includes an optimal frequency adjustment unit and a graph neural network and prediction archive correction unit;

[0077] In the frequency regulation optimization, the optimal frequency regulation unit considers the power deviation between the power input and output as the minimum optimization objective, as described below:

[0078] ;

[0079] in and They represent the first Total regulated power input and actual power output, Number of frequency adjustment control intervals for each service period;

[0080] In frequency regulation, the generator output rate constraint, power limitation constraint, power balance constraint, and regulation direction constraint can be written in the following form:

[0081]

[0082] in Represents the ramp of the k-th regulating resource. No. -1 Total regulating power input and actual power output, , and This represents the power command and downlink / uplink regulation capability of the k-th regulation resource within the t-th time control interval.

[0083] Optionally, the real-time frequency adjustment optimization model includes an optimal frequency adjustment unit and a graph neural network and prediction archive correction unit;

[0084] The spatiotemporal graph neural network used in the graph neural network and prediction archive correction unit includes an input layer, a graph convolutional layer, a fully connected layer, an LSTM layer, and an output layer. Besides the input features, the input layer also includes an adjacency matrix. The adjacency matrix directly reveals the relationships between different input features. In graph theory, the relationships between resources in frequency regulation can be guaranteed to be a graph. The larger the weight value of the adjacency matrix, the higher the correlation between two vertices. In real-time frequency regulation optimization, the adjacency matrix is ​​measured by utilizing the similarity of the average frequency regulation capacity, average frequency regulation mileage, response delay score, and ramp score of each unit in day-ahead scheduling. The construction of the weighted adjacency matrix is ​​as follows:

[0085] During the forward propagation of the graph convolutional layer, the feature output is calculated using the weight parameters, adjacency matrix, and degree matrix. The calculation formula is as follows:

[0086] ;

[0087] in and This represents the input and output characteristics of the graph convolutional layer; I is an identity matrix; , and These represent the distance matrix, the adjacency matrix, and the adjacency matrix with self-connections, respectively. This represents the element in the k-th row and j-th column of the adjacency matrix with self-joins;

[0088] and These represent the network weights and biases of the graph convolutional layer, respectively.

[0089] The program for the new frequency regulation task can be generated directly from the total regulation power command, mileage history data, and the regulation capability based on a trained spatiotemporal graph neural network, as follows:

[0090] ;

[0091] in The spatiotemporal graph neural network used for real-time frequency regulation is represented in the first... dimensional input features These represent the historical sequences of total power command, mileage command, operating power, upper limit adjustment capability, and lower limit adjustment capability, respectively. It is a 1×K vector, and the other input features are w×K matrices, where w is the length of the history sequence; The optimal scheduling scheme; This represents a trained spatiotemporal graphical neural network. and The optimal weight and bias set; It is a power modification scheme for adjusting resources; This represents the optimal function considering Euclidean distance.

[0092] The beneficial effects of this invention are as follows:

[0093] Existing scheduling methods mainly focus on optimization at a single time scale or independent resource scheduling, making it difficult to achieve multi-time scale collaborative optimization and multi-energy complementarity. Therefore, this invention addresses the operational characteristics of isolated energy systems by studying system scheduling schemes from a multi-time scale collaborative perspective. By integrating the tiered regulation of thermal power units, the rapid response of energy storage systems, and the flexible adjustment capabilities of electricity-to-hydrogen production, it improves the level of renewable energy absorption and the stability of system operation.

[0094] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing embodiments of the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0095] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0096] Figure 1 This is a schematic diagram of the power system structure described in the embodiments of the present invention;

[0097] Figure 2 This is a comparison of the power deviation between the method proposed in the embodiments of the present invention and the traditional case;

[0098] Figure 3 This is a flowchart illustrating the real-time frequency adjustment solution as described in an embodiment of the present invention. Detailed Implementation

[0099] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0100] It should be noted that similar reference numerals or letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0101] Example 1:

[0102] like Figure 1 As shown, this embodiment provides a power system dispatch optimization method based on graph neural networks, applicable to power systems, where the power includes wind and photovoltaic generators, hydrogen generators, energy storage and thermal power units, and the method includes steps S100, S200 and S300.

[0103] Step S100: Operating parameters and load power of wind and photovoltaic generators, hydrogen generators, energy storage and thermal power units in the initial power system;

[0104] Step S200: Construct a day-ahead scheduling optimization model based on the stepped ramp of thermal power units, and input the initialized operating parameters of wind power and photovoltaic power generation units, hydrogen generators, energy storage and thermal power units as well as the load power into the model to obtain the adjustable resource capacity information for the day. The adjustable resource capacity information is used to characterize the upper and lower limits of the operating parameters of wind power and photovoltaic power generation units, hydrogen generators, energy storage and thermal power units in the power system.

[0105] Step S300: Input the adjustable resource capacity information into the trained real-time frequency regulation optimization model to obtain the optimized power regulation scheme. The real-time frequency regulation optimization model is a neural network model, including an optimal frequency regulation unit and a graph neural network and prediction file correction unit.

[0106] The project includes constructing a day-ahead scheduling optimization model based on the stepped ramping of thermal power units. This model comprises a stepped ramping sub-model for thermal power units, a stepped ramping constraint sub-model considering both upward and downward reserve capacity, and a stepped frequency regulation performance sub-model. The stepped ramping sub-model for thermal power units is used to characterize the ramping capability of units in different peak-shaving operating ranges in a stepped ramping manner, ensuring that the day-ahead planned power meets achievable upward and downward variations between adjacent time periods.

[0107] A step ramping constraint sub-model that takes into account both uplink and downlink reserve capacity is used to provide the system with callable uplink and downlink reserves while ensuring that the unit's daily output meets the operating constraints, and to limit the reserve capacity margin and step ramping capability of the unit.

[0108] A tiered frequency regulation performance sub-model is used to evaluate mileage and comprehensive regulation capabilities based on unit regulation performance, segmented by oil injection depth peak shaving, deep peak shaving, and conventional peak shaving. Combining uplink and downlink frequency capacity and frequency regulation costs, it generates comprehensive performance values ​​for day-ahead dispatching and frequency regulation capacity allocation. The power system allocates frequency regulation capacity to units with uplink or downlink regulation capabilities based on the response speed and adjustable upper and lower limits of different regulation resources.

[0109] (1) The method for constructing the stepped ramp sub-model of thermal power units is as follows:

[0110] The output of thermal power units is managed in segments according to the deep peak shaving stage. Each different peak shaving stage includes conventional peak shaving, deep peak shaving and oil injection deep peak shaving. Each different peak shaving stage corresponds to different up and down ramp rates and output upper and lower limits. Secondly, the current peak shaving stage and the corresponding feasible ramp range are determined by combining the start-up status of thermal power units and the corresponding deep peak shaving boundary.

[0111] (2) The method for constructing the stepped ramp constraint sub-model that takes into account both uplink and downlink reserve capacity is as follows:

[0112] Based on the stepped ramp sub-model of thermal power units, the uplink and downlink frequency capacity reserves that thermal power units can provide in each time period are calculated. At the same time, thermal power units are subject to the joint constraints of the stage ramp rate, unit capacity, start-up and shutdown status and deep peak shaving boundary, thus forming the day-ahead adjustable resource capacity information, which is used as the safety boundary for frequency reserve capacity allocation.

[0113] (3) The method for constructing the stepped frequency regulation performance sub-model is as follows:

[0114] Establish segmented regulation performance evaluation standards for different peak shaving stages, and give a weighted score by combining mileage regulation performance and overall regulation performance; when in the deep peak shaving or deep oil injection peak shaving stage, introduce different regulation performance performance and regulation performance score.

[0115] Based on the weighted scoring or regulation performance score, the corresponding uplink and downlink capacity, response delay and ramp rate characteristics are obtained, and the power regulation scheme is coordinated and optimized based on the uplink and downlink capacity, response delay and ramp rate characteristics corresponding to each peak shaving stage.

[0116] (4) Constructing a day-ahead scheduling optimization model based on the step-by-step ramping of thermal power units also includes constructing a flexible resource operation sub-model of the system, wherein the flexible resource operation sub-model of the system includes a hydrogen production and storage sub-model and an energy storage sub-model.

[0117] (4.1) The method for constructing the hydrogen production and storage sub-model in the flexible resource operation sub-model of the system is as follows:

[0118] Excess electricity is prioritized for hydrogen production, which is then injected into storage tanks and sold to the market. The operating power of the hydrogen generator is constrained by the equipment's maximum power and minimum stable output ratio, and the corresponding uplink / downlink frequency adjustment capability is generated based on the operating point of the hydrogen generator in the system.

[0119] Secondly, the capacity of the hydrogen storage tank is limited by upper and lower limits and charging and discharging conditions, and the initial and final hydrogen storage amounts are assumed to be equal; thus forming an optimization model for hydrogen production revenue and available backup capacity.

[0120] (4.2) The method for constructing the energy storage sub-model in the flexible resource operation sub-model of the system is as follows:

[0121] The energy storage state is constrained by preset upper and lower limits; only charging or discharging states are allowed during any optimization period; taking into account charging / discharging efficiency, power limits and equivalent uplink / downlink frequency regulation capabilities, it can perform the function of rapid power regulation while meeting safety boundaries;

[0122] Furthermore, the consistency condition of the initial and final states of charge is applied during the day-to-day rolling process to ensure the sustainability of its compensation plan.

[0123] (4.3) Based on the hydrogen production and storage sub-model and the energy storage sub-model, construct a flexible backup and planned output optimization scheme to enable renewable energy to be absorbed, thereby reducing wind and solar energy curtailment, and at the same time alleviating the pressure of thermal power units on ramping and deep peak shaving under load fluctuation and frequency regulation demand.

[0124] Secondly, the adjustable resource capacity information formed by its flexible backup and planned output optimization scheme serves as a security boundary for real-time frequency allocation, thereby supporting the comprehensive optimization of the system's economy, security, and flexibility.

[0125] Example 2:

[0126] This embodiment, based on Embodiment 1, provides a specific method for constructing a day-ahead scheduling optimization model based on the stepped ramping of thermal power units:

[0127] 1. Day-ahead scheduling optimization model:

[0128] The power system includes wind and solar renewable energy, hydrogen generators, batteries, and thermal power units, with an overall structure as follows: Figure 1 As shown.

[0129] 1.1 Stepped Climbing of Thermal Power Units:

[0130] Thermal power is typically used as a peak-shaving resource to compensate for the inherent instability of renewable energy and provide a stable power supply to electricity users. With the increasing penetration rate of renewable energy, thermal power units are needed to improve system flexibility and reduce renewable energy consumption. This trend increases the operational burden on thermal power units that are in deep peak-shaving mode when necessary. However, previous frequency capacity allocation or frequency response did not fully consider the impact of thermal power in deep peak-shaving mode, and the implementation methods were as follows:

[0131] 1.1.1 Ramp Rate Limit: The rate of change of the regulated power output for each regulated resource should be limited to the allowable range, as shown below:

[0132] ;

[0133] ;

[0134] in , , and These represent the output state, ramping state, deep peak shaving state, and startup state of the nth thermal power unit at time t, respectively.

[0135] , , and These represent the output state, ramping state, deep peak shaving state, and startup state of the nth thermal power unit at time t-1, respectively.

[0136] The ramp value represents the peak-shaving value under the oil injection depth condition; Represents the ramp value under deep peak shaving conditions; This represents the ramp value under normal peak-shaving conditions.

[0137] This represents the lower limit of power output under peak-shaving conditions at the depth of oil injection. This represents the lower limit of power under normal peak-shaving conditions; This represents the lower limit of power under deep peak shaving conditions.

[0138] This represents the ramp value under peak shaving conditions; if If the output power of the nth thermal power unit at time t is in a normal deep peak-shaving state, then... The climbing state is ;if Thermal power output is at Its climbing state is ;if The unit power is And, the climbing stage is ; It is the deep peak-shaving power boundary, , and There are two peak-shaving states. This indicates that the changes in the unit's peak-shaving state before and after are considered nearest neighbors, i.e. ; This indicates that when the unit's peak-shaving state changes twice, then... ;

[0139] ;

[0140] in and These represent the uplink and downlink frequency regulation capabilities of the nth thermal power unit in the t-th optimal interval, respectively.

[0141] 1.1.2 Constructing the Maintenance Cost of Thermal Power Plants: The maintenance cost of thermal power units can be the sum of coal consumption costs and oil costs during deep peak shaving, and its calculation formula is as follows:

[0142] ;

[0143] in , and These represent the maintenance cost, rotor cracking cycle, and fuel consumption of the nth thermal power unit during the tth optimal time interval, respectively. This represents the squared output value of the nth generator unit at time t; , and These are the preset coal consumption characteristic function parameters for the nth thermal power unit; , and These represent the investment price, carbon price, and oil price of the nth thermal power unit, respectively. and These represent the operating loss coefficients of the nth thermal power unit under oil-filled and oil-free depth adjustment, respectively;

[0144] 1.1.3 Construction of the Regulation Performance Model for Thermal Power Units: In frequency regulation, the stepped performance of the TPU (Temporary Power Unit) must be considered. The load power status of the thermal power unit will affect the regulation mileage performance and frequency regulation performance score of the TPU participating in frequency regulation. When the unit is in a deep peak-shaving state, both the regulation mileage performance and regulation performance score of the thermal power unit will decrease, as detailed below:

[0145] ;

[0146] ;

[0147] in Let be the comprehensive value of the frequency regulation performance of the nth thermal power unit at time t. , It refers to the frequency regulation cost and peak-shaving status of the nth thermal power unit with uplink frequency capacity. and This represents the mileage regulation performance and total regulation performance score of the nth thermal power unit in the current interval; and This indicates the mileage regulation performance and overall regulation performance score when the frequency regulation is upward; and These represent the mileage regulation performance of the nth thermal power unit under deep peak shaving and conventional peak shaving when the frequency regulation is upward; and This indicates the mileage regulation performance of the nth thermal power unit when the frequency regulation is upward, encompassing both deep peak shaving and conventional peak shaving. This indicates the boundary of the peak-shaving state. Similarly, the corresponding frequency regulation mileage and frequency regulation cost for the TPU can be calculated based on this. Considering both peak-shaving and frequency regulation, the relationship between load power and the thermal elevator's graded ramp type is such that all three states correspond to participation in peak-shaving. If the unit's power and standby position... Then the output power of the nth TPU in the t-th interval represents the peak shaving state at the oil injection depth. , and .if Then the output power of the nth TPU in the t-th interval is represented as being in the normal deep peak shaving state. , and .if Then the output power of the nth TPU in the t-th interval is considered to be in the normal peak-shaving state. The mileage adjustment performance level when the frequency adjustment capability increases And the overall regulation performance score when the frequency regulation capability increases .

[0148] 1.2 System Flexible Resource Operation Constraints:

[0149] Thermal power is typically used as a peak-shaving resource to compensate for the inherent instability of renewable energy and provide a stable power supply to electricity users. With the increasing penetration rate of renewable energy, thermal power units are needed to improve system flexibility and reduce renewable energy consumption. This trend increases the operational burden on thermal power units operating under deep peak-shaving conditions when necessary. However, previous frequency capacity allocation or frequency response did not fully consider the impact of thermal power under deep peak-shaving conditions.

[0150] A day-ahead scheduling optimization model based on the stepped ramping of thermal power units is constructed, including the construction of flexible resource operation constraints for the system. The implementation method is as follows:

[0151] 1.2.1: Calculation formula for P2H hydrogen production constraints:

[0152] ;

[0153] in The amount of hydrogen stored in the hydrogen storage tank during the (t+1)th optimal time interval. , and These represent the hydrogen storage capacity of the hydrogen storage tank, the operating power of the P2H equipment, and the flow rate of hydrogen sold from the hydrogen storage tank to the hydrogen market during the t-th optimal time interval, respectively. and Indicates the efficiency of P2H equipment and hydrogen storage tanks; This indicates the energy conversion rate between electrical energy and hydrogen energy;

[0154] ;

[0155] in This represents the maximum power of the P2H device; This represents the minimum operating power ratio of P2H equipment; and This represents the frequency adjustment capability of the P2H device in the t-th optimal range.

[0156] In the current scheduling process, the capacity of the hydrogen storage tank in the termination optimization interval should be equal to the capacity of the initial optimization interval, as detailed below:

[0157] ;

[0158] in and These represent the capacity status of the hydrogen storage tank at the initial optimization and the end of the optimization, respectively.

[0159] The constraint calculation formula for the energy storage model is as follows:

[0160] ;

[0161] in , and These represent the SOC, charging power, and discharging power of BESS during the t-th optimal time interval, respectively. This represents the SOC of BESS during the (t+1)th optimal time interval; It is BESS's maximum SOC; and This represents the charge and discharge efficiency of BESS;

[0162] Secondly, within an optimization range, the energy storage system can only be in one of two states of charge or discharge, therefore capacity constraints must also be met. The calculation method is as follows:

[0163] ;

[0164] in and It is a binary variable representing the discharge / charge state of the BESS within the t-th optimal interval; and These represent the uplink and downlink frequency regulation capabilities of BESS at the t-th optimal interval, respectively.

[0165] In daytime dispatching, the SOC of the energy storage system in the termination optimization interval should be equal to the SOC of the initial optimization interval. The calculation formula is as follows:

[0166] ;

[0167] in and These represent the SOC of BESS at the initial and final intervals, respectively.

[0168] 1.3 Construct a day-ahead scheduling model, implemented as follows:

[0169] 1.3.1 For day-ahead energy balance, the sum of regulating resources should equal the load demand for each time control interval, as shown below:

[0170] ;

[0171] in , and These represent the optimal wind power output, solar power output, and load power for the t-th optimal interval, respectively.

[0172] During the current optimization phase, the resource optimization methods for peak shaving and frequency regulation capabilities are as follows:

[0173] ;

[0174] in , , and These represent the hydrogen production profit, the start-up and operation cost of thermal power units, the renewable energy reduction, and the frequency cost for the t-th optimal time interval, respectively.

[0175] 1.3.2 By selling the hydrogen from the hydrogen tanks to the hydrogen market, the following profits can be obtained. The profit calculation method for hydrogen production is as follows:

[0176] ;

[0177] in That is the selling price of hydrogen, and It is the density of hydrogen gas;

[0178] 1.3.3 Renewable Energy Reduction Penalty: To curb the waste of wind and solar energy, this constraint is added to reduce the consumption of renewable energy. The calculation formula is as follows:

[0179] ;

[0180] in This represents the penalty factor for each kilowatt-hour of renewable energy reduction; Represents the minimum reduction rate for wind power. This represents the minimum reduction rate for photovoltaic power. and Predicted power output representing solar and wind energy; and This represents the uplink frequency adjustment threshold for wind power and photovoltaic power at time t.

[0181] 1.3.4 Adjustment Capacity Constraint: The frequency adjustment capacity of resources should be higher than the minimum frequency capacity constraint of the day-ahead demand, and the total mileage should also meet the demand, as shown in the table below:

[0182] ;

[0183] in This represents the maximum ratio of the frequency capacity of the k-th regulating resource to its power capacity. This represents the mileage regulation performance of the k-th regulation resource. , , and These represent the uplink and downlink frequency modulation capacity and frequency modulation mileage of the k-th frequency modulation resource within the t-th optimal interval, respectively. , , and This represents the uplink / downlink capacity requirements and FM mileage requirements for the day.

[0184] 2. Real-time frequency adjustment optimization model:

[0185] The real-time frequency adjustment optimization model includes an optimal frequency adjustment unit and a graph neural network and prediction file correction unit.

[0186] 2.1 Optimal Frequency Adjustment Unit:

[0187] For this system, real-time frequency regulation aims to improve the overall system's response to power disturbances while satisfying various operating constraints. Therefore, in the frequency regulation optimization step, the power deviation between the total regulated power input and output is considered as the minimum optimization objective, as described below:

[0188] ;

[0189] in and They represent the first Total regulated power input and actual power output, Number of frequency adjustment control intervals for each service period;

[0190] In frequency regulation, the generator output rate constraint, power limitation constraint, power balance constraint, and regulation direction constraint can be written in the following form:

[0191] ;

[0192] in Represents the ramp of the k-th regulating resource. No. -1 to the total regulated power input and actual power output. , and This represents the power command and downlink / uplink regulation capability of the k-th regulation resource within the t-th time control interval.

[0193] 2.2 Graph Neural Network and Predictive Archive Correction Unit:

[0194] The spatiotemporal graph neural network employed includes an input layer, graph convolutional layers, fully connected layers, LSTM layers, and an output layer. In addition to input features, the input layer also includes an adjacency matrix, which directly reveals the relationships between different input features. In graph theory, the relationships between resources in frequency regulation can be guaranteed to be a graph; the larger the weight value of the adjacency matrix, the higher the correlation between two vertices. In real-time frequency regulation optimization, the similarity of the average frequency regulation capacity, average frequency regulation mileage, response delay score, and ramp score of each unit in day-ahead scheduling can be used to measure the adjacency matrix. The construction of the weighted adjacency matrix is ​​as follows:

[0195] During the forward propagation of the graph convolutional layer, the feature output can be calculated using the weight parameters, adjacency matrix, and degree matrix. The calculation formula is as follows:

[0196] ;

[0197] in and This represents the input and output characteristics of the graph convolutional layer; I is an identity matrix; , and These represent the distance matrix, the adjacency matrix, and the adjacency matrix with self-connections, respectively. This represents the element in the k-th row and j-th column of the adjacency matrix with self-joins; and These represent the network weights and biases of the graph convolutional layer, respectively.

[0198] The program for the new frequency regulation task can be generated directly from the total regulation power command, mileage history data, and the regulation capability based on a trained spatiotemporal graph neural network, as follows:

[0199] ;

[0200] in = indicates that the spatiotemporal graph neural network used for real-time frequency adjustment is in the first... dimensional input features These represent the historical sequences of total power command, mileage command, operating power, upper limit adjustment capability, and lower limit adjustment capability, respectively. It is a 1×K vector, and the other input features are w×K matrices, where w is the length of the history sequence; The optimal scheduling scheme; This represents a trained spatiotemporal graphical neural network. and The optimal weight and bias set; It is a power modification scheme for adjusting resources; This represents the optimal function considering Euclidean distance.

[0201] 3. Solution Method

[0202] like Figure 3 As shown, the entire algorithm for solving the day-ahead real-time scheduling problem consists of three steps. Steps 1 to 3 perform day-ahead optimization, and steps 4 to 7 perform data collection and network training (i.e., using the newly collected dataset as new historical training data to train the real-time frequency regulation optimization model). The remaining steps perform online application of real-time frequency regulation. The specific solution process is as follows: Figure 3 As shown.

[0203] Example 3:

[0204] This embodiment is based on Embodiment 1, using a specific power system as an example for analysis. This example is based on a simulation algorithm built on the MATLAB / Python platform, and uses the Gurobi solver and the PyTorch graph neural network framework for solving.

[0205] The planning cost results for the example are shown in Table 1; the specific optimization results of each method in the example are compared below. Figure 2 As shown.

[0206] Table 1 Real-time prediction performance under the example

[0207]

[0208] It is evident that the prediction method considering spatiotemporal characteristics proposed in this application can effectively reduce power deviation and frequency deviation, more safely cope with system disturbances, and enable the system to allocate frequency scheduling schemes more quickly and efficiently.

[0209] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A graph neural network-based power system dispatch optimization method, suitable for a power system, the power system comprising wind energy and photovoltaic generators, hydrogen production machines, energy storage, and thermal power generators, characterized in that, The method comprises: initializing the operation parameters of wind power and photovoltaic generators, hydrogen production machines, energy storage and thermal power generators in the power system and load power; a day-ahead scheduling optimization model based on stepwise climbing of thermal power generators is constructed, and the initialized operation parameters of wind power and photovoltaic generators, hydrogen production machines, energy storage and thermal power generators and load power are brought into the model to obtain adjustable resource capacity information of the day, which is used to represent the upper and lower limits of the operation parameters of wind power and photovoltaic generators, hydrogen production machines, energy storage and thermal power generators in the power system; the adjustable resource capacity information is brought into the trained real-time frequency regulation optimization model to obtain an optimized power regulation scheme, wherein the real-time frequency regulation optimization model is a neural network model, comprising an optimal frequency regulation unit and a graph neural network and prediction archive correction unit; wherein the day-ahead scheduling optimization model based on stepwise climbing of thermal power generators comprises a thermal power generator stepwise climbing sub-model, a stepwise climbing constraint sub-model considering uplink and downlink reserve capacity and a stepwise frequency regulation performance sub-model; the thermal power generator stepwise climbing sub-model is constructed in the following manner: the output of the thermal power generator is managed by depth peak regulation stage, and each different depth peak regulation stage comprises conventional peak regulation, depth peak regulation and oil injection depth peak regulation, and each different peak regulation stage corresponds to different uplink and downlink climbing rates and output upper and lower limits; secondly, the current peak regulation stage and the corresponding feasible climbing interval are determined in combination with the start state of the thermal power generator and the corresponding depth peak regulation boundary, and the specific implementation manner is as follows: a slope rate limit is constructed: wherein and θ n,t respectively represent the output state, climbing state, deep peak regulation state and starting state of the nth thermal power unit at the tth time. and θ n,t-1 respectively represent the output state, climbing state, deep peak regulation state and starting state of the nth thermal power unit at t-1 time. representing the ramp-up value in the oil injection depth peak shaving state; representing the ramp-up value in the depth peak shaving state; representing the ramp-up value in the normal peak shaving state; representing the power lower limit in the oil injection depth peak regulation state; representing the power lower limit in the normal peak regulation state; representing the power lower limit in the deep peak regulation state; represent the ramping value in the peak regulation state, if the output power of the nth thermal power unit at the tth moment is in the normal depth regulation state, that is the ramping state is if the thermal power is in its ramping state is if the unit power is and the ramping stage is as the depth regulation power boundary, and as two regulation states, represent the ramping upper limit in the normal regulation state, represent when the unit regulation state changes from before to after the neighbor, that is represent when the unit regulation state changes from before to after two, the stepwise climbing constraint sub-model considering uplink and downlink reserve capacity is constructed in the following manner: based on the thermal power generator stepwise climbing sub-model, the uplink frequency capacity reserve and the downlink frequency capacity reserve that can be provided by the thermal power generator in each period are calculated, and meanwhile the thermal power generator is jointly constrained by the stage climbing rate, the unit capacity, the start-stop state and the depth peak regulation boundary, thereby forming the day-ahead adjustable resource capacity information, which is used as a safety boundary for frequency reserve capacity distribution; the specific implementation manner is as follows: wherein and respectively represent the up-regulation and down-regulation frequency regulation capability of the nth thermal power unit in the tth optimal interval; and respectively represent the up-regulation and down-regulation frequency regulation capability of the nth thermal power unit in the t-1th optimal interval; a thermal power maintenance cost is constructed, and the calculation formula is as follows: wherein L n,t and respectively represent the maintenance cost, rotor cracking period and oil consumption of the nth thermal power unit in the tth optimal time interval; represent the output square value of the nth unit at the tth time;a n , b n and c n are preset coal consumption characteristic function parameters of the nth thermal power unit; ρ c and ρ o respectively represent the investment price, carbon price and oil price of the nth thermal power unit; and respectively represent the operation loss coefficient of the nth thermal power unit under oil and oil-free deep regulation. the stepwise frequency regulation performance sub-model is constructed in the following manner: segmented regulation performance evaluation standards are established for different peak regulation stages, and the mileage regulation performance and the total regulation performance are weighted and scored; when in the depth peak regulation or oil injection depth peak regulation stage, different regulation performance and regulation performance scores are introduced; the corresponding uplink and downlink capacity, response delay and climbing rate characteristics are obtained based on the weighted scores or the regulation performance scores, and the coordinated optimization power regulation scheme is obtained based on the uplink and downlink capacity, response delay and climbing rate characteristics corresponding to each peak regulation stage; the specific implementation manner is as follows: wherein is the frequency regulation performance comprehensive value of the nth thermal power unit at the tth time, is the frequency regulation cost and the peak regulation state of the nth thermal power unit with uplink frequency capacity, and represent the mileage regulation performance and the total regulation performance score of the nth thermal power unit in the current interval; and represent the mileage regulation performance and the total regulation performance score when the frequency regulation is upward; and respectively represent the mileage regulation performance of the nth thermal power unit in deep peak regulation and regular peak regulation when the frequency regulation is upward; and represent the mileage regulation performance of the nth thermal power unit in deep peak regulation and regular peak regulation when the frequency regulation is upward; represent the boundary of the peak regulation state.

2. The power system dispatch optimization method based on graph neural network according to claim 1, wherein, the day-ahead scheduling optimization model based on stepwise climbing of thermal power generators further comprises a system flexible resource operation sub-model, wherein the system flexible resource operation sub-model comprises a hydrogen production and storage sub-model and an energy storage sub-model; wherein the day-ahead scheduling model is constructed in the following manner: For the day-ahead energy balance, the sum of the adjustment resources should be equal to the load demand of each time control interval, as follows: wherein P t WT , P t PV , P t L , P t P2H , P t dis and P t chr are the solar power output, the wind power output, the load power, the operation power of the P2H device, the discharging power of the energy storage and the charging power of the energy storage at time t, respectively, and the P2H device is a hydrogen production device. In the day-ahead optimization stage, the adjustment resource optimization mode of peak shaving and frequency modulation capacity is as follows: wherein and respectively represent the hydrogen production profit of the tth optimal time interval, the start-up operation cost of the nth thermal power unit, the renewable energy curtailment and the frequency cost. Hydrogen production profit calculation mode: where p H is the sales price of hydrogen, and σ H is the hydrogen density, is the sale of hydrogen from the hydrogen storage tank to the hydrogen market at time t Gas flow rate; Renewable energy reduction penalty calculation formula: where ρ cu represents the penalty coefficient for each kW·h of renewable energy curtailment; ω PV represents the minimum curtailment ratio for wind power, ω WT represents the minimum curtailment ratio for photovoltaic power; P t PV,pre and P t WT,pre represent the predicted power of solar and wind power; F t U,PV and F t U,WT represent the up-regulation threshold of wind power and photovoltaic power at the tthmoment; Adjustment capacity constraint condition: where r k represents the maximum ratio of the frequency capacity of the kth adjustment resource in the power capacity, represents the mileage adjustment performance of the kth adjustment resource, and respectively represent the uplink and downlink frequency modulation capacities and the frequency modulation mileage of the kth frequency modulation resource in the tth optimal interval, F t U,min , F t D,min , and are the daily frequency modulation uplink / downlink capacity demand and the frequency modulation mileage demand.

3. The power system dispatch optimization method based on graph neural network according to claim 2, characterized in that, The hydrogen production and hydrogen storage sub-model in the construction of the system flexible resource operation sub-model is as follows: Firstly, the excess electricity is preferentially used for electric hydrogen production, hydrogen production is injected into the hydrogen storage tank and sold on the market; wherein the hydrogen production machine operation power is constrained by the maximum power and minimum stable output ratio of the equipment, and the corresponding uplink / downlink frequency regulation capacity is generated combined with the operation point of the hydrogen production machine in the system; Secondly, the hydrogen storage tank capacity is limited by the upper and lower limits and the charging and discharging conditions, and the initial and final hydrogen storage amounts are equal; and then an optimization model of hydrogen production benefit and available standby capacity is formed; The specific implementation mode is as follows: wherein the hydrogen storage amount of the hydrogen storage tank in the t+1th optimal time interval, P t P2H and respectively the hydrogen storage amount of the hydrogen storage tank in the tth optimal time interval, the operating power of the P2H device and the flow rate of hydrogen sold from the hydrogen storage tank to the hydrogen market; η p and η h represent the efficiencies of the P2H device and the hydrogen storage tank; γ h represents the energy conversion rate of electric energy and hydrogen energy. wherein represents the maximum power of the P2H device; γ p represents the minimum operating power ratio of the P2H device; F t U,P2H and F t D,P2H represents the up and down frequency adjustment capability of the P2H device in the tth optimal interval.

4. The power system dispatch optimization method based on graph neural network according to claim 3, characterized in that, The energy storage sub-model in the construction of the system flexible resource operation sub-model is as follows: The energy storage state is constrained by the preset upper and lower limits; only charging state or discharging state is allowed in any optimization period; the charging / discharging efficiency, power limit and equivalent uplink / downlink frequency regulation capacity are considered, so that it can meet the safety boundary while undertaking the function of rapid power regulation; And in the day-ahead rolling process, the consistency condition of the initial and final state of charge is applied, so as to ensure the sustainability of the compensation plan, and the specific implementation mode is as follows: In the day-ahead scheduling, the hydrogen storage tank capacity of the termination optimization interval should be equal to the capacity of the initial optimization interval, as follows: wherein and respectively the capacity state of the hydrogen storage tank at the beginning of the optimization and at the end of the optimization. The constraint calculation formula of the energy storage model is as follows: wherein P t chr and P t dis SOCt, Pt, Pt+1represent the SOC, charging power and discharging power of the BESS in the tth optimal time interval, respectively; SOCtrepresents the SOC of the BESS in the tth optimal time interval; is the maximum SOC of the BESS; η C and η d represents the charge-discharge efficiency of the BESS; Secondly, in an optimization interval, the charging and discharging state of the energy storage system can only be one of the two, so it should also meet the capacity constraint, and the calculation mode is as follows: wherein and are binary variables representing the discharge / charge state of the BESS within the tth optimal interval; F t U,BESS and F t D,BESS represent the up and down frequency regulation capability of the BESS at the tth optimal interval, respectively. In the day-ahead scheduling, the SOC of the energy storage system in the termination optimization interval should be equal to the SOC in the initial optimization interval, and the calculation formula is as follows: wherein and BESS at the initial interval and the termination interval, respectively.

5. The power system dispatch optimization method based on graph neural network according to claim 4, characterized in that, Based on the hydrogen production and hydrogen storage sub-model and the energy storage sub-model, the flexible standby and planned output optimization scheme is constructed, so that the renewable energy can be consumed, thereby reducing the abandoned wind and light energy, and at the same time, the climbing and deep peak shaving pressure of the thermal power unit is relieved under the load fluctuation and frequency modulation demand; Secondly, the adjustable resource capacity information formed by the flexible standby and planned output optimization scheme is used as the safety boundary of real-time frequency allocation, thereby supporting the comprehensive optimization of system economy, safety and flexibility; The specific implementation mode is as follows: the power deviation between the adjustment power input and output is regarded as the minimum optimization target, and the specific description is as follows: wherein and Pτ and Pτ,actual represent the total regulated power input and actual power output of the τth, respectively, is the number of frequency regulation control intervals per service period; In frequency regulation, the generation rate constraint, power limit constraint, power balance constraint and adjustment direction constraint can be written as follows: wherein a slope representing the kth regulating resource, the kth regulating resource, -1 a total regulating power input and an actual power output, and a power instruction of the kth regulating resource in the tth time control interval, a downlink / uplink regulating capability; The space-time graph neural network used by the graph neural network and the prediction archive correction unit comprises an input layer, a graph convolution layer, a full connection layer, an LSTM layer and an output layer. In addition to input features, the input layer also comprises an adjacency matrix, which directly reveals the relationship between different input features. For graph theory, the relationship between each resource in the frequency adjustment can be ensured to be a graph. The greater the weight value of the adjacency matrix, the higher the correlation between the two vertices. In real-time frequency adjustment optimization, the adjacency matrix is measured by using the similarity of the average frequency modulation capacity, the average frequency modulation mileage, the response delay score and the slope score of each unit in the day-ahead scheduling. The construction of the weight adjacency matrix is as follows: When performing the forward propagation of the graph convolution layer, the feature output is calculated by using the weight parameter, the adjacency matrix and the degree matrix. The calculation formula is as follows: where H in and H out denote the input and output characteristics of the graph convolution layer; I is a unit matrix; A and represent the distance matrix, the adjacency matrix and the adjacency matrix with self-connection, respectively; represents the element of the kth row and jth column of the adjacency matrix with self-connection; W c and B c represent the network weight and the bias value of the graph convolution layer, respectively; The program of the new frequency adjustment task can be directly generated from the total adjustment power instruction, the mileage historical data and the adjustment capacity based on the trained space-time graph neural network. The specific mode is as follows: wherein represents the input feature of the spatio-temporal graph neural network for real-time frequency regulation in the τth dimension, respectively represent the historical sequence of total power instruction, mileage instruction, operating power, upper limit adjustment capacity and lower limit adjustment capacity, is a 1×K vector, and other input features are w×K matrices, w is the length of the historical sequence; is the optimal scheduling scheme; F GCN-LSTM represents the trained spatio-temporal graph neural network, and B is the optimal weight and bias set; is the modified power scheme of the adjustment resource; f od represents the optimal function considering the Euclidean distance.

Images