Thermal power unit deep peak shaving and primary frequency modulation, power generation control system optimization method

Abstract

The application provides a thermal power unit deep peak regulation and primary frequency modulation and power generation control system optimization method, and relates to the technical field of power system control and smart grid. The thermal power unit deep peak regulation and primary frequency modulation and power generation control system optimization method comprises the following steps: S1: based on historical load data, weather data and holiday characteristic data of a power grid, original data is cleaned, normalized and time series reconstructed, and a load prediction data set is constructed; on this basis, a long short-term memory neural network prediction method is used to establish a power grid short-term load prediction model. Through joint arrangement of historical load, weather information and holiday characteristics of the power grid and formation of a time series structure, the load change law is more completely described, the power grid short-term load trend forms a continuous prediction curve, and the scheduling decision has an advance; on the basis of the load change trend, historical operation data and equipment operation parameters of the unit are introduced.

Description

Technical Field

[0001] This invention relates to the field of power system control and smart grid technology, specifically to a method for optimizing the control system of deep peak shaving and primary frequency regulation of thermal power units and power generation. Background Technology

[0002] The field of power system control and smart grid technology involves the comprehensive management of power system dispatching, control, protection, and optimization using modern information technology, automation control, communication technology, and intelligent equipment. This field primarily addresses how to achieve efficient, reliable, and safe operation of the power grid, especially in the face of fluctuating load demand, the volatility of renewable energy, and the integrated utilization of various power resources. Specifically, deep peak shaving and primary frequency regulation optimization methods for thermal power units refer to the dynamic operation modeling and optimization of thermal power units to coordinate the generation of thermal power units with the grid frequency, thereby achieving multiple objectives such as grid stability, load forecasting, and improved unit efficiency.

[0003] Traditional power dispatching relies primarily on historical load curves and manual experience for dispatching decisions. Load change trend analysis often remains at the level of simple statistics or short-cycle experience. External factors such as weather changes and holiday behavior characteristics are often not systematically analyzed, leading to potential biases in load forecasting. When sudden load fluctuations occur in the power grid, dispatching responses lack lead time. For example, during hot summer months, when air conditioning loads surge, experience-based forecasting can lead to delayed judgments, causing short-term frequency fluctuations in unit load regulation. On the other hand, thermal power unit operation control typically revolves around single output or single frequency control. The coupling relationship between boiler combustion, turbine steam parameters, and generator output lacks a unified description at the dispatching level. The unit output regulation process struggles to accurately reflect the dynamic characteristics of the equipment. When units are under low load or in deep peak-shaving ranges, combustion stability and steam parameter fluctuations are prone to occur, consequently affecting frequency regulation performance and unit efficiency. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method for optimizing the control system of deep peak shaving, primary frequency regulation, and power generation of thermal power units. This method solves the problems of traditional power dispatching and operation methods that mainly rely on historical load curves and manual experience for dispatching judgments. Load change trend analysis is mostly limited to simple statistics or short-cycle experience judgments. External factors such as weather changes and holiday behavior characteristics are often not systematically analyzed, resulting in deviations in load forecasting results and a lack of lead time in dispatching response when the power grid experiences sudden load fluctuations.

[0005] To achieve the above objectives, the present invention provides a method for optimizing the deep peak shaving, primary frequency regulation, and power generation control system of thermal power units, comprising the following steps:

[0006] S1: Based on historical power grid load data, meteorological data, and holiday characteristic data, the raw data is cleaned, normalized, and reconstructed using time series to construct a load forecast dataset. On this basis, a short-term power grid load forecast model is established using a long short-term memory neural network forecasting method, and the trend of power grid load change within a preset time window is predicted and analyzed to generate a short-term load forecast curve.

[0007] S2: Based on the short-term load forecast curve generated by S1, as well as the historical operating data of thermal power units, equipment operating parameters and unit output data, a dynamic operating model of thermal power units is established using the support vector machine modeling method. The dynamic coupling relationship between the boiler system, turbine system and generator system is modeled and analyzed to generate a dynamic operating model of thermal power units.

[0008] S3: Based on the short-term load forecast curve generated in S1 and the dynamic operation model of thermal power units generated in S2, a multi-objective optimization function for deep peak shaving and primary frequency regulation is constructed. The optimization function includes the objectives of minimizing grid frequency deviation, minimizing unit fuel consumption, and maximizing unit operation stability. The multi-objective optimization function is solved using a non-dominated sorting genetic optimization method to generate a multi-objective optimization model.

[0009] S4: Based on the dynamic operation model of the thermal power unit generated in S2 and the multi-objective optimization model generated in S3, combined with the minimum technical output constraint, ramp rate constraint and grid frequency deviation constraint, the model predictive control method is used to calculate the coordinated control strategy of deep peak shaving and primary frequency regulation of the thermal power unit, so as to predict the changes in unit output within the future time window and perform dynamic optimization adjustment, and generate the coordinated control strategy of unit peak shaving and frequency regulation.

[0010] S5: Based on the unit peak shaving and frequency regulation coordinated control strategy generated by S4, construct the power control loop of the thermal power unit, and use the particle swarm optimization method to optimize the parameters of the adaptive proportional-integral-derivative controller in order to improve the power regulation response speed and stability of the unit under load fluctuation conditions, and generate an optimized power control parameter set;

[0011] S6: Based on the optimized power control parameter set generated by S5 and the real-time frequency monitoring data of the power grid, a frequency regulation strategy learning environment is constructed, and a deep Q-network learning method is used to train and optimize the primary frequency regulation control strategy of the unit, so that the unit can adaptively adjust the power output according to the frequency deviation change and generate an adaptive frequency regulation control strategy.

[0012] S7: Based on the dynamic operation model of the thermal power unit generated by S2 and the adaptive frequency regulation control strategy generated by S6, and combined with the real-time operation data collected by the unit's field sensors, a digital twin operation model of the thermal power unit is constructed, and dynamic prediction and simulation analysis of the future operation status of the unit are performed to generate predicted data of the unit's operation status.

[0013] S8: Based on the short-term load forecast curve generated by S1, the unit peak-shaving and frequency regulation coordinated control strategy generated by S4, and the unit operation status forecast data generated by S7, combined with the new energy power generation output data and energy storage system operation data, the mixed integer linear programming optimization method is used to optimize the coordinated scheduling of thermal power units and new energy systems, and generate a multi-energy coordinated scheduling scheme.

[0014] S9: Based on the optimized power control parameter set generated in S5, the adaptive frequency regulation control strategy generated in S6, and the multi-energy collaborative dispatch scheme generated in S8, the real-time feedback control method is used to optimize and adjust the power generation control system of thermal power units online, so as to realize the dynamic control of power generation of thermal power units under deep peak shaving and primary frequency regulation operation conditions, and generate the optimized power generation control strategy.

[0015] Preferably, step S1 includes the following steps:

[0016] S101: Based on historical load data of the power grid, meteorological data and holiday characteristic data, data cleaning and normalization methods are adopted. The original load data is preprocessed by imputing missing values, removing outliers and linear normalization to generate a cleaned and normalized load dataset.

[0017] S102: Based on the cleaned and normalized load dataset, a time series reconstruction method is used to form the input and output sequences of the prediction model through sliding window segmentation and feature construction, thereby generating a time series load dataset.

[0018] S103: Based on the time series load dataset, a long short-term memory neural network prediction method is used for model training and prediction to obtain the power grid load change trend within a preset time window in the future and generate a short-term load prediction curve.

[0019] Preferably, step S2 includes the following steps:

[0020] S201: Based on historical operating data of thermal power units, equipment operating parameters and unit output data, data cleaning and feature extraction methods are used to standardize the unit operating indicators, and key features such as boiler temperature, turbine speed and generator power are extracted to generate a standardized unit feature dataset.

[0021] S202: Based on the standardized unit feature dataset and the short-term load forecast curve, a dynamic operation model of the thermal power unit is established using the support vector machine modeling method to describe the dynamic coupling relationship between the boiler system, turbine system and generator system, and to generate the dynamic operation model of the thermal power unit.

[0022] Preferably, step S3 includes the following steps:

[0023] S301: Based on the short-term load forecast curve and the dynamic operation model of the thermal power unit, construct a multi-objective optimization function that includes minimizing grid frequency deviation, minimizing unit fuel consumption, and maximizing unit operation stability, and generate the multi-objective optimization function.

[0024] S302: Based on the multi-objective optimization function, a non-dominated sorting genetic algorithm is used to solve the multi-objective optimization problem. The optimal solution set is obtained through population initialization, crossover mutation, and non-dominated sorting selection, thereby generating a multi-objective optimization model.

[0025] Preferably, step S4 includes the following steps:

[0026] S401: Based on the dynamic operation model of the thermal power unit and the multi-objective optimization model, and combined with the minimum technical output constraint, ramp rate constraint and grid frequency deviation constraint, construct the model predictive control problem and generate the model predictive control problem model;

[0027] S402: Based on the model predictive control problem model, the model predictive control method is used to perform rolling optimization calculations to determine the unit output regulation strategy within the future time window and generate a unit peak-shaving and frequency regulation coordinated control strategy.

[0028] Preferably, step S5 includes the following steps:

[0029] S501: Based on the unit peak shaving and frequency regulation coordinated control strategy, construct a unit power control loop model containing a proportional-integral-derivative control structure, and generate a power control loop model;

[0030] S502: Based on the power control loop model, the particle swarm optimization method is used to optimize the proportional, integral and derivative parameters to improve the power regulation response speed and stability of the unit, and generate an optimized power control parameter set.

[0031] Preferably, step S6 includes the following steps:

[0032] S601: Based on the optimized power control parameter set and real-time grid frequency monitoring data, construct a frequency regulation strategy learning environment, including state space, action space and reward function, and generate a frequency regulation strategy training environment;

[0033] S602: Based on the frequency modulation strategy training environment, a deep Q-network learning method is used for iterative training of the strategy. The strategy is optimized through experience playback and target network update to generate an adaptive frequency modulation control strategy.

[0034] Preferably, step S7 includes the following steps:

[0035] S701: Based on the dynamic operation model of the thermal power unit and the adaptive frequency regulation control strategy, and combined with the real-time operation data collected by the unit's field sensors, a digital mapping relationship between the unit's physical system and control system is constructed to generate a digital twin unit model.

[0036] S702: Based on the aforementioned digital twin unit model, a simulation prediction method is used to simulate and analyze the future operating status of the unit, generating predicted data for the unit's operating status.

[0037] Preferably, step S8 includes the following steps:

[0038] S801: Based on the short-term load forecast curve, the unit peak-shaving and frequency regulation coordinated control strategy, and the unit operation status forecast data, and combined with the new energy power generation output data and energy storage system operation data, a coordinated scheduling optimization model is constructed to generate a multi-energy coordinated scheduling optimization model.

[0039] S802: Based on the multi-energy coordinated scheduling optimization model, a mixed integer linear programming method is used to optimize and solve the problem, generating a multi-energy coordinated scheduling scheme.

[0040] Preferably, step S9 includes the following steps:

[0041] S901: Based on the optimized power control parameter set, the adaptive frequency regulation control strategy, and the multi-energy collaborative dispatch scheme, construct a dynamic control model for the power generation of thermal power units and generate a dynamic control model for power generation.

[0042] S902: Based on the aforementioned dynamic power generation control model, a real-time feedback control method is used to adjust the unit's output power online, generating an optimized power generation control strategy.

[0043] This invention provides a method for optimizing the deep peak shaving, primary frequency regulation, and power generation control system of thermal power units. It has the following beneficial effects:

[0044] This invention, by jointly organizing historical load data, meteorological information, and holiday characteristics of the power grid into a time series structure, provides a more complete characterization of load change patterns and forms a continuous forecast curve for short-term power grid load trends, enabling dispatch decisions to have advance notice. Based on load change trends, it introduces historical unit operating data and equipment operating parameters to quantify the dynamic coupling relationship between boilers, turbines, and generators, forming a calculable expression for unit output changes and establishing a quantifiable correlation between unit operating status and external load fluctuations. Based on load forecast results and unit dynamic behavior, it constructs a joint optimization objective including frequency deviation, fuel consumption, and operational stability, and solves this multi-objective problem using a population evolutionary search method, achieving a coordinated balance between different operational objectives and avoiding operational bias caused by a single adjustment index. Under unit output and frequency constraints, it predicts future time windows and dynamically adjusts them to ensure consistent unit output. The system possesses forward-looking regulation capabilities, forming a synergistic response structure between deep peak shaving and primary frequency regulation. By performing group search optimization of power control parameters and combining them with real-time frequency changes to form an adaptive frequency regulation strategy, the power regulation speed and stability are improved simultaneously. By constructing an operation mapping structure based on real-time operating data and predicting and analyzing future operating states, the system continuously monitors unit operating trends and equipment status changes. With the participation of load forecasting information, unit output change trends, and new energy and energy storage operation data, collaborative scheduling optimization is carried out, forming a dynamic complementary relationship between thermal power units and multi-energy systems. Under the condition of real-time frequency data feedback, the power generation control strategy is continuously corrected, enabling the units to maintain a stable output change trajectory in deep peak shaving and frequency regulation environments. This achieves a continuous synergistic relationship between load forecasting, unit dynamic modeling, multi-objective scheduling, and real-time feedback control, simultaneously improving grid frequency stability, unit operating economy, and regulation response speed. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of the main steps of the present invention;

[0046] Figure 2 This is a detailed schematic diagram of S1 of the present invention;

[0047] Figure 3 This is a detailed schematic diagram of S2 of the present invention;

[0048] Figure 4 This is a detailed schematic diagram of S3 of the present invention;

[0049] Figure 5 This is a detailed schematic diagram of S4 of the present invention;

[0050] Figure 6 This is a detailed schematic diagram of S5 of the present invention;

[0051] Figure 7This is a detailed schematic diagram of S6 of the present invention;

[0052] Figure 8 This is a detailed schematic diagram of S7 of the present invention;

[0053] Figure 9 This is a detailed schematic diagram of S8 of the present invention;

[0054] Figure 10 This is a detailed schematic diagram of S9 of the present invention. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] Example:

[0057] like Figure 1-10 As shown, this embodiment of the invention provides a method for optimizing the deep peak shaving, primary frequency regulation, and power generation control system of thermal power units, including the following steps:

[0058] S1: Based on historical power grid load data, meteorological data, and holiday characteristic data, the raw data is cleaned, normalized, and reconstructed using time series to construct a load forecast dataset. On this basis, a short-term power grid load forecast model is established using a long short-term memory neural network forecasting method, and the trend of power grid load change within a preset time window is predicted and analyzed to generate a short-term load forecast curve.

[0059] S2: Based on the short-term load forecast curve generated by S1, as well as the historical operating data of thermal power units, equipment operating parameters and unit output data, a dynamic operating model of thermal power units is established using the support vector machine modeling method. The dynamic coupling relationship between the boiler system, turbine system and generator system is modeled and analyzed to generate a dynamic operating model of thermal power units.

[0060] S3: Based on the short-term load forecast curve generated in S1 and the dynamic operation model of thermal power units generated in S2, a multi-objective optimization function for deep peak shaving and primary frequency regulation is constructed. The optimization function includes the objectives of minimizing grid frequency deviation, minimizing unit fuel consumption, and maximizing unit operation stability. The multi-objective optimization function is solved using a non-dominated sorting genetic optimization method to generate a multi-objective optimization model.

[0061] S4: Based on the dynamic operation model of the thermal power unit generated in S2 and the multi-objective optimization model generated in S3, combined with the minimum technical output constraint, ramp rate constraint and grid frequency deviation constraint, the model predictive control method is used to calculate the coordinated control strategy of deep peak shaving and primary frequency regulation of the thermal power unit, so as to predict the changes in unit output within the future time window and perform dynamic optimization adjustment, and generate the coordinated control strategy of unit peak shaving and frequency regulation.

[0062] S5: Based on the unit peak shaving and frequency regulation coordinated control strategy generated by S4, construct the power control loop of the thermal power unit, and use the particle swarm optimization method to optimize the parameters of the adaptive proportional-integral-derivative controller in order to improve the power regulation response speed and stability of the unit under load fluctuation conditions, and generate an optimized power control parameter set;

[0063] S6: Based on the optimized power control parameter set generated by S5 and the real-time frequency monitoring data of the power grid, a frequency regulation strategy learning environment is constructed, and a deep Q-network learning method is used to train and optimize the primary frequency regulation control strategy of the unit, so that the unit can adaptively adjust the power output according to the frequency deviation change and generate an adaptive frequency regulation control strategy.

[0064] S7: Based on the dynamic operation model of the thermal power unit generated by S2 and the adaptive frequency regulation control strategy generated by S6, and combined with the real-time operation data collected by the unit's field sensors, a digital twin operation model of the thermal power unit is constructed, and dynamic prediction and simulation analysis of the future operation status of the unit are performed to generate predicted data of the unit's operation status.

[0065] S8: Based on the short-term load forecast curve generated by S1, the unit peak-shaving and frequency regulation coordinated control strategy generated by S4, and the unit operation status forecast data generated by S7, combined with the new energy power generation output data and energy storage system operation data, the mixed integer linear programming optimization method is used to optimize the coordinated scheduling of thermal power units and new energy systems, and generate a multi-energy coordinated scheduling scheme.

[0066] S9: Based on the optimized power control parameter set generated in S5, the adaptive frequency regulation control strategy generated in S6, and the multi-energy collaborative dispatch scheme generated in S8, the real-time feedback control method is used to optimize and adjust the power generation control system of thermal power units online, so as to realize the dynamic control of power generation of thermal power units under deep peak shaving and primary frequency regulation operation conditions, and generate the optimized power generation control strategy.

[0067] S1 includes the following steps:

[0068] S101: Based on historical load data of the power grid, meteorological data and holiday characteristic data, data cleaning and normalization methods are adopted. The original load data is preprocessed by imputing missing values, removing outliers and linear normalization to generate a cleaned and normalized load dataset.

[0069] First, raw data is extracted from historical load data, meteorological data, and holiday characteristic data of the power grid. This data often contains missing values ​​and outliers. To process this data, data cleaning methods are first used to preprocess the raw data, including missing value imputation and outlier removal. Missing value imputation can be handled by linear interpolation. Assuming that the load data at a certain moment is missing, a reasonable value for that moment can be calculated from the load data at the preceding and following moments. For example, if the load value at a certain moment is missing, it can be filled by averaging the load values ​​at the preceding and following moments. Outlier removal is done by calculating the standard deviation of each data point and considering data points exceeding the preset standard deviation range as outliers and removing them. Next, linear normalization is used to normalize all load data, so that the range of data values ​​is unified to a standard interval, such as [0,1]. Implementation, in which This is the original data. and These are the minimum and maximum values ​​in the dataset, respectively. After the above processing, the cleaned load dataset is obtained, which can be used as the input dataset for further modeling.

[0070] S102: Based on the cleaned and normalized load dataset, a time series reconstruction method is used to form the input and output sequences of the prediction model through sliding window segmentation and feature construction, thereby generating a time series load dataset.

[0071] Based on a cleaned and normalized load dataset, a time series reconstruction method is used for feature construction. The entire dataset is divided into multiple time periods using a sliding window segmentation method. The window size is typically set according to actual needs; assuming a 24-hour window is chosen, the data within each window is used to predict subsequent load changes. For example, assuming we have data from the past week, a 24-hour sliding window can be used to obtain the load change characteristics for each day, with each window forming an input-output pair, where the input is the data from the past 24 hours and the output is the upcoming load value. Next, feature construction is performed on the data within each time window, such as extracting statistical features like daily load peak, average, and variance, as well as meteorological data (e.g., temperature, humidity) and holiday information, ultimately forming a structured time series dataset. The core purpose of this process is to transform historical data into a format suitable for predictive models, making subsequent training and prediction more efficient and accurate.

[0072] S103: Based on the time series load dataset, a long short-term memory neural network prediction method is used for model training and prediction to obtain the power grid load change trend within a preset time window in the future and generate a short-term load prediction curve.

[0073] A Long Short-Term Memory (LSTM) neural network is used to train and predict time-series load datasets. First, the time-series dataset is divided into training and validation sets. The training set is used for model learning, and the validation set is used to test the model's performance. LSTM neural networks effectively identify long-term trends and short-term fluctuations by capturing the time dependencies of the data. Specifically, the network weights are adjusted using the backpropagation algorithm, enabling the model to learn patterns of load changes from historical load data. During training, the loss function of the LSTM network can be optimized using the mean squared error (MSE), as shown in the formula: ,in For the true value, For predicted values, The number of data points is denoted by . By iteratively optimizing the network weights, an LSTM model capable of predicting future load changes is obtained. The output of this model is the load forecast value within a preset time window, thus generating a short-term load forecast curve.

[0074] S2 includes the following steps:

[0075] S201: Based on historical operating data of thermal power units, equipment operating parameters and unit output data, data cleaning and feature extraction methods are used to standardize the unit operating indicators, and key features such as boiler temperature, turbine speed and generator power are extracted to generate a standardized unit feature dataset.

[0076] The data sources for thermal power unit historical monitoring records, equipment status data, and unit output capacity information are analyzed. First, multiple monitoring sequence data, including boiler-side temperature records, turbine shaft rotation records, and generator-end power records generated within a continuous operating cycle, are retrieved from the power plant operation information platform. A unified data table structure is then established in the data management software environment according to unit number and time stamp. During the data cleaning phase, each collected record is checked by setting anomaly identification threshold range; for example, temperature ranges are constructed within the boiler main steam temperature sequence. ,in Represents the temperature value at a specific sampling time, with the subscript... Represents a time series index. and The range is determined by long-term statistical intervals of similar generating units. When a temperature record exceeds this range, it is either removed or interpolated near the nearest neighbor in the data processing program. A linear expression is used in the interpolation calculation. Supplement missing records, including and These represent adjacent sampled values, respectively. Then, the operating indicators of different dimensions are standardized, and the standardized expressions are used in the computing environment. To achieve scale uniformity, among which This is the original monitoring data. This represents the mean of similar monitoring sequences. The standard deviation is represented by the mean and dispersion interval calculated using statistical software within the sample window. For example, if the mean of the boiler temperature series is within the normal operating range and the fluctuation range is stable, then feature extraction is performed on the standardized variables. During the feature selection stage, the boiler temperature change rate, turbine speed stability, and generator power fluctuation range are calculated. The temperature change rate can be obtained through... get, The sampling time interval is represented by the comparison of multiple time segment change rate sequences and the selection of representative index combinations according to the stable interval, which ultimately forms a unified set of unit operation characteristic data.

[0077] S202: Based on the standardized unit feature dataset and the short-term load forecast curve, a dynamic operation model of the thermal power unit is established using the support vector machine modeling method to describe the dynamic coupling relationship between the boiler system, turbine system and generator system, and to generate the dynamic operation model of the thermal power unit.

[0078] The existing set of unit operation characteristic data and grid load forecast curves are decomposed and processed. First, the load change sequence within the forecast time interval is read from the dispatch data platform and then matched and integrated with the unit characteristic data according to the time index to form a training sample matrix. During the sample matrix construction process, boiler temperature characteristics, turbine speed characteristics, and generator power characteristics are used as input variable vectors. subscript This represents the number of characteristics, while using the unit output power or load response value as the target variable. Subsequently, during the modeling phase, a mapping relationship is established through the support vector machine calculation process. In the calculation steps, the kernel function matrix is ​​first constructed. ,in and Represents the feature vectors of any two samples. Represents Euclidean distance, parameters These are the kernel function coefficients, whose values ​​are determined through an interval search method, for example, by gradually adjusting them over a range of orders of magnitude and using an error function. For comparison, in the formula This is the actual output value. For model calculation results, subscript The sample number is used to calculate the error range under different parameter combinations, and the coefficient combination with the smaller error range is selected as the modeling parameter. During the constraint solution stage, a quadratic programming expression is used. Perform iterative solution, where For the model weight vector, Indicates the penalty coefficient. This represents a slack variable, which is adjusted gradually over several intervals. The error distribution of the training samples is calculated to complete the parameter setting. After obtaining the parameter combination, the boiler system variables, turbine system variables and generator system variables are input into the model to calculate their response results. The mathematical description structure of unit operation that reflects the coupling relationship of the three types of systems is formed through iterative training using continuous operating samples.

[0079] S3 includes the following steps:

[0080] S301: Based on the short-term load forecast curve and the dynamic operation model of the thermal power unit, construct a multi-objective optimization function that includes minimizing grid frequency deviation, minimizing unit fuel consumption, and maximizing unit operation stability, and generate the multi-objective optimization function.

[0081] After obtaining the short-timescale load change curve and the mathematical expression structure of the dynamic response of the thermal power unit, the power sequence of adjacent sampling periods in the curve is discretized, and the load power is denoted as... ,in Indicates the first Each scheduling time node This represents the system power demand value corresponding to this node, and the dynamic operation relationship of the unit is set as follows: ,in Indicates the unit's output power. Indicates the unit speed variable. This represents the control input. Load data from several consecutive scheduling cycles is selected as an example sample in a power grid operation scenario. For instance, the load sequence of a certain regional power grid during the early morning to mid-morning period is in the range of approximately several hundred megawatts to nearly one gigawatt. The system frequency deviation is represented as... ,in Indicates the first The system frequency measurement at time [time]. This represents the reference frequency value, which is typically set in a stable range near the rated frequency in the scheduling rules. It is calculated by taking the absolute value of the deviation for each scheduling period. Obtain the frequency deviation index and introduce a quadratic function into the fuel consumption item. , where the coefficient , , Obtained by fitting historical combustion efficiency curves, the example is set with parameter combinations in the decimal range, and a state change rate index is established by assessing the unit's operational stability. To indicate the degree of power fluctuation, the system reads the unit output records from the scheduling database during calculation. For example, if the output is in the range of hundreds of megawatts for several consecutive periods, the values ​​are substituted into the formula to calculate the differences and sum them. The three types of indicators are then combined using a weighted combination formula. This represents the comprehensive evaluation function, where the weights are... According to the scheduling strategy document, the proportions are set and their sum is limited to a unit range. The scheduling engineers select the range of weights based on frequency stability requirements, fuel cost curves, and unit oscillation records. For example, the frequency deviation weight is in a high range, the fuel weight is in a medium range, and the stability weight is in a medium-high range. The corresponding example combination is allocated a proportion between zero and one and meets the summation condition. Then, the above-mentioned index combination expressions are organized into a calculable multi-objective evaluation structure and recorded in the scheduling calculation module as the input structure for subsequent solutions.

[0082] S302: Based on the multi-objective optimization function, a non-dominated sorting genetic algorithm is used to solve the multi-objective optimization problem. The optimal solution set is obtained through population initialization, crossover mutation, and non-dominated sorting selection, thereby generating a multi-objective optimization model.

[0083] After forming a combined evaluation structure containing multiple evaluation indicators, the unit output sequence is represented as a set of decision variables. First, a population initialization operation is performed, generating several sets of candidate solution vectors in the computing environment. Each set of vectors contains the unit output values ​​for each scheduling period. The initialization data is generated by a random function within the allowable output range of the unit. This range is set according to the unit's technical manual; for example, the minimum output is located near the lower bound of the several hundred megawatt range, and the maximum output is located near the upper bound of the same range. After generating sample solutions, each set of solutions is substituted into the aforementioned evaluation function for calculation. The indicators are used to determine non-dominance relationships through comparison rules. The determination method is that if a solution is not inferior to another solution in all indicators and has at least one indicator that is superior, then a dominance relationship is defined. In the computer program, a loop structure is used to compare pairs of solutions to complete the judgment and form a ranking table. During the crossover operation phase, the two sets of solutions with the higher ranking are selected as parent solutions, and new solutions are generated through linear combination. For example, executing... ,in and These represent the two parent solutions at time [time]. The output value, This represents the crossover coefficient, which is a random number between zero and one. In the example operation, a proportion of values ​​in the middle range can be used to calculate a new candidate output sequence. Subsequently, a mutation operation is performed to add a small perturbation to the output values ​​at certain time points. ,in The value range is set to a small percentage range based on the unit's ramp rate limit, for example, not exceeding a very small percentage of the unit's rated capacity. The three evaluation indicators are recalculated using the updated solution, and non-dominated sorting is performed again. The sorting results are stored in a set structure, and several groups of candidate solutions with better rankings are selected. The selection rule is based on a joint judgment of ranking level and congestion distance. The congestion distance is calculated by measuring the distance between adjacent solutions in the target space. Obtain, among which and This represents the value of the objective function among neighboring individuals. By iterating through several generations, the population is updated and a stable solution set is gradually formed. Finally, the structure of this solution set is recorded in the scheduling system and used as a set of candidate schemes for subsequent unit scheduling decisions.

[0084] S4 includes the following steps:

[0085] S401: Based on the dynamic operation model of the thermal power unit and the multi-objective optimization model, and combined with the minimum technical output constraint, ramp rate constraint and grid frequency deviation constraint, construct the model predictive control problem and generate the model predictive control problem model;

[0086] First, basic parameters such as rated output, minimum technical output, ramp rate, unit inertia, and frequency regulation response characteristics of the generating units are collected. Statistical analysis of historical unit operating data is performed to calculate the response curves of each unit under different load levels. For example, if the rated output is set to 1000 MW, the minimum technical output to 400 MW, and the ramp rate to 10 MW / min, the output power over time curves are fitted under different loads. These curves are used to construct a dynamic unit model. Next, economic dispatch, peak shaving and frequency regulation, and environmental constraints are used as multi-objective optimization functions, each assigned a weight coefficient (e.g., 0.5 for economic objective, 0.3 for frequency regulation, and 0.2 for emission constraints). A set of optimization objectives is formed through linear weighted summation or Pareto front analysis. Simultaneously, considering the minimum technical output constraint, ramp rate constraint, and grid frequency deviation constraint, a sliding time window method is used to predict the possible output power of the unit every minute, calculating whether the constraints are met. For example, for the ramp rate constraint, the formula ΔP / Δt ≤ 10... MW / min, ΔP represents the power change between adjacent time points, and Δt is the time interval of 1 minute. By traversing the power adjustment at each time point within the future prediction window, a model for predictive control problem is generated.

[0087] S402: Based on the model predictive control problem model, the model predictive control method is used to perform rolling optimization calculations to determine the unit output regulation strategy within the future time window and generate a unit peak-shaving and frequency regulation coordinated control strategy.

[0088] First, within each rolling time window, the predicted load curve, grid frequency deviation, and unit status are input into the model. Using control variables such as unit power increment ΔP and start / stop status S, the model calculates the change in unit output for each minute in the future, using the following formula: Where P(t) represents the unit power at the current moment, and ΔP(t) is the power adjustment calculated by rolling optimization. For the unit's minimum technical output constraint, if the calculated P(t+1) is lower than 400MW, then ΔP(t) is adjusted to make P(t+1) = 400MW. Simultaneously, the ramp rate constraint is checked. The ΔP(t) is decomposed into multi-step adjustments through linear interpolation, with each step not exceeding 10 MW / min. Further, combined with the grid frequency deviation prediction, if the deviation of f(t+1)-50 Hz exceeds 0.2 Hz, ΔP(t) is adjusted appropriately to ensure that the frequency is within 50±0.2 Hz. All calculations are completed through manual data tables or simple script iterations. At the same time, the rolling prediction is updated after each window ends, and the above calculations are repeated to finally generate the unit output regulation strategy within the future time window, which is then summarized into the unit peak-shaving and frequency regulation coordinated control strategy.

[0089] S5 includes the following steps:

[0090] S501: Based on the unit peak shaving and frequency regulation coordinated control strategy, construct a unit power control loop model containing a proportional-integral-derivative control structure, and generate a power control loop model;

[0091] First, the signal paths in the unit's power regulation loop are decomposed, defining the frequency deviation signal, power command signal, and actual output power signal as input variables. The frequency deviation is calculated from the difference between the real-time system frequency and the rated frequency. For example, the frequency deviation Δf is obtained by reading the current grid frequency from the dispatch monitoring system and calculating the difference with the unit's rated frequency. This deviation is written into the control model variable table through the control system interface. Subsequently, three independent calculation modules—proportional, integral, and derivative—are established in the control modeling software. The proportional term multiplies the proportional coefficient with the current deviation to obtain the proportional output. The integral term accumulates the deviation data in the time series to obtain the integral quantity, which can be expressed as follows: Where I(t) represents the integral output, t represents the time variable, the integration process is completed by gradually accumulating the deviation within a discrete time step, and the differential term is obtained by calculating the rate of change of the deviation in the continuous sampling period, and its discrete form is expressed as: Δt represents the sampling time interval. The deviation change rate is calculated from historical sampling data. Then, the proportional output, integral output, and derivative output are linearly superimposed at the control node to generate a control signal. For example, the three control quantities are multiplied by their corresponding control coefficients and summed to obtain the unit power regulation command signal. This signal is then written to the power regulation module through the unit control interface. During the simulation operation, the relationship between the control signal and the unit output power is recorded, and the model signal path is repeatedly verified based on the actual operating curve of the unit, thus forming a unit power control loop model that includes proportional, integral, and derivative components.

[0092] S502: Based on the power control loop model, the particle swarm optimization method is used to optimize the proportional, integral and derivative parameters to improve the power regulation response speed and stability of the unit, and generate an optimized power control parameter set.

[0093] Based on the established power control loop model, when performing particle swarm optimization on the proportional, integral, and derivative parameters, the first step is to define the value ranges of the three control parameters in the parameter space. The upper and lower bounds of the proportional, integral, and derivative parameters are set according to the unit regulation system design specifications. This range is then discretized into several candidate particles, each containing a set of three control parameter values. Subsequently, power response simulation calculations are performed on each set of particle parameters in the simulation control environment. Evaluation indices are constructed by reading the difference sequence between the unit power output curve and the target power command. These evaluation indices are calculated using the cumulative sum of squared errors, and their calculation expression is as follows: Where J represents the evaluation function value, Pref(t) represents the power command value at a certain moment, Pout(t) represents the output power value of the unit at the same moment, and Σ represents the cumulative calculation over the entire simulation time. The evaluation value is obtained by substituting each set of particle parameters into the control model and running the simulation. The evaluation value is then used as a particle fitness index for ranking and comparison. During particle update, velocity and position are updated according to the particle's current position, historical best position, and group best position. The update form can be expressed as: Where v(i) represents the particle velocity variable, x(i) represents the particle's current position parameter vector, and w represents the inertia coefficient. Represents the learning coefficient. The random coefficients are calculated and updated iteratively. After each iteration, the unit power simulation is run again to calculate the evaluation value. The calculation stops when the change range of the evaluation function is within a preset convergence range in multiple consecutive iterations. The combination of proportional parameters, integral parameters, and differential parameters corresponding to the optimal particle in the final swarm is recorded as the control parameter set, forming the optimized power control parameter set.

[0094] S6 includes the following steps:

[0095] S601: Based on the optimized power control parameter set and real-time grid frequency monitoring data, construct a frequency regulation strategy learning environment, including state space, action space and reward function, and generate a frequency regulation strategy training environment;

[0096] First, the variables in the power regulation parameter set are broken down and recorded. These variables include the rated output ratio coefficient of the unit. Frequency deviation response coefficient and power change rate coefficient The power grid frequency measurement sequence is read through the data acquisition interface in the dispatch and monitoring platform. subscript The sampling time is indicated by setting the sampling interval to a period of several seconds within a stable range, and then using a frequency reference value. Constructing frequency deviation ,in The state vector is determined based on the median interval of historical statistical frequency distribution. For example, a statistical center value is selected as a reference within a small interval above and below the rated frequency. During the state space construction phase, the state vector is defined as... ,in This indicates the ratio of the current unit's output power to its rated power. This represents the rate of frequency change within the most recent time window, expressed by the formula... Obtain, in the formula This represents the adjacent sampling interval. In practical examples, if the frequency measured in a certain sampling period is within a small range below the rated value and the frequency of the previous period is slightly higher, then the calculated rate of change falls into the negative change range. Subsequently, an action space set is established. Each action represents a different power adjustment ratio, such as increasing the power ratio within a small range of the rated power, maintaining the original output, or decreasing it within a small range. The reward function is constructed by combining the frequency deviation and the power adjustment amount, and the calculation formula is written as follows: The weighting coefficient Based on historical frequency modulation records, for example, when the absolute value of the frequency deviation is in the middle range, the corresponding weight is set to be slightly higher than the power change weight. By importing the daily running sample data into the scheduling simulation platform for calculation, when the frequency deviation is in a narrow range near the rated value and the power change amplitude is small, the corresponding reward value is obtained and recorded in the training sample set, forming a data structure containing state vector, action selection and reward value, and finally generating a frequency modulation learning environment for policy training.

[0097] S602: Based on the frequency modulation strategy training environment, a deep Q-network learning method is used for iterative training of the strategy. The strategy is optimized through experience playback and target network update to generate an adaptive frequency modulation control strategy.

[0098] First, establish the state-action value function within the deep Q-network framework. ,in This represents the state vector in the state space. This represents a power adjustment action in the action space. The system randomly reads a set of state samples from the training environment. And obtain the corresponding action value estimate through neural network. During the action selection phase, a probabilistic exploration mechanism is used to filter actions. When a random number falls within a preset exploration probability range, a random action is selected; otherwise, the action with the larger current estimated value is selected. This exploration probability parameter... As the training rounds gradually decrease, the value range gradually shrinks from an initial higher exploration range to a lower range, thus establishing an experience sample structure during experience replay processing. ,in The system stores sample records within a certain time window in the experience cache pool to obtain the reward value calculated by the reward function in the previous stage. When the cache size reaches a set capacity range, a number of samples are extracted using random sampling for batch training. During the target network update stage, the target Q value is calculated. ,in The discount factor is set within a moderate attenuation range based on historical frequency modulation cycle statistics, for example, a value close to but slightly less than one. The loss function is constructed by the difference between the predicted Q value and the target Q value. Gradient updates are performed during the training process. For example, when the frequency deviation in a training sample is in the moderate deviation range and the power corresponding to the action increases by a small proportion, the above formula is substituted to obtain a new target value estimate and the network parameters are updated. After multiple training iterations, the changes in the action sequence output by the strategy are compared for several consecutive rounds. When the change in the strategy enters the stable range, the current set of strategy parameters is recorded to form an adaptive frequency modulation control strategy that can select power adjustment actions according to the real-time frequency state.

[0099] S7 includes the following steps:

[0100] S701: Based on the dynamic operation model of the thermal power unit and the adaptive frequency regulation control strategy, and combined with the real-time operation data collected by the unit's field sensors, a digital mapping relationship between the unit's physical system and control system is constructed to generate a digital twin unit model.

[0101] First, the operating variables of the boiler, turbine, and generator sides of the unit are decomposed and extracted, and the monitored variables such as temperature, pressure, steam flow, speed, and electrical power are divided into a set of state parameters. With control parameter set The set of state parameters represents the physical state variables of the equipment, and the set of control parameters represents the frequency regulation command variables. In the data acquisition phase, the unit's on-site distributed control system periodically collects sensor signals and forms time-series data. Subsequently, the collected signals are discretized, and the monitored quantity for each sampling period is represented as... The relationship between control variables and state variables is established through a simple proportional mapping formula, for example, by using a relational expression. ,in Indicates the first The state variable and the first Mapping weights between control variables This represents the control gain coefficient, which is set within the stable operating experience range based on the unit's frequency regulation experience parameter range, and the midpoint of the range is taken as an example value. During the calculation process, for example, when the steam flow monitoring quantity is at the middle level of the typical operating range of the unit, it is multiplied with the valve opening control coefficient to obtain the corresponding mapping weight. By performing cycle-by-cycle calculations on multiple sets of time series data in industrial simulation software, the obtained mapping matrix is ​​sorted by time index. Then, the mapping matrices of the boiler thermal subsystem, the turbine power subsystem, and the generator electrical subsystem are combined to form a unified mapping structure, forming a digital correspondence structure between the unit's physical equipment status and control logic, and outputting the unit's virtual simulation structure model.

[0102] S702: Based on the aforementioned digital twin unit model, a simulation prediction method is used to simulate and analyze the future operating status of the unit, generating predicted data for the unit's operating status.

[0103] Based on the established virtual simulation structure of the unit, its set of state variables is extracted and a time series prediction input matrix is ​​constructed. Variables such as boiler pressure, main steam temperature, rotor speed, and electric power are arranged into a matrix according to the sampling time order. Then, a time-step recursive operation is performed on the matrix to obtain future trends, where the prediction calculation uses a recursive formula. In the formula Indicates the state quantity at the next prediction time. Indicates the change between adjacent sampling times. This represents the time prediction weighting coefficient. This coefficient is calculated based on an empirical range selected from the unit's inertial time constant interval, using example values ​​from that range. In the actual calculation example, the difference between adjacent samples is calculated by reading historical monitoring sequences. The difference is then multiplied by the prediction weight to obtain the state increment, which is then added to the current state quantity to obtain the predicted state quantity for the next moment. Through repeated recursive calculations, the state sequence for multiple future time steps is gradually obtained. During the calculation process, the interval determination operation is also required for the prediction sequence to divide each state variable into a stable interval, a fluctuating interval, and an offset interval. The stable interval is defined as the variable being in the middle of the historical operating statistical range, the fluctuating interval is defined as the interval close to the upper and lower limits, and the offset interval is defined as the interval exceeding the normal statistical range. Through time-by-time calculation and interval determination, a future time series prediction matrix is ​​formed, and the set of future operating state data of the unit is output.

[0104] S8 includes the following steps:

[0105] S801: Based on the short-term load forecast curve, the unit peak-shaving and frequency regulation coordinated control strategy, and the unit operation status forecast data, and combined with the new energy power generation output data and energy storage system operation data, a coordinated scheduling optimization model is constructed to generate a multi-energy coordinated scheduling optimization model.

[0106] Based on short-term load forecast curves, unit peak-shaving and frequency regulation coordinated control strategies, and unit operating status forecast data, combined with renewable energy generation output data and energy storage system operating data, a coordinated scheduling optimization model is constructed. First, short-term load forecast curves need to be collected and processed. These curves are typically based on historical load data and predicted using specific models (such as time series analysis and regression models). Accurate data input combined with algorithms yields the load demand for a future period. Second, according to the unit peak-shaving and frequency regulation coordinated control strategy, the operating status of different units in the current system needs to be analyzed to assess the response capability and peak-shaving and frequency regulation capability of each unit. Further optimization of control strategies is needed to minimize energy consumption. Simultaneously, acquiring predictive data on unit operating status requires real-time monitoring of unit operating parameters (such as power, load, and failure rate) and predictive model analysis to understand the reliability and adaptability of each unit. Furthermore, combining renewable energy generation output data and energy storage system operating data allows for the assessment of resource availability within a given timeframe by collecting data on wind, solar, and other renewable energy sources, as well as information on the discharge and charging status of energy storage systems. Finally, based on this data, different energy dispatch strategies are optimized and merged through model construction, ultimately forming a multi-energy collaborative dispatch optimization model. This model effectively coordinates the dispatch of various energy resources, enabling the system to meet load demands while minimizing costs and optimizing energy utilization.

[0107] S802: Based on the multi-energy coordinated scheduling optimization model, a mixed integer linear programming method is used to optimize and solve the problem, generating a multi-energy coordinated scheduling scheme.

[0108] By employing mixed-integer linear programming (MIBLP) for optimization, firstly, after determining the optimization objective, a MIBLP model needs to be established, including input variables and constraints for all energy resources (such as wind power, photovoltaics, battery storage, and the power grid). By setting output and demand models for these resources, the optimal solution is found while meeting scheduling requirements. To make the scheduling scheme more realistic, some control parameters must be set in the model, such as energy conversion efficiency, energy storage device charging and discharging efficiency, and unit start-up and shutdown constraints. These constraints ensure the feasibility of the optimization scheme. The objective function in MIBLP typically involves cost minimization or optimal resource allocation, as shown in the following formula:

[0109]

[0110] in, For total cost, Let i be the unit cost of the i-th energy source. Let be the scheduling amount of the i-th energy source. To solve this model, optimization algorithms (such as the simplex method, branch and bound method, etc.) can be used to solve it step by step to obtain a suitable scheduling scheme. During the solution process, parameters such as the allocation ratio of different energy resources and the running time are gradually optimized to generate the final multi-energy collaborative scheduling scheme. This scheme provides the system with a reasonable energy scheduling arrangement and takes into account the timeliness and complementarity of different energy sources.

[0111] S9 includes the following steps:

[0112] S901: Based on the optimized power control parameter set, the adaptive frequency regulation control strategy, and the multi-energy collaborative dispatch scheme, construct a dynamic control model for the power generation of thermal power units and generate a dynamic control model for power generation.

[0113] Based on the verified set of power control parameters, on-site unit frequency regulation rules, and multi-energy unit output allocation table obtained in the previous steps, historical operation records and real-time measurement data are first imported into the dispatch workstation to form an input data table. This table includes boiler main steam pressure range values, unit rated capacity range values, wind power and photovoltaic real-time output range values, and system frequency deviation records. In the monitoring software environment, the data is organized according to the time series and a calculation matrix is ​​formed. Subsequently, the power control parameters are broken down item by item; for example, the adjustment coefficient is denoted as... The load distribution factor is denoted as The frequency response coefficient is denoted as Its initial settings are determined based on historical frequency modulation (FM) record intervals. For example, by statistically analyzing multiple adjustment amplitudes in the FM record table and taking the median interval as an example, the interval is set to approximately 0.2 to 0.4. During model construction, this is achieved through calculation... Estimating the instantaneous output of the unit, among which This represents the target power of the generator unit at time t. Indicates the basic load range value. This represents the system frequency deviation value. Indicates the change in power grid load. This represents the fluctuation in renewable energy output. The parameters are obtained through the real-time data interface of the dispatching platform. For example, if the grid frequency deviation is about a small negative deviation range and the renewable energy output fluctuation is in the medium range at a certain moment, the midpoint of the corresponding range is substituted into the formula to calculate the target output range value of the unit. The result is then input into the dynamic simulation module. In the simulation module, the output power sequence is iteratively updated step by step according to the time step. After multiple iterations, a set of unit power change curves is formed. After being organized, this set of curves constitutes a complete data structure for the dynamic control model of unit power generation.

[0114] S902: Based on the aforementioned dynamic power generation control model, a real-time feedback control method is used to adjust the unit's output power online, generating an optimized power generation control strategy.

[0115] After completing the dynamic model of the unit's power generation, the model is imported into the real-time calculation module of the unit's control system and connected to the measurement signal stream, which includes the unit's current output power, system frequency deviation, steam pressure, and recorded load demand values. A real-time feedback calculation table is established on the control terminal, first reading the current power measurement value and recording it as... Then read the model's predicted power and record it as... Then, using the difference formula The power deviation is obtained, where For the output power difference parameter, when the difference is in a small range, such as a tiny range close to zero, the original adjustment step size is maintained; when the difference falls into a medium range, the proportional coefficient is adjusted. A correction calculation is performed, with the proportional gain set in the range of approximately 0.3 to 0.5 based on historical adjustment stability records. This is achieved through the calculation formula. A new power regulation command is obtained, in which This represents the updated power command value. Subsequently, the scheduling software reads real-time data cyclically according to the control cycle and re-executes the above calculation process. For example, if the actual power of the unit is detected to be slightly lower than the median of the model's predicted power range during a certain scheduling cycle, the difference between the two is substituted into the formula to calculate the new command power range value. The command is then sent to the unit load control module for execution. The deviation calculation and parameter update are repeated for multiple consecutive control cycles, and the power change sequence is recorded. Finally, a continuously updated set of unit power generation regulation rules is formed and organized into a power generation control strategy data table.

[0116] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for optimizing the deep peak shaving, primary frequency regulation, and power generation control system of thermal power units, characterized in that, Includes the following steps: S1: Based on historical power grid load data, meteorological data, and holiday characteristic data, the raw data is cleaned, normalized, and reconstructed using time series to construct a load forecast dataset. On this basis, a short-term power grid load forecast model is established using a long short-term memory neural network forecasting method, and the trend of power grid load change within a preset time window is predicted and analyzed to generate a short-term load forecast curve. S2: Based on the short-term load forecast curve generated by S1, as well as the historical operating data of thermal power units, equipment operating parameters and unit output data, a dynamic operating model of thermal power units is established using the support vector machine modeling method. The dynamic coupling relationship between the boiler system, turbine system and generator system is modeled and analyzed to generate a dynamic operating model of thermal power units. S3: Based on the short-term load forecast curve generated in S1 and the dynamic operation model of thermal power units generated in S2, a multi-objective optimization function for deep peak shaving and primary frequency regulation is constructed. The optimization function includes the objectives of minimizing grid frequency deviation, minimizing unit fuel consumption, and maximizing unit operation stability. The multi-objective optimization function is solved using a non-dominated sorting genetic optimization method to generate a multi-objective optimization model. S4: Based on the dynamic operation model of the thermal power unit generated in S2 and the multi-objective optimization model generated in S3, combined with the minimum technical output constraint, ramp rate constraint and grid frequency deviation constraint, the model predictive control method is used to calculate the coordinated control strategy of deep peak shaving and primary frequency regulation of the thermal power unit, so as to predict the changes in unit output within the future time window and perform dynamic optimization adjustment, and generate the coordinated control strategy of unit peak shaving and frequency regulation. S5: Based on the unit peak shaving and frequency regulation coordinated control strategy generated by S4, construct the power control loop of the thermal power unit, and use the particle swarm optimization method to optimize the parameters of the adaptive proportional-integral-derivative controller in order to improve the power regulation response speed and stability of the unit under load fluctuation conditions, and generate an optimized power control parameter set; S6: Based on the optimized power control parameter set generated by S5 and the real-time frequency monitoring data of the power grid, a frequency regulation strategy learning environment is constructed, and a deep Q-network learning method is used to train and optimize the primary frequency regulation control strategy of the unit, so that the unit can adaptively adjust the power output according to the frequency deviation change and generate an adaptive frequency regulation control strategy. S7: Based on the dynamic operation model of the thermal power unit generated by S2 and the adaptive frequency regulation control strategy generated by S6, and combined with the real-time operation data collected by the unit's field sensors, a digital twin operation model of the thermal power unit is constructed, and dynamic prediction and simulation analysis of the future operation status of the unit are performed to generate predicted data of the unit's operation status. S8: Based on the short-term load forecast curve generated by S1, the unit peak-shaving and frequency regulation coordinated control strategy generated by S4, and the unit operation status forecast data generated by S7, combined with the new energy power generation output data and energy storage system operation data, the mixed integer linear programming optimization method is used to optimize the coordinated scheduling of thermal power units and new energy systems, and generate a multi-energy coordinated scheduling scheme. S9: Based on the optimized power control parameter set generated in S5, the adaptive frequency regulation control strategy generated in S6, and the multi-energy collaborative dispatch scheme generated in S8, the real-time feedback control method is used to optimize and adjust the power generation control system of thermal power units online, so as to realize the dynamic control of power generation of thermal power units under deep peak shaving and primary frequency regulation operation conditions, and generate the optimized power generation control strategy.

2. The method for optimizing the deep peak shaving, primary frequency regulation, and power generation control system of thermal power units according to claim 1, characterized in that: S1 includes the following steps: S101: Based on historical load data of the power grid, meteorological data and holiday characteristic data, data cleaning and normalization methods are adopted. The original load data is preprocessed by imputing missing values, removing outliers and linear normalization to generate a cleaned and normalized load dataset. S102: Based on the cleaned and normalized load dataset, a time series reconstruction method is used to form the input and output sequences of the prediction model through sliding window segmentation and feature construction, thereby generating a time series load dataset. S103: Based on the time series load dataset, a long short-term memory neural network prediction method is used for model training and prediction to obtain the power grid load change trend within a preset time window in the future and generate a short-term load prediction curve.

3. The method for optimizing the deep peak shaving, primary frequency regulation, and power generation control system of thermal power units according to claim 1, characterized in that: S2 includes the following steps: S201: Based on historical operating data of thermal power units, equipment operating parameters and unit output data, data cleaning and feature extraction methods are used to standardize the unit operating indicators, and key features such as boiler temperature, turbine speed and generator power are extracted to generate a standardized unit feature dataset. S202: Based on the standardized unit feature dataset and the short-term load forecast curve, a dynamic operation model of the thermal power unit is established using the support vector machine modeling method to describe the dynamic coupling relationship between the boiler system, turbine system and generator system, and to generate the dynamic operation model of the thermal power unit.

4. The method for optimizing the deep peak shaving, primary frequency regulation, and power generation control system of thermal power units according to claim 1, characterized in that: S3 includes the following steps: S301: Based on the short-term load forecast curve and the dynamic operation model of the thermal power unit, construct a multi-objective optimization function that includes minimizing grid frequency deviation, minimizing unit fuel consumption, and maximizing unit operation stability, and generate the multi-objective optimization function. S302: Based on the multi-objective optimization function, a non-dominated sorting genetic algorithm is used to solve the multi-objective optimization problem. The optimal solution set is obtained through population initialization, crossover mutation, and non-dominated sorting selection, thereby generating a multi-objective optimization model.

5. The method for optimizing the deep peak shaving, primary frequency regulation, and power generation control system of thermal power units according to claim 1, characterized in that: S4 includes the following steps: S401: Based on the dynamic operation model of the thermal power unit and the multi-objective optimization model, and combined with the minimum technical output constraint, ramp rate constraint and grid frequency deviation constraint, construct the model predictive control problem and generate the model predictive control problem model; S402: Based on the model predictive control problem model, the model predictive control method is used to perform rolling optimization calculations to determine the unit output regulation strategy within the future time window and generate a unit peak-shaving and frequency regulation coordinated control strategy.

6. The method for optimizing the deep peak shaving, primary frequency regulation, and power generation control system of thermal power units according to claim 1, characterized in that: S5 includes the following steps: S501: Based on the unit peak shaving and frequency regulation coordinated control strategy, construct a unit power control loop model containing a proportional-integral-derivative control structure, and generate a power control loop model; S502: Based on the power control loop model, the particle swarm optimization method is used to optimize the proportional, integral and derivative parameters to improve the power regulation response speed and stability of the unit, and generate an optimized power control parameter set.

7. The method for optimizing the deep peak shaving, primary frequency regulation, and power generation control system of thermal power units according to claim 1, characterized in that: S6 includes the following steps: S601: Based on the optimized power control parameter set and real-time grid frequency monitoring data, construct a frequency regulation strategy learning environment, including state space, action space and reward function, and generate a frequency regulation strategy training environment; S602: Based on the frequency modulation strategy training environment, a deep Q-network learning method is used for iterative training of the strategy. The strategy is optimized through experience playback and target network update to generate an adaptive frequency modulation control strategy.

8. The method for optimizing the deep peak shaving, primary frequency regulation, and power generation control system of thermal power units according to claim 1, characterized in that: S7 includes the following steps: S701: Based on the dynamic operation model of the thermal power unit and the adaptive frequency regulation control strategy, and combined with the real-time operation data collected by the unit's field sensors, a digital mapping relationship between the unit's physical system and control system is constructed to generate a digital twin unit model. S702: Based on the aforementioned digital twin unit model, a simulation prediction method is used to simulate and analyze the future operating status of the unit, generating predicted data for the unit's operating status.

9. The method for optimizing the deep peak shaving, primary frequency regulation, and power generation control system of thermal power units according to claim 1, characterized in that: S8 includes the following steps: S801: Based on the short-term load forecast curve, the unit peak-shaving and frequency regulation coordinated control strategy, and the unit operation status forecast data, and combined with the new energy power generation output data and energy storage system operation data, a coordinated scheduling optimization model is constructed to generate a multi-energy coordinated scheduling optimization model. S802: Based on the multi-energy coordinated scheduling optimization model, a mixed integer linear programming method is used to optimize and solve the problem, generating a multi-energy coordinated scheduling scheme.

10. The method for optimizing the deep peak shaving, primary frequency regulation, and power generation control system of thermal power units according to claim 1, characterized in that: S9 includes the following steps: S901: Based on the optimized power control parameter set, the adaptive frequency regulation control strategy, and the multi-energy collaborative dispatch scheme, construct a dynamic control model for the power generation of thermal power units and generate a dynamic control model for power generation. S902: Based on the aforementioned dynamic power generation control model, a real-time feedback control method is used to adjust the unit's output power online, generating an optimized power generation control strategy.

Images