Multi-power-supply cooperative frequency modulation control method

Abstract

The invention provides a multi-power-supply cooperative frequency modulation control method, and relates to the field of power grid dispatching, and the method comprises the steps: obtaining the real-time operation data of a power grid, the real-time operation data of the power grid comprises system frequency deviation, load prediction information, and the real-time output and adjustment capability data of multiple power supplies such as wind power, photovoltaic, hydroelectric and thermal power; according to the real-time operation data of the power grid, through a mixed integer linear programming algorithm, taking power generation cost minimization as a target function, and taking multi-power output upper and lower limits, a climbing rate, a system frequency deviation limit value and the like as constraint conditions, calculating an optimal output distribution scheme of multiple power supplies; and converting the optimal output distribution scheme into a power supply output adjustment instruction. Cooperative frequency modulation of multiple power supplies is realized, the system frequency stability and economical efficiency are improved, the renewable energy source grid connection and consumption capability is enhanced, and an effective solution is provided for large-scale new energy source access to a power grid.

Description

Technical Field

[0001] The present invention relates to the field of power grid dispatching, and in particular to a multi-power supply coordinated frequency modulation control method. Background Art

[0002] In the coordinated frequency regulation control of wind, light, water and thermal power sources, the global coordination layer needs to formulate the output allocation plan of multiple power sources based on information such as system frequency deviation and load forecast to achieve the global optimal frequency regulation effect. However, the response speed and frequency regulation capability of different types of power sources vary greatly. How to take into account the economy and reliability of each power source while meeting the frequency stability requirements of the power grid is a complex technical problem. Wind power and photovoltaic power generation have large capacity and fast response speed, but they are highly volatile due to weather and difficult to accurately predict; hydropower has strong regulation capacity, but is limited by reservoir capacity and downstream ecological flow; thermal power has a wide regulation range, but the response speed is relatively slow. How to reasonably formulate the base power and regulation margin of each power source according to the power source characteristics and power grid needs, coordinate the output response of each power source on different time scales, avoid contradictions or offsets between multiple power source control instructions, and take into account the start and stop status of the unit and the spare capacity to achieve optimal frequency regulation control under multiple time scales and multiple constraints are key technical issues that need to be solved urgently. In addition, in actual engineering applications, it is also necessary to consider the impact of factors such as communication delay and control lag on the coordinated control performance of multiple power sources, comprehensively balance the frequency regulation performance and the difficulty of engineering implementation, and ensure the reliability and effectiveness of the coordinated frequency regulation control strategy of multiple power sources. Summary of the invention

[0003] The present invention provides a multi-power supply coordinated frequency modulation control method, which mainly includes:

[0004] Acquire real-time operation data of the power grid, wherein the real-time operation data of the power grid includes system frequency deviation, load forecast information, and real-time output and regulation capacity data of multiple power sources such as wind power, photovoltaic power, hydropower and thermal power; calculate the optimal output allocation plan of multiple power sources based on the real-time operation data of the power grid through a mixed integer linear programming algorithm, with minimization of power generation cost as the objective function, and upper and lower limits of multiple power output, climbing rate and system frequency deviation limit as constraints; convert the optimal output allocation plan into power output adjustment instructions, and issue them to each power control system through a power grid dispatching automation system; continuously monitor the operating status and performance indicators of the multi-power coordinated frequency regulation control system, and continuously optimize the multi-power coordinated frequency regulation control strategy through online learning and strategy iteration to improve the effect of power grid frequency control.

[0005] Furthermore, after obtaining the real-time operation data of the power grid, it also includes: preprocessing and feature extraction of the real-time operation data of the power grid, calculating key indicators such as the system frequency deviation mean and load forecast error, and converting multi-power supply data into a standardized input format.

[0006] Furthermore, before calculating the optimal output allocation plan for multiple power sources based on the real-time operation data of the power grid, it also includes: using a decision tree algorithm to classify the operation status of the power grid to determine whether the current status meets the frequency stability and economic operation requirements. If not, triggering the multi-power output optimization adjustment process.

[0007] Furthermore, after converting the optimal output allocation plan into a power output adjustment instruction, it also includes: during the power output adjustment process, using power simulation software to monitor the power grid operation status and system frequency changes in real time. If the frequency deviation exceeds a preset threshold, an emergency frequency regulation strategy based on fuzzy PID control is triggered, and the system frequency stability is quickly restored by adjusting power output and load power limiting and other measures.

[0008] Furthermore, after continuously monitoring the operating status and performance indicators of the multi-power supply coordinated frequency regulation control system, it also includes: using a support vector machine algorithm to train historical communication status data, control effect data and optimization parameters, establishing a prediction model, and using the prediction model to achieve autonomous optimization of the star topology structure and PID control parameters without human intervention.

[0009] Furthermore, the real-time operation data of the power grid is obtained, including: system frequency deviation and real-time output data of each power source, and the power source regulation capability data and load forecast result data are obtained at the same time. The real-time operation data of the power grid, the power source regulation capability data and the load forecast result data are cleaned and feature extracted, and comprehensive evaluation data is generated through data fusion technology; the comprehensive evaluation data is input into the Q-learning reinforcement learning model, and the power grid operation optimization model is updated regularly by iteratively updating the Q value table; according to the power source operation status evaluation result, the optimization objective function and constraint conditions of the multi-power source coordinated frequency regulation control are determined, and the optimal frequency regulation control strategy of each power source is obtained by solving the optimization problem, and the optimal frequency regulation control strategy includes AGC instructions and primary frequency regulation power allocation.

[0010] Furthermore, when continuously optimizing the multi-power source coordinated frequency regulation control strategy, the strategy gradient REINFORCE algorithm is used for online learning. Through continuous trial and error and strategy iteration, a continuously optimized multi-power source coordinated frequency regulation control strategy is obtained, the frequency regulation power allocation and control instruction issuance plan are determined, and the frequency of the grid connection point is quickly and accurately adjusted.

[0011] Furthermore, the real-time operating status data of the multi-power supply cooperative control system is obtained, including power supply status, load demand, etc., and used as the input of the particle swarm optimization algorithm; the standard particle swarm optimization algorithm is adopted, and the star topology parameters and PID controller parameters are used as optimization variables to establish an optimization model; according to the optimization results, the fuzzy control method is used to dynamically adjust the control period and control gain; if the system performance indicators do not meet the requirements for multiple consecutive control cycles, the current status data is re-input into the particle swarm optimization algorithm for optimization and adjustment again.

[0012] Furthermore, a communication network based on the IEC61850 protocol is used between the power grid dispatching automation system and each power supply control system to achieve real-time transmission and interaction of power grid operation information and control instructions.

[0013] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0014] The present invention discloses a method for coordinated frequency regulation control of multiple power sources. In view of the characteristics of multiple power sources such as wind power, photovoltaic power, hydropower and thermal power in the power grid, a mixed integer linear programming algorithm is used to calculate the optimal output allocation plan to ensure the stability and economy of the system frequency. For the volatility of wind and solar power generation, support vector machine prediction and sliding average smoothing control are adopted. In view of the fact that hydropower is limited by reservoir capacity, dynamic programming optimization scheduling is used. Deep peak regulation technology is used for thermal power to expand the regulation range. In the coordination of multiple power sources, a hierarchical control architecture is adopted, the base point power and regulation margin are formulated at the global level, and a genetic algorithm is used for multi-objective optimization. In order to reduce the impact of communication delay, Kalman filter predictive control and adaptive algorithm are adopted. Taking into account the performance and implementation difficulty, the control strategy is optimized through simulation and actual verification. The present invention realizes the coordinated frequency regulation of multiple power sources, improves the system frequency stability and economy, enhances the grid connection and absorption capacity of renewable energy, and provides an effective solution for large-scale new energy access to the power grid. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 The present invention is a flowchart of a multi-power supply coordinated frequency modulation control method. DETAILED DESCRIPTION

[0016] In order to further understand the content of the present invention, the present invention is described in detail in conjunction with the accompanying drawings and embodiments. The present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It is understood that the specific embodiments described herein are only used to explain the relevant inventions, rather than to limit the invention. It is also necessary to explain that, for ease of description, only the parts related to the invention are shown in the accompanying drawings.

[0017] like Figure 1 In this embodiment, a multi-power supply coordinated frequency modulation control method may specifically include:

[0018] S101. Based on the real-time operation data of the power grid, the system frequency deviation and load forecast information are obtained. In combination with the real-time output and regulation capabilities of multiple power sources such as wind power, photovoltaic power, hydropower and thermal power, the optimal output allocation plan of multiple power sources is calculated through a mixed integer linear programming algorithm to ensure system frequency stability and power economy.

[0019] Obtain real-time grid operation data, including system frequency deviation, load forecast information, and real-time output and regulation capacity data of multiple power sources such as wind power, photovoltaic, hydropower and thermal power. Preprocess and extract features of the acquired data, calculate key indicators such as the mean value of system frequency deviation and load forecast error, and convert the multi-power data into a standardized input format. Use the decision tree algorithm to classify the grid operation status to determine whether the current status meets the requirements of frequency stability and economic operation. If not, trigger the multi-power output optimization adjustment process. Construct a mixed integer linear programming model, with the minimization of power generation cost as the objective function, and the upper and lower limits of multi-power output, ramp rate and system frequency deviation limit as constraints to solve the optimal output allocation plan of multiple power sources. Convert the optimal output allocation plan into power output adjustment instructions, and send them to each power control system through the power grid dispatching automation system. During the power output adjustment process, use power simulation software to monitor the grid operation status and system frequency changes in real time. If the frequency deviation exceeds the preset threshold, the emergency frequency regulation strategy based on fuzzy PID control is triggered, and the system frequency stability is quickly restored by adjusting power output and load power restriction. After the adjustment is completed, the economic indicators such as the grid operation status and power generation cost before and after the adjustment of the multi-power output are compared and analyzed to evaluate the optimization adjustment effect. The evaluation results are used to correct the objective function coefficients and constraint parameters in the mixed integer linear programming model to achieve adaptive optimization of the model. Based on the historical operation data of the power grid and the optimization results, the support vector machine regression algorithm is used to establish a prediction model for the output distribution of multiple power sources. The model is used to predict the grid operation status and power output distribution in the future, provide auxiliary decision support for optimized scheduling, and improve the system frequency stability and economic operation level.

[0020] For example, obtaining real-time grid operation data is like a doctor examining a patient and obtaining various physiological indicators of the patient. For example, the system frequency deviation is like whether the heartbeat is stable, the load forecast information is like the prediction of the patient's future activity, and the real-time output and regulation capacity data of multiple power sources such as wind power, photovoltaic, hydropower and thermal power are like the operation and potential of various organs in the human body. These data can be obtained through sensors and smart meters deployed at various nodes of the grid. For example, smart meters can monitor the user's electricity consumption in real time, and the system load data is collected together. The wind turbines in the wind farm are equipped with sensors that can measure wind speed, wind direction, etc., and then calculate the output of the wind turbine. Data preprocessing and feature extraction are like a doctor analyzing the examination results. The obtained raw data is cleaned and denoised, such as removing abnormal load peak data. Calculating key indicators such as the mean of system frequency deviation and load forecast error is like calculating the average heart rate and blood pressure of a patient. Converting multiple power data into a standardized input format is like unifying various indicators of a patient into a standard unit of measurement. For example, converting the output of different types of power sources into MW. Using the decision tree algorithm to classify the grid operation status is like a doctor judging the condition of a patient based on various indicators. The decision tree classifies the grid operation status into frequency stability, mild frequency fluctuation, severe frequency fluctuation and other categories based on indicators such as system frequency deviation and load forecast error. For example, if the frequency deviation exceeds a certain threshold, it is judged as a frequency fluctuation state. This can help quickly determine whether the grid operation status meets the requirements of frequency stability and economic operation. Triggering the multi-power output optimization adjustment process is like a doctor formulating a treatment plan based on the condition. If the grid operation status does not meet the requirements, it is necessary to adjust the output of multiple power sources. For example, when the system frequency is low, it is necessary to increase the output of the generator; otherwise, it is necessary to reduce the output. Constructing a mixed integer linear programming model is like a doctor formulating a medication plan, considering the dosage, interaction and tolerance of multiple drugs. Taking the minimization of power generation cost as the objective function is like minimizing the cost of treatment. Constraints such as the upper and lower limits of multi-power output, ramp rate and system frequency deviation limit are like medication dosage, drug contraindications and patient physical condition restrictions. For example, the output of a thermal power unit cannot exceed its maximum output limit, nor can it be lower than its minimum technical output limit; its output change speed cannot exceed its ramp rate limit. By solving this model, the optimal output allocation plan for each power source can be obtained. The optimal output allocation plan is converted into a power output adjustment instruction, just like a doctor converts a medication plan into a specific prescription. It is sent to each power control system through the power grid dispatching automation system, just like a pharmacist dispenses medicine according to a prescription. For example, the output adjustment instruction of the thermal power unit is sent to the control system of the thermal power plant, and the control system will automatically adjust the fuel supply and turbine speed to change the output of the generator.Using power simulation software to monitor the grid operation status and system frequency changes in real time is like a doctor continuously monitoring the patient's vital signs during treatment. If the frequency deviation exceeds the preset threshold, it is like some of the patient's indicators are abnormal. Triggering the emergency frequency regulation strategy based on fuzzy PID control is like a doctor taking emergency measures to stabilize the patient's condition. By adjusting the power output and load power restriction and other measures, the system frequency stability can be quickly restored, just like controlling the patient's condition through drugs or surgery. For example, when the system frequency suddenly drops a lot, the fast-response pumped storage unit or gas turbine unit can be started to increase the output, and some non-important loads can be restricted to power, so as to quickly increase the system frequency. After the adjustment is completed, the economic indicators such as the grid operation status and power generation cost before and after the adjustment of the multi-power output are compared and analyzed to evaluate the optimization adjustment effect, just like a doctor evaluates the treatment effect. For example, the fluctuation degree of the system frequency before and after the adjustment, the level of power generation cost, etc., are compared to determine whether the adjustment is effective. The evaluation results are used to modify the objective function coefficients and constraint parameters in the mixed integer linear programming model to achieve adaptive optimization of the model, just like a doctor adjusts the medication plan according to the treatment effect. For example, if it is found that the actual output adjustment effect of a certain power source is not as good as expected, the output cost coefficient or ramp rate constraint of the power source in the model can be adjusted. Based on the historical operation data of the power grid and the optimization results, the support vector machine regression algorithm is used to establish a prediction model for the output distribution of multiple power sources, just like a doctor predicts the progression of a disease based on the patient's medical history and treatment records. The model is used to predict the operation status of the power grid and the output distribution of power sources in the future, providing auxiliary decision support for optimized scheduling, just like a doctor predicts the disease and makes preparations in advance to prevent the disease from worsening. This can improve the operation efficiency and safety of the power grid.

[0021] S102. In view of the volatility and forecasting difficulty of wind power and photovoltaic power generation, a short-term power forecasting algorithm based on support vector machine and a real-time power smoothing control strategy based on sliding average are adopted to reduce the impact of wind and solar power fluctuations on the grid frequency. Based on the forecast results and real-time power data, the base power and regulation margin of wind and solar power sources are dynamically adjusted to improve the ability and reliability of wind and solar power sources to participate in system frequency regulation.

[0022] The historical power data and meteorological data of wind farms and photovoltaic power stations are obtained, including parameters such as wind speed, wind direction, irradiance, and temperature. Data cleaning and anomaly detection algorithms are used to remove abnormal and invalid data and normalize the data. According to the timestamps of power data and meteorological data, they are time-aligned and synchronized. Meteorological features with high correlation with power, such as wind speed and irradiance, are selected to construct the input feature vector of the prediction model. The support vector regression (SVR) algorithm is used to establish a nonlinear mapping relationship between the input feature vector and the output power. Through grid search and cross-validation, the hyperparameters of the SVR model, such as the kernel function type, penalty coefficient, and kernel function parameter, are optimized. The optimized SVR model is used to predict the power of real-time meteorological data of wind farms and photovoltaic power stations. According to the grid dispatching requirements, the prediction time scale is set, such as 15 minutes, 1 hour, etc. The input feature data is updated every time step, and the power prediction value for a period of time in the future is output. The predicted power sequence is smoothed using the moving average filter algorithm. The initial smoothing window size is set to 10 time steps and the weight coefficient is 6. According to the peak regulation demand of the power grid and the fluctuation characteristics of wind and solar resources, the smoothing window size and weight coefficient are adaptively adjusted to balance the smoothness and response speed of power output. The smoothed power output is used as the base power of wind farms and photovoltaic power stations. According to the frequency regulation demand of the power grid, the regulation margin of wind and solar power sources is optimized in real time. When the grid frequency is high, the base power is reduced to increase the regulation margin; when the grid frequency is low, the base power is increased to reduce the regulation margin. When the grid frequency deviates from the rated value, the wind farm and photovoltaic power station start the primary and secondary frequency regulation respectively. The wind farm adopts a combination of variable pitch control and variable speed control to quickly adjust the active power output according to the frequency deviation and regulation margin. The photovoltaic power station adopts a combination of power reservation and energy storage to smoothly adjust the active power output according to the frequency deviation and regulation margin. The effect of wind farms and photovoltaic power stations participating in frequency regulation is continuously monitored, and indicators such as frequency regulation contribution, frequency regulation accuracy, and frequency regulation response time are used for evaluation. Using the evaluation results, the feature selection and hyperparameter setting of the SVR prediction model are optimized, the smoothing window and weight parameters of the moving average filter algorithm are optimized, the regulation margin and control strategy of wind and solar power sources are optimized, and the reliability and economy of wind and solar power sources participating in grid frequency regulation are improved.

[0023] For example, the output of wind farms and photovoltaic power stations fluctuates greatly, and accurate prediction of their output power is crucial to the stable operation of the power grid. Obtaining historical power data and meteorological data is the basis for prediction. For example, a wind farm records the wind speed, wind direction, temperature and corresponding power generation every 15 minutes, and the photovoltaic power station records the irradiance, temperature and corresponding power generation every 15 minutes. These data constitute the training set of the prediction model. In the data preprocessing stage, cleaning and anomaly detection are required. For example, the wind speed data of a certain day has an abnormal spike due to sensor failure, which needs to be removed or interpolated with surrounding data. Power data may also be missing due to equipment maintenance and other reasons, and needs to be supplemented. Normalization processing can unify data of different dimensions to the same scale, such as normalizing data such as wind speed, temperature, and irradiance to between 0 and 1 to avoid excessive influence of certain features on the model. Time alignment and synchronization are also crucial to ensure that meteorological data and power data correspond one to one. For example, the timestamp of wind speed data is the hour, and the timestamp of power data is 10 minutes after the hour, and the two need to be aligned to the same timestamp. Selecting appropriate input features is crucial to model accuracy. Generally speaking, wind speed and irradiance are most correlated with wind and solar output. In addition, factors such as wind direction and temperature will also have a certain impact on the output. Features can be screened using methods such as correlation analysis to construct input feature vectors. For example, by analyzing historical data, it is found that wind speed and wind direction have a greater impact on wind farm output, while temperature has a smaller impact. In this case, wind speed and wind direction can be used as input features. The support vector regression (SVR) algorithm can establish a nonlinear mapping relationship between input feature vectors and output power, and is suitable for processing nonlinear relationships and small sample data. For example, an SVR model is trained using historical data to learn the complex relationship between wind speed, wind direction and wind farm output. Grid search and cross-validation can optimize the hyperparameters of the SVR model and improve the generalization ability of the model. For example, the optimal penalty coefficient and kernel function parameters can be found through grid search to avoid overfitting or underfitting of the model. The trained SVR model can be used to predict power in the future. For example, based on the weather forecast data for the next 24 hours, the output of wind farms and photovoltaic power stations in the next 24 hours can be predicted. The prediction time scale can be adjusted according to the scheduling requirements. For example, for intraday scheduling, a prediction time scale of 15 minutes can be set; for long-term planning, a prediction time scale of 1 hour or even longer can be set. Sliding update of input feature data can ensure the real-time performance of the prediction model. For example, meteorological data is updated every 15 minutes, and the output for the next 15 minutes is predicted. Moving average filtering can smooth the predicted power sequence and reduce the fluctuations caused by prediction errors. For example, using a moving average filter with a window size of 10 time steps can make the predicted power curve smoother. Adaptive adjustment of the smoothing window size and weight coefficient can balance smoothness and response speed.For example, when the wind and solar power output fluctuates greatly, the window size can be reduced to improve the response speed; when the output fluctuation is small, the window size can be increased to improve smoothness. The smoothed power output is used as the base power of the wind and solar power source for subsequent frequency regulation. The participation of wind and solar power sources in frequency regulation can improve the stability of the power grid. When the grid frequency is high, the base power is reduced to increase the regulation margin, and vice versa. For example, when the grid frequency increases by 0.1Hz, the wind farm reduces the base power by 5MW, and the photovoltaic power station reduces the base power by 3MW, reserving more output for frequency regulation. Primary frequency regulation is the primary means of regulating the grid frequency and requires fast response. The wind farm can quickly adjust the output through variable pitch control and variable speed control. For example, when the grid frequency decreases, the wind turbine increases the pitch angle or speed to increase the output. The photovoltaic power station can participate in primary frequency regulation by coordinating power reservation and energy storage. For example, when the grid frequency decreases, the reserved power is released or the power is output from the energy storage system. The secondary frequency regulation has a relatively slow response speed and is used to restore the frequency to the rated value. Wind and solar power sources can smoothly adjust their output according to the frequency deviation and regulation margin and participate in secondary frequency regulation. Continuously monitoring and evaluating the effect of wind and solar power sources in frequency regulation can continuously optimize the control strategy. For example, statistics on indicators such as the frequency regulation contribution, frequency regulation accuracy, and frequency regulation response time of wind and solar power sources can be used to analyze their response characteristics under different frequency deviations. According to the evaluation results, the SVR prediction model, moving average filtering algorithm, and the regulation margin and control strategy of wind and solar power sources can be adjusted to improve the reliability and economy of wind and solar power sources in participating in grid frequency regulation. For example, if it is found that the frequency regulation response time of a wind farm is too long, its control strategy can be optimized to improve the response speed.

[0024] S103. In view of the problem that the hydropower regulation capacity is limited by the reservoir capacity and downstream ecological flow, a dynamic programming reservoir optimization dispatching model is adopted to meet the grid frequency regulation needs while taking into account the long-term operation benefits of the reservoir and the protection of the ecological environment. Based on the reservoir optimization dispatching results, the output distribution and start-stop status of the hydropower units are optimized to improve the flexibility and economy of hydropower frequency regulation.

[0025] First, according to the real-time water level and inflow of the reservoir, combined with the downstream ecological flow demand, the dynamic programming algorithm is used to determine the optimal scheduling strategy of the reservoir in the future period, and the outflow process of the reservoir is obtained. The state variable of dynamic programming is defined as the reservoir water level, and the decision variable is the outflow. The state transition equation is: water level (t+1) = water level (t) + inflow (t) - outflow (t) - evaporation (t) leakage (t). The objective function is to maximize the hydropower generation in the future period and meet the constraints of the downstream ecological flow demand. Then, according to the frequency regulation demand of the power grid, combined with the outflow obtained by the optimal scheduling of the reservoir, the mixed integer programming algorithm is used to optimize the output distribution and start-stop status of the hydropower unit. The decision variables include the output and start-stop status of each unit. The constraints include the outflow constraint of the reservoir, the unit output limit, and the start-stop time constraint of the unit. The objective function is to maximize the long-term operating benefits of the hydropower station, for example, maximizing the power generation income or minimizing the fuel cost. The unit combination scheme is determined, and the units are divided into several levels according to the unit capacity. For example, a certain number of units are selected in each level of small units, medium units and large units to form multiple optional unit combination schemes to improve the flexibility of hydropower frequency regulation. In mixed integer programming, the output distribution and start-stop status of different unit combination schemes are considered to meet the frequency regulation requirements and maximize long-term benefits. By setting the minimum duration constraint of unit start-stop, frequent start-stop is avoided and unit wear is reduced. Finally, according to the long-term operation data of the reservoir, the Q-learning algorithm is used to adaptively adjust the various parameters in the dynamic programming model, such as the weight coefficient in the objective function and the evaporation leakage coefficient in the state transfer equation, so that the optimization scheduling strategy can adapt to the changes in hydrological conditions and ensure the maximization of the long-term benefits of the reservoir. The state space is defined as the reservoir water level and inflow, the action space is defined as the outflow, the reward function is defined as the power generation, and the learning rate and discount factor are set according to the hydrological conditions and operation experience.

[0026] For example, the optimal dispatching strategy of a reservoir is related to the benefits of hydropower generation and the guarantee of downstream ecological water demand. Dynamic programming algorithms can effectively solve such multi-stage decision-making problems. Taking a reservoir as an example, suppose that the outflow flow strategy for the next week needs to be determined. First, the reservoir water level is divided into multiple discrete states. For example, several water level states are divided from the dead water level to the flood limit water level at intervals of 0.5 meters. Then, the outflow flow is also divided into multiple discrete values. For example, several outflow flow values ​​are divided from the minimum ecological flow to the maximum flood discharge flow at intervals of 10 cubic meters per second. The state transition equation describes the change of the reservoir water level over time, that is, the water level at the next moment is jointly determined by the current water level, inflow flow, outflow flow, evaporation and leakage. For example, if the current water level is 100 meters, the inflow flow is 50 cubic meters per second, the outflow flow is 30 cubic meters per second, and the evaporation and leakage is 2 cubic meters per second, then the water level at the next moment is 100+50-30-2=118 meters. The objective function aims to maximize the hydropower generation in the next week while meeting the constraints of the downstream ecological flow demand. For example, if the minimum downstream ecological flow is set to 20 cubic meters per second, the outflow at any time during the optimization process cannot be lower than this value. Through the dynamic programming algorithm, the optimal outflow corresponding to each moment in the next week can be obtained to guide the scheduling and operation of the reservoir. The output distribution and start-stop status of the hydropower unit directly affect the operating efficiency of the hydropower station. The mixed integer programming algorithm can optimize the operating status of the unit while meeting various constraints. Assume that the hydropower station has 4 units, and the output limit and start-stop time constraints of each unit are different. For example, the output range of unit 1 is 50 to 100 MW, and the minimum start-stop time is 2 hours; the output range of unit 2 is 30 to 80 MW, and the minimum start-stop time is 1 hour. The outflow of the reservoir determines the total output of the hydropower station. For example, according to the results of the reservoir optimization scheduling, the total output of the hydropower station corresponding to the outflow at a certain moment in the future is 200 MW. The objective function of mixed integer programming can be set to maximize power generation revenue or minimize fuel costs. Constraints include reservoir outflow constraints, that is, the total output of the four units must be equal to 200 MW; unit output limits, that is, the output of each unit must be within its output range; and unit start-stop time constraints, that is, the start-stop time of each unit must meet its minimum start-stop time limit. By solving the mixed integer programming problem, the optimal output and start-stop state of each unit can be obtained, thereby maximizing the operating benefits of the hydropower station. In order to meet the frequency regulation needs of the power grid, it is necessary to consider the output distribution and start-stop state of different unit combination schemes. Assume that the four units are divided into two groups, units 1 and 2 are one group, and units 3 and 4 are one group. In this way, two unit combination schemes can be formed: Scheme 1, operating units 1 and 2; Scheme 2, operating units 3 and 4.In mixed integer programming, the output distribution and start-stop status of the two schemes are considered respectively, and the optimal scheme is selected according to the frequency regulation demand and long-term benefits. For example, when the grid frequency decreases, the hydropower station needs to increase the output quickly. At this time, a unit combination scheme with faster start-up speed and stronger regulation ability can be selected. By setting the minimum duration constraint of the unit start-up and shutdown, for example, stipulating that each unit must run for at least 2 hours after starting, it is possible to avoid frequent start-up and shutdown of the unit, reduce unit wear and extend the service life. The long-term operation of the reservoir will be affected by many factors, such as water inflow conditions and climate change. The Q-learning algorithm can adaptively adjust the parameters in the dynamic programming model according to the long-term operation data of the reservoir, thereby improving the adaptability of the scheduling strategy. The reservoir water level and inflow flow are used as state variables, the outflow flow is used as the action variable, and the power generation is used as the reward function. For example, when the reservoir water level is high and the inflow flow is large, the outflow flow can be appropriately increased to increase the power generation; when the reservoir water level is low and the inflow flow is small, the outflow flow needs to be reduced to ensure water supply safety. Through continuous learning and adjustment, the Q-learning algorithm can find the optimal outflow strategy, thereby maximizing the long-term benefits of the reservoir. The learning rate and discount factor are important parameters in the Q-learning algorithm and need to be set according to the hydrological conditions and operating experience. For example, the learning rate determines the speed at which the algorithm learns new knowledge, and the discount factor determines the degree of influence of future rewards on current decisions.

[0027] S104. In view of the fact that thermal power has a wide regulation range but a relatively slow response speed, deep peak regulation technology for thermal power units is adopted to expand the regulation range of thermal power units and improve the response speed. According to the regulation characteristics of thermal power units, the load distribution and spare capacity of thermal power units are optimized to improve the ability and reliability of thermal power units to participate in system frequency regulation.

[0028] The real-time operating parameters of the thermal power unit are obtained, including the current load, upper and lower limits of the regulation range, and response speed, to form a parameter set. Based on the parameter set, the BP neural network algorithm is used to establish the regulation characteristic model of the thermal power unit to obtain the optimized regulation range and response speed threshold. If the regulation range and response speed of the thermal power unit do not reach the optimization threshold, the deep peak regulation control is triggered, and the regulation range is expanded and the response speed is improved by optimizing the combustion and steam-water systems. According to the system load forecast and spare capacity requirements, the particle swarm optimization algorithm is used to calculate the optimal load distribution plan of the thermal power unit to balance the unit load rate and spare capacity. The load distribution plan is sent to the thermal power unit control system, and the unit output is adjusted through the automatic generation control (AGC) system to improve the unit's ability to participate in system frequency regulation. In the frequency regulation process, the PID control algorithm is used to dynamically optimize the frequency regulation parameters to improve the frequency regulation accuracy and reliability. According to the unit frequency regulation performance evaluation results, the PID control parameters in the frequency regulation process are adjusted and optimized in real time to continuously improve the regulation performance of the thermal power unit and the system frequency regulation level. By continuously optimizing the deep peak-shaving strategy and load distribution plan, a closed-loop optimization is formed, which continuously improves the regulation performance of thermal power units and the system frequency regulation level.

[0029] Exemplarily, the real-time operating parameters of thermal power units are the basis of deep peak-shaving control and load distribution. Through sensors and data acquisition systems, key parameters such as the current load, upper and lower limits of the regulation range, and response speed of each thermal power unit can be obtained in real time and summarized into a parameter set. For example, a 600MW thermal power unit has a current load of 500MW, an upper limit of the regulation range of 550MW, a lower limit of 400MW, and a response speed of 2MW / min. Collecting these parameters can help understand the current state and regulation capability of the unit. Using the collected parameter set, a BP neural network algorithm can be used to establish a thermal power unit regulation characteristic model. BP neural network is a multi-layer feedforward neural network that can establish a mapping relationship between unit parameters and regulation characteristics by learning a large amount of historical data. For example, historical load, regulation range, response speed and other data can be used as input, and actual peak-shaving performance indicators can be used as output to train the BP neural network model. Through training, the model can predict the optimal regulation range and response speed threshold of the unit under different load and parameter conditions. Assume that through model calculation, the optimal regulation range of the 600MW unit under a load of 500MW is 450MW to 580MW, and the response speed threshold is 4MW / min. If the actual regulation range and response speed of the thermal power unit do not reach the optimization threshold, deep peak-shaving control needs to be triggered. Deep peak-shaving control refers to further expanding the regulation range and improving the response speed of the unit by optimizing the combustion system (such as adjusting the burner inclination angle, changing the fuel ratio) and the steam-water system (such as optimizing steam temperature and pressure control). For example, for the above 600MW unit, the combustion system can be optimized to increase its response speed from 2MW / min to 3MW / min, which is closer to the optimization threshold of 4MW / min. At the same time, by improving the steam-water system control strategy, the lower limit of its regulation range can be extended from 400MW to 380MW. In order to balance the unit load rate and spare capacity, it is necessary to use the particle swarm optimization algorithm to calculate the optimal load distribution plan of the thermal power unit based on the system load forecast and spare capacity requirements. The particle swarm optimization algorithm is an intelligent optimization algorithm that simulates the foraging behavior of bird flocks and can effectively search for the optimal solution. For example, assuming that the system load forecast for the next hour is 1000MW, the spare capacity requirement is 200MW, and there are two 600MW thermal power units available, the particle swarm optimization algorithm can be used to calculate the optimal load distribution for each unit, for example, one unit bears 550MW load and the other bears 450MW load, which not only meets the system load demand but also leaves enough spare capacity. The calculated optimal load distribution plan needs to be sent to the thermal power unit control system, and the unit output adjustment is realized through the automatic generation control (AGC) system. The AGC system can automatically adjust the output of each unit according to the load distribution plan to achieve the target value.For example, if the load assigned to a unit is 550MW, the AGC system will automatically control the output of the unit to stabilize it near 550MW, thereby improving the unit's ability to participate in system frequency regulation. In the frequency regulation process, in order to improve the accuracy and reliability of frequency regulation, it is necessary to use the PID control algorithm to dynamically optimize the frequency regulation parameters. The PID control algorithm is a classic control algorithm that can achieve precise control of the controlled object through the control of three links: proportional, integral, and differential. For example, when the system frequency drops, the PID controller will adjust the output of the thermal power unit in real time according to the frequency deviation, so that it can quickly increase the output and restore the stability of the system frequency. In order to continuously improve the regulation performance of the thermal power unit and the level of system frequency regulation, it is necessary to adjust and optimize the PID control parameters in the frequency regulation process in real time according to the unit frequency regulation performance evaluation results. For example, if it is found that the frequency regulation response speed of the unit is slow, the proportional coefficient of the PID controller can be appropriately increased to improve the response speed. By continuously optimizing the deep peak regulation strategy and load distribution plan, and combining the dynamic adjustment of PID parameters to form a closed-loop optimization, the regulation performance of the thermal power unit and the level of system frequency regulation can be continuously improved to better meet the needs of the power system.

[0030] S105. In the coordinated frequency regulation control of multiple power sources, a hierarchical coordination control architecture is adopted. The global coordination layer formulates the base power and regulation margin of multiple power sources based on the system frequency deviation, load forecast information, and the real-time output and regulation capabilities of each power source, and coordinates the output response of each power source on different time scales. A multi-objective optimization algorithm based on genetic algorithm is adopted to meet the constraints of grid frequency stability while taking into account the economy and reliability of each power source, so as to avoid contradictions or offsets between the control instructions of multiple power sources.

[0031] According to the real-time operation data of the power grid, the system frequency deviation, the real-time output and regulation capacity of each power source and other information are obtained, combined with the load forecast results, and input into the global coordination control layer. The global coordination control layer adopts the NSGA-II multi-objective optimization algorithm, takes the power grid frequency stability, power economy and reliability as the optimization objectives, and generates the optimization results of the base point power and regulation margin of multiple power sources. The base point power and regulation margin of multiple power sources generated by the global coordination control layer are sent to each power control unit as their operation constraints. Each power control unit determines its own output regulation strategy through the sequential quadratic programming method based on the received base point power and regulation margin combined with its own operation status. During the real-time operation process, each power control unit uses the PID control algorithm to dynamically adjust the output according to the instructions issued by the global coordination control layer and its own optimization results, so as to realize the coordinated frequency regulation of multiple power sources. The coordinated frequency regulation effect of multiple power sources is judged by real-time monitoring of the power grid frequency. If the frequency deviation exceeds the preset threshold, the early warning system triggers the global coordination control layer to re-optimize the multiple power sources and generate new control instructions. According to the load change trend obtained by the power grid load forecasting model and the evaluation results of each power supply operation status evaluation system, the adaptive dynamic programming algorithm is used to regularly update the optimization model objective function and constraint conditions in the global coordination control layer to adapt to the dynamic changes of power grid operation and ensure the continuous optimization of multi-power supply coordinated frequency regulation control.

[0032] Exemplarily, the global coordination control layer is responsible for coordinating the output of all power sources to maintain the stability of the grid frequency. It receives real-time operation data from the grid, such as system frequency deviation, real-time output and regulation capacity of each power source (such as thermal power, hydropower, photovoltaic, etc.), and load forecast results. For example, at a certain moment, the system frequency deviation is -0.1Hz, indicating that the frequency is slightly lower than the standard value; the current output of a thermal power unit is 500MW, and the adjustable range is 400MW to 550MW; the current output of a hydropower unit is 200MW, and the adjustable range is 100MW to 300MW; the load forecast results show that the load will increase by 100MW in the next hour. All this information will be input into the global coordination control layer. The global coordination control layer adopts the NSGA-II multi-objective optimization algorithm. The NSGA-II algorithm can simultaneously consider multiple optimization objectives, such as grid frequency stability, power economy and reliability. It will calculate a set of optimal base point power and regulation margins based on the input real-time data and load forecast results, and allocate them to each power source. For example, in order to cope with the load increase of 100MW in the next hour, the global coordination control layer may increase the base power of the thermal power unit to 520MW, and keep the regulation margin at 50MW; increase the base power of the hydropower unit to 230MW, and keep the regulation margin at 70MW. This can not only meet the demand for load growth, but also maintain a certain frequency stability margin, while taking into account the economy and reliability of different power sources. The base power and regulation margin of multiple power sources calculated by the global coordination control layer will be sent to each power control unit. For example, the thermal power unit control unit will receive an instruction of a base power of 520MW and a regulation margin of 50MW; the hydropower unit control unit will receive an instruction of a base power of 230MW and a regulation margin of 70MW. These instructions serve as constraints for the operation of each power source and guide its output regulation. After receiving the instructions from the global coordination control layer, each power control unit will determine its own output regulation strategy through the sequential quadratic programming method based on its own operating status, such as current output, adjustable range, response speed, etc. For example, the control unit of a thermal power unit may formulate a strategy to gradually increase the output, such as increasing it by 2MW per minute until it reaches 520MW, based on the base power of 520MW and the regulation margin of 50MW, as well as its own operating status. During real-time operation, each power control unit will dynamically adjust the output using the PID control algorithm based on the instructions of the global coordination control layer and its own optimization results. The PID control algorithm is a classic control algorithm that can dynamically adjust the output of the power source according to the real-time frequency deviation to achieve precise control. For example, if the system frequency is lower than the standard value, the PID controller will automatically increase the output of the power source according to the size of the frequency deviation until the frequency returns to the standard value. By real-time monitoring of the grid frequency, the effect of coordinated frequency regulation of multiple power sources can be judged.If the frequency deviation exceeds the preset threshold, such as ±0.2Hz, the early warning system will trigger the global coordination control layer to re-optimize multiple power sources and generate new control instructions. For example, if the system frequency is detected to continue to drop and exceeds the threshold of -0.2Hz, the early warning system will immediately notify the global coordination control layer, which will re-evaluate the current grid operation status and load forecast results, and recalculate the base power and adjustment margin of each power source, and generate new control instructions to send to each power control unit. In order to adapt to the dynamic changes in grid operation, the optimization model in the global coordination control layer needs to be updated regularly. Using the adaptive dynamic programming algorithm, the objective function and constraints in the global coordination control layer can be regularly updated according to the load change trend obtained by the grid load forecasting model and the evaluation results of the operation status of each power source. For example, if it is predicted that the load will continue to grow in the future, the objective function of the optimization model can be adjusted to pay more attention to frequency stability; if the operating status of a certain power source changes, such as the adjustable range is reduced, the corresponding constraints need to be updated. Through continuous optimization, it can be ensured that the coordinated frequency regulation control of multiple power sources is always in the optimal state to better meet the needs of the power system.

[0033] S106. Obtain the system frequency deviation and the real-time output of each power source based on the real-time operation data of the power grid. Combined with the power regulation capability and load forecast results, regularly update the global coordination control layer optimization model through an adaptive dynamic programming algorithm, judge the power operation status evaluation results, determine the optimization objective function and constraints, and achieve continuous optimization of multi-power coordinated frequency regulation control to adapt to the dynamic changes in power grid operation.

[0034] Obtain real-time grid operation data, including system frequency deviation and real-time output data of each power source, and obtain power source regulation capability data and load forecast result data at the same time. Perform data cleaning and feature extraction on real-time grid operation data, power source regulation capability data and load forecast result data, and generate comprehensive evaluation data through data fusion technology such as Kalman filtering. Adopt the Q-learning reinforcement learning algorithm in the adaptive dynamic programming algorithm, input the comprehensive evaluation data into the Q-learning model, and regularly update the grid operation optimization model of the global coordination control layer by continuously iteratively updating the Q value table. Through the grid operation optimization model, comprehensively consider multiple objectives such as grid safety, economy, and low carbon, and construct a multi-objective optimization function to judge the current power supply operation status and obtain the power supply operation status evaluation result. According to the power supply operation status evaluation result, determine the optimization objective function of multi-power coordinated frequency regulation control, such as minimizing frequency deviation, minimizing frequency regulation cost, etc., as well as the corresponding power output constraints, climbing rate constraints, etc., as the basis for coordinated frequency regulation control. The optimization objective function and constraints are input into the multi-power coordinated frequency control model. This model is based on rolling optimization and model predictive control methods. By solving the optimization problem, the optimal frequency control strategy of each power source is obtained, including AGC instructions and primary frequency control power allocation. The optimal frequency control strategy is sent to each power control unit, including thermal power, hydropower, wind power, photovoltaic power, etc., to control each power source to adjust its output according to the frequency control instruction to achieve rapid recovery and stability of the grid frequency. Through rolling optimization and feedforward correction, the multi-power coordinated frequency control model is continuously optimized using the real-time operation data of the grid, and the frequency control strategy is adaptively adjusted to adapt to the dynamic changes of grid operation conditions such as load fluctuations and changes in new energy output, ensuring that the grid frequency is stable near the rated value for a long time.

[0035] Exemplarily, real-time operation data of the power grid, such as system frequency deviation and real-time output data of each power source (thermal power, hydropower, photovoltaic, wind power, etc.), are the basis for coordinated frequency regulation control of multiple power sources. The system frequency deviation reflects the real-time supply and demand balance of the power grid. For example, a frequency deviation of -0.1Hz indicates that the system power supply is insufficient. The real-time output data of each power source reflects the current operating status of each power source. For example, the output of a thermal power unit is 500MW and the output of a hydropower unit is 200MW. In addition to the real-time operation data of the power grid, it is also necessary to obtain power regulation capability data, such as the adjustable range of a thermal power unit is 400MW to 550MW, and load forecast result data, such as predicting that the load will increase by 100MW in the next hour. These data constitute the input information of coordinated frequency regulation control of multiple power sources. After obtaining the original data, data cleaning and feature extraction are required. For example, abnormal data can be eliminated, the data can be smoothed, and features such as frequency change trend and load change trend can be extracted. Then, through data fusion technology, such as Kalman filtering, data from multiple sources are fused into comprehensive evaluation data. For example, the system frequency deviation, load forecast results and the output data of each power source can be fused into a comprehensive indicator reflecting the real-time operation status of the power grid. The Q-learning reinforcement learning algorithm in the adaptive dynamic programming algorithm can be used to regularly update the power grid operation optimization model of the global coordinated control layer. Taking Q-learning as an example, it learns an optimal action strategy by continuously interacting with the environment. The comprehensive evaluation data is input into the Q-learning model, and the model will select an action according to the current state, such as adjusting the output of a certain power source. Then, the model will update the Q value table according to the feedback from the environment (such as the new frequency deviation) and continuously optimize the control strategy. The power grid operation optimization model needs to comprehensively consider multiple objectives such as power grid security, economy, and low carbon. For example, a multi-objective optimization function can be constructed, considering the minimization of frequency deviation, the minimization of frequency regulation cost, and the minimization of carbon emissions. By solving this optimization function, the power supply operation status evaluation results can be obtained, such as the evaluation of the economy, reliability, and low carbon of each power source. According to the power supply operation status evaluation results, the optimization objective function and constraints of multi-power supply coordinated frequency regulation control can be determined. For example, if the current frequency deviation is large, minimizing the frequency deviation is given priority; if the system operating cost is high, minimizing the frequency regulation cost is given priority. At the same time, power output constraints and climbing rate constraints need to be considered. For example, the output of a thermal power unit cannot exceed its maximum output limit, and its output change rate cannot exceed its climbing rate limit. The multi-power source collaborative frequency regulation control model is based on rolling optimization and model predictive control methods. It obtains the optimal frequency regulation control strategy for each power source by solving the optimization problem based on the optimization objective function and constraints, including AGC (automatic generation control) instructions and primary frequency regulation power allocation. For example, the model may issue instructions requiring thermal power units to increase their output by 20MW and hydropower units to reduce their output by 10MW.After receiving the optimal frequency control strategy, each power control unit (thermal power, hydropower, wind power, photovoltaic, etc.) will adjust the output according to the instructions. For example, the thermal power unit control unit will increase the amount of fuel according to the AGC instruction to increase the output; the hydropower unit control unit will adjust the turbine opening according to the instruction to reduce the output. The rolling optimization and feedforward correction mechanism enables the multi-power collaborative frequency control model to adaptively adjust the frequency control strategy. For example, when a sudden increase in load is detected, the model will re-optimize according to the latest grid operation data and generate new control instructions to quickly restore grid frequency stability. Through continuous optimization, the grid frequency can be stabilized near the rated value for a long time.

[0036] S107. In view of the impact of communication delay and control time lag on the coordinated control performance of multiple power sources, a predictive control strategy and adaptive control algorithm based on Kalman filtering are used to reduce the impact of communication delay and control time lag on system performance. According to the real-time communication status and control effect, the communication network topology and control parameters are optimized to improve the real-time performance and reliability of the coordinated control of multiple power sources.

[0037] The real-time communication status data and control effect data of the multi-power system are obtained, and the preset delay and time lag thresholds are used to determine whether they exceed the allowable range. If they exceed the range, the optimization process is triggered. The Kalman filter algorithm is used to estimate and predict the state of the multi-power system to obtain the predicted state value in the future. The predicted state value is used as the input of the collaborative control to make control decisions and execution in advance to reduce the impact of communication delay and control time lag. According to the real-time and reliability requirements of the multi-power collaborative control, the optimization objective function of the star topology and PID control parameters is determined. The particle swarm optimization algorithm is used to adjust and optimize the number of nodes, connection mode and PID control parameters of the star topology online to minimize the objective function. The optimized star topology and PID control parameters are obtained, and the configuration of the multi-power collaborative control system is updated to improve the real-time and reliability of the system. The communication status and control effect of the multi-power system are continuously monitored. If the communication delay or control time lag is detected to exceed the threshold again, the optimization process is re-triggered to form a closed-loop control. The support vector machine algorithm is used to train the historical communication status data, control effect data and optimization parameters to establish a prediction model. The prediction model is used to realize the autonomous optimization of the star topology and PID control parameters without manual intervention. According to the optimized star topology and PID control parameters, the control cycle and control gain of the multi-power cooperative control system are dynamically adjusted to ensure efficient and reliable cooperative control in the presence of communication delay and control lag. When the communication delay and control lag decrease, the control cycle is shortened and the control gain is reduced accordingly; when the communication delay and control lag increase, the control cycle is extended and the control gain is increased accordingly. The stability and dynamic performance of the multi-power cooperative control system are maintained by adaptively adjusting the control cycle and control gain.

[0038] For example, in the coordinated frequency regulation control of multiple power sources, communication delay and control lag will affect the control effect. In order to solve this problem, a series of optimization measures can be taken. First, it is necessary to obtain the communication status data and control effect data of the multi-power system in real time, such as the output data of each power source, frequency deviation data, communication delay time, control instruction execution time, etc. Set the threshold of delay and lag, for example, the communication delay does not exceed 10ms, and the control lag does not exceed 50ms. If the monitored data exceeds the threshold, the optimization process is triggered. The Kalman filter algorithm can be used to estimate and predict the state of the multi-power system. For example, the system frequency deviation in the future period can be predicted based on the historical frequency deviation data, power output data and load change data. Assuming that the system frequency deviation is -0.1Hz at the current moment, it is predicted that the frequency deviation will drop to -0.15Hz in the next 1 minute. This predicted value will be used as the input of the coordinated control to adjust the output of each power source in advance, thereby reducing the impact of communication delay and control lag. For example, the output of the hydropower unit can be increased in advance to cope with the predicted frequency drop. In order to improve the real-time and reliability of the coordinated control of multiple power sources, it is necessary to optimize the star topology and PID control parameters. The star topology refers to a network topology that connects multiple power nodes with a central node as the center. PID control parameters include proportional coefficient, integral coefficient and differential coefficient. The optimization goal is to make the system respond to the frequency deviation as quickly as possible and control the frequency deviation within the allowable range while meeting the stability requirements. For example, the objective function can be set as the weighted sum of the sum of the square of the frequency deviation and the sum of the square of the control instruction change. The particle swarm optimization algorithm can be used to find the optimal number of nodes, connection mode and PID control parameters of the star topology. Assume that the particle swarm optimization algorithm determines that the optimal number of nodes is 5, the connection mode is fully connected, and the PID control parameters are proportional coefficient 0.8, integral coefficient 0.1, and differential coefficient 0.05. After obtaining the optimized parameters, the configuration of the multi-power collaborative control system needs to be updated. For example, set the communication interface of the central node to 5, and set the parameters of the PID controller to the optimized values. Continuously monitor the communication status and control effect of the system. If the communication delay or control lag exceeds the threshold again, the optimization process is re-triggered to form a closed-loop control. In order to achieve autonomous optimization, the support vector machine algorithm can be used to train historical data and establish a prediction model. For example, the historical communication status data, control effect data and optimization parameters can be used as input, and the optimized star topology and PID control parameters can be used as output to train a support vector machine model. Using this model, the optimal star topology and PID control parameters can be predicted based on the current system status without manual intervention. Assume that the prediction model predicts that the optimal PID parameters are a proportional coefficient of 0.9, an integral coefficient of 0.12, and a differential coefficient of 0.06. The system will automatically update the parameters of the PID controller.According to the optimized star topology and PID control parameters, the control cycle and control gain are dynamically adjusted. The control cycle refers to the time interval for the controller to perform a control operation. The control gain refers to the response sensitivity of the controller to the frequency deviation. When the communication delay and control lag are large, the control cycle and control gain need to be increased to ensure the control effect. On the contrary, the control cycle and control gain need to be reduced to improve the control efficiency. For example, when the communication delay increases to 15ms, the control cycle is adjusted from 10ms to 20ms, and the control gain is adjusted from 1.0 to 1.2. By adaptively adjusting the control cycle and control gain, it can be ensured that the system can maintain stability and dynamic performance under different communication delays and control lags.

[0039] S108. According to the real-time and reliability requirements of the coordinated control of multiple power sources, a particle swarm optimization algorithm is used to perform online adjustment and optimization on the number of nodes, connection mode and PID control parameters of the star topology structure, obtain the optimized star topology structure and PID control parameters, dynamically adjust the control period and control gain of the coordinated control system of multiple power sources, and maintain the stability and dynamic performance of the coordinated control system of multiple power sources by adaptively adjusting the control period and control gain.

[0040] According to the real-time and reliability requirements of the multi-power supply cooperative control system, the operating status data of the current system, including power supply status, load demand, etc., are obtained and used as the input of the particle swarm optimization algorithm. The standard particle swarm optimization algorithm (PSO) is used to establish the optimization model with the number of nodes, connection mode, and proportional, integral, and differential parameters of the PID controller in the star topology as optimization variables. The fitness function is set to consider the stability margin and dynamic response indicators of the system, such as overshoot and adjustment time. Through iterative optimization, the optimal number of nodes, connection mode, and PID control parameter combination are searched to obtain the optimized star topology and control parameters. The optimization results are applied to the multi-power supply cooperative control system. A star communication network is constructed according to the optimized number of nodes and connection mode, and the PID controller parameters are set to the optimized values. According to the optimization results, the fuzzy control method is used to dynamically adjust the control cycle and control gain. A fuzzy rule base between the control cycle and gain and the system stability margin and dynamic performance indicators is established. According to the real-time operation status of the system, the control cycle and gain values ​​are dynamically calculated through fuzzy reasoning to adaptively maintain the stability and dynamic performance of the system. During the adjustment process, the system's operating status is continuously monitored to obtain data such as power supply status and load demand to determine whether the system's stability margin and dynamic performance indicators meet the preset threshold requirements. If the system performance indicators do not meet the requirements for multiple consecutive control cycles, the current status data is re-entered into the particle swarm optimization algorithm for optimization and adjustment again. Set the maximum number of iterative optimization and dynamic adjustment limits. When the maximum number is reached, if the system performance indicators still do not meet the requirements, an alarm message is issued for manual intervention and disposal. If the system performance indicators meet the requirements continuously, it is considered that the optimization and adjustment have reached stability, and the iterative optimization and dynamic adjustment process is stopped, and the normal monitoring mode is switched to.

[0041] For example, the real-time and reliability of the multi-power cooperative control system are crucial. To achieve this goal, a control strategy that can dynamically adjust itself according to the system operation status is required. Here, the particle swarm optimization algorithm is combined with fuzzy control to achieve this goal. First, it is necessary to obtain the real-time operation status data of the system. These data include the output power, voltage, current of each power supply, and the load demand of the entire system, such as the total active power and reactive power. For example, assume a system consisting of three power supplies, power supply A outputs 10MW, power supply B outputs 15MW, power supply C outputs 5MW, and the current total system load demand is 28MW. These data will be used as input to the particle swarm optimization algorithm. The goal of the particle swarm optimization algorithm is to find the optimal combination of star topology and PID control parameters. The star topology refers to a network topology that connects all power supply nodes with a central control node as the center. The connection mode can be fully connected or partially connected. The parameters of the PID controller include proportional coefficient, integral coefficient and differential coefficient. These parameters determine the response speed and stability of the controller to system deviations. For example, a possible optimization goal is to minimize frequency deviation and control energy consumption while meeting the system stability requirements. The design of the fitness function is crucial, as it reflects the optimization goal. Usually, the fitness function takes into account the stability margin and dynamic response indicators of the system, such as overshoot and adjustment time. Assume that the stability margin of the system is required to be greater than 0.1, the overshoot is less than 5%, and the adjustment time is less than 1 second. Then a fitness function can be designed to sum these indicators after multiplying them with weights. Through the iterative search of the particle swarm optimization algorithm, the optimal number of nodes, connection method, and PID control parameter combination of the star topology can be found. For example, the final optimization result may be: a fully connected star topology with 5 nodes, and the PID parameters are proportional coefficient 0.8, integral coefficient 0.1, and differential coefficient 0.05. After obtaining the optimization results, they need to be applied to the actual system. First, the star communication network is rebuilt according to the optimized number of nodes and connection method. Then, the PID controller parameters of each power supply are set to the optimized values. For example, the proportional coefficient of the PID controller of all power supplies is set to 0.8, the integral coefficient is set to 0.1, and the differential coefficient is set to 0.05. In order to further improve the adaptability of the system, the fuzzy control method is used to dynamically adjust the control cycle and control gain. The core of fuzzy control is the fuzzy rule base, which describes the relationship between the control period and gain and the system stability margin and dynamic performance indicators. For example, a fuzzy rule can be: if the system stability margin is low and the dynamic response is slow, increase the control gain and shorten the control period. Assume that the system stability margin is 0.08, which is less than the preset threshold of 0.1, and the dynamic response time is 1.2 seconds, which is greater than the preset threshold of 1 second. Based on the fuzzy rule base, the fuzzy inference engine will calculate the appropriate control period and gain adjustment value.For example, shorten the control cycle from 10ms to 8ms, and increase the control gain from 1.0 to 1.2. During the adjustment process, it is necessary to continuously monitor the operating status of the system. If the system performance indicators do not meet the requirements for multiple consecutive control cycles, re-optimization is required. For example, if the system stability margin is lower than 0.1 for 3 consecutive cycles, the current state data is re-input into the particle swarm optimization algorithm to search for the optimal parameters again. In order to avoid infinite loops, it is necessary to set a maximum number of iterative optimization and dynamic adjustment limits. For example, set the maximum number of iterations to 10 times. If the system performance still does not meet the requirements after reaching the maximum number of times,.

[0042] If the system performance indicators meet the requirements continuously, the system is considered to have reached a stable state, and iterative optimization and dynamic adjustment are stopped.

[0043] S109. Based on the communication optimization results, in the design of the multi-power coordinated frequency regulation control strategy, the frequency regulation performance and the difficulty of engineering implementation are comprehensively considered, and the control strategy parameters and implementation scheme are optimized by combining simulation testing and actual engineering verification. By real-time monitoring of the operating status and performance indicators of the multi-power coordinated frequency regulation control system, the control strategy is continuously optimized to ensure its reliability and effectiveness.

[0044] According to the results of communication optimization, the network topology and communication link quality parameters of the multi-power coordinated frequency modulation control system are obtained, and the objectives and constraints of control strategy optimization are determined, such as frequency modulation response time, frequency modulation accuracy, communication delay, etc. Multiple alternative control strategy parameter combinations are designed, including parameters such as frequency modulation power allocation ratio, frequency modulation dead zone, and frequency modulation rate. For each control strategy parameter combination, a simulation test platform such as MATLAB / Simulink is used to build a simulation model of the multi-power coordinated frequency modulation control system, simulate the operation process of the system under different working conditions, and obtain the simulation results of each frequency modulation performance indicator. Analyze the simulation test results, evaluate the frequency modulation performance of each control strategy parameter combination, and determine whether it meets the optimization objectives and constraints. Add the parameter combinations that meet the requirements to the candidate set, and eliminate the combinations that do not meet the requirements. Among the candidate control strategy parameter combinations, combined with the results of engineering verification, the difficulty of implementing the strategy and the frequency regulation performance are comprehensively considered. The multi-objective optimization algorithm NSGA-II is used to generate the Pareto optimal solution set with the frequency regulation response time and frequency regulation accuracy as the optimization objectives. The optimal parameter combination that balances the difficulty and performance is selected as the final parameter configuration and implementation scheme of the multi-power coordinated frequency regulation control strategy. The optimized multi-power coordinated frequency regulation control strategy is applied to coordinate the frequency regulation power allocation and control command issuance between multiple power sources to achieve fast and accurate adjustment of the grid connection point frequency. A communication network based on the IEC61850 protocol is deployed to monitor the grid connection point frequency, active power and other key indicators of each power source in real time. When the frequency deviation exceeds ±2Hz or the active power fluctuation rate exceeds 5%, the re-optimization process of the control strategy is triggered. In the re-optimization process, according to the real-time operation conditions of the power grid and the latest communication optimization results, the strategy gradient REINFORCE algorithm is used for online learning. Through continuous trial and error and strategy iteration, the multi-power coordinated frequency regulation control strategy is continuously optimized to improve the robustness and dynamic performance of the system and realize the adaptive update of the frequency regulation strategy.

[0045] Exemplarily, the result of communication optimization will provide the network topology and communication link quality parameters of the multi-power coordinated frequency modulation control system. For example, the topology may be a star, ring or mesh structure, and the communication link quality parameters include bandwidth, delay and packet loss rate. These parameters will affect the formulation of the control strategy. The goal of control strategy optimization is to improve the frequency modulation performance, for example, to shorten the frequency modulation response time to less than 200ms, control the frequency modulation accuracy to less than 0.1Hz, and the communication delay to be less than 50ms. The constraint condition is to ensure system stability and consider the limitations of communication resources. Multiple alternative control strategy parameter combinations can be designed. For example, combination 1: the frequency modulation power allocation ratio is hydropower: thermal power: wind power = 4:3:3, the frequency modulation dead zone is 0.02Hz, and the frequency modulation rate is 1MW / s; combination 2: the frequency modulation power allocation ratio is 3:4:3, the frequency modulation dead zone is 0.01Hz, the frequency modulation rate is 2MW / s, etc. These parameter combinations will be evaluated in subsequent simulations. Use MATLAB / Simulink to build a simulation model of a multi-power coordinated frequency modulation control system. The model includes dynamic models of each power source, communication network model and control algorithm module. The operation process of the system under different working conditions is simulated, such as load step change, power output fluctuation, etc. The simulation results of various frequency regulation performance indicators, such as frequency regulation response time, frequency regulation accuracy, communication delay, etc. are recorded and analyzed. Assume that the frequency regulation response time of combination 1 is 180ms, the frequency regulation accuracy is 0.08Hz, and the communication delay is 40ms, which meets the optimization objectives and constraints; while the frequency regulation response time of combination 2 is 150ms, but the frequency regulation accuracy is 0.12Hz, which exceeds the set target, then combination 2 is eliminated. Among the candidate control strategy parameter combinations, a comprehensive evaluation is conducted in combination with the results of engineering verification, such as test data in the actual power grid. Assume that combination 1 performs stably in actual tests and has low implementation difficulty, while the performance improvement of other candidate combinations is limited, but the implementation difficulty is high. The multi-objective optimization algorithm NSGA-II is used to generate the Pareto optimal solution set with frequency regulation response time and frequency regulation accuracy as optimization objectives. For example, one solution in the solution set is a frequency modulation response time of 170ms and a frequency modulation accuracy of 0.09Hz; another solution is a frequency modulation response time of 190ms and a frequency modulation accuracy of 0.07Hz. Finally, combination 1 with a frequency modulation response time of 180ms and a frequency modulation accuracy of 0.08Hz was selected as the final parameter configuration scheme because it balances the implementation difficulty and frequency modulation performance. The optimized multi-power coordinated frequency modulation control strategy is applied. When the system detects that the frequency deviation exceeds ±2Hz or the active power fluctuation rate exceeds 5%, the re-optimization process is triggered. For example, at a certain moment, the load suddenly increases, causing the frequency to drop to 49.8Hz, triggering the re-optimization process. In the re-optimization process, the policy gradient REINFORCE algorithm is used for online learning. For example, the algorithm will try different frequency modulation power allocation ratios and reward or punish according to the changes in frequency deviation.Through continuous trial and error and strategy iteration, the algorithm will gradually learn the optimal control strategy, such as increasing the frequency modulation power allocation ratio of hydropower to 50% to restore frequency faster. Continuously optimize the multi-power coordinated frequency modulation control strategy, improve the robustness and dynamic performance of the system, and realize the adaptive update of the frequency modulation strategy.

[0046] S1010. Based on the real-time operating conditions of the power grid and the latest communication optimization results, the strategy gradient REINFORCE algorithm is used for online learning. Through continuous trial and error and strategy iteration, a continuously optimized multi-power coordinated frequency regulation control strategy is obtained, the frequency regulation power allocation and control command issuance plan are determined, and fast and accurate frequency regulation of the grid connection point is achieved.

[0047] The real-time operation status data of the power grid and the communication optimization results are obtained as the input of the reinforcement learning algorithm to initialize the policy network parameters. The policy network adopts a multi-layer perceptron structure, with the input as the power grid status characteristics and the output as the probability distribution of the multi-power coordinated frequency regulation control strategy. According to the current policy network parameters, a batch of multi-power coordinated frequency regulation control strategies are sampled and generated. Each strategy contains information such as the frequency regulation power allocation and control instruction issuance time of each power source. This batch of strategies are tested in the power grid simulation environment, and evaluation indicators such as the frequency deviation of the grid connection point and the frequency regulation response time are obtained to calculate the reward value of each strategy. The reward value considers factors such as the sum of the absolute values ​​of the frequency deviation and the weighted average of the frequency regulation response time. The policy network parameters are updated using the policy gradient method. According to the reward value and gradient estimation value of each strategy, the policy gradient is calculated, and the weight of the policy network is updated using the gradient ascent method to obtain the optimized policy network. Repeat steps 2-4 for multiple rounds of policy iterative optimization until the policy network converges or reaches the preset number of iterations. The convergence judgment is based on the variance of the policy reward value being less than the preset threshold. The optimal multi-power coordinated frequency regulation control strategy is sampled from the converged strategy network and applied to the actual operation environment of the power grid. The actual control effect of the strategy is evaluated by monitoring indicators such as the frequency deviation of the grid connection point. If the actual effect does not meet expectations, that is, the frequency deviation exceeds the allowable range or the frequency regulation time is too long, the real-time operation status data of the power grid and the evaluation results are fed back to the reinforcement learning model to trigger a new round of online learning and strategy optimization. Repeat steps 1-6 to make the control strategy continuously adapt to changes in the operation conditions of the power grid.

[0048] Exemplarily, reinforcement learning is a machine learning method that allows an agent to learn optimal behavior by interacting with its environment. The agent receives rewards or penalties to guide its learning process. In this example, reinforcement learning is used to optimize the multi-power coordinated frequency regulation control strategy. First, the real-time operation status data of the power grid and the communication optimization results are obtained. For example, the real-time operation status data of the power grid includes the current frequency, load power, and the output of each power source. The communication optimization results include network topology, communication bandwidth, delay, and packet loss rate. These data will be used as inputs to the reinforcement learning algorithm. At the same time, the policy network parameters are initialized. The policy network adopts a multi-layer perceptron structure, whose input is the power grid state characteristics, and the output is the probability distribution of the multi-power coordinated frequency regulation control strategy. For example, the input of the policy network can be frequency deviation, load change rate, power output, etc. The output can be the frequency regulation power allocation ratio of different power sources. Next, according to the current policy network parameters, a batch of multi-power coordinated frequency regulation control strategies are sampled and generated. Each strategy contains information such as the frequency regulation power allocation of each power source and the time when the control instruction is issued. For example, one strategy can be that the frequency modulation power allocation ratio of hydropower: thermal power: wind power is 4:3:3, and the control instruction issuance time is the current moment plus 5 milliseconds. Another strategy can be that the frequency modulation power allocation ratio of hydropower: thermal power: wind power is 3:4:3, and the control instruction issuance time is the current moment plus 10 milliseconds, and so on. Then, test this batch of strategies in the power grid simulation environment. The power grid simulation environment is a software platform that simulates the operation of a real power grid. It can simulate various operating conditions, such as load step changes, power output fluctuations, etc. In the simulation environment, the performance of each strategy can be evaluated. Obtain evaluation indicators such as grid connection point frequency deviation and frequency modulation response time, and calculate the reward value of each strategy. The reward value considers factors such as the sum of the absolute values ​​of the frequency deviation and the weighted average of the frequency modulation response time. For example, if a strategy makes the frequency deviation smaller and the frequency modulation response time shorter, its reward value is higher. Conversely, if a strategy makes the frequency deviation larger and the frequency modulation response time longer, its reward value is lower. Afterwards, the policy gradient method is used to update the policy network parameters. The policy gradient method is a gradient-based optimization algorithm. It calculates the policy gradient according to the reward value and gradient estimate of each policy, and updates the weight of the policy network using the gradient ascent method to obtain the optimized policy network. This is done to enable the policy network to generate better policies. Repeat the policy iteration optimization process and perform multiple rounds of policy iteration optimization until the policy network converges or reaches the preset number of iterations. The convergence judgment is based on the variance of the policy return value being less than the preset threshold. This means that the policy network has learned a stable policy. The optimal multi-power coordinated frequency regulation control strategy is sampled from the converged policy network and applied to the actual operation environment of the power grid. The actual control effect of the strategy is evaluated by monitoring indicators such as the frequency deviation of the grid connection point.If the actual effect does not meet expectations, that is, the frequency deviation exceeds the allowable range or the frequency regulation time is too long, the real-time operation status data of the power grid and the evaluation results are fed back to the reinforcement learning model to trigger a new round of online learning and strategy optimization. This is a process of continuous learning and optimization to ensure that the control strategy can adapt to changes in the operating conditions of the power grid. For example, suppose that the load of the power grid suddenly increases at a certain moment, causing the frequency to drop. At this time, the reinforcement learning model will generate a new frequency regulation control strategy based on the current power grid state. This strategy may increase the output of hydropower to quickly restore the frequency. If the actual control effect of the strategy is good, the reinforcement learning model will retain the strategy. If the actual control effect is not good, the reinforcement learning model will re-learn and optimize to generate a new strategy. This process is repeated, and finally an optimal control strategy that can adapt to various operating conditions is obtained.

[0049] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. A multi-power supply coordinated frequency modulation control method, characterized in that: include: Acquire real-time grid operation data, including system frequency deviation, load forecast information, and real-time output and regulation capacity data of wind power, photovoltaic power, hydropower, and thermal power; According to the real-time operation data of the power grid, the optimal output allocation plan of multiple power sources is calculated by a mixed integer linear programming algorithm, with minimization of power generation cost as the objective function and upper and lower limits of multiple power outputs, ramp rate and system frequency deviation limit as constraints; Convert the optimal output allocation plan into a power output adjustment instruction, and send it to each power control system through the power grid dispatching automation system; Continuously monitor the operating status and performance indicators of the multi-power coordinated frequency regulation control system, and continuously optimize the multi-power coordinated frequency regulation control strategy through online learning and strategy iteration to improve the effect of grid frequency control.

2. The multi-power coordinated frequency modulation control method according to claim 1, characterized in that: After obtaining the real-time operation data of the power grid, the method further includes: The real-time operation data of the power grid is preprocessed and feature extracted, key indicators such as the system frequency deviation mean and load forecast error are calculated, and multi-power data is converted into a standardized input format.

3. The multi-power source coordinated frequency modulation control method according to claim 1, characterized in that: Before calculating the optimal output distribution scheme of multiple power sources according to the real-time operation data of the power grid, the method further includes: The decision tree algorithm is used to classify the operating status of the power grid to determine whether the current status meets the requirements of frequency stability and economic operation. If not, the multi-power output optimization adjustment process is triggered.

4. The multi-power coordinated frequency modulation control method according to claim 1, characterized in that: After converting the optimal output allocation scheme into a power output adjustment instruction, the method further includes: During the power output adjustment process, power simulation software is used to monitor the grid operation status and system frequency changes in real time. If the frequency deviation exceeds the preset threshold, the emergency frequency regulation strategy based on fuzzy PID control is triggered. The system frequency stability is quickly restored by adjusting the power output and load power limitation.

5. The multi-power coordinated frequency modulation control method according to claim 1, characterized in that: After the continuous monitoring of the operating status and performance indicators of the multi-power supply coordinated frequency modulation control system, the method further includes: The support vector machine algorithm is used to train the historical communication status data, control effect data and optimization parameters to establish a prediction model. The prediction model is used to achieve autonomous optimization of the star topology structure and PID control parameters without human intervention.

6. The multi-power source coordinated frequency modulation control method according to claim 1, characterized in that: Acquiring real-time operation data of the power grid, including: System frequency deviation and real-time output data of each power source, as well as power source regulation capability data and load forecast result data, perform data cleaning and feature extraction on the real-time operation data of the power grid, power source regulation capability data and load forecast result data, and generate comprehensive evaluation data through data fusion technology; Inputting the comprehensive evaluation data into a Q-learning reinforcement learning model, and regularly updating the power grid operation optimization model by iteratively updating the Q value table; According to the power supply operation status evaluation results, the optimization objective function and constraint conditions of the multi-power supply coordinated frequency modulation control are determined, and by solving the optimization problem, the optimal frequency modulation control strategy of each power supply is obtained. The optimal frequency modulation control strategy includes AGC instructions and primary frequency modulation power allocation.

7. The multi-power coordinated frequency modulation control method according to claim 1, characterized in that: When continuously optimizing the multi-power source coordinated frequency regulation control strategy, the strategy gradient REINFORCE algorithm is used for online learning. Through continuous trial and error and strategy iteration, a continuously optimized multi-power source coordinated frequency regulation control strategy is obtained, the frequency regulation power allocation and control instruction issuance plan are determined, and fast and accurate adjustment of the grid connection point frequency is achieved.

8. The multi-power source coordinated frequency modulation control method according to claim 1, characterized in that: Obtain real-time operating status data of multi-power coordinated control systems, including; Power supply status, load demand, etc., which are used as inputs to the particle swarm optimization algorithm; The standard particle swarm optimization algorithm is used to establish the optimization model with the star topology parameters and PID controller parameters as optimization variables; According to the optimization results, the fuzzy control method is used to dynamically adjust the control period and control gain; If the system performance indicators do not meet the requirements for multiple consecutive control cycles, the current state data will be re-input into the particle swarm optimization algorithm for optimization and adjustment again.

9. The multi-power source coordinated frequency modulation control method according to any one of claims 1 to 8, characterized in that: The power grid dispatching automation system and each power supply control system use a communication network based on the IEC61850 protocol to achieve real-time transmission and interaction of power grid operation information and control instructions.

Images