Thermal power unit grid-related system online analysis method for power grid stability analysis

Abstract

The present application relates to the field of power system safety and stability analysis and control technology, especially to the online analysis method of thermal power unit grid-related system for power grid stability analysis. The technical scheme is as follows: the online analysis method of thermal power unit grid-related system for power grid stability analysis comprises the following steps: S1. Multi-level grid-related system state perception and data acquisition: a distributed state perception network covering the core grid-connected subsystem of coal-fired units is built to realize synchronous acquisition and preprocessing of multi-source data and form standardized data flow; S2. Grid-related system dynamic characteristic extraction and state identification: based on the standardized data flow, dynamic characteristics of multiple time scales are extracted, and signal decomposition algorithm and state identification rule base are used to automatically identify the operating conditions of the grid-related system. The present application realizes second-level perception, minute-level analysis and millisecond-level strategy issuing of the grid-related system state, and the time effectiveness is improved by multiple orders of magnitude compared with the traditional offline analysis mode.

Description

Technical Field

[0001] This invention relates to the field of power system safety and stability analysis and control technology, and in particular to an online analysis method for grid-related systems of thermal power units for grid stability analysis. Background Technology

[0002] With the advancement of technology, my country's power system is undergoing profound changes, with high-proportion renewable energy integration and high-proportion application of power electronic equipment becoming prominent features of the new power system. Against this backdrop, the grid connection performance of synchronous generator units (especially thermal power units) plays an increasingly important role in supporting the safe and stable operation of the power grid.

[0003] The grid-related systems of thermal power units mainly include: ① Turbine speed control system: responds to frequency changes and participates in primary frequency regulation; ② Excitation system: maintains stable terminal voltage and provides reactive power support; ③ Generator: generates active power and provides damping power for the unit speed control system; ④ Power system stabilizer (PSS): suppresses low-frequency oscillations and provides positive damping; ⑤ Automatic generation control (AGC): tracks dispatch commands and participates in secondary frequency regulation; ⑥ Coordinated control system (CCS): realizes coordination between the boiler and turbine and ensures regulation capability.

[0004] The existing technology has the following shortcomings: 1. Insufficient timeliness in analysis: Traditional network performance analysis relies on offline tests and periodic inspections (such as primary frequency regulation tests and PSS parameter measurements), with time spans measured in months or quarters, failing to reflect the real-time status of the units. When the performance of the network system deteriorates or parameters drift, the dispatching terminal cannot detect it in a timely manner, posing a stability risk.

[0005] 2. Lack of multi-system coupling analysis: Existing methods mostly analyze individual grid-connected systems independently, failing to fully consider the strong coupling characteristics between excitation, speed regulation, PSS, and AGC. In actual operation, the adjustment actions of one system can affect other systems through electromechanical coupling, making it difficult to accurately assess the unit's comprehensive support capability for the power grid from a single perspective.

[0006] 3. Inadequate for new power system scenarios: With the increasing penetration of new energy sources, the frequency / voltage fluctuation characteristics of the power grid have changed significantly (increased fluctuation amplitude and faster fluctuation rate). Traditional analysis methods based on linearization assumptions are difficult to adapt to scenarios with strong uncertainty, and there is an urgent need to introduce new technologies such as artificial intelligence to improve the level of intelligent analysis.

[0007] 4. Control strategy optimization is lagging behind. Existing grid-connected system control parameters are mostly fixed values ​​or based on offline simulation tuning, failing to adaptively adjust according to the real-time stability requirements of the power grid. In grid emergency situations, generating units struggle to reach their maximum stability support potential. Summary of the Invention

[0008] This invention proposes an online analysis method for grid-related systems of thermal power units for grid stability analysis, which solves the problem that traditional methods in the prior art cannot meet the needs of new power systems for real-time, accurate, comprehensive evaluation and adaptive control of the grid-related performance of thermal power units.

[0009] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: An online analysis method for grid-connected systems of thermal power units for power grid stability analysis includes the following steps: S1. Multi-level grid-connected system state perception and data acquisition: Constructing a distributed state perception network covering the core grid-connected subsystem of the coal-fired unit to achieve synchronous acquisition and preprocessing of multi-source data, forming a standardized data stream; S2. Grid-connected system dynamic feature extraction and state identification: Based on the standardized data stream, extracting dynamic features at multiple time scales, and using a signal decomposition algorithm and a state identification rule base to automatically identify the operating conditions of the grid-connected system; S3. Parallel system simulation and prediction based on a hybrid model: Constructing a hybrid model including the grid-connected subsystem and related subsystems, and performing simulation on the system. The simulation is designed to construct an end-to-end correlation model to evaluate the comprehensive support efficiency of the generating units for grid stability in real time; S4. Grid performance index mapping model: Establish a quantitative evaluation index system and use a nonlinear regression model to map multi-dimensional grid performance indexes to grid stability margin indices; S5. Online diagnosis, early warning and optimized control strategy generation: Execute hierarchical early warning based on the analysis results of steps S3 and S4, and automatically generate optimized control strategies for grid-connected systems under alarm conditions; S6. Closed-loop verification and strategy distribution: After virtual verification of the generated optimized control strategies through a parallel simulation platform, distribute them to the unit's distributed control system (DCS) for execution through a security authentication mechanism.

[0010] Furthermore, in step S1, the core grid-connected subsystem includes a turbine speed control system, an excitation control system, a synchronous generator, a power system stabilizer (PSS), an automatic generation control (AGC), a primary frequency regulation system, and a coordinated control system (CCS); the synchronous acquisition is achieved through a unified clock synchronization protocol; the data includes electrical measurement data, mechanical measurement data, control command data, and status identification data; the preprocessing is completed at the edge computing node, including outlier filtering, data alignment, and compression encoding.

[0011] Furthermore, in step S2, the signal decomposition algorithm is an improved variational mode decomposition (VMD) algorithm, which is used to decompose non-stationary signals and extract time-frequency domain features in combination with wavelet packet transform; the dynamic features of the multiple time scales include: millisecond-level features reflecting the rapid dynamics of the system, second-level features reflecting the regulation performance, and minute-level features reflecting the long-term trend.

[0012] Furthermore, in step S3, the hybrid model adopts a parallel architecture of mechanistic model and data-driven model, and the outputs of the two are integrated through weighted fusion, residual correction or hierarchical decision-making; the end-to-end correlation model is a deep reinforcement learning model.

[0013] Furthermore, in step S4, the quantitative evaluation index system includes frequency support index, voltage support index, and power angle support index; the nonlinear regression model is a Gaussian process regression (GPR) model, and its model parameters are updated online through Bayesian optimization.

[0014] Furthermore, in step S5, the graded early warning mechanism includes a normal state, a warning state, and an alarm state; the generation of the optimized control strategy adopts a multi-objective optimization algorithm to take into account both the grid stability requirements and the unit's safe operation constraints.

[0015] Furthermore, in step S1, the distributed state awareness network adopts a three-layer architecture of end-edge-cloud: the end layer consists of intelligent sensing terminals deployed in the control cabinets of each network-related system; the edge layer consists of edge computing nodes deployed in the unit's electronics room that integrate field-programmable gate array (FPGA) acceleration cards; and the cloud layer consists of a cloud analysis platform deployed at the scheduling end.

[0016] Furthermore, the improved VMD algorithm adaptively adjusts the number of modes K and the penalty factor α, calculates the center frequency of each intrinsic mode function (IMF) component, merges modes with a center frequency difference less than a threshold, dynamically adjusts α based on the signal reconstruction error, and uses the mutual information criterion to screen effective IMF components.

[0017] Furthermore, the state space of the deep reinforcement learning model integrates the features of the excitation system, speed regulation system, PSS, AGC, and the grid operation state. Its action space is defined as the adjustment amount of control parameters of multiple grid-related systems. Its reward function is designed as a weighted combination of the grid frequency, voltage, power angle stability margin improvement amount, unit regulation cost, and constraint violation penalty.

[0018] Furthermore, the Gaussian process regression (GPR) model employs a multi-kernel function combination strategy, including radial basis function (RBF) kernel, periodic kernel (PER) kernel, and linear kernel (LIN) kernel; the generation of the optimized control strategy uses an improved third-generation non-dominated sorting genetic algorithm (NSGA-III); in step S6, the parallel simulation platform is a hybrid model parallel simulation platform built based on digital twin and hardware-in-the-loop (HIL) technology; the security authentication mechanism uses the national cryptographic algorithm SM2 to verify the signature of control commands.

[0019] The positive effects of this invention are: Multi-level collaborative perception architecture: Construct a network for sensing the status of network-connected systems covering multiple time scales from milliseconds to seconds to minutes, enabling microsecond-level synchronous acquisition and edge preprocessing, and solving the problem of multi-source heterogeneous data fusion.

[0020] Intelligent Feature Extraction Engine: This paper proposes an improved VMD algorithm combined with deep feature learning to automatically extract non-stationary operating features of network-connected systems and achieve accurate identification of operating conditions.

[0021] End-to-end stability analysis model: The first end-to-end mapping model of "grid state-grid stability" based on deep reinforcement learning is created, which breaks through the limitation of traditional mechanism modeling in poor adaptability to complex scenarios.

[0022] Online optimization control strategy: Establish a quantitative mapping relationship between grid performance indicators and grid stability margin to achieve adaptive optimization of control parameters, taking into account both grid demand and unit safety constraints.

[0023] Real-time performance improvement: Enables second-level perception, minute-level analysis, and millisecond-level policy delivery of network-connected system status, improving timeliness by three orders of magnitude compared to traditional offline analysis modes.

[0024] Enhanced accuracy: By establishing an end-to-end mapping model through deep reinforcement learning, the stability assessment accuracy is >95% in the context of strong uncertainty in new power systems, which is better than the traditional mechanism model (accuracy of about 80%).

[0025] Economic optimization: By optimizing control strategies online, the grid connection potential of the generating units is fully explored, and the amount of wind and solar power curtailed is reduced by about 3-5% while ensuring grid stability, thereby improving the capacity for renewable energy consumption.

[0026] Security Assurance: Digital twin verification and security authentication mechanisms ensure the reliability of the strategy and avoid the risks associated with direct online trial and error. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the overall architecture of the online analysis method for grid-related systems of thermal power units for grid stability analysis according to the present invention. Figure 2 This is a topology diagram of the state-aware network of the multi-level network-connected system in this invention; Figure 3 This is a flowchart of the feature extraction process of the improved VMD algorithm in this invention; Figure 4 This is a network structure diagram of the deep reinforcement learning analysis engine in this invention; Figure 5 This is a schematic diagram illustrating the mapping relationship between grid performance indicators and grid stability margin in this invention; Figure 6 This is a flowchart of the online diagnosis, early warning, and strategy generation process in this invention; Figure 7 This is a diagram of the digital twin closed-loop verification system architecture in this invention. Detailed Implementation

[0028] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0029] Example 1

[0030] An online analysis method for grid-related systems of thermal power units for power grid stability analysis includes the following steps: S1. Status Awareness and Data Acquisition of Multi-Level Network-Interconnected Systems A distributed state awareness network was established, covering the turbine speed control system, excitation control system, synchronous generator, power system stabilizer (PSS), automatic generation control (AGC), primary frequency regulation system, and collaborative control system (CCS) of coal-fired power units. Microsecond-level synchronous data acquisition was achieved through a unified clock synchronization protocol (IEEE 1588). Acquired data included: electrical measurement data (terminal voltage, current, active power, reactive power), mechanical measurement data (speed, valve opening, main steam pressure), control command data (excitation voltage reference value, governor command, PSS output signal), and status identification data (system operation status, limiter action indicator). Edge computing nodes were used for data preprocessing, including outlier filtering, data alignment, and compression encoding, to form a standardized data stream. S2. Dynamic Feature Extraction and State Identification of Network-related Systems Based on the standardized data stream obtained in step S1, a multi-timescale feature extraction framework is constructed: Millisecond-level characteristics: excitation system response delay, speed control system hydraulic motor time constant, PSS phase compensation characteristics; Second-level characteristics: AGC adjustment rate, primary frequency regulation response characteristics, and unit ramp-up capability; Minute-level characteristics: Network system parameter drift trends and equipment aging characteristics; An improved variational mode decomposition (VMD) algorithm is used to decompose non-stationary signals, and wavelet packet transform is used to extract time-frequency domain features. Through a preset state identification rule base, the operating conditions of the network system (normal adjustment state, saturation limit state, fault ride-through state, and abnormal parameter state) are automatically identified.

[0031] S3. Parallel System Simulation and Prediction Based on Hybrid Model A hybrid model of the grid-connected subsystem and related subsystems is constructed, adapting to different unit structures and renewable energy penetration scenarios. Simulation of each subsystem and core data extrapolation are completed simultaneously. Multi-system feature vectors are extracted from the model to form a high-dimensional state representation; the control space is defined as the set of adjustment instructions for the control parameters of the grid-connected system. The benefit function is designed as a weighted combination of the improvement in grid security and stability margin and the unit control cost; an end-to-end correlation model of "grid-connected system status - grid security and stability response" is constructed; and the hybrid model is combined to output the comprehensive support performance evaluation value of the unit for grid frequency stability, voltage stability and power angle stability in real time.

[0032] S4. Modeling the mapping between grid performance indicators and grid stability margin Establish a quantitative evaluation indicator system: ① Frequency support indicators: primary frequency modulation contribution rate, AGC tracking accuracy, and frequency modulation dead zone dynamic optimization capability; ② Voltage support indicators: excitation system strong excitation ratio, low excitation limit dynamic boundary, reactive power regulation response speed; ③ Power angle support indicators: PSS effective damping ratio, equivalent inertia time constant, and oscillation mode suppression capability; A nonlinear mapping model is constructed using Gaussian process regression (GPR) to map multidimensional grid-related performance indicators into a grid stability margin index. The mapping model parameters are updated online through Bayesian optimization to adapt to the time-varying characteristics of grid operation.

[0033] S5. Online diagnostics, early warning, and optimization control strategy generation Based on the analysis results of steps S3 and S4, a tiered early warning mechanism is implemented: Normal state: Continuous monitoring, periodic output of network performance assessment reports; Status alert: Identifies performance degradation trends and triggers preventative maintenance prompts; Alarm status: When the network system parameters are detected to be out of limit or the stability margin is insufficient, an optimized control strategy is automatically generated. The optimized control strategy includes: dynamic adjustment of the excitation system reference voltage, adaptive optimization of PSS gain and time constant, online correction of the speed regulation system droop coefficient, and reconfiguration of AGC and primary frequency regulation coordinated control parameters. The strategy generation adopts a multi-objective optimization algorithm to take into account both grid stability requirements and unit safe operation constraints.

[0034] S6. Closed-loop verification and policy distribution The generated optimized control strategy is virtually verified through a hybrid model parallel simulation platform to evaluate the grid stability improvement effect after the strategy is implemented. After the verification is passed, it is sent to the unit's DCS system for execution through a security authentication mechanism to realize online closed-loop optimization of the grid-connected system control parameters.

[0035] The distributed state-aware network described in step S1 adopts a three-layer architecture of "device-edge-cloud": End layer: Intelligent sensing terminals deployed in the control cabinets of various network-connected systems, supporting IEC 61850-MMS and IEEE C37.118 synchronous phasor transmission protocols; Edge layer: Edge computing nodes deployed in the unit's electronics room, integrating FPGA acceleration cards to achieve millisecond-level data preprocessing; Cloud layer: The cloud analytics platform deployed at the scheduling end undertakes large-scale data storage, model simulation, and prediction tasks; the three layers adopt a hybrid network of 5G private network and fiber optic channel to ensure the real-time performance and reliability of data transmission.

[0036] A hybrid modeling approach is introduced to build models of the grid-connected core subsystem and related subsystems. These models are adaptable to different unit structures and renewable energy penetration scenarios. Simulation results are used to realize operational data models and extrapolations. Effective modal components containing core dynamic characteristics are automatically selected. Hilbert transforms are applied to the selected IMF components to extract instantaneous amplitude, instantaneous frequency, and instantaneous phase, thus establishing a dynamic characteristic identifier for the grid-connected system.

[0037] The hybrid modeling approach described in step S3 employs a parallel structure of mechanistic and data-driven models. In this architecture, the mechanistic model (a white-box model built based on first principles of physics, chemistry, or biology) and the data-driven model (a black-box / grey-box model that uses machine learning or statistical methods to extract patterns from data) operate synchronously and complement each other. The mechanistic model provides highly interpretable and extrapolable physical constraints and prior knowledge, ensuring the model's rationality under boundary conditions and extreme operating conditions. The data-driven model captures complex nonlinear relationships, compensates for unmodeled dynamics resulting from mechanistic simplification, and improves the fitting accuracy and adaptability to real-world data. The parallel outputs of both can be integrated through weighted fusion, residual correction, or hierarchical decision-making, preserving the reliability of domain knowledge while maintaining the flexibility of data learning.

[0038] The Gaussian process regression mapping model described in step S4 employs a multi-kernel function combination strategy: The radial basis function (RBF) is selected as the basic kernel function to capture the global nonlinear trend. The periodic kernel function (PER) is superimposed to handle the periodic impact caused by the fluctuation of new energy power output. The linear kernel function (LIN) is introduced to characterize the linear drift characteristics of the grid-related parameters. The computational complexity is reduced by a sparsification approximation algorithm to meet the real-time requirements of online analysis (single inference time <100ms). The model uncertainty quantification output is used to guide the adaptive adjustment of the warning threshold in step S5.

[0039] The multi-objective optimization algorithm described in step S5 uses an improved NSGA-III algorithm: To address the high-dimensionality (decision variable dimension > 20) of the grid-connected system control parameter optimization problem, an adaptive adjustment mechanism for reference points is introduced to dynamically change the distribution density of reference points according to the unit's operating conditions. An individual ranking strategy based on constraint violation degree is adopted to ensure that the generated strategy strictly meets the constraints of unit thermal protection settings, environmental emission limits, and equipment life loss. The Pareto front solution set is output for the dispatcher to make decisions based on the current grid priority (frequency stability priority or voltage stability priority).

[0040] The method for constructing a hybrid model parallel simulation platform as described in step S6: Based on the actual design parameters of the generator set and the measured model of the grid-connected system, a 1:1 high-fidelity digital twin was built in the STS real-time simulation environment. Hardware-in-the-loop (HIL) technology was used to connect to the actual DCS controller to verify the execution effect of the control strategy on the real hardware. Multiple fault scenario libraries were set up to evaluate the robustness of the strategy. After the simulation verification was passed, control commands were issued through a security authentication mechanism to prevent malicious tampering.

[0041] Based on the above method, an online analysis device for grid-related systems of thermal power units for grid stability analysis is proposed, comprising: Data acquisition module: Configured with multi-protocol adapter interface, supporting IEC 61850, IEEE C37.118, ModbusTCP, OPC UA protocols, to realize data interface with the unit's DCS, DEH, excitation regulator, and PMU device; Edge computing module: integrates ARM+FPGA heterogeneous computing architecture, runs an embedded Linux system, and deploys data preprocessing and lightweight feature extraction algorithms; Cloud analytics module: Configures GPU server clusters to run deep reinforcement learning models, Gaussian process regression models, and multi-objective optimization algorithms; Human-computer interaction module: provides a visual interface for the status of network-connected systems, an interface for pushing early warning information, and an interface for confirming and issuing control strategies; Security authentication module: Integrates national cryptographic algorithms SM2 / SM3 to achieve data transmission encryption, command signature verification, and operation log auditing; Each module is interconnected with the 5G network via industrial Ethernet to form a complete online analysis system.

[0042] Example 2

[0043] Based on Example 1, the above method will be further explained, such as... Figure 1 As shown, the method of the present invention includes the following steps: Step S1: Multi-level network system status awareness and data acquisition refer to Figure 2A distributed state-aware network is deployed at the thermal power unit site. Taking a 600MW supercritical unit as an example: Terminal layer deployment: Configure an IEC 61850-MMS interface on the excitation regulator (ABB UNITROL 6000) to collect terminal voltage (sampling rate 4kHz), excitation current, and firing angle; configure an OPC UA interface on the DEH system (OVATION) to collect speed, valve position command, and actual opening degree; configure an IEEE C37.118 interface on the PMU device (NARI PCS-996) to collect synchronization phasor data (100 frames per second).

[0044] Edge deployment: Edge computing nodes (NVIDIA Jetson AGX Xavier) are configured in the unit's electronics room to run data preprocessing algorithms. Implemented functions include: outlier detection based on a sliding window (3σ criterion), data clock alignment based on the PTP protocol, and data compression based on the LZ4 algorithm (compression ratio > 5:1).

[0045] Cloud deployment: A GPU server (8×NVIDIA A100) is configured in the provincial dispatch center to undertake model training and large-scale data analysis tasks.

[0046] By using a hybrid network of 5G private network (uRLLC slicing, air interface latency <10ms) and fiber channel (bandwidth 10Gbps), real-time and reliable data transmission is ensured.

[0047] Step S2: Dynamic Feature Extraction and State Identification of Network-connected Systems Taking the excitation system as an example, the feature extraction process is explained as follows: Collect excitation system step response data (e.g., a step change of +5% in the terminal voltage reference value), and decompose it using an improved VMD algorithm, such as... Figure 3 As shown. The standard VMD algorithm requires a preset number of modes K and a penalty factor α. This invention introduces an adaptive adjustment mechanism: ① Initialize K = 2~10, α = 100~10000; ② Calculate the center frequency of each IMF component. If the difference in center frequency between adjacent IMFs is less than the threshold (e.g., 10Hz), then merge the modes. ③ Dynamically adjust α based on the signal reconstruction error to ensure the error is less than 1%; ④ Use mutual information criteria to screen effective IMFs, with a threshold set at 0.3; ⑤ Perform Hilbert transform on the effective IMF to extract instantaneous features.

[0048] Through the above processing, key features such as the time constant Te, gain Ke, and phase lag of the excitation system can be extracted and compared with the parameters of the standard model to achieve parameter drift detection.

[0049] Step S3: Multi-system coupling stability analysis based on deep reinforcement learning, such as... Figure 4 As shown.

[0050] Building a deep reinforcement learning environment: State space design: Integrating excitation system characteristics (4-dimensional), speed regulation system characteristics (5-dimensional), PSS state (3-dimensional), AGC performance (4-dimensional), and grid operation status (new energy output fluctuation rate, tie line power, system frequency deviation, etc., 6-dimensional), a 22-dimensional state vector is formed.

[0051] Motion space design: defined as the adjustment of 7 control parameters, including excitation system AVR gain (±20%), PSS gain (±30%), PSS time constant (±15%), speed control system droop coefficient (±0.5%), AGC adjustment rate limit (±10%), etc., with 7 motion dimensions.

[0052] Reward function design:

[0053] in, , , These represent the improvements in frequency stability margin, voltage stability margin, and power angle stability margin, respectively. The cost of unit regulation (such as the number of valve actions and the amount of excitation current overshoot); To constrain violations and penalties; These are the weighting coefficients.

[0054] The PPO algorithm was used for training, with safety layer constraints set as follows: shaft torsional stress < allowable value, excitation current < 1.5 times rated current, and main steam pressure fluctuation < ±0.5 MPa. After training, the model can output in real time the expected impact of adjusting each control parameter on grid stability under the current grid connection status.

[0055] Step S4: Model the mapping between grid performance indicators and grid stability margin, such as... Figure 5 As shown.

[0056] Construct a quantitative evaluation indicator system: Frequency support: primary frequency modulation contribution rate, actual power contribution / theoretical power contribution within 15 seconds; AGC tracking accuracy, RMSE of (actual output - target output) / rated capacity; Voltage support: excitation multiple, excitation peak voltage / rated excitation voltage. Reactive response speed: the time required for reactive power output to increase from 50% to 100%; Power angle support: effective damping ratio, based on mode identification results from PMU data; Equivalent inertial time constant, based on online identification of frequency change rate.

[0057] A mapping model is established using Gaussian process regression:

[0058] Among them, the comprehensive stability margin index of the power grid (0-100 points) is the power grid comprehensive stability margin index. , , These are the frequency, voltage, and power angle support index vectors, respectively.

[0059] By employing the sparsity approximation (FITC method) and selecting 1000 induction points, the single inference time is reduced to less than 100ms, thus meeting the requirements for online analysis.

[0060] Step S5: Online diagnosis, early warning, and optimization control strategy generation, such as... Figure 6 As shown.

[0061] Establish a tiered early warning mechanism: Green (Normal): δgrid > 80 points, all network-related indicators are in the excellent range, and the system generates a performance evaluation report every 15 minutes; Yellow (Attention): 60 points < δgrid ≤ 80 points, or a single indicator shows a deteriorating trend (declining for 30 consecutive minutes). The system prompts attention and suggests preventive checks. Red (Alarm): δgrid≤60 points, or an abnormality in the network-related system is detected (such as PSS exit, excitation limiter action), the system will automatically trigger the generation of optimization strategies.

[0062] The optimization strategy generation uses an improved NSGA-III algorithm: ① Dynamically adjust the density of reference points based on the current weak links in power grid stability (frequency / voltage / power angle); ② Set constraints: thermal protection settings, environmental emission limits, and equipment life loss < threshold; ③ Generate Pareto solution set, containing 5-8 optimization schemes; ④ The dispatcher selects a scheme based on the current priority, or enables automatic mode (the scheme with the largest stability margin is selected by default).

[0063] Step S6: Closed-loop verification and policy distribution Build a digital twin verification platform, such as Figure 7 As shown: ① Model building: A detailed model of the 600MW unit was built in MATLAB / Simulink, including subsystems such as turbine, boiler, excitation system, speed control system, and PSS. The accuracy of the model was verified by field tests (error <3%). ② HIL Verification: Connect the actual DCS controller (such as EMERSON Ovation) to the simulation loop through the IO interface to verify the execution effect of the control strategy on the real hardware; ③ Scenario testing: Set up fault scenarios such as new energy grid disconnection (loss of 1000MW) and DC single-pole blocking to verify the robustness of the strategy; ④ Secure distribution: After successful verification, the control command is signed with the national cryptographic SM2 algorithm and distributed to the unit's DCS through the scheduling data network to achieve closed-loop optimization.

[0064] The above-described embodiments are detailed and specific, illustrating preferred embodiments of the present invention. They are only used to illustrate the technical ideas and features of the present invention, with the aim of enabling those skilled in the art to understand the content of the present invention and implement it accordingly. However, they are not limited to the present invention, and the patent scope of the present invention cannot be limited by this embodiment alone. That is, any equivalent changes or modifications made to the spirit disclosed in the present invention, without departing from the structure of the present invention, such as local improvements within the system and modifications or transformations between subsystems, are still within the patent scope of the present invention.

Claims

1. An online analysis method for grid-related systems of thermal power units for power grid stability analysis, characterized in that, Includes the following steps: S1. Multi-level grid-connected system status awareness and data acquisition: Build a distributed status awareness network covering the core grid-connected subsystem of coal-fired power units to realize the synchronous acquisition and preprocessing of multi-source data and form a standardized data stream; S2. Dynamic feature extraction and state identification of grid-connected systems: Based on the standardized data stream, dynamic features at multiple time scales are extracted, and signal decomposition algorithms and state identification rule bases are used to automatically identify the operating conditions of grid-connected systems; S3. Parallel system simulation and prediction based on hybrid model: A hybrid model including grid-connected subsystems and related subsystems is constructed to simulate and extrapolate the system, and an end-to-end correlation model is constructed to evaluate the comprehensive support efficiency of the units for grid stability in real time; S4. Grid performance index mapping model: Establish a quantitative evaluation index system and use a nonlinear regression model to map multidimensional grid performance indexes to grid stability margin index; S5. Online diagnosis, early warning and optimization control strategy generation: Execute hierarchical early warning based on the analysis results of steps S3 and S4, and automatically generate optimization control strategies for grid-connected systems in the alarm state; S6. Closed-loop verification and strategy distribution: After the generated optimized control strategy is virtually verified through a parallel simulation platform, it is distributed to the unit's distributed control system (DCS) for execution via a security authentication mechanism.

2. The online analysis method for grid-related systems of thermal power units oriented towards grid stability analysis according to claim 1, characterized in that, In step S1, the core grid-connected subsystem includes a turbine speed control system, an excitation control system, a synchronous generator, a power system stabilizer (PSS), an automatic generation control (AGC), a primary frequency regulation system, and a coordinated control system (CCS). The synchronous acquisition is achieved through a unified clock synchronization protocol. The data includes electrical measurement data, mechanical measurement data, control command data, and status identification data. The preprocessing is completed at the edge computing node, including outlier filtering, data alignment, and compression encoding.

3. The online analysis method for grid-related systems of thermal power units oriented towards grid stability analysis according to claim 1, characterized in that, In step S2, the signal decomposition algorithm is an improved variational mode decomposition (VMD) algorithm, which is used to decompose non-stationary signals and extract time-frequency domain features by combining wavelet packet transform. The dynamic features across multiple time scales include: millisecond-level features reflecting the system's rapid dynamics, second-level features reflecting the regulation performance, and minute-level features reflecting long-term trends.

4. The online analysis method for grid-related systems of thermal power units oriented towards grid stability analysis according to claim 1, characterized in that, In step S3, the hybrid model adopts a parallel architecture of mechanistic model and data-driven model, and the outputs of the two are integrated through weighted fusion, residual correction or hierarchical decision-making; the end-to-end correlation model is a deep reinforcement learning model.

5. The online analysis method for grid-related systems of thermal power units oriented towards grid stability analysis according to claim 1, characterized in that, In step S4, the quantitative evaluation index system includes frequency support index, voltage support index, and power angle support index; the nonlinear regression model is a Gaussian process regression (GPR) model, and its model parameters are updated online through Bayesian optimization.

6. The online analysis method for grid-related systems of thermal power units for grid stability analysis according to claim 1, characterized in that, In step S5, the graded early warning mechanism includes a normal state, a warning state, and an alarm state; the generation of the optimized control strategy adopts a multi-objective optimization algorithm to take into account both the grid stability requirements and the unit's safe operation constraints.

7. The online analysis method for grid-related systems of thermal power units oriented towards grid stability analysis according to claim 2, characterized in that, In step S1, the distributed state awareness network adopts a three-layer architecture of end-edge-cloud: the end layer consists of intelligent sensing terminals deployed in the control cabinets of each network-related system; the edge layer consists of edge computing nodes deployed in the unit's electronics room that integrate field-programmable gate array (FPGA) acceleration cards; and the cloud layer consists of a cloud analysis platform deployed at the scheduling end.

8. The online analysis method for grid-related systems of thermal power units for grid stability analysis according to claim 3, characterized in that, The improved VMD algorithm adaptively adjusts the number of modes K and the penalty factor α, calculates the center frequency of each intrinsic mode function (IMF) component, merges modes with a center frequency difference less than a threshold, dynamically adjusts α based on the signal reconstruction error, and uses the mutual information criterion to screen effective IMF components.

9. The online analysis method for grid-related systems of thermal power units for grid stability analysis according to claim 4, characterized in that, The state space of the deep reinforcement learning model integrates the features of the excitation system, speed regulation system, PSS, AGC and the grid operation state. Its action space is defined as the adjustment amount of control parameters of multiple grid-related systems. Its reward function is designed as a weighted combination of the grid frequency, voltage, power angle stability margin improvement amount and the unit regulation cost and constraint violation penalty.

10. The online analysis method for grid-related systems of thermal power units for grid stability analysis according to claim 5, characterized in that, The Gaussian process regression (GPR) model employs a multi-kernel function combination strategy, including radial basis function (RBF) kernel, periodic kernel (PER) kernel, and linear kernel (LIN) kernel; the optimized control strategy is generated using an improved third-generation non-dominated sorting genetic algorithm (NSGA-III); in step S6, the parallel simulation platform is a hybrid model parallel simulation platform built based on digital twin and hardware-in-the-loop (HIL) technology; the security authentication mechanism uses the national cryptographic algorithm SM2 to verify the signature of control commands.

Images