Layered optimization scheduling method for deep peak regulation of thermal power generating unit

Abstract

The invention discloses a hierarchical optimization scheduling method for deep peak regulation of thermal power generating units, and the method comprises the following steps: system layering: dividing a power system into a plurality of subsystems, each subsystem comprising a thermal power generating unit cluster and differentiated load demands; subsystem-level prediction: generating a basic scheduling plan; dynamic preference-driven multi-target collaborative optimization: receiving multi-target data from each subsystem, carrying out multi-target collaborative optimization, modeling three targets of a power grid company, an environmental protection department and a terminal user as game participants, quantifying the priority of each party by adopting a fuzzy membership function, generating a dynamic game solution through a Nash equilibrium solver, and carrying out multi-target collaborative optimization; screening out a Pareto optimal solution giving consideration to interests of multiple parties; performing online verification on the dynamic security domain; virtual synchronous machine cooperative support: simulating operation characteristics of a synchronous generator; and global coordination and iterative optimization: uploading the optimal scheduling scheme of each subsystem to a scheduling center through a message queue telemetry transmission protocol, and performing global constraint verification.

Description

Technical Field

[0001] The present invention relates to the technical field of thermal power peak regulation, and in particular to a hierarchical optimization scheduling method for deep peak regulation of thermal power units. Background Art

[0002] As the proportion of renewable energy generation continues to increase and grid load characteristics become more complex, traditional thermal power peak-shaving models face challenges such as delayed response and insufficient coordination. New power systems must strike a multi-dimensional balance between renewable energy absorption, load tracking, and operational safety. Traditional centralized dispatching methods struggle to adapt to the demands of time-space decoupling. A layered optimization architecture based on the Internet of Things and edge computing achieves data-driven dynamic decoupling of unit clusters and differentiated loads. Combined with AI prediction and game-based optimization mechanisms, it can precisely match the uncertainties of both supply and demand.

[0003] In the existing technology, when dealing with multi-objective optimization problems, the complex relationships between different stakeholders may not be fully considered, resulting in a certain degree of subjectivity in the selection of Pareto front solution sets. Therefore, a hierarchical optimization scheduling method for deep peak regulation of thermal power units is proposed. Summary of the Invention

[0004] The purpose of the present invention is to solve the shortcomings of the prior art and to propose a hierarchical optimization scheduling method for deep peak regulation of thermal power units.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] A hierarchical optimization scheduling method for deep peak regulation of thermal power units comprises the following steps:

[0007] System layering: Leveraging IoT technology to deploy sensor networks, real-time data collection is used to collect unit operating parameters, grid load demands, and external meteorological data. Edge computing nodes are used to cleanse and extract features from this data, dividing the power system into several subsystems. Each subsystem contains a cluster of thermal power units and differentiated load demands. The subsystem boundaries are dynamically determined by the grid topology and power flow distribution.

[0008] Subsystem-level forecasting: Based on historical data and long short-term memory (LSTM) networks, we build subsystem-level load forecasting and renewable energy output forecasting models to generate basic dispatch plans. We use an improved multi-objective particle swarm optimization (MOPSO) algorithm for preliminary optimization to generate Pareto frontiers for multiple objectives, including economic efficiency, carbon emissions, and peak-shaving margins.

[0009] Dynamic preference-driven multi-objective collaborative optimization: This system receives multi-objective data on economic, environmental, and social benefits from each subsystem, adjusts the weights of each objective through dynamic preference settings, and performs multi-objective collaborative optimization. While meeting system requirements, it models the objectives of the power grid company, environmental protection department, and end users as game participants, quantifies the priorities of each party using a fuzzy membership function, and generates dynamic game solutions using a Nash equilibrium solver, screening out the Pareto optimal solution that balances the interests of multiple parties.

[0010] Dynamic safety domain online verification: This system receives real-time unit operating data, such as key parameters like main steam pressure, temperature, and turbine shaft vibration, and uses a graph neural network (GNN) to dynamically verify the safety domain boundaries, promptly identifying potential safety hazards. Based on historical fault data and reinforcement learning, it pre-generates a library of rescheduling strategies for N-1 fault scenarios. When a fault occurs, it selects the optimal strategy through dynamic time warping (DTW).

[0011] Virtual synchronous machine collaborative support: simulates the operating characteristics of synchronous generators, receives unit operating status data such as frequency, voltage, power factor, etc., optimizes control system strategies and adjusts equipment parameters;

[0012] Global collaboration and iterative optimization: The optimal scheduling plan for each subsystem is uploaded to the dispatch center through the Message Queue Telemetry Transport (MQTT) protocol for global constraint verification (such as cross-regional interconnection line power limit). If there is a conflict, the conflicting constraints are decomposed through the Lagrangian relaxation method, and the subsystem is returned for iterative optimization to form an overall scheduling plan.

[0013] The above further includes:

[0014] Furthermore, the power system is divided into several subsystems, each subsystem includes a cluster of thermal power units and differentiated load requirements, and the subsystem boundaries are dynamically determined by the grid topology and power flow distribution, including the following steps:

[0015] Data collection and preprocessing: Collect grid topology (node-branch connection relationships), thermal power unit parameters (output range, ramp rate), load demand time series curves (categorized by type, such as industrial / residential load), line flow limits, and real-time operating data. Load demand is clustered by type, and the unit output range is normalized to the [0, 1] range.

[0016] Initial topology partitioning based on graph theory: construct a weighted adjacency matrix, use the inverse of line reactance as the edge weight, reflect the electrical distance between nodes, and calculate modularity. The calculation formula is expressed as Among them, A ij is the adjacency matrix element, k i =∑ j A ij is the node degree, is the total edge weight, δ(c i , c j ) is the community indicator function (1 when nodes i and j are in the same community, otherwise 0), and the Louvain algorithm is used to iteratively optimize the modularity and divide the power grid into an initial set of subsystems;

[0017] Subsystem boundary correction under power flow constraints: Power balance constraints are established based on the AC power flow equation. The power flow equation constraint is expressed as in, To contribute to the unit, is the load demand, G ij +jB ij It is an admittance matrix element and dynamically adjusts the subsystem boundaries by setting the tie-line power flow threshold and voltage over-limit detection criteria. For example, when the tie-line power flow exceeds the limit, the adjacent subsystems are merged, or when the voltage exceeds the limit, the subsystem is split, thereby ensuring the power balance within the subsystem and meeting the safe operation boundary.

[0018] Differentiated load demand matching: By calculating the load demand fluctuation coefficient and the unit regulation capacity coefficient, the load demand fluctuation coefficient calculation formula is expressed as The calculation formula of the unit regulation capacity coefficient is expressed as follows: A clustering algorithm is used to match high-fluctuation loads with high-regulation-capacity units to the same subsystem, achieving coordinated optimization of load characteristics and unit regulation capabilities, and improving the subsystem's adaptability to deep peak-shaving needs.

[0019] Dynamic boundary adjustment mechanism: By real-time monitoring of the subsystem self-sufficiency rate and the proportion of interconnection line power, and setting threshold trigger rules, if the subsystem self-sufficiency rate is <0.8 and the interconnection line power is >0.9, the subsystem merger is triggered; if the subsystem self-sufficiency rate is >1.2 and there is an interruptible load, the subsystem is split. A closed-loop feedback adjustment mechanism is established, which enables the subsystem division to respond to real-time changes in operating status, ensuring the timeliness and economy of the division.

[0020] Furthermore, the specific steps of the subsystem level prediction are:

[0021] Data preprocessing: Min-Max normalization is used to standardize historical load, renewable energy output, and meteorological data, and a sliding window technique is used to construct time series input samples.

[0022] LSTM model construction: Utilizing the forget gate, input gate, output gate, and cell state update mechanism of a two-layer LSTM network (containing 64 hidden units), the model dynamically adjusts the ratio of historical information retention to new information writing through nonlinear activation functions (such as Sigmoid and Tanh). This model extracts in-depth features of power load time series patterns (such as daily periodicity and holiday effects) and renewable energy output volatility (such as the wind speed-power mapping relationship). Finally, the fully connected layer outputs a joint forecast of future load and renewable energy output, providing high-precision boundary conditions for scheduling.

[0023] Improved multi-objective particle swarm optimization (MOPSO): First, three objective functions are defined: economy (the weighted sum of coal consumption cost and start-up and shutdown cost of thermal power units), carbon emissions (quadratic function model of unit output), and peak-shaving margin (the difference between the system's maximum adjustable capacity and load demand). Power balance, upper and lower limits of unit output, and ramp rate constraints are introduced. Subsequently, dynamic inertia weight (which decays linearly with the number of iterations) and adaptive Gaussian mutation (adjusting the mutation amplitude based on the solution set congestion) are used to enhance the algorithm's global search capability and avoid premature convergence. Ultimately, a Pareto frontier solution set is generated (such as three typical solutions: economy first, low carbon emissions, and high peak-shaving capacity), providing decision makers with multi-dimensional trade-off options.

[0024] Furthermore, the method receives multi-objective data such as economic benefits, environmental benefits, and social benefits from each subsystem, adjusts the weight of each objective through dynamic preference settings, performs multi-objective collaborative optimization, models the objectives of the power grid company, the environmental protection department, and the end user as game participants, quantifies the priorities of each party using a fuzzy membership function, generates a dynamic game solution through a Nash equilibrium solver, and screens out a Pareto optimal solution that takes into account the interests of multiple parties, including the following steps:

[0025] Multi-objective function modeling: Quantify the economic benefits of the power grid company, the environmental benefits of the environmental protection department, and the social benefits of the end users. The calculation formula of the quantified economic benefits of the power grid company is expressed as Among them, min F1 is the minimum total operating cost, is the fuel cost of unit i, is the start-up and shutdown cost, and the calculation formula of the quantified environmental benefits of the environmental protection department is expressed as Among them, min F2 is the minimized CO2 emission, α i , β i , γ i is the emission coefficient of unit i, and the calculation formula of the quantified social benefits of the end user is expressed as Among them, max F3 is, Sat(λ t ) is the electricity price satisfaction function, Reliability t It is an indicator of power supply reliability;

[0026] Dynamic preference quantification: Utilizing fuzzy membership functions, the priorities of the three parties' objectives are dynamically quantified, mapping the real-time demands for economic, environmental, and social benefits into continuous weights in the [0, 1] interval. Through historical extreme value normalization, preferences are adaptively adjusted based on system status (e.g., fluctuations in renewable energy output and carbon emission quotas), ensuring that weight allocation is strongly coupled with actual operating scenarios.

[0027] Game theory modeling: The three parties are modeled as game participants through non-cooperative game theory. The subsystem output adjustment is used as the strategy variable, and the weighted objective function is used as the benefit function. A game model with power balance constraints is constructed to simulate the strategic interaction process of the three parties under conflicting interests. The power balance constraint is expressed as ∑ΔP i =0;

[0028] Nash equilibrium solution: The Nash equilibrium solver is used to iteratively solve the game equilibrium. The equilibrium condition is expressed as Generate a stable strategy portfolio under the conditions of satisfying system constraints, ensuring that no party can improve its own benefits by unilaterally changing the strategy, thereby ensuring the stability and global feasibility of the strategy;

[0029] Pareto optimal solution screening: Eliminate inferior solutions from the Nash equilibrium solution set. The inferior solution is represented by a solution that is better in at least one objective and not inferior in other objectives. Combined with dynamic weight verification, the matching of the solution with the current system state is finally output to balance economy, environmental protection, and fairness.

[0030] Furthermore, the real-time receiving of unit operation data, the use of graph neural networks (GNN) to dynamically verify the safety domain boundary, and the pre-generation of a rescheduling strategy library under N-1 fault scenarios based on historical fault data and reinforcement learning, include the following steps:

[0031] Real-time data reception and graph structure construction: A distributed sensor network collects unit operating data (such as main steam pressure, temperature, and turbine shaft vibration) in real time, and constructs a graph structure representing the unit's operating status. Physical equipment (boilers, turbines, and generators) are abstracted as nodes, and energy or signal flows between devices are abstracted as edges. The node feature matrix integrates multi-source parameters (such as pressure, temperature, and vibration displacement), and the adjacency matrix defines the physical connection relationships between devices.

[0032] Dynamic safety domain boundary verification based on graph neural network (GNN): Graph Attention Network (GAT) is used to perform time series modeling on graph structure data. The attention mechanism aggregates neighbor node information and dynamically updates the safety domain boundary parameters. The safety domain is defined as a set that satisfies multiple constraints (such as the upper limit of the main steam pressure and the vibration threshold of the turbine shaft system), which is expressed as Among them, f i(·) is the i-th safety constraint function, θ i (t) is the time-varying parameter output by the graph neural network (GNN). The graph neural network (GNN) verifies in real time whether the current state violates the safety constraints by learning the safety boundary patterns in historical operation data;

[0033] Generation of a rescheduling strategy library: A Markov decision process (MDP) is constructed using historical fault data. The unit output adjustment amount is defined as the action space, and the degree of safety domain violation and load demand are defined as the state space. The optimal strategy is trained through a deep Q-network (DQN). The reward function is designed to balance safety (penalizing safety domain violations) and economy (penalizing excessive output adjustments). Ultimately, a rescheduling strategy library covering various fault scenarios (such as unit tripping and pipeline rupture) is generated.

[0034] Furthermore, the specific steps of selecting the optimal strategy through dynamic time warping (DTW) are:

[0035] Real-time fault feature sequence extraction: Extracting the time series features that characterize the fault evolution from the fault signal output by the dynamic safety domain verification module;

[0036] Pre-stored fault scenario library: Build a standardized feature sequence library based on historical fault data, and record the complete evolution pattern (such as the time series change of vibration displacement) of each typical fault type (such as sudden change in turbine shaft vibration);

[0037] Dynamic Time Warping (DTW) matching: By constructing a cumulative distance matrix and using a recursive formula to calculate the minimum matching distance between the real-time sequence and the pre-stored sequence, the recursive formula is expressed as:

[0038]

[0039] Where D(i, j) represents the minimum matching distance. Its core purpose is to overcome differences in fault development rates (such as the timing misalignment between fast-bursting faults and slowly deteriorating faults) through flexible time alignment. This ensures accurate pattern recognition even when the sampling interval of the fault feature sequence fluctuates due to scheduling instructions.

[0040] Optimal strategy selection and execution: Based on the dynamic time warping (DTW) matching results, the response plan with the highest similarity in the pre-stored strategy library is selected and directly transmitted to the unit actuator through closed-loop control. This constructs a response chain from fault detection to strategy execution under the complex operating conditions of deep peak regulation of thermal power units.

[0041] Furthermore, the specific steps of simulating the operating characteristics of the synchronous generator, receiving the unit operating status data, optimizing the control system strategy and adjusting the equipment parameters are as follows:

[0042] Synchronous generator characteristic modeling: Based on the second-order synchronous generator model, the virtual rotor motion equation and voltage regulation equation are constructed. Using real-time collected unit frequency, voltage, power factor, active power, and reactive power, the rotor inertia of the synchronous generator is simulated (kinetic energy buffering is achieved through the dynamic balance equation of angular velocity deviation and mechanical-electromagnetic torque difference) and frequency and voltage regulation characteristics (voltage and reactive power coupling control is achieved through a PI controller). Therefore, when the deep peak regulation of the thermal power unit causes a sudden load drop, the virtual inertia can slow the frequency change rate and avoid parameter oscillation caused by rapid load reduction.

[0043] Control strategy optimization and dynamic parameter adjustment: Using droop control optimization and virtual damping enhancement technology, the active output is dynamically adjusted according to the frequency deviation and low-frequency oscillation is suppressed. The droop control optimization is expressed as P = P0 + k f (f0-f), where P0 is the reference power, k f is the frequency droop coefficient, f0 is the rated frequency, and the load rate adaptive parameter setting is used to increase the virtual inertia constant under low load conditions to improve the unit's resistance to power fluctuations;

[0044] Multi-time scale collaborative control: Achieve primary frequency modulation through droop control, combine global collaboration and iterative optimization modules to modify the scheduling plan, expressed as P VSG =P primary +γ∫(P ref -P VSG )dt, where γ is the integral gain, P ref The power command is issued, and the safety domain verification module is linked to actively limit power under fault conditions to avoid equipment damage.

[0045] Furthermore, the specific steps of the global collaboration and iterative optimization are:

[0046] Subsystem optimal solution upload and global constraint verification: Each subsystem uploads independent optimization results (such as unit output plan and tie line power distribution) to the dispatch center through the MQTT protocol. The dispatch center builds a global optimization model, which is expressed as:

[0047]

[0048] Among them, C i (·) is the operating cost of subsystem i, is the transmission limit of tie line (i, j), S is the set of subsystems, and L is the set of tie lines;

[0049] Lagrangian relaxation decomposes conflicting constraints: When a global constraint conflict is detected (e.g., inter-subsystem tie line power exceeding limits), Lagrangian relaxation is used to convert coupling constraints (e.g., the tie line power balance equation) into decomposable penalty terms. By constructing a dual problem involving Lagrangian multipliers, the original problem is decomposed into independent subsystem-level optimization models. Each subsystem focuses on a local cost function and a weighted multiplier term (e.g., the optimization objective of subsystem A becomes its own operating cost plus the product of the tie line multiplier and the power plan). This transforms the global coordination problem into an iterative solution process.

[0050] Subsystem iterative optimization and multiplier update: Dynamically update the Lagrange multiplier through the subgradient method, driving the subsystem to gradually modify the dispatch plan (such as reducing the tie line power plan or adjusting the unit output) until the convergence conditions are met (such as the duality gap is less than 1% or the multiplier update amplitude is lower than the threshold);

[0051] Generation of an overall scheduling plan: After the iteration converges, the scheduling center merges the optimization results of each subsystem to generate a global optimal scheduling plan. It generates an overall optimal scheduling instruction that takes into account both economy (such as minimizing the total operating cost) and safety (such as ensuring that the power of the interconnection line does not exceed the limit), and sends it to the execution agencies of each subsystem through the MQTT protocol.

[0052] The present invention has the following beneficial effects:

[0053] In this paper, by integrating fuzzy logic and game theory, the objectives of different stakeholders are considered as game participants. Fuzzy membership functions are used to quantify the priority of each objective, enabling dynamic screening and collaborative optimization of Pareto solutions. This helps to find more reasonable and acceptable scheduling solutions while protecting the interests of all parties. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 This is a step diagram of a hierarchical optimization scheduling method for deep peak regulation of thermal power units proposed by the present invention. DETAILED DESCRIPTION

[0055] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0056] See also Figure 1 As shown, the present invention is a hierarchical optimization scheduling method for deep peak regulation of thermal power units, comprising the following steps:

[0057] System layering: Leveraging IoT technology to deploy sensor networks, real-time data collection is used to collect unit operating parameters, grid load demands, and external meteorological data. Edge computing nodes are used to cleanse and extract features from this data, dividing the power system into several subsystems. Each subsystem contains a cluster of thermal power units and differentiated load demands. The subsystem boundaries are dynamically determined by the grid topology and power flow distribution.

[0058] Subsystem-level forecasting: Based on historical data and long short-term memory (LSTM) networks, we build subsystem-level load forecasting and renewable energy output forecasting models to generate basic dispatch plans. We use an improved multi-objective particle swarm optimization (MOPSO) algorithm for preliminary optimization to generate Pareto frontiers for multiple objectives, including economic efficiency, carbon emissions, and peak-shaving margins.

[0059] Dynamic preference-driven multi-objective collaborative optimization: This system receives multi-objective data on economic, environmental, and social benefits from each subsystem, adjusts the weights of each objective through dynamic preference settings, and performs multi-objective collaborative optimization. While meeting system requirements, it models the objectives of the power grid company, environmental protection department, and end users as game participants, quantifies the priorities of each party using a fuzzy membership function, and generates dynamic game solutions using a Nash equilibrium solver, screening out the Pareto optimal solution that balances the interests of multiple parties.

[0060] Dynamic safety domain online verification: This system receives real-time unit operating data, such as key parameters like main steam pressure, temperature, and turbine shaft vibration, and uses a graph neural network (GNN) to dynamically verify the safety domain boundaries, promptly identifying potential safety hazards. Based on historical fault data and reinforcement learning, it pre-generates a library of rescheduling strategies for N-1 fault scenarios. When a fault occurs, it selects the optimal strategy through dynamic time warping (DTW).

[0061] Virtual synchronous machine collaborative support: simulates the operating characteristics of synchronous generators, receives unit operating status data such as frequency, voltage, power factor, etc., optimizes control system strategies and adjusts equipment parameters;

[0062] Global collaboration and iterative optimization: The optimal scheduling plan for each subsystem is uploaded to the dispatch center through the Message Queue Telemetry Transport (MQTT) protocol for global constraint verification (such as cross-regional interconnection line power limit). If there is a conflict, the conflicting constraints are decomposed through the Lagrangian relaxation method, and the subsystem is returned for iterative optimization to form an overall scheduling plan.

[0063] In one embodiment, the power system is divided into several subsystems, each subsystem includes a cluster of thermal power units and differentiated load requirements, and the subsystem boundaries are dynamically determined based on the grid topology and power flow distribution, including the following steps:

[0064] Data collection and preprocessing: Collect grid topology (node-branch connection relationships), thermal power unit parameters (output range, ramp rate), load demand time series curves (categorized by type, such as industrial / residential load), line flow limits, and real-time operating data. Load demand is clustered by type, and the unit output range is normalized to the [0, 1] range.

[0065] Initial topology partitioning based on graph theory: construct a weighted adjacency matrix, use the inverse of line reactance as the edge weight, reflect the electrical distance between nodes, and calculate modularity. The calculation formula is expressed as Among them, A ij is the adjacency matrix element, k i =∑ j A ij is the node degree, is the total edge weight, δ(c i , c j ) is the community indicator function (1 when nodes i and j are in the same community, otherwise 0), and the Louvain algorithm is used to iteratively optimize the modularity and divide the power grid into an initial set of subsystems;

[0066] Subsystem boundary correction under power flow constraints: Power balance constraints are established based on the AC power flow equation. The power flow equation constraint is expressed as in, To contribute to the unit, is the load demand, G ij +jB ij It is an admittance matrix element and dynamically adjusts the subsystem boundaries by setting the tie-line power flow threshold and voltage over-limit detection criteria. For example, when the tie-line power flow exceeds the limit, the adjacent subsystems are merged, or when the voltage exceeds the limit, the subsystem is split, thereby ensuring the power balance within the subsystem and meeting the safe operation boundary.

[0067] Differentiated load demand matching: By calculating the load demand fluctuation coefficient and the unit regulation capacity coefficient, the load demand fluctuation coefficient calculation formula is expressed as The calculation formula of the unit regulation capacity coefficient is expressed as follows: A clustering algorithm is used to match high-fluctuation loads with high-regulation-capacity units to the same subsystem, achieving coordinated optimization of load characteristics and unit regulation capabilities, and improving the subsystem's adaptability to deep peak-shaving needs.

[0068] Dynamic boundary adjustment mechanism: By real-time monitoring of the subsystem self-sufficiency rate and the proportion of interconnection line power, and setting threshold trigger rules, if the subsystem self-sufficiency rate is <0.8 and the interconnection line power is >0.9, the subsystem merger is triggered; if the subsystem self-sufficiency rate is >1.2 and there is an interruptible load, the subsystem is split. A closed-loop feedback adjustment mechanism is established, which enables the subsystem division to respond to real-time changes in operating status, ensuring the timeliness and economy of the division.

[0069] In one embodiment, the specific steps of the subsystem-level prediction are:

[0070] Data preprocessing: Min-Max normalization is used to standardize historical load, renewable energy output, and meteorological data, and a sliding window technique is used to construct time series input samples.

[0071] LSTM model construction: Utilizing the forget gate, input gate, output gate, and cell state update mechanism of a two-layer LSTM network (containing 64 hidden units), the model dynamically adjusts the ratio of historical information retention to new information writing through nonlinear activation functions (such as Sigmoid and Tanh). This model extracts in-depth features of power load time series patterns (such as daily periodicity and holiday effects) and renewable energy output volatility (such as the wind speed-power mapping relationship). Finally, the fully connected layer outputs a joint forecast of future load and renewable energy output, providing high-precision boundary conditions for scheduling.

[0072] Improved multi-objective particle swarm optimization (MOPSO): First, three objective functions are defined: economy (the weighted sum of coal consumption cost and start-up and shutdown cost of thermal power units), carbon emissions (quadratic function model of unit output), and peak-shaving margin (the difference between the system's maximum adjustable capacity and load demand). Power balance, upper and lower limits of unit output, and ramp rate constraints are introduced. Subsequently, dynamic inertia weight (which decays linearly with the number of iterations) and adaptive Gaussian mutation (adjusting the mutation amplitude based on the solution set congestion) are used to enhance the algorithm's global search capability and avoid premature convergence. Ultimately, a Pareto frontier solution set is generated (such as three typical solutions: economy first, low carbon emissions, and high peak-shaving capacity), providing decision makers with multi-dimensional trade-off options.

[0073] In one embodiment, the method of receiving multi-objective data such as economic benefits, environmental benefits, and social benefits from each subsystem, adjusting the weight of each objective through dynamic preference settings, performing multi-objective collaborative optimization, modeling the objectives of the power grid company, the environmental protection department, and the end user as game participants, quantifying the priorities of each party using a fuzzy membership function, generating a dynamic game solution through a Nash equilibrium solver, and screening out a Pareto optimal solution that takes into account the interests of multiple parties includes the following steps:

[0074] Multi-objective function modeling: Quantify the economic benefits of the power grid company, the environmental benefits of the environmental protection department, and the social benefits of the end users. The calculation formula of the quantified economic benefits of the power grid company is expressed as Among them, min F1 is the minimum total operating cost, is the fuel cost of unit i, is the start-up and shutdown cost, and the calculation formula of the quantified environmental benefits of the environmental protection department is expressed as Among them, min F2 is the minimized CO2 emission, α i , βi , γ i is the emission coefficient of unit i, and the calculation formula of the quantified social benefits of the end user is expressed as Among them, max F3 is, Sat(λ t ) is the electricity price satisfaction function, Reliability t It is an indicator of power supply reliability;

[0075] Dynamic preference quantification: Utilizing fuzzy membership functions, the priorities of the three parties' objectives are dynamically quantified, mapping the real-time demands for economic, environmental, and social benefits into continuous weights in the [0, 1] interval. Through historical extreme value normalization, preferences are adaptively adjusted based on system status (e.g., fluctuations in renewable energy output and carbon emission quotas), ensuring that weight allocation is strongly coupled with actual operating scenarios.

[0076] Game theory modeling: The three parties are modeled as game participants through non-cooperative game theory. The subsystem output adjustment is used as the strategy variable, and the weighted objective function is used as the benefit function. A game model with power balance constraints is constructed to simulate the strategic interaction process of the three parties under conflicting interests. The power balance constraint is expressed as ∑ΔP i =0;

[0077] Nash equilibrium solution: The Nash equilibrium solver is used to iteratively solve the game equilibrium. The equilibrium condition is expressed as Generate a stable strategy portfolio under the conditions of satisfying system constraints, ensuring that no party can improve its own benefits by unilaterally changing the strategy, thereby ensuring the stability and global feasibility of the strategy;

[0078] Pareto optimal solution screening: Eliminate inferior solutions from the Nash equilibrium solution set. The inferior solution is represented by a solution that is better in at least one objective and not inferior in other objectives. Combined with dynamic weight verification, the matching of the solution with the current system state is finally output to balance economy, environmental protection, and fairness.

[0079] In one embodiment, the real-time reception of unit operation data, the use of a graph neural network (GNN) to dynamically verify the safety domain boundary, and the pre-generation of a rescheduling strategy library under an N-1 fault scenario based on historical fault data and reinforcement learning include the following steps:

[0080] Real-time data reception and graph structure construction: A distributed sensor network collects unit operating data (such as main steam pressure, temperature, and turbine shaft vibration) in real time, and constructs a graph structure representing the unit's operating status. Physical equipment (boilers, turbines, and generators) are abstracted as nodes, and energy or signal flows between devices are abstracted as edges. The node feature matrix integrates multi-source parameters (such as pressure, temperature, and vibration displacement), and the adjacency matrix defines the physical connection relationships between devices.

[0081] Dynamic safety domain boundary verification based on graph neural network (GNN): Graph Attention Network (GAT) is used to perform time series modeling on graph structure data. The attention mechanism aggregates neighbor node information and dynamically updates the safety domain boundary parameters. The safety domain is defined as a set that satisfies multiple constraints (such as the upper limit of the main steam pressure and the vibration threshold of the turbine shaft system), which is expressed as Among them, f i (·) is the i-th safety constraint function, θ i (t) is the time-varying parameter output by the graph neural network (GNN). The graph neural network (GNN) verifies in real time whether the current state violates the safety constraints by learning the safety boundary patterns in historical operation data;

[0082] Generation of a rescheduling strategy library: A Markov decision process (MDP) is constructed using historical fault data. The unit output adjustment amount is defined as the action space, and the degree of safety domain violation and load demand are defined as the state space. The optimal strategy is trained through a deep Q-network (DQN). The reward function is designed to balance safety (penalizing safety domain violations) and economy (penalizing excessive output adjustments). Ultimately, a rescheduling strategy library covering various fault scenarios (such as unit tripping and pipeline rupture) is generated.

[0083] In one embodiment, the specific steps of selecting the optimal strategy through dynamic time warping (DTW) are:

[0084] Real-time fault feature sequence extraction: Extracting the time series features that characterize the fault evolution from the fault signal output by the dynamic safety domain verification module;

[0085] Pre-stored fault scenario library: Build a standardized feature sequence library based on historical fault data, and record the complete evolution pattern (such as the time series change of vibration displacement) of each typical fault type (such as sudden change in turbine shaft vibration);

[0086] Dynamic Time Warping (DTW) matching: By constructing a cumulative distance matrix and using a recursive formula to calculate the minimum matching distance between the real-time sequence and the pre-stored sequence, the recursive formula is expressed as:

[0087]

[0088] Where D(i, j) represents the minimum matching distance. Its core purpose is to overcome differences in fault development rates (such as the timing misalignment between fast-bursting faults and slowly deteriorating faults) through flexible time alignment. This ensures accurate pattern recognition even when the sampling interval of the fault feature sequence fluctuates due to scheduling instructions.

[0089] Optimal strategy selection and execution: Based on the dynamic time warping (DTW) matching results, the response plan with the highest similarity in the pre-stored strategy library is selected and directly transmitted to the unit actuator through closed-loop control. This constructs a response chain from fault detection to strategy execution under the complex operating conditions of deep peak regulation of thermal power units.

[0090] In one embodiment, the specific steps of simulating the operating characteristics of a synchronous generator, receiving the operating status data of the unit, optimizing the control system strategy and adjusting the equipment parameters are as follows:

[0091] Synchronous generator characteristic modeling: Based on the second-order synchronous generator model, the virtual rotor motion equation and voltage regulation equation are constructed. Using real-time collected unit frequency, voltage, power factor, active power, and reactive power, the rotor inertia of the synchronous generator is simulated (kinetic energy buffering is achieved through the dynamic balance equation of angular velocity deviation and mechanical-electromagnetic torque difference) and frequency and voltage regulation characteristics (voltage and reactive power coupling control is achieved through a PI controller). Therefore, when the deep peak regulation of the thermal power unit causes a sudden load drop, the virtual inertia can slow the frequency change rate and avoid parameter oscillation caused by rapid load reduction.

[0092] Control strategy optimization and dynamic parameter adjustment: Using droop control optimization and virtual damping enhancement technology, the active output is dynamically adjusted according to the frequency deviation and low-frequency oscillation is suppressed. The droop control optimization is expressed as P = P0 + k f (f0-f), where P0 is the reference power, k f is the frequency droop coefficient, f0 is the rated frequency, and the load rate adaptive parameter setting is used to increase the virtual inertia constant under low load conditions to improve the unit's resistance to power fluctuations;

[0093] Multi-time scale collaborative control: Achieve primary frequency modulation through droop control, combine global collaboration and iterative optimization modules to modify the scheduling plan, expressed as P VSG =P primary +γ∫(P ref -P VSG )dt, where γ is the integral gain, P ref The power command is issued, and the safety domain verification module is linked to actively limit power under fault conditions to avoid equipment damage.

[0094] In one embodiment, the specific steps of the global collaboration and iterative optimization are:

[0095] Subsystem optimal solution upload and global constraint verification: Each subsystem uploads independent optimization results (such as unit output plan and tie line power distribution) to the dispatch center through the MQTT protocol. The dispatch center builds a global optimization model, which is expressed as:

[0096]

[0097] Among them, C i (·) is the operating cost of subsystem i, is the transmission limit of tie line (i, j), S is the set of subsystems, and L is the set of tie lines;

[0098] Lagrangian relaxation decomposes conflicting constraints: When a global constraint conflict is detected (e.g., inter-subsystem tie line power exceeding limits), Lagrangian relaxation is used to convert coupling constraints (e.g., the tie line power balance equation) into decomposable penalty terms. By constructing a dual problem involving Lagrangian multipliers, the original problem is decomposed into independent subsystem-level optimization models. Each subsystem focuses on a local cost function and a weighted multiplier term (e.g., the optimization objective of subsystem A becomes its own operating cost plus the product of the tie line multiplier and the power plan). This transforms the global coordination problem into an iterative solution process.

[0099] Subsystem iterative optimization and multiplier update: Dynamically update the Lagrange multiplier through the subgradient method, driving the subsystem to gradually modify the dispatch plan (such as reducing the tie line power plan or adjusting the unit output) until the convergence conditions are met (such as the duality gap is less than 1% or the multiplier update amplitude is lower than the threshold);

[0100] Generation of an overall scheduling plan: After the iteration converges, the scheduling center merges the optimization results of each subsystem to generate a global optimal scheduling plan. It generates an overall optimal scheduling instruction that takes into account both economy (such as minimizing the total operating cost) and safety (such as ensuring that the power of the interconnection line does not exceed the limit), and sends it to the execution agencies of each subsystem through the MQTT protocol.

[0101] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A hierarchical optimization scheduling method for deep peak regulation of thermal power units, characterized in that: The following steps are involved: System layering: Leveraging IoT technology to deploy sensor networks, real-time data collection is used to collect unit operating parameters, grid load demands, and external meteorological data. Edge computing nodes are used to cleanse and extract features from this data, dividing the power system into several subsystems. Each subsystem contains a cluster of thermal power units and differentiated load demands. The subsystem boundaries are dynamically determined by the grid topology and power flow distribution. Subsystem-level forecasting: Based on historical data and long-short-term memory networks, a subsystem-level load forecasting and renewable energy output forecasting model is constructed to generate a basic dispatch plan. An improved multi-objective particle swarm algorithm is used for preliminary optimization to generate a multi-objective Pareto frontier. Dynamic preference-driven multi-objective collaborative optimization: This system receives multi-objective data from each subsystem, adjusts the weights of each objective through dynamic preference settings, and performs multi-objective collaborative optimization. While meeting system requirements, it models the objectives of the power grid company, environmental protection department, and end users as game participants, quantifies the priorities of each party using a fuzzy membership function, and generates dynamic game solutions using a Nash equilibrium solver, ultimately selecting the Pareto optimal solution that balances the interests of all parties. Dynamic safety domain online verification: Receives unit operating data in real time, uses graph neural networks to dynamically verify safety domain boundaries, and pre-generates a library of rescheduling strategies for fault scenarios based on historical fault data and reinforcement learning. When a fault occurs, the optimal strategy is selected through dynamic time warping. Virtual synchronous machine collaborative support: simulates the operating characteristics of synchronous generators, receives unit operating status data, optimizes control system strategies and adjusts equipment parameters; Global collaboration and iterative optimization: The optimal scheduling plan for each subsystem is uploaded to the scheduling center through the message queue telemetry transmission protocol for global constraint verification. If there is a conflict, the conflicting constraints are decomposed using the Lagrangian relaxation method, and the subsystem is returned for iterative optimization to form an overall scheduling plan.

2. A hierarchical optimization scheduling method for deep peak regulation of thermal power units according to claim 1, characterized in that: The power system is divided into several subsystems, each of which includes a cluster of thermal power units and differentiated load requirements. The subsystem boundaries are dynamically determined based on the grid topology and power flow distribution, including the following steps: Data collection and preprocessing: Collect grid topology, thermal power unit parameters, load demand time series curves, line power flow limits, and real-time operating data. Cluster load demand by type and normalize the unit output range to the [0, 1] interval. Initial topological partitioning based on graph theory: construct a weighted adjacency matrix, use the inverse of the line reactance as the edge weight, and calculate the modularity. The calculation formula is expressed as Among them, A ij is the adjacency matrix element, k i =∑ j A ij is the node degree, is the total edge weight, δ(c i , c j ) is the community indicator function, and the Louvain algorithm is used to iteratively optimize the modularity and divide the power grid into an initial subsystem set; Subsystem boundary correction under power flow constraints: Power balance constraints are established based on the AC power flow equation. The power flow equation constraint is expressed as in, To contribute to the unit, is the load demand, G ij +jB ij It is an admittance matrix element and dynamically adjusts the subsystem boundary by setting the tie line power flow threshold and voltage over-limit detection criteria; Differentiated load demand matching: By calculating the load demand fluctuation coefficient and the unit regulation capacity coefficient, the load demand fluctuation coefficient calculation formula is expressed as The calculation formula of the unit regulation capacity coefficient is expressed as follows: A clustering algorithm is used to match high-fluctuation loads and high-regulation-capacity units to the same subsystem; Dynamic boundary adjustment mechanism: By real-time monitoring of the subsystem self-sufficiency rate and interconnection line power ratio, and setting threshold trigger rules, a closed-loop feedback adjustment mechanism is established.

3. The hierarchical optimization scheduling method for deep peak regulation of thermal power units according to claim 1 is characterized in that: The specific steps of the subsystem-level prediction are: Data preprocessing: Min-Max normalization is used to standardize historical load, renewable energy output, and meteorological data, and a sliding window technique is used to construct time series input samples. LSTM model construction: Utilizing the forget gate, input gate, output gate, and cell state update mechanism of a two-layer LSTM network, a nonlinear activation function is used to dynamically adjust the ratio of historical information retention to new information writing. This allows for deep feature extraction of power load timing patterns and renewable energy output volatility, ultimately outputting a combined forecast of future load and renewable energy output through a fully connected layer. Improved multi-objective particle swarm optimization: First, the three objective functions of economy, carbon emissions and peak margin are defined, and power balance, upper and lower limits of unit output and ramp rate constraints are introduced. Then, the global search capability of the algorithm is enhanced through dynamic inertia weight and adaptive Gaussian mutation, and finally the Pareto front solution set is generated.

4. The hierarchical optimization scheduling method for deep peak regulation of thermal power units according to claim 1 is characterized in that: The method receives multi-objective data from each subsystem, adjusts the weight of each objective through dynamic preference settings, performs multi-objective collaborative optimization, models the objectives of the power grid company, the environmental protection department, and the end user as game participants, quantifies the priorities of each party using a fuzzy membership function, generates a dynamic game solution through a Nash equilibrium solver, and screens out a Pareto optimal solution that takes into account the interests of multiple parties, including the following steps: Multi-objective function modeling: quantify the economic benefits of the power grid company, the environmental benefits of the environmental protection department, and the social benefits of the end users. The calculation formula of the quantified economic benefits of the power grid company is expressed as Among them, minF1 is the minimum total operating cost, is the fuel cost of unit i, is the start-up and shutdown cost, and the calculation formula of the quantified environmental benefits of the environmental protection department is expressed as Among them, minF2 is the minimized CO2 emission, α i , β i , γ i is the emission coefficient of unit i, and the calculation formula of the quantified social benefits of the end user is expressed as Among them, maxF3 is, Sat(λ t ) is the electricity price satisfaction function, Reliability t It is an indicator of power supply reliability; Dynamic preference quantification: Utilizes fuzzy membership functions to dynamically quantify the priorities of the three parties' objectives, mapping the real-time demands for economic, environmental, and social benefits into continuous weights in the [0,1] interval. Through historical extreme value normalization, preferences are adaptively adjusted according to system status. Game theory modeling: The three parties are modeled as game participants through non-cooperative game theory. The subsystem output adjustment is used as the strategy variable, and the weighted objective function is used as the benefit function. A game model with power balance constraints is constructed to simulate the strategic interaction process of the three parties under conflicting interests. The power balance constraint is expressed as ∑ΔP i =0; Nash equilibrium solution: The Nash equilibrium solver is used to iteratively solve the game equilibrium. The equilibrium condition is expressed as Generate a stable strategy portfolio while satisfying system constraints; Pareto optimal solution screening: Eliminate inferior solutions from the Nash equilibrium solution set, and combine dynamic weight verification to verify the compatibility of the solution with the current system state, and ultimately output a scheduling plan that takes into account economy, environmental protection, and fairness.

5. The hierarchical optimization scheduling method for deep peak regulation of thermal power units according to claim 1 is characterized in that: The real-time receiving of unit operation data, the use of graph neural networks to dynamically verify the safety domain boundary, and the pre-generation of a rescheduling strategy library under the N-1 fault scenario based on historical fault data and reinforcement learning, include the following steps: Real-time data reception and graph structure construction: The distributed sensor network collects unit operating data in real time and constructs a graph structure representation of the unit's operating status. Physical devices are abstracted as nodes, and energy or signal flows between devices are abstracted as edges. The node feature matrix integrates multi-source parameters, and the adjacency matrix defines the physical connection relationships between devices. Dynamic security domain boundary verification based on graph neural network: Graph attention network is used to perform time series modeling on graph structure data. Neighbor node information is aggregated through the attention mechanism to dynamically update the security domain boundary parameters. The security domain is defined as a set that satisfies multiple constraints and is expressed as Among them, f i (·) is the i-th safety constraint function, θ i (t) is the time-varying parameter output by the graph neural network. The graph neural network verifies in real time whether the current state violates the safety constraints by learning the safety boundary patterns in historical operation data; Generation of a re-dispatch strategy library: Utilize historical fault data to construct a Markov decision process, define the unit output adjustment as the action space, and the safety domain violation degree and load demand as the state space. The optimal strategy is trained through a deep Q network, and the reward function is designed to balance safety and economy. Ultimately, a re-dispatch strategy library covering various fault scenarios is generated.

6. A hierarchical optimization scheduling method for deep peak regulation of thermal power units according to claim 5, characterized in that: The specific steps of selecting the optimal strategy through dynamic time warping are: Real-time fault feature sequence extraction: Extract the time series features that characterize the fault evolution from the fault signal output by the dynamic safety domain verification module; Pre-stored fault scenario library: Build a standardized feature sequence library based on historical fault data, and record the complete evolution pattern of each typical fault type; Dynamic time warping matching: By constructing a cumulative distance matrix and using a recursive formula to calculate the minimum matching distance between the real-time sequence and the pre-stored sequence, the recursive formula is expressed as: Where D(i, j) represents the minimum matching distance; Optimal strategy selection and execution: Based on the dynamic time warping matching results, the response plan with the highest similarity in the pre-stored strategy library is selected and directly sent to the unit actuator through closed-loop control.

7. The hierarchical optimization scheduling method for deep peak regulation of thermal power units according to claim 1, characterized in that: The specific steps of simulating the operating characteristics of the synchronous generator, receiving the unit operating status data, optimizing the control system strategy and adjusting the equipment parameters are as follows: Synchronous generator characteristic modeling: Based on the second-order synchronous generator model, the virtual rotor motion equation and voltage regulation equation are constructed. The rotor inertia and frequency and voltage regulation characteristics of the synchronous generator are simulated using the real-time collected unit frequency, voltage, power factor, active power, and reactive power. Control strategy optimization and dynamic parameter adjustment: Using droop control optimization and virtual damping enhancement technology, the active output is dynamically adjusted according to the frequency deviation and low-frequency oscillation is suppressed. The droop control optimization is expressed as P = P0 + k f (f0-f), where P0 is the reference power, k f is the frequency droop coefficient, f0 is the rated frequency, and the virtual inertia constant is increased under low load conditions by setting the load rate adaptive parameters; Multi-time scale collaborative control: Achieve primary frequency modulation through droop control, combine global collaboration and iterative optimization modules to modify the scheduling plan, expressed as P VSG =P primary +γ∫(P ref -P VSG )dt, where γ is the integral gain, P ref The power command is issued, and the safety domain verification module is linked to actively limit the power under fault conditions.

8. The hierarchical optimization scheduling method for deep peak regulation of thermal power units according to claim 1, characterized in that: The specific steps of the global collaboration and iterative optimization are: Subsystem optimal solution upload and global constraint verification: Each subsystem uploads the independent optimization results to the dispatch center through the MQTT protocol. The dispatch center builds a global optimization model, which is expressed as: Among them, C i (·) is the operating cost of subsystem i, is the transmission limit of tie line (i, j), S is the set of subsystems, and L is the set of tie lines; Lagrangian relaxation method to decompose conflicting constraints: When a global constraint conflict is detected, Lagrangian relaxation method is used to convert the coupled constraints into decomposable penalty terms. By constructing a dual problem involving Lagrangian multipliers, the original problem is decomposed into independent subsystem-level optimization models, allowing each subsystem to focus on the local cost function and multiplier weighting terms. Subsystem iterative optimization and multiplier update: Dynamically update the Lagrange multiplier through the subgradient method, driving the subsystem to gradually correct the scheduling plan until the convergence conditions are met; Generation of overall scheduling plan: After the iteration converges, the scheduling center merges the optimization results of each subsystem to generate a global optimal scheduling plan, generates the overall optimal scheduling instructions that take into account both economy and safety, and sends them to the execution agencies of each subsystem through the MQTT protocol.

Images