Virtual power plant frequency modulation transaction optimization method and system based on distributed robust model predictive control

Abstract

This invention provides a method and system for optimizing frequency regulation trading in virtual power plants (VPPs) based on distributed robust optimization model predictive control, belonging to the field of distributed robust optimization technology for VPPs. By combining spatiotemporal graph neural networks to dynamically model the correlation of multi-source energy, a robust optimization framework incorporating uncertainty constraints is constructed to achieve real-time adjustment of reserve capacity and maximize economic efficiency. Simultaneously, a quantum genetic algorithm is introduced for global optimization of wind, solar, and energy storage coordinated scheduling, and robustness trade-off coefficients are dynamically adjusted through reinforcement learning to adapt to changes in market signals. Ultimately, this method aims to improve the reserve capacity coverage and operating cost efficiency of VPPs in frequency regulation trading, ensuring a balance between high flexibility, strong robustness, and commercial value in complex and uncertain environments.

Description

Technical Field

[0001] This invention belongs to the field of distributed robust optimization technology for virtual power plants, specifically relating to a method and system for optimizing frequency regulation transactions in virtual power plants based on distributed robust model predictive control. Background Technology

[0002] With the rapid increase in installed capacity of new energy sources, the stability of power system frequency faces severe challenges. Traditional centralized unit frequency regulation models can no longer meet the dual requirements of rapid response and economic efficiency. Virtual Power Plants (VPPs), by aggregating distributed power sources, energy storage, and adjustable loads, have the potential to participate in the frequency regulation ancillary services market. However, VPPs still face many challenges in frequency regulation trading. For example, the uncertainty of renewable energy output makes it difficult to accurately determine frequency regulation capacity and pricing; frequent market price fluctuations make traditional static pricing models unable to balance returns and risks; prediction errors and time lags exist in actual frequency regulation execution, reducing frequency regulation performance; conventional models have fixed predictive control time steps and single feedback, resulting in low efficiency and insufficient accuracy in real-time optimization calculations. Therefore, a new method is needed to achieve virtual power plant frequency regulation trading optimization and real-time adaptive control under uncertain conditions.

[0003] To address these challenges, existing optimization methods for virtual power plant frequency regulation trading using Bluebar model predictive control primarily rely on uncertainty modeling, dynamic optimization objectives, and algorithm design. First, by generating scenarios and employing a Bluebar optimization framework, the uncertainties of renewable energy output, load demand, and market prices are modeled, constructing constraints that include worst-case distributions. Second, optimization objectives typically include economic efficiency (e.g., minimizing operating costs, maximizing revenue) and robustness (e.g., reserve capacity coverage). Some studies achieve a balance between these two by dynamically adjusting tradeoff coefficients. Finally, algorithm design combines spatiotemporal graphical neural networks (ST-GNN) and multi-objective optimization techniques (e.g., quantum genetic algorithms, particle swarm optimization), and incorporates real-time market signals (e.g., electricity price volatility) for online learning and adjustment to improve model adaptability and computational efficiency.

[0004] Significant progress has been made in current research, but technical challenges remain. Traditional methods suffer from insufficient reserve capacity or reduced economic efficiency when dealing with high-uncertainty scenarios. For example, DRMPC-based methods significantly improve the robustness and economy of VPP frequency regulation by introducing spatial difference compensation terms and dynamic robustness adjustment mechanisms. However, these methods have high computational complexity (e.g., second-order cone programming or semidefinite programming for demand solutions), and most schemes rely on static electricity price models, making it difficult to flexibly handle real-time market game dynamics. Furthermore, the real-time performance and multi-source data collaboration capabilities of the algorithms still need further optimization to adapt to the dynamic requirements of large-scale VPP systems. Existing solutions mainly include three core paths: (i) The spatiotemporal correlation of wind and solar power is extracted by using a spatiotemporal graph neural network, and robust constraints are constructed by combining a stochastic optimization framework to achieve coordinated optimization of reserve capacity and economy; (ii) By dynamically adjusting the robustness trade-off coefficients through a deep Q-network (DQN), the system can be updated in real time with market price fluctuations, thereby improving its responsiveness. (iii) Incorporate physical nodes such as wind farms and photovoltaic power stations into a dynamic game framework (such as Nash equilibrium), achieve the global optimal solution through multi-objective particle swarm optimization (MOPSO) or quantum genetic algorithm (QGA), and enhance transaction transparency by combining with blockchain technology. These solutions optimize VPP frequency regulation performance from the perspectives of prediction accuracy, dynamic response, and collaborative control, respectively.

[0005] The investigation revealed several shortcomings in existing virtual power plant frequency regulation trading optimization technologies. For example, there is a lack of a unified framework for balancing economic efficiency and robustness; stochastic optimization prioritizes profitability, while robust optimization prioritizes safety, failing to address both simultaneously. Frequency regulation pricing lacks dynamic adaptability; most models do not correct for real-time forecast errors after determining the price a day-ahead. Real-time control response is insufficient; traditional MPC control has a fixed step size and singular feedback, making it unable to cope with multi-source uncertain disturbances.

[0006] Existing methods are mostly based on confidence intervals on a single time scale, making it difficult to capture the correlation between geographically distributed differences (such as wind field location and topographical influence) and prediction errors, leading to unreasonable reserve capacity allocation. Furthermore, traditional centralized scheduling strategies lag in responding to market fluctuations and equipment anomalies, failing to dynamically adjust resource allocation weights to adapt to real-time demand changes. In existing research, robustness indicators are mostly manually set, lacking the ability to adaptively adjust to dynamic market environments (such as fluctuations in electricity and gas prices), making it difficult to achieve synergistic optimization of cost minimization and system robustness. Summary of the Invention

[0007] The purpose of this invention is to overcome the performance limitations of traditional frequency regulation strategies in handling scenarios such as wind and solar power output fluctuations, load uncertainties, and electricity price fluctuations. It provides a virtual power plant frequency regulation trading optimization method and system based on a split-blown robust model predictive control. By integrating a spatiotemporal graphical neural network (ST-GNN) with a multi-objective robust optimization framework, and combining a dynamic weight adjustment mechanism with a multi-agent game strategy, it achieves accurate quantification of wind and solar forecasting errors and synergistic optimization of system robustness and economy.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a virtual power plant frequency regulation trading optimization method based on split-blown bar model predictive control, comprising the following steps: Collect wind and solar data and geospatial parameters from wind farms and photovoltaic stations; Based on the collected wind and light data and geospatial parameters, a spatiotemporal graph neural model is constructed and trained to generate a spatiotemporal confidence interval for rotating reserve capacity. A dual-objective optimization model is established with the day-ahead total operating cost and robustness index as optimization objectives. The spatiotemporal confidence interval of the generated spinning reserve capacity is used as a constraint input into the constructed dual-objective optimization model to generate the day-ahead scheduling scheme. Based on the generated day-ahead scheduling scheme, wind farms and photovoltaic stations are divided into multiple agents corresponding to the three markets of electricity, gas and heat. Under the operating boundary determined by the day-ahead scheduling scheme, multi-agent intraday real-time game optimization is performed to generate intraday real-time scheduling instructions. An adaptive step size adjustment mechanism and a multi-timescale collaborative optimization interface are constructed to dynamically manage the intraday real-time game optimization process and generate frequency adjustment and energy trading instructions. Based on the generated frequency regulation and energy trading instructions, the system outputs time-period output plans and frequency regulation capacity quotations for the corresponding thermal power units, combined heat and power plants, electric boilers, electric-to-gas equipment, and thermal storage tanks for electricity, gas, and heat intelligent agents.

[0009] The method for constructing and training a spatiotemporal graph neural model based on collected wind and solar data and geospatial parameters to generate the spatiotemporal confidence interval of rotating reserve capacity is as follows: A spatiotemporal graph neural network model is constructed using wind farms and photovoltaic stations as graph nodes and geospatial parameters and the correlation between adjacent equipment as edge weights. Based on the collected historical wind and solar data and geospatial parameters, a spatiotemporal graph neural network model is trained to generate a joint spatiotemporal confidence band for prediction errors, thereby obtaining the upper and lower limits of prediction errors for wind power and photovoltaic power at each time point. Based on the upper and lower limits of the prediction errors for wind power and photovoltaic power at each time point, the spatial difference compensation term is calculated, and combined with the confidence level parameter, the spatiotemporal confidence interval of the spinning reserve capacity is generated.

[0010] The formula for calculating the spatiotemporal confidence interval of spinning reserve capacity based on the upper and lower limits of the prediction errors for wind and solar power at each time point, combined with the confidence level parameters, is expressed as follows: (1) (2) (3) Among them, R down R(t) represents the minimum reserve capacity with the upper limit of prediction error at time t. up (t) represents the minimum reserve capacity at time t, which is the lower limit of the error. This represents the upper limit of the wind power forecast value at time t. The upper limit of the photovoltaic forecast value at time t. This represents the lower limit of the wind power prediction value at time t. This represents the lower limit of the photovoltaic prediction value at time t. This represents the spatial difference compensation term. and , respectively, represent the historical average prediction error of the k-th wind farm or photovoltaic power station. and Let K be the actual output deviation of the k-th device at time t, where K is the number of devices.

[0011] The method for establishing a bi-objective optimization model with the day-ahead total operating cost and robustness index as optimization objectives, and using the generated spatiotemporal confidence interval of the spinning reserve capacity as a constraint input to the constructed bi-objective optimization model to generate the day-ahead scheduling scheme is as follows: A dual-objective optimization model is established with the day-ahead total operating cost and robustness index as optimization objectives. The day-ahead total operating cost includes electricity, gas, and heat dispatch costs as well as carbon trading costs. The generated spatiotemporal confidence interval of the spinning reserve capacity is input as a constraint into the bi-objective optimization model. A hybrid quantum genetic algorithm is used to solve the established bi-objective optimization model. The hybrid quantum genetic algorithm integrates quantum genetic algorithm and dynamic Pareto front search strategy to generate a day-ahead scheduling scheme that takes into account both economy and robustness. The scheme includes spinning reserve capacity configuration and 24-hour power generation, energy storage and frequency regulation plans.

[0012] The robustness index is embodied by a robustness evaluation function, the formula of which is as follows:

[0013] in, These are dynamic weighting coefficients; This represents the worst-case opportunity cost or penalty cost at time t, related to system uncertainty. This represents the opportunity cost or penalty cost in the best-case scenario related to system uncertainty at time t.

[0014] The generated day-ahead scheduling scheme divides wind farms and photovoltaic stations into multiple agents corresponding to the electricity, gas, and heat markets. Under the operational boundaries determined by the day-ahead scheduling scheme, the step of generating day-ahead real-time scheduling instructions through multi-agent intraday real-time game optimization is as follows: The optimal strategy among agents is learned through a deep Q-network. The state space is defined as the current wind and solar power output, load demand, market price and equipment status, the action space is the equipment output adjustment amount, and the reward function is the weighted sum of the day-ahead total operating cost and robustness. Under the boundary constraints determined by the current scheduling scheme, multiple agents engage in real-time multi-agent game optimization through continuous interaction and trial and error with the environment. The joint actions generated after the strategies of each agent converge constitute the system's real-time, coordinated, and near-optimal intraday real-time scheduling instructions.

[0015] In the steps of constructing an adaptive step size adjustment mechanism and a multi-timescale collaborative optimization interface, dynamically managing the intraday real-time game optimization process, and generating frequency regulation and energy trading instructions, the adaptive step size adjustment mechanism generates a comprehensive decision index based on real-time wind and solar power output error and operating cost error, and dynamically adjusts the rolling optimization time domain length and optimization trigger frequency accordingly. The multi-timescale collaborative optimization interface is used to pass the dynamic weight coefficients of the robustness evaluation function. Spatial difference compensation item The weights of the reward functions among the multiple agents are updated based on market signals, and the agents' game strategies are corrected based on the backtracking day-ahead plan deviations.

[0016] The specific method of the adaptive step size adjustment mechanism is as follows: Obtain real-time power output data from wind farms and solar power stations; The real-time output data of wind farms and photovoltaic stations are compared with the predicted values ​​of the generated intraday real-time dispatch instructions to obtain the real-time wind and solar output error; at the same time, based on the actual operating status and market prices, the current actual operating cost is calculated and compared with the expected cost in the dispatch plan to obtain the operating cost error. A comprehensive decision-making index is calculated based on real-time output error and operating cost error. Based on comprehensive decision indicators and preset coefficients of variation, the change in time step is calculated and the length of the rolling time domain is dynamically adjusted to achieve a balance between accuracy and efficiency.

[0017] The specific working method of the multi-timescale collaborative optimization interface is as follows: Dynamic weight coefficients of the robustness evaluation function based on multi-timescale collaborative optimization interface transmission Spatial difference compensation item ; Based on market signals, dynamic weight parameters are generated and passed to the reward function of the game agent to update the weights of the reward functions among the multiple agents, thereby adjusting the agent's behavioral priority and strategy. By reviewing the deviations of the previous optimization scheme, and combining the adjusted behavioral priorities and strategies of the agent, the game strategy is corrected. Based on the updated reward function and the revised game strategy, the system drives multiple agents to perform real-time game optimization within a dynamically adjusted time step, generating and outputting frequency modulation and energy trading instructions.

[0018] Secondly, the present invention provides a virtual power plant frequency regulation trading optimization system based on sub-Brooker model predictive control, comprising: The data acquisition module is used to collect wind and solar data and geospatial parameters from wind farms and photovoltaic stations; The spatiotemporal confidence interval generation module is used to build and train a spatiotemporal graph neural model based on collected wind and light data and geospatial parameters, and generate the spatiotemporal confidence interval of rotating reserve capacity. The day-ahead scheduling scheme generation module is used to establish a bi-objective optimization model with the day-ahead total operating cost and robustness index as optimization objectives. The generated spatiotemporal confidence interval of the spinning reserve capacity is used as a constraint input to the constructed bi-objective optimization model to generate the day-ahead scheduling scheme. The intraday real-time dispatch instruction generation module is used to divide wind farms and photovoltaic stations into multiple agents corresponding to the three markets of electricity, gas and heat based on the generated day-ahead dispatch scheme. Under the operating boundary determined by the day-ahead dispatch scheme, the module performs intraday real-time game optimization of the multiple agents to generate intraday real-time dispatch instructions. The frequency modulation and energy trading instruction generation module is used to construct an adaptive step size adjustment mechanism and a multi-timescale collaborative optimization interface, dynamically manage the intraday real-time game optimization process, and generate frequency modulation and energy trading instructions. The output and execution module is used to output the time-period output plan and frequency regulation capacity quotation of the corresponding thermal power units, combined heat and power plants, electric boilers, electric-to-gas equipment and thermal storage tanks based on the generated frequency regulation and energy trading instructions.

[0019] Compared with the prior art, the present invention has the following beneficial effects: This invention provides a virtual power plant frequency regulation trading optimization method based on a split-blown model predictive control. It accurately predicts the spatiotemporal confidence interval of wind and solar power output through a spatiotemporal graph neural network, providing a more realistic rotating reserve constraint for day-ahead optimization. This avoids the excessive or insufficient reserves caused by traditional deterministic or simple interval predictions, thereby reducing reserve costs and lowering total operating costs while ensuring safety. The constructed bi-objective optimization model explicitly considers system robustness and incorporates a high-reliability spatiotemporal reserve confidence interval as a constraint, enabling the generated day-ahead scheduling scheme to possess strong anti-interference capabilities to cope with wind and solar uncertainties. The wind and solar power plants are divided into intelligent agents corresponding to the electricity, gas, and heat markets, and various flexible components such as combined cogeneration, electricity-to-gas conversion, electric boilers, and thermal storage are introduced to construct a true multi-energy market game environment. This facilitates the exploration of the conversion, storage, and complementarity potential between different energy networks, forming a spatiotemporal multi-dimensional regulation capability. The constructed two-level architecture of "day-ahead optimization - intraday real-time game," and the multi-timescale collaborative optimization interface connecting the two, achieves effective coordination between long-term planning and short-term adjustments, and between centralized optimization and distributed decision-making. The current solution sets a safety boundary for intraday game theory, and the real-time information from intraday game feedback can provide data correction for future predictions and optimizations, forming a closed-loop optimization. The final output, consisting of time-period output plans and frequency regulation capacity quotations for diverse equipment such as thermal power units, combined heat and power plants, electric boilers, power-to-gas conversion equipment, and thermal storage tanks, constitutes a complete dispatch instruction set covering energy and ancillary services, spanning multiple systems including electricity, gas, and heat. This solution, through deep integration of advanced forecasting technologies, multi-objective robust optimization, multi-agent game theory, and multi-timescale collaboration, constructs an adaptive, highly flexible, and cost-effective multi-energy system dispatch framework. It not only effectively addresses the challenges of the randomness and volatility of new energy sources but also releases the value of diverse and flexible resources through market mechanisms, providing a key technical path and practical solution for building a clean, low-carbon, safe, and controllable new energy system.

[0020] Furthermore, this invention constructs a spatiotemporal graph neural network to accurately model the spatial correlation between multiple distributed wind farms and photovoltaic stations by calculating spatial difference compensation terms. This solves the problem that traditional ARIMA time series methods cannot capture spatial heterogeneity, and significantly improves the accuracy of wind and solar power output prediction and reserve capacity configuration.

[0021] Furthermore, the HQGA hybrid quantum genetic algorithm achieves a Pareto optimal balance among multiple objectives such as operating cost, system robustness, renewable energy utilization, and carbon emissions during the day-ahead optimization phase by dynamically adjusting the weight coefficients. Compared with the traditional NSGA-II genetic algorithm, it has a faster convergence speed and stronger global optimization ability, effectively reducing multiple operating costs such as thermal power fuel costs, carbon trading costs, and equipment start-up and shutdown costs.

[0022] Furthermore, DQN deep reinforcement learning enables real-time adaptive decision-making for the proportion of thermal power output. Through continuous interactive learning with the environment, it has stronger intelligence and flexibility compared to the fixed rules of traditional MPC model predictive control. It significantly reduces the operation and maintenance costs of renewable energy, energy storage loss costs and ancillary service fees, while shortening the system response time and improving the control effect of wind and solar curtailment rates.

[0023] Furthermore, the adaptive step size adjustment mechanism dynamically optimizes the reserve capacity configuration for the next moment based on the prediction error feedback, avoiding the waste of resources due to excessive reservation or the threat of power supply problems caused by insufficient reservation due to the fixed margin in the traditional solution, thus further improving the robustness and economy of the system. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the uncertainty modeling and confidence interval generation process in this invention; Figure 3 This is a schematic diagram of the intraday optimization process in this invention; Figure 4 This is a flowchart of the adaptive step size and multi-scale coordination process in this invention; Figure 5 This is a schematic diagram comparing the operating costs in Embodiment 2 of the present invention; Figure 6 This is a schematic diagram comparing the robustness of the system in Embodiment 2 of the present invention; Figure 7 This is a comparison chart of renewable energy utilization rates in Embodiment 2 of the present invention; Figure 8 This is a comparison chart of carbon emissions in Example 2 of the present invention; Figure 9 This is a comparison chart of rotating reserve capacity in Embodiment 2 of the present invention; Figure 10 This is a radar chart comparing the overall performance in Embodiment 2 of the present invention. Detailed Implementation

[0025] To further understand the content of this invention, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments are merely illustrative and not limiting of the invention.

[0026] Example 1 like Figure 1 As shown, a virtual power plant frequency regulation trading optimization method based on split-blown bar model predictive control includes the following steps: S1: Collect wind and solar data and geospatial parameters from wind farms and photovoltaic stations; S2: Based on the collected wind and light data and geospatial parameters, a spatiotemporal neural model is constructed and trained to generate the spatiotemporal confidence interval of the rotating reserve capacity; S3: Establish a bi-objective optimization model with the day-ahead total operating cost and robustness index as optimization objectives. Input the generated spatiotemporal confidence interval of the spinning reserve capacity as a constraint into the constructed bi-objective optimization model to generate the day-ahead scheduling scheme. S4: Based on the generated day-ahead scheduling scheme, wind farms and photovoltaic stations are divided into multiple agents corresponding to the three markets of electricity, gas and heat. Under the operating boundary determined by the day-ahead scheduling scheme, multi-agent intraday real-time game optimization is performed to generate intraday real-time scheduling instructions. S5: Construct an adaptive step size adjustment mechanism and a multi-timescale collaborative optimization interface to dynamically manage the intraday real-time game optimization process and generate frequency adjustment and energy trading instructions; S6: Based on the generated frequency regulation and energy trading instructions, output the time-by-time output plan and frequency regulation capacity quotation for the corresponding thermal power units, combined heat and power plants, electric boilers, electric-to-gas equipment and thermal storage tanks of the power, gas and heat intelligent agents.

[0027] Specifically, in S1, wind and solar data, as well as geospatial parameters, are collected from wind farms and photovoltaic stations. The wind and solar data include historical forecast data, actual power output data, wind field distribution, and regional differences in solar irradiance. The geospatial parameters include the altitude of wind farms and photovoltaic stations.

[0028] Specifically, in S2, a spatio-temporal graph neural network (ST-GNN) model is constructed and trained based on collected historical landscape data and geospatial parameters. The specific method is as follows: S21: Using wind farms and photovoltaic stations as graph nodes and geospatial parameters and the correlation between adjacent equipment as edge weights, construct a spatiotemporal graph neural network model; S22: Based on the collected historical wind and solar data and geospatial parameters, train a spatiotemporal graph neural network model to generate a joint spatiotemporal confidence band for prediction errors, and obtain the upper limit and lower limit of prediction errors for wind power and photovoltaic power at each time. S23: Based on the obtained upper and lower bounds of prediction errors for wind and solar power at each time point, calculate the spatial difference compensation term, and combine it with the confidence level parameters to generate the spatiotemporal confidence interval for the spinning reserve capacity. The specific formula is expressed as follows: (1) (2) Among them, R up(t) represents the upward spin-up reserve capacity (i.e., the minimum reserve capacity to cope with the lower bound of wind and solar power forecast errors) that the system needs to configure at time t to cope with the uncertainty of renewable energy output being lower than expected; R down (t) represents the downspin-off reserve capacity that the system needs to configure at time t to cope with the uncertainty of renewable energy output exceeding expectations (i.e., the minimum reserve capacity to cope with the upper limit of prediction error).

[0029] This represents the upper limit of the wind power forecast at time t, i.e., the maximum forecast error that the actual wind power output may reach at that time point. Specifically, This represents the positive deviation of the actual wind power output from the predicted value, reflecting the potential over-generation under favorable wind and solar conditions (such as wind speed exceeding the predicted value).

[0030] : Represents the upper limit of the photovoltaic prediction value at time t, that is, the maximum prediction error that the actual photovoltaic output may reach at that time point.

[0031] : Represents the lower limit of the wind power forecast at time t, i.e., the minimum forecast error that the actual wind power output may be lower than the forecast value. represents the negative deviation of the actual wind power output from the forecast value, reflecting the impact of extreme situations (such as insufficient wind speed) on the system's supply and demand balance.

[0032] This represents the lower limit of the photovoltaic (PV) forecast at time t, i.e., the minimum prediction error that the actual PV output may be lower than the forecast. This parameter is used to quantify the impact of adverse factors such as insufficient sunlight on system stability.

[0033] : Represents the spatial variability compensation term, used to correct for the neglect of geographical location differences of equipment in traditional single-time-scale confidence intervals. Its calculation is as follows: (3) in and , respectively, represent the historical average prediction error of the k-th wind farm or photovoltaic power station. and Let K be the actual output deviation of the k-th device at time t, where K is the number of devices. This compensation term improves the accuracy of robust interval generation by capturing the impact of geographical location differences (such as terrain, climate zone, and device layout) on prediction errors.

[0034] Preferably, the Transformer model can be used instead of ST-GNN, and the long-term dependence and spatial correlation of wind and solar power output can be captured through the self-attention mechanism to generate customized error confidence intervals.

[0035] Specifically, in S3, a bi-objective optimization model is established with the day-ahead total operating cost and robustness index as optimization objectives. The generated spatiotemporal confidence interval of the spinning reserve capacity is used as a constraint input to the constructed bi-objective optimization model to generate the day-ahead scheduling scheme. The details are as follows: A bi-objective optimization model is established with day-ahead total operating cost and robustness index as optimization objectives. The day-ahead total operating cost includes electricity, gas, and heat dispatch costs as well as carbon trading costs. The robustness evaluation function is defined by the following formula: (4) in, This is a dynamic weighting coefficient that is adjusted based on real-time market volatility. This represents the opportunity cost or penalty cost in the worst-case scenario related to system uncertainty at time t. This represents the opportunity cost or penalty cost in the best-case scenario at time t, related to system uncertainty. In a bi-objective optimization model, It is typically used as an objective function to measure the robustness of a system. The structure of this formula aims to maximize the robustness benefit of the system or minimize its robustness cost.

[0036] Furthermore, based on the generated spatiotemporal confidence interval of the spinning reserve capacity, it is input as a constraint into the bi-objective optimization model. The established bi-objective optimization model is solved using a hybrid quantum genetic algorithm (HQGA). The hybrid quantum genetic algorithm integrates the quantum genetic algorithm (QGA) and the dynamic Pareto front search strategy (DPFSS) to generate a day-ahead dispatch scheme that balances economy and robustness. This scheme includes the spinning reserve capacity configuration and 24-hour power generation, energy storage and frequency regulation plans, including unit start-up and shutdown plans, power base point and the spatiotemporal confidence interval of reserve capacity as rigid operating boundaries.

[0037] Specifically, in S4, based on the generated day-ahead scheduling scheme, wind farms and photovoltaic stations are divided into multiple agents corresponding to the three markets of electricity, gas and heat. Under the operating boundary determined by the day-ahead scheduling scheme, multi-agent intraday real-time game optimization is performed to generate real-time scheduling instructions.

[0038] The optimization objectives of the three market agents are as follows: Electricity market: maximizing frequency regulation revenue and minimizing start-up and shutdown losses.

[0039] Gas Market: Balancing the costs of electricity-to-gas conversion with the benefits of carbon sequestration.

[0040] Heat market: Optimize the charging and discharging strategy of thermal storage tanks while meeting heat load demand.

[0041] Furthermore, the optimization method for intraday real-time multi-agent game is as follows: The optimal policy among agents is learned through a Deep Q-Network (DQN). The state space (S) is defined as the current wind and solar power output, load demand, market price and equipment status, the action space (A) is the equipment output adjustment amount, and the reward function (R) is the weighted sum of the day-ahead total operating cost and robustness.

[0042] The unit start-up and shutdown plans, power baselines, and spatiotemporal confidence intervals of reserve capacity determined by the current scheduling scheme serve as rigid operating boundaries. Under these boundary constraints, multiple DQN agents continuously interact with the environment and engage in trial and error to optimize the system through real-time multi-agent game theory. The joint actions generated after the strategies of each agent converge constitute the system's real-time, coordinated, and near-optimal intraday real-time scheduling instructions.

[0043] Specifically, in S5, an adaptive step size adjustment mechanism and a multi-timescale collaborative optimization interface are constructed to dynamically manage the intraday real-time game optimization process.

[0044] 1) The adaptive step size adjustment mechanism generates a comprehensive decision index based on real-time wind and solar power output error and operating cost error, and dynamically adjusts the rolling optimization time domain length and optimization trigger frequency accordingly. Details are as follows: Obtain real-time power output data from wind farms and solar power stations; The real-time output data of wind farms and photovoltaic stations are compared with the predicted values ​​of the generated intraday real-time dispatch instructions to obtain the real-time wind and solar output error; at the same time, based on the actual operating status and market prices, the current actual operating cost is calculated and compared with the expected cost in the dispatch plan to obtain the operating cost error. A comprehensive decision-making index is calculated based on real-time output error and operating cost error.

[0045] in, As a comprehensive decision-making indicator, As the first weighting coefficient, This is the predicted power value. This is the actual value of the power. This is the second weighting coefficient. This is a predicted value for operating costs. This represents the actual value of the operating cost. and It is dynamically adjusted based on the spatiotemporal error distribution.

[0046] Based on comprehensive decision indicators and preset coefficients of variation, the change in time step is calculated and the length of the rolling time domain is dynamically adjusted to achieve a balance between accuracy and efficiency. The specific formula is as follows:

[0047]

[0048] in, The change in time step, The preset coefficient of variation, The adjusted rolling time domain length, The reference length is .

[0049] 2) The multi-timescale collaborative optimization interface is used to pass the dynamic weight coefficients of the robustness evaluation function. Spatial difference compensation item The weights of the reward functions among the multiple agents are updated based on market signals, and the agents' game strategies are corrected based on the backtracking day-ahead plan deviations.

[0050] Dynamic weight coefficients of the robustness evaluation function based on multi-timescale collaborative optimization interface transmission Spatial difference compensation item ; Based on market signals (such as electricity price fluctuations and carbon price changes), dynamic weight parameters are generated and transmitted to the reward function of the game agents to update the weights of the reward functions among the multiple agents, thereby adjusting the behavioral priorities and strategies of the agents; for example, frequency regulation of thermal storage tanks is prioritized during periods of high electricity prices, and carbon sequestration mechanism is triggered during periods of low wind power. By reviewing the deviations of the previous optimization scheme, and combining the adjusted behavioral priorities and strategies of the agent, the game strategy is corrected. Based on the updated reward function and the revised game strategy, the system drives multiple agents to perform real-time game optimization within a dynamically adjusted time step, generating and outputting frequency modulation and energy trading instructions.

[0051] Preferably, the cooperative strategy between agents can be trained using a policy gradient algorithm to further improve the efficiency of real-time game playing.

[0052] Specifically, in S6, based on the generated frequency regulation and energy trading instructions, the system outputs the time-period output plans and frequency regulation capacity quotations for the corresponding thermal power units, combined heat and power plants, electric boilers, electric-to-gas equipment, and thermal storage tanks for electricity, gas, and heat. The system then executes these plans, uploads the scheduling scheme to the power grid dispatch center, and triggers the ancillary service market trading process to ensure consistency between actual scheduling and market mechanisms.

[0053] Example 2 This invention establishes a simulation model for virtual power plant frequency regulation trading, with a simulation cycle of 24 hours. It primarily simulates the output characteristics of various energy sources, including wind power, photovoltaic power, thermal power, and energy storage, while considering multi-dimensional factors such as market electricity prices, carbon emissions, and load fluctuations. The system operates in a Python-based environment. The performance of the reinforcement learning (DQN) algorithm is compared with that of the traditional optimization algorithm (ARIMA+NSGA-II+MPC) to verify the comprehensive performance improvement effect of the reinforcement learning method in frequency regulation trading. Some specific simulation data are presented below.

[0054] In the simulation environment, the benchmark electricity price is set at 0.12 yuan / kWh, and the gas price is set at 0.05 yuan / kWh to characterize the fuel cost of thermal power units; the carbon emission price is set at 50 yuan / ton CO2 to calculate carbon trading costs. The simulation duration is 24 hours, with a time resolution of 1 hour. The virtual power plant consists of multiple energy sources, including three wind turbines and two photovoltaic power plants. The rated capacities of the wind turbines are 100MW, 80MW, and 120MW, respectively; the photovoltaic power plants have capacities of 60MW and 50MW, respectively; the rated capacity of the thermal power units is 150MW, used for primary frequency regulation and compensation tasks; and the energy storage unit has a capacity of 50MW, mainly used for auxiliary frequency regulation and backup support.

[0055] The following diagrams present the results of the method described in this invention, based on the set simulation data. In the operation of a virtual power plant, cost is a significant factor, as it directly impacts the final economic benefits. For example... Figure 5 As shown, this solution, through comprehensive optimization using DQN reinforcement learning, achieved a significant reduction of 15.2% in total operating costs over a 24-hour operating cycle. This cost includes six major cost items: thermal power fuel (coal), carbon emission trading, equipment start-up and shutdown, renewable energy operation and maintenance, energy storage losses, and ancillary services, providing a more comprehensive reflection of the real operating expenses of the virtual power plant. The curves show that the total cost of both solutions reaches 8,000-9,000 yuan during peak electricity consumption periods (0-2 am and 6-10 pm), dropping to 1,000-3,000 yuan during the midday off-peak period (8-12 pm), demonstrating a significant peak-to-valley difference. The blue curve of this solution consistently remains below the red curve of the traditional solution, thanks to DQN's targeted optimization of each cost item: a 5% increase in thermal power fuel efficiency, a 12% reduction in carbon costs, a 35% reduction in start-up and shutdown frequency, a 20% reduction in renewable energy operation and maintenance costs, a 33% reduction in energy storage losses, and a 17% reduction in ancillary service costs. This synergistic effect results in daily savings of several thousand yuan in operating costs, and annual savings of over one million yuan.

[0056] like Figure 6The robustness comparison shows that the robustness index of this scheme remains stable at a high level of 78-80%, which is about 37 percentage points higher than the fixed 42.5% of the traditional scheme, representing a performance advantage of nearly double. The robustness index reflects the system's ability to maintain stable operation when facing uncertainties such as wind and solar power output fluctuations, load surges, and equipment failures; a higher value indicates a more reliable system. This scheme uses an ST-GNN spatiotemporal graph neural network to accurately model the spatial differences between multiple distributed wind farms and photovoltaic stations. By dynamically adjusting the reserve capacity coverage ratio (alpha_upper=1.15 and alpha_lower=1.08) and the weight coefficient ξ (which changes in real time according to market electricity price volatility), the system can adaptively cope with various disturbances. Traditional solutions use static robustness parameters (gamma=0.85, 42.5% after normalization), which cannot adjust strategies according to real-time operating status. This can easily lead to insufficient power supply or excessive backup when wind and solar power fluctuate drastically. In contrast, the high robustness of this solution ensures that the virtual power plant can operate safely and stably under various extreme conditions, significantly reducing the risk of power outages and emergency response costs.

[0057] Then compare the utilization rate of new energy sources, such as Figure 7 The proposed scheme increases the renewable energy utilization rate to 97.8%, a 5.6 percentage point improvement compared to the traditional scheme's 92.2%. This means that the wind and solar curtailment rate is significantly reduced from 8.5% to 2.3%, achieving near-complete absorption of clean energy. Both curves remain nearly horizontal and stable over 24 hours because the virtual power plant mitigates the volatility of wind and solar output through various means such as energy storage systems, thermal power peak shaving, and demand-side response. The blue line in this scheme consistently remains higher because the DQN agent has learned the optimal thermal power output ratio control strategy through thousands of simulations, maximizing the absorption of clean energy such as wind and solar while ensuring power supply reliability. Specifically, DQN's five actions (thermal power output ratio of 0.6, 0.7, 0.8, 0.9, and 1.0) can be dynamically selected based on real-time wind and solar forecasts, load demand, electricity price signals, and historical costs, avoiding the dilemma of traditional schemes that either over-rely on thermal power leading to wind and solar curtailment or excessively reducing thermal power leading to power supply risks. Every 1 percentage point increase in utilization rate is equivalent to reducing tens of megawatt-hours of clean energy waste per day. This not only improves economic efficiency (reducing electricity purchase costs) but also aligns with carbon neutrality policy guidelines, which is of great significance for achieving the carbon peak in 2030 and carbon neutrality in 2060.

[0058] exist Figure 8The diagram shows a comparison of the carbon emission costs of this invention with traditional methods. This solution reduces carbon emission costs by 17.0%, and with a carbon price of 50 yuan / ton, daily savings in carbon trading fees range from several hundred to several thousand yuan. Unlike the total operating cost diagram, this diagram only shows the CO2 trading fees generated by carbon emissions and does not include other costs such as fuel, start-up, shutdown, and maintenance. Therefore, the value ranges from 1,000 to 9,000 yuan, accounting for approximately 30-40% of the total operating cost. The curve shows a high correlation between carbon costs and load demand: during peak electricity consumption periods (0-2 am and 18-22 pm), increased thermal power output leads to a surge in carbon emissions, causing carbon costs to soar to 7,000-9,000 yuan; during off-peak periods (8-12 pm), renewable energy is sufficient to meet load demand, thermal power output decreases or even drops to zero, and carbon costs fall to 1,000-2,000 yuan. The blue curve of this solution consistently falls below the red curve of the traditional solution, primarily due to two optimizations: First, the carbon cost was optimized by a factor of 0.88 (reducing it by 12%), simulating the effect of lowering the unit price of carbon trading through carbon trading market strategies (such as purchasing carbon allowances and participating in carbon financial derivatives). Second, DQN reduces the total output of thermal power plants through intelligent dispatch, thereby reducing total carbon emissions at the source (each reduction of 1 MWh of thermal power output is equivalent to a reduction of 0.8 tons of CO2 emissions). In the current market environment with a carbon price of 50 yuan / ton, this translates to daily savings of several hundred yuan in carbon costs, resulting in annual savings of hundreds of thousands of yuan. As carbon prices rise in the future (expected to reach 100-150 yuan / ton by 2030), the economic advantages of this solution will become even more significant, while also providing technical support for enterprises to fulfill their social responsibilities and cope with stricter carbon emission policies.

[0059] Figure 9 The rotating reserve capacity comparison chart shows that the reserve capacity configuration curve of this scheme exhibits dynamic characteristics that change in tandem with wind and solar power output and load demand, with configuration accuracy controlled within ±5%, while the configuration error of the traditional scheme is as high as ±12%, representing a 58% improvement in accuracy. Reserve capacity refers to the emergency power generation capacity reserved to cope with uncertain events such as wind and solar power output prediction errors, sudden equipment failures, and sudden load increases. Excessive configuration will lead to resource waste (idling losses of thermal power units), while insufficient configuration may lead to the risk of power outages.

[0060] As shown in the graph, the two curves reach a peak of 120-130MW during the midday period (8-12 pm). This is because photovoltaic output is highest during this period but also most volatile (affected by cloud cover), requiring more reserve capacity to smooth out fluctuations. At night and in the early morning (0-6 am and 18-24 pm), the output drops to 30-60MW, at which time power supply mainly relies on relatively stable wind and thermal power. This patented solution, through the spatial difference compensation term calculated by ST-GNN (improving efficiency by 35%), can accurately capture the spatial correlation between three wind farms and two photovoltaic stations. For example, when wind farm A's output decreases due to weather conditions, wind farm B, which is farther away, may maintain normal output. By modeling this spatial complementarity, the conservative assumption of "all power sources fluctuating simultaneously" in traditional solutions can be avoided. This patented solution also adopts an adaptive step size adjustment mechanism, which dynamically adjusts the reserve capacity at the next moment based on the prediction error feedback (error_ratio), realizing intelligent configuration of "increasing reserve when the error is large and decreasing reserve when the error is small", ultimately improving the configuration accuracy to 95% (compared to only 88% in traditional solutions). This ensures power supply reliability (reserve adequacy rate close to 100%) and avoids resource waste (reserve capacity is reduced by an average of 5-12% compared to traditional solutions).

[0061] Figure 10 The comprehensive performance radar chart uses a 0-100 score system to normalize the scores across five key dimensions. This solution outperforms the traditional solution (60, 43, 92, 65, and 30 points respectively) in all five dimensions: cost optimization (78 points), robustness (78 points), renewable energy utilization rate (98 points), carbon emission reduction (91 points), and response speed (51 points). The blue area completely envelops the red area, forming a significant performance advantage envelope effect. The overall score is 30-83% higher, fully demonstrating that this solution (ST-GNN+HQGA+DQN) is significantly superior to the traditional solution (ARIMA+NSGA-II+MPC) in multiple key indicators such as economy, stability, environmental protection, accuracy, and real-time performance. It represents a comprehensive technological breakthrough in the field of virtual power plant frequency regulation trading.

[0062] Table 1 illustrates the six core operating cost items of virtual power plant frequency regulation trading and their comparative optimization effects, comprehensively reflecting the technical advantages of this scheme (ST-GNN+HQGA+DQN) in cost control compared to the traditional scheme (ARIMA+NSGA-II+MPC). The thermal power fuel cost (0.35 yuan / MWh) represents the cost of coal consumption per megawatt-hour of electricity produced. This invention reduces this cost by 5% to 0.3325 yuan / MWh through DQN optimization of combustion efficiency. The carbon emission cost is the allowance payment for carbon dioxide generated by thermal power in the carbon trading market. This patent reduces this cost by 12% to 44 yuan / ton through HQGA carbon market strategy optimization and DQN emission reduction technology. The start-up and shutdown cost (50 yuan / time) covers equipment wear and tear, start-up fuel, and labor costs for each start-up or shutdown of the thermal power unit. This invention reduces the start-up and shutdown frequency by 35% through DQN predictive scheduling, lowering the cost to 32.5 yuan / time. The renewable energy operation and maintenance cost (0.025 yuan / MWh) includes the inspection, maintenance, and monitoring costs of wind and solar power turbines. This invention reduces this cost by 20% to 0.02 yuan / MWh through ST-GNN accurate prediction and preventative maintenance. The energy storage loss cost (0.015 yuan / MWh) reflects the efficiency loss, lifespan degradation, and thermal management energy consumption during battery charging and discharging. This invention reduces this cost by 33% to 0.01 yuan / MWh through DQN optimal charging and discharging strategies. The ancillary service cost (120 + 25 × sin(t) yuan) is the cost of providing services such as frequency and voltage regulation to the grid, and it fluctuates over time. This invention reduces the benchmark cost by 17% to 100 + 20 × sin(t) yuan through rapid response and high robustness.

[0063] The synergistic optimization of these six cost components creates a positive feedback loop: improved thermal power efficiency directly reduces carbon emissions, fewer start-ups and shutdowns extend equipment lifespan, precise renewable energy dispatch reduces maintenance requirements, and intelligent energy storage management reduces losses. As a result, overall system stability is improved, reducing reliance on ancillary services. Ultimately, the average daily total operating cost is reduced from approximately RMB 4,130 to RMB 3,500, achieving a comprehensive cost reduction of 15.2% and annual savings of approximately RMB 6.68 million. This fully demonstrates the enormous economic value of artificial intelligence technology in the multi-dimensional synergistic optimization of energy systems and provides a solid cost competitiveness foundation for the commercial operation of virtual power plants in a carbon-neutral context.

[0064] Table 1. Cost Item Comparison Chart

[0065] The entire simulation process comprises five stages. First, virtual data on wind and solar power, load, and electricity prices are generated, and spatiotemporal discrepancy compensation is calculated using ST-GNN. Then, a DQN agent is used for stepwise decision-making and training, followed by HQGA for adaptive optimization of price fluctuations. Subsequently, various performance indicators are calculated and compared with results from traditional algorithms. Finally, comparative charts are generated, including operating costs, carbon costs, robustness, utilization rate, reserve capacity, and overall performance, achieving full-process optimization of virtual power plant frequency regulation trading.

[0066] Example 3 A virtual power plant frequency regulation trading optimization system based on split-blown bar model predictive control includes: The data acquisition module is used to collect wind and solar data and geospatial parameters from wind farms and photovoltaic stations; The spatiotemporal confidence interval generation module is used to build and train a spatiotemporal graph neural model based on collected wind and light data and geospatial parameters, and generate the spatiotemporal confidence interval of rotating reserve capacity. The day-ahead scheduling scheme generation module is used to establish a bi-objective optimization model with the day-ahead total operating cost and robustness index as optimization objectives. The generated spatiotemporal confidence interval of the spinning reserve capacity is used as a constraint input to the constructed bi-objective optimization model to generate the day-ahead scheduling scheme. The intraday real-time dispatch instruction generation module is used to divide wind farms and photovoltaic stations into multiple agents corresponding to the three markets of electricity, gas and heat based on the generated day-ahead dispatch scheme. Under the operating boundary determined by the day-ahead dispatch scheme, the module performs intraday real-time game optimization of the multiple agents to generate intraday real-time dispatch instructions. The frequency modulation and energy trading instruction generation module is used to construct an adaptive step size adjustment mechanism and a multi-timescale collaborative optimization interface, dynamically manage the intraday real-time game optimization process, and generate frequency modulation and energy trading instructions. The output and execution module is used to output the time-period output plan and frequency regulation capacity quotation of the corresponding thermal power units, combined heat and power plants, electric boilers, electric-to-gas equipment and thermal storage tanks based on the generated frequency regulation and energy trading instructions.

[0067] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A virtual power plant frequency regulation trading optimization method based on split-blown bar model predictive control, characterized in that, Includes the following steps: Collect wind and solar data and geospatial parameters from wind farms and photovoltaic stations; Based on the collected wind and light data and geospatial parameters, a spatiotemporal graph neural model is constructed and trained to generate a spatiotemporal confidence interval for rotating reserve capacity. A dual-objective optimization model is established with the day-ahead total operating cost and robustness index as optimization objectives. The spatiotemporal confidence interval of the generated spinning reserve capacity is used as a constraint input into the constructed dual-objective optimization model to generate the day-ahead scheduling scheme. Based on the generated day-ahead scheduling scheme, wind farms and photovoltaic stations are divided into multiple agents corresponding to the three markets of electricity, gas and heat. Under the operating boundary determined by the day-ahead scheduling scheme, multi-agent intraday real-time game optimization is performed to generate intraday real-time scheduling instructions. An adaptive step size adjustment mechanism and a multi-timescale collaborative optimization interface are constructed to dynamically manage the intraday real-time game optimization process and generate frequency adjustment and energy trading instructions. Based on the generated frequency regulation and energy trading instructions, the system outputs time-period output plans and frequency regulation capacity quotations for the corresponding thermal power units, combined heat and power plants, electric boilers, electric-to-gas equipment, and thermal storage tanks for electricity, gas, and heat intelligent agents.

2. The virtual power plant frequency regulation trading optimization method based on split-blob bar model predictive control according to claim 1, characterized in that, The method for constructing and training a spatiotemporal graph neural model based on collected wind and solar data and geospatial parameters to generate the spatiotemporal confidence interval of rotating reserve capacity is as follows: A spatiotemporal graph neural network model is constructed using wind farms and photovoltaic stations as graph nodes and geospatial parameters and the correlation between adjacent equipment as edge weights. Based on the collected historical wind and solar data and geospatial parameters, a spatiotemporal graph neural network model is trained to generate a joint spatiotemporal confidence band for prediction errors, thereby obtaining the upper and lower limits of prediction errors for wind power and photovoltaic power at each time point. Based on the upper and lower limits of the prediction errors for wind power and photovoltaic power at each time point, the spatial difference compensation term is calculated, and combined with the confidence level parameter, the spatiotemporal confidence interval of the spinning reserve capacity is generated.

3. The virtual power plant frequency regulation trading optimization method based on split-blob bar model predictive control according to claim 2, characterized in that, The formula for calculating the spatiotemporal confidence interval of spinning reserve capacity based on the upper and lower limits of the prediction errors for wind and solar power at each time point, combined with the confidence level parameters, is expressed as follows: (1) (2) (3) Among them, R down R(t) represents the minimum reserve capacity with the upper limit of prediction error at time t. up (t) represents the minimum reserve capacity at time t, which is the lower limit of the error. This represents the upper limit of the wind power forecast value at time t. The upper limit of the photovoltaic forecast value at time t. This represents the lower limit of the wind power prediction value at time t. This represents the lower limit of the photovoltaic prediction value at time t. This represents the spatial difference compensation term. and , respectively, represent the historical average prediction error of the k-th wind farm or photovoltaic power station. and Let K be the actual output deviation of the k-th device at time t, where K is the number of devices.

4. The virtual power plant frequency regulation trading optimization method based on split-blob bar model predictive control according to claim 3, characterized in that, The method for establishing a bi-objective optimization model with the day-ahead total operating cost and robustness index as optimization objectives, and using the generated spatiotemporal confidence interval of the spinning reserve capacity as a constraint input to the constructed bi-objective optimization model to generate the day-ahead scheduling scheme is as follows: A dual-objective optimization model is established with the day-ahead total operating cost and robustness index as optimization objectives. The day-ahead total operating cost includes electricity, gas, and heat dispatch costs as well as carbon trading costs. The generated spatiotemporal confidence interval of the spinning reserve capacity is input as a constraint into the bi-objective optimization model. A hybrid quantum genetic algorithm is used to solve the established bi-objective optimization model. The hybrid quantum genetic algorithm integrates quantum genetic algorithm and dynamic Pareto front search strategy to generate a day-ahead scheduling scheme that takes into account both economy and robustness. The scheme includes spinning reserve capacity configuration and 24-hour power generation, energy storage and frequency regulation plans.

5. The virtual power plant frequency regulation trading optimization method based on split-blob bar model predictive control according to claim 4, characterized in that, The robustness index is embodied by a robustness evaluation function, the formula of which is as follows: in, These are dynamic weighting coefficients; This represents the worst-case opportunity cost or penalty cost at time t, related to system uncertainty. This represents the opportunity cost or penalty cost in the best-case scenario related to system uncertainty at time t.

6. The virtual power plant frequency regulation trading optimization method based on split-blown bar model predictive control according to claim 5, characterized in that, The generated day-ahead scheduling scheme divides wind farms and photovoltaic stations into multiple agents corresponding to the electricity, gas, and heat markets. Under the operational boundaries determined by the day-ahead scheduling scheme, the step of generating day-ahead real-time scheduling instructions through multi-agent intraday real-time game optimization is as follows: The optimal strategy among agents is learned through a deep Q-network. The state space is defined as the current wind and solar power output, load demand, market price and equipment status, the action space is the equipment output adjustment amount, and the reward function is the weighted sum of the day-ahead total operating cost and robustness. Under the boundary constraints determined by the current scheduling scheme, multiple agents engage in real-time multi-agent game optimization through continuous interaction and trial and error with the environment. The joint actions generated after the strategies of each agent converge constitute the system's real-time, coordinated, and near-optimal intraday real-time scheduling instructions.

7. The virtual power plant frequency regulation trading optimization method based on split-blob bar model predictive control according to claim 6, characterized in that, In the steps of constructing an adaptive step size adjustment mechanism and a multi-timescale collaborative optimization interface, dynamically managing the intraday real-time game optimization process, and generating frequency regulation and energy trading instructions, the adaptive step size adjustment mechanism generates a comprehensive decision index based on real-time wind and solar power output error and operating cost error, and dynamically adjusts the rolling optimization time domain length and optimization trigger frequency accordingly. The multi-timescale collaborative optimization interface is used to pass the dynamic weight coefficients of the robustness evaluation function. Spatial difference compensation item The weights of the reward functions among the multiple agents are updated based on market signals, and the agents' game strategies are corrected based on the backtracking day-ahead plan deviations.

8. The virtual power plant frequency regulation trading optimization method based on split-blown bar model predictive control according to claim 7, characterized in that, The specific method of the adaptive step size adjustment mechanism is as follows: Obtain real-time power output data from wind farms and solar power stations; The real-time output data of wind farms and photovoltaic stations are compared with the predicted values ​​of the generated intraday real-time dispatch instructions to obtain the real-time wind and solar output error; at the same time, based on the actual operating status and market prices, the current actual operating cost is calculated and compared with the expected cost in the dispatch plan to obtain the operating cost error. A comprehensive decision-making index is calculated based on real-time output error and operating cost error. Based on comprehensive decision indicators and preset coefficients of variation, the change in time step is calculated and the length of the rolling time domain is dynamically adjusted to achieve a balance between accuracy and efficiency.

9. The virtual power plant frequency regulation trading optimization method based on split-blown bar model predictive control according to claim 8, characterized in that, The specific working method of the multi-timescale collaborative optimization interface is as follows: Dynamic weight coefficients of the robustness evaluation function based on multi-timescale collaborative optimization interface transmission Spatial difference compensation item ; Based on market signals, dynamic weight parameters are generated and passed to the reward function of the game agent to update the weights of the reward functions among the multiple agents, thereby adjusting the agent's behavioral priority and strategy. By reviewing the deviations of the previous optimization scheme, and combining the adjusted behavioral priorities and strategies of the agent, the game strategy is corrected. Based on the updated reward function and the revised game strategy, the system drives multiple agents to perform real-time game optimization within a dynamically adjusted time step, generating and outputting frequency modulation and energy trading instructions.

10. A virtual power plant frequency regulation trading optimization system based on sub-Brønsted bar model predictive control, comprising the virtual power plant frequency regulation trading optimization method based on sub-Brønsted bar model predictive control as described in any one of claims 1 to 9, characterized in that, include: The data acquisition module is used to collect wind and solar data and geospatial parameters from wind farms and photovoltaic stations; The spatiotemporal confidence interval generation module is used to build and train a spatiotemporal graph neural model based on collected wind and light data and geospatial parameters, and generate the spatiotemporal confidence interval of rotating reserve capacity. The day-ahead scheduling scheme generation module is used to establish a bi-objective optimization model with the day-ahead total operating cost and robustness index as optimization objectives. The generated spatiotemporal confidence interval of the spinning reserve capacity is used as a constraint input to the constructed bi-objective optimization model to generate the day-ahead scheduling scheme. The intraday real-time dispatch instruction generation module is used to divide wind farms and photovoltaic stations into multiple agents corresponding to the three markets of electricity, gas and heat based on the generated day-ahead dispatch scheme. Under the operating boundary determined by the day-ahead dispatch scheme, the module performs intraday real-time game optimization of the multiple agents to generate intraday real-time dispatch instructions. The frequency modulation and energy trading instruction generation module is used to construct an adaptive step size adjustment mechanism and a multi-timescale collaborative optimization interface, dynamically manage the intraday real-time game optimization process, and generate frequency modulation and energy trading instructions. The output and execution module is used to output the time-period output plan and frequency regulation capacity quotation of the corresponding thermal power units, combined heat and power plants, electric boilers, electric-to-gas equipment and thermal storage tanks based on the generated frequency regulation and energy trading instructions.

Images