A method and system for autonomous decision-making and control of coal-fired power units under all operating conditions

Abstract

This invention discloses a method and system for autonomous decision-making control of coal-fired power units under all operating conditions. The method includes: real-time acquisition and intelligent fusion of multi-source heterogeneous data; identification of monitoring blind spots and execution of blind spot data completion or validity verification; construction of an intelligent control system architecture covering all operating conditions of the unit, decentralizing the functions of the plant-level monitoring information system to the production control area; acquisition of electromagnetic radiation interference characteristic parameters and environmental interference characteristic parameters, and implementation of electromagnetic radiation interference suppression or environmental interference correction based on judgment conditions; fusion of a large-scale artificial intelligence model with a dynamic feature decoupling network and a decision knowledge graph to form a quasi-expert fuzzy decision-making system, executing an integrated control strategy of intelligent control, cruise control, and monitoring; and construction of a high-fidelity digital twin parallel system for offline verification and iterative optimization of the control strategy. This invention achieves autonomous decision-making and near-zero human intervention for coal-fired power units under all operating conditions, improving operational safety and economy.

Description

Technical Field

[0001] This invention belongs to the field of intelligent control technology for coal-fired power generating units, and particularly relates to a method and system for autonomous decision-making control of coal-fired power generating units under all operating conditions. Background Technology

[0002] As the global energy structure accelerates its transition to a low-carbon model, the proportion of renewable energy sources such as wind and solar power connected to the grid continues to rise. The role of coal-fired power plants is shifting from baseload power to a supporting power source capable of deep peak shaving and flexible operation. Against this backdrop, coal-fired power plants generally face multiple challenges, including efficiency improvement, pollution control, and intelligent upgrading. However, the current level of automation and intelligence in coal-fired power plants remains low, relying heavily on manual operation and monitoring, resulting in low energy efficiency, high operation and maintenance costs, and significant safety risks. Analysis reveals that the most prominent shortcomings of the existing control system are: First, the problem of data silos is severe. Distributed control systems (DCS), video surveillance, and inspection robots generate heterogeneous data from multiple sources, scattered across different systems and belonging to different security zones such as the field control area and the information zone. DCS generates time-series data at the millisecond to second level, video surveillance generates frame-level image stream data, and unmanned inspection systems generate event-based data at the minute to hour level. These data differ significantly in time scale, data format, and communication protocols, resulting in a severe deficiency in real-time acquisition and intelligent fusion capabilities. Traditional plant-level monitoring information systems (SIS) are mainly used for historical data storage and offline analysis, exhibiting significant delays of several seconds to minutes in the real-time control link, thus failing to directly support real-time intelligent decision-making in the control process.

[0003] Second, the system architecture is fragmented. Traditional SIS (System-Independent System) and DCS (Distributed Control System) have fixed boundaries, preventing SIS functions from reaching the field control area. This creates a disconnect where data is viewed from the upper level while control is implemented from the lower level. Online performance calculations and operating condition identification cannot be fed back to the controller in real time, resulting in control strategies not being adjusted promptly according to changes in unit performance. When facing all operating conditions—including unit start-up and shutdown, routine operation, load regulation, and anomaly handling—there is a lack of a unified autonomous decision-making control architecture. Control strategies under different operating conditions are independent, and switching relies on manual intervention.

[0004] Third, it lacks full-condition autonomous control capabilities. Especially for complex conditions such as the transition between dry and wet states, automatic control is difficult and still requires significant manual intervention. During the dry-wet transition, the outlet temperature of the steam-water separator and the water level in the storage tank exhibit strongly coupled and nonlinear dynamic characteristics, which traditional PID-based control strategies struggle to adapt to. Existing control strategies are mostly based on mechanistic models or simple rules, lacking fuzzy decision-making systems with expert-like reasoning capabilities, making it difficult to cope with complex and changing operating conditions.

[0005] Fourth, the level of intelligence is insufficient. The application of artificial intelligence technology in the power generation industry is still in the exploratory stage, and the deployment of large models in the field of real-time control faces multiple challenges such as computing power, latency, and interpretability. Existing methods mostly adopt simple logic such as static filtering and fixed threshold alarms, without forming a multi-source data cross-validation and dynamic calibration mechanism, making it difficult to fundamentally solve the problem of data distortion. In addition, existing alarm systems are based on fixed thresholds and cannot achieve early warning, often issuing alarms only after parameters exceed limits, leaving insufficient time for operators to respond.

[0006] Therefore, developing a control method that enables autonomous decision-making under all operating conditions of coal-fired power units, reduces human intervention, and improves operational safety and economy is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] To address the shortcomings of the prior art, this invention provides a method and system for autonomous decision-making control of coal-fired power units under all operating conditions. It aims to solve the technical problems of scattered multi-source heterogeneous data, system fragmentation, insufficient utilization of control information, insufficient accuracy of automatic control under complex operating conditions, reliance on human experience for response, lack of adaptive and expert-like reasoning capabilities in traditional control strategies, and frequent human intervention, making it impossible to achieve single-person operation and near-zero intervention.

[0008] To achieve the above objectives, the present invention proposes the following technical solution: a method for autonomous decision-making and control of a coal-fired power unit under all operating conditions, comprising the following steps: S1. Establish a unified data analysis platform in the production control area to collect and intelligently integrate multi-source heterogeneous data from distributed control systems, video surveillance systems, and unmanned inspection systems in real time; determine whether there are monitoring blind spots: obtain the grid sensor straight-line distance value and sensor signal strength deviation value based on grid division; if the grid sensor straight-line distance value is greater than the preset monitoring radius threshold and the sensor signal strength deviation value is less than the preset signal strength deviation threshold, mark it as a blind spot unit and perform blind spot data completion; otherwise, mark it as a monitoring unit and perform monitoring data validity verification. S2. Construct an intelligent control system architecture that covers all operating conditions, including unit start-up and shutdown, normal operation, load regulation and abnormal handling, and push the functions of the plant-level monitoring information system down to the production control area; determine whether to perform data multi-interference identification: obtain the environmental parameters of the monitoring unit, and if the environmental parameters do not meet the preset qualified conditions, perform data multi-interference identification; otherwise, jump to the dynamic calculation of the wind network. S3. In the data multi-interference identification, electromagnetic radiation interference characteristic parameters and environmental interference characteristic parameters are obtained. If the electromagnetic radiation interference judgment condition is met, electromagnetic radiation interference suppression is adopted. If the environmental interference judgment condition is met, environmental interference correction is adopted. Otherwise, the process jumps to the wind network dynamic solution. S4. Introduce a large-scale artificial intelligence model into the decision-making layer and integrate it with a dynamic feature decoupling network based on curve morphology topology and a decision knowledge graph to form a fuzzy decision-making system with expert-like capabilities, output control decisions, and execute intelligent control, cruise and monitoring integrated control strategies. S5. Construct a high-fidelity digital twin parallel system, perform closed-loop verification and strategy iteration optimization of the control decisions in an offline environment, and then send the optimized strategy to the actual control system.

[0009] Furthermore, the blind spot data completion in step S1 includes the following sub-steps: S101. Obtain a preset number of sensor monitoring data sequences that are closest to the blind zone unit and meet the blind zone data completion conditions as mine reference data. The blind zone data completion conditions indicate that the data acquisition success rate is greater than a preset data acquisition success threshold and the data fluctuation coefficient is less than a preset data fluctuation threshold. S102. Data fusion and completion are performed based on the Kriging interpolation method to obtain fused and completed data; S103. Based on the correlation analysis method, the fused and completed data is correlated with the monitoring data of the neighboring monitoring units to verify the correlation index. If the correlation index is greater than the preset high validity threshold, the completion is confirmed to be effective. If the correlation index is less than the preset low validity threshold, a ventilation abnormality prompt is sent and manual intervention is performed.

[0010] Furthermore, the full operating conditions in step S2 include at least the dry-wet transition point and the low-load operating condition where the unit load is less than 30% of the rated load; the system architecture is divided into a perception layer, a decision layer, and an execution layer, the perception layer is the unified data analysis platform, the decision layer is the decision engine driven by the artificial intelligence big model, and the execution layer is the control loop of the underlying execution mechanism; the function sinking specifically means: after the online performance calculation, real-time operating condition identification, and trend prediction function modules in the plant-level monitoring information system are refactored in a lightweight manner, they are deployed to the decision layer of the production control area.

[0011] Furthermore, the electromagnetic radiation interference suppression in step S3 includes the following sub-steps: S301. Obtain the mine signal spectrum distribution based on Fast Fourier Transform, extract discrete spectrum interference components, and generate an interference spectrum feature set; S302. Dynamically construct a digital band-stop filter bank based on the interference spectrum feature set, using the center frequency of the interference component as the center stopband frequency, and multiplying the bandwidth of the interference component by a preset bandwidth expansion coefficient as the stopband bandwidth. S303. Input the mine data sequence into the digital band-stop filter bank for filtering. After filtering, obtain the signal-to-noise ratio of the mine data. If the signal-to-noise ratio is greater than the preset signal-to-noise ratio threshold, mark it as qualified data and jump to the dynamic solution of the ventilation network. Otherwise, take environmental interference correction.

[0012] Furthermore, the environmental interference correction in step S3 includes the following sub-steps: S311. Input the environmental interference characteristic parameters and the humidity of the monitoring unit into the preset environmental correction mapping set, and query the environmental correction coefficient of the mine sensor. S312. Set the amplitude corresponding to the environmental correction coefficient of the mine sensor as the adjustment step size, gradually adjust the sensor data correction amount in the opposite direction of data drift, and superimpose the sensor data correction amount into the original monitoring data; S313. Continuously monitor the validity value of the data correction. If the validity value of the data correction is greater than the preset validity threshold, it is marked as qualified data and the process is redirected to the dynamic calculation of the wind network. If the validity value is still not greater than the preset validity threshold when the sensor data correction amount is greater than the preset maximum data correction threshold, an environmental correction processing failure prompt is sent.

[0013] Furthermore, the construction of the dynamic feature decoupling network in step S4 includes: treating the time-series variation curve of each key parameter as a topological object, and extracting the rising slope, falling slope, inflection point position, extreme point distribution, and oscillation frequency as morphological features; constructing a graph neural network, where the nodes of the graph neural network represent the features of different parameters, the edges of the graph neural network represent the physical association and delay time weight between parameters, and capturing the dynamic coupling relationship between the features of different parameters through message passing of the graph neural network to form an interpretable feature distillation module.

[0014] Furthermore, the intelligent control in step S4 employs a predictive control algorithm or an adaptive control algorithm to generate control commands for the actuators of the coal feeder, coal mill, blower, and feedwater pump; the predictive control algorithm employs a predictive controller based on a state-space model; and the adaptive control algorithm employs a model reference adaptive controller based on recursive least squares. The cruise mode is specifically defined as follows: when the unit is in steady-state operation, the system automatically switches to the cruise mode to automatically maintain the unit at the optimal operating point obtained in the pre-optimization process; the monitoring panel is implemented based on an intelligent human-machine interaction engine, which integrates a speech recognition module, a natural language understanding module based on BERT or GPT architecture, an active early warning module, and a video linkage module.

[0015] Furthermore, it also includes a system fault self-healing step: when a blower trips or a feedwater pump trips, the system automatically calls the preset self-healing strategy library to perform fault isolation and system reconstruction; the self-healing strategy library includes a rapid load reduction strategy and a standby equipment interlocking start strategy, the rapid load reduction strategy reduces the unit load to below 50% of the rated load at a rate of not less than 50% of the rated load / minute.

[0016] This invention also proposes an autonomous decision-making control system for all operating conditions of a coal-fired power unit, employing the aforementioned autonomous decision-making control method for all operating conditions of a coal-fired power unit. The system includes: A unified data fusion platform is deployed in the production control area, comprising a hyperconverged server cluster, a distributed time-series database, and a streaming data processing engine; The full-condition autonomous decision-making intelligent control engine integrates a large artificial intelligence model, a decision knowledge graph, and a dynamic feature decoupling network, and is deployed on an artificial intelligence inference server. The integrated control execution module includes predictive control and adaptive control algorithms deployed in the distributed control system controller, as well as a communication interface connecting the decision layer and the control layer; The digital twin parallel verification system operates independently of the production system and is built based on a hybrid model of high-fidelity mechanism and data-driven model. The intelligent human-computer interaction engine is deployed at the operator station and integrates a speech recognition module, a natural language understanding module, an active early warning module, and a video linkage module.

[0017] Furthermore, the dynamic feature decoupling network is constructed by treating the temporal variation curves of each key parameter as topological objects and extracting morphological features, and by constructing a graph neural network to capture the dynamic coupling relationship between different parameter features; the decision knowledge graph is constructed based on the interpretable feature distillation module output by the dynamic feature decoupling network, combined with operational experience and control rules, and is used to verify and correct the output of the large model.

[0018] Beneficial effects: 1. By building a unified data analysis platform in the production control area, and using OPC UA and MQTT protocols to collect multi-source heterogeneous data, the unmanned inspection data is converted into standardized time-series data through event-driven time window alignment and exponential decay weighted fusion technology. This solves the data silo problem and provides a low-latency, highly integrated data foundation for real-time decision-making.

[0019] 2. A three-tier architecture of perception, decision-making, and execution was constructed, and the key functional modules of SIS were restructured in a lightweight manner and moved down to the production control area, breaking the traditional boundary between SIS and DCS and realizing the integrated fusion of analysis and control.

[0020] 3. By introducing the Transformer large model into the core control and combining it with a dynamic feature decoupling network based on curve morphology topology and a decision knowledge graph, a quasi-expert fuzzy decision system with physical interpretability was constructed, which solved the black box problem of the large model and improved the safety and reliability of decision-making.

[0021] 4. By adopting predictive control and adaptive control algorithms with parameters such as specific control cycle, prediction time domain, and forgetting factor, combined with cruise mode and intelligent monitoring with advanced warning and automatic focusing functions, manual intervention is significantly reduced, achieving single-person operation and near-zero intervention.

[0022] 5. By constructing a digital twin parallel system based on a high-fidelity mechanism model and a neural network proxy model, offline security verification and rapid iteration of control strategies were achieved, avoiding the risks that online optimization may bring.

[0023] 6. By pre-setting a self-healing strategy with specific action rates and thresholds, rapid self-healing under typical fault scenarios is achieved, significantly improving the safety of unit operation. Attached Figure Description

[0024] Figure 1 This is a flowchart of an autonomous decision-making and control method for coal-fired power units under all operating conditions.

[0025] Figure 2 This is a block diagram of an autonomous decision-making control system for a coal-fired power unit under all operating conditions. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention. The invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments described herein are only for explaining the invention and are not intended to limit the invention.

[0027] Example 1 like Figure 1 As shown, this invention proposes an autonomous decision-making control method for coal-fired power units under all operating conditions, including the following steps S1 to S5.

[0028] S1. Construction of Multi-Source Heterogeneous Data and Judgment of Monitoring Blind Spots A unified data analysis platform is built in the production control area. This platform includes a cluster of at least five hyperconverged servers, a distributed time-series database, and a streaming data processing engine. Through industrial isolation gateways and industrial firewalls, a bidirectional data communication interface is established with the DCS using the OPC UA protocol, and with the video surveillance system and unmanned inspection system using the MQTT protocol. Real-time data acquisition is performed on real-time measurement points from the DCS, H.264 or H.265 encoded image streams from the video surveillance system, and infrared thermal images and vibration spectrum data from the unmanned inspection system. The discontinuous inspection data generated by the unmanned inspection system is converted into standardized time-series data using an event-driven time window alignment method. Historical inspection information is then fused using an exponential decay weighting method, enabling the converted time-series data to be spatiotemporally aligned and fused with the DCS time-series data, forming a unified data foundation.

[0029] Then, it is determined whether there is a monitoring blind zone: Based on functional requirements, the unit monitoring area is divided into several equal grid units. The three-dimensional coordinates of the center point of each grid unit and the straight-line distance value of the grid sensor are obtained, and the sensor signal strength deviation value is obtained. If the straight-line distance value of the grid sensor is greater than a preset monitoring radius threshold and the sensor signal strength deviation value is less than a preset signal strength deviation threshold, it is marked as a blind zone unit and blind zone data completion is performed; otherwise, it is marked as a monitoring unit and monitoring data validity verification is performed.

[0030] S2, Architecture Construction and Multi-Interference Identification and Judgment A comprehensive intelligent control system architecture for autonomous decision-making covering all operating conditions, including unit start-up and shutdown, routine operation, load regulation, and anomaly handling, is constructed. The comprehensive operating conditions include at least the dry-wet transition point and low-load conditions where the unit load is below 30% of the rated load. This architecture is divided into three layers: a perception layer, a decision layer, and an execution layer. The perception layer is the unified data analysis platform, the decision layer is the decision engine driven by a large artificial intelligence model, and the execution layer is the control loop of the underlying actuators. The online performance calculation, real-time operating condition identification, and trend prediction modules originally deployed in the Information Zone 3 SIS are lightweighted and restructured. The prototype code based on Matlab or Python is converted to C++ or Go, and the algorithm complexity is optimized. This code is then deployed to the decision layer in the production control zone, enabling it to participate in the real-time control link with a cycle of no more than one second, forming a unified autonomous decision-making framework oriented towards the control execution layer.

[0031] Then, it is determined whether to perform data multi-interference identification: acquire the environmental parameters of the monitoring unit, including the distance value of the interference source and the humidity of the monitoring unit. If the environmental parameters do not meet the preset qualified conditions (i.e., the distance value of the interference source is less than the preset safe distance threshold and the humidity is less than the preset maximum humidity threshold), then data multi-interference identification is performed; otherwise, proceed to the subsequent dynamic calculation of the wind network.

[0032] S3, Multi-interference identification and targeted suppression / correction In the data multi-interference identification process, electromagnetic radiation interference characteristic parameters and environmental interference characteristic parameters are acquired. Electromagnetic radiation interference characteristic parameters include wind speed signal distortion and fan harmonic intensity; environmental interference characteristic parameters include sensor data drift and mine data fluctuation quantification index. If the electromagnetic radiation interference judgment condition is met (wind speed signal distortion greater than a preset distortion threshold and fan harmonic intensity greater than a preset harmonic threshold), electromagnetic radiation interference suppression is implemented; otherwise, if the environmental interference judgment condition is met (sensor data drift greater than a preset drift threshold and mine data fluctuation quantification index greater than a preset interference fluctuation threshold), environmental interference correction is implemented; otherwise, the process jumps to dynamic calculation of the wind network.

[0033] S4, Large Model Decision Making and Integrated Control A large-scale model based on the Transformer architecture is introduced into the decision-making layer. Self-supervised pre-training is performed on the large-scale model using historical operating data from the past 3 to 5 years. The pre-training task is to predict parameter changes for the next 5 to 30 minutes given data from the past 30 to 120 minutes. Supervised fine-tuning is then performed using high-value data from the most recent 6 to 12 months (including start-up and shutdown data, load disturbance data, and typical fault records). The power plant's operating procedures, expert experience rule base, and accident handling plans are injected into the large-scale model using knowledge distillation techniques based on logical tensors to form prior knowledge constraints.

[0034] A dynamic feature decoupling network based on curve morphology topology is constructed. This network treats the time-series variation curves of each key parameter as topological objects, extracting morphological features including rising slope, falling slope, inflection point location, extreme point distribution, and oscillation frequency. A graph neural network captures the dynamic coupling relationships between different parameter features. Nodes in the graph neural network represent features of different parameters, and edges represent physical associations and time delay weights between parameters. Based on the interpretable feature distillation module output by the dynamic feature decoupling network, and combined with operational experience and control rules, a decision knowledge graph stored in a graph database is constructed. Node types include operating condition nodes, state nodes, and operation nodes, and edge types include conditional relationships, causal relationships, and preventive relationships. The large model, the dynamic feature decoupling network, and the decision knowledge graph together constitute a fuzzy decision-making system with expert-like capabilities. This system comprehensively analyzes unit status, operational objectives, and real-time data, outputting control decisions including control mode switching, setpoint adjustments, and actuator commands.

[0035] An integrated intelligent control, cruise, and monitoring strategy is implemented. Intelligent control, based on the setpoints output by the fuzzy decision system, employs predictive or adaptive control algorithms to directly generate control commands for the actuators of the coal feeder, coal mill, blower, and feedwater pump during the control cycle. When the unit is in steady-state operation, the system automatically switches to cruise mode, maintaining the unit at the pre-optimized operating point and automatically handling minor disturbances. An intelligent human-machine interaction engine integrating large-scale model natural language processing capabilities is constructed, integrating a speech recognition module, a BERT or GPT-based natural language understanding module, an active warning module, and a video linkage module. This provides active warnings and automatic focus on alarm screens for the entire plant's operating status and supports natural language command parsing. Through these strategies, a continuous, autonomous, and closed-loop control process is formed for the unit's full-load operating conditions, enabling the automatic completion of automatic power generation control commands.

[0036] S5, Digital Twin Parallel Verification and Strategy Iteration A high-fidelity digital twin parallel system driven by artificial intelligence is constructed. This system is based on a hybrid construction of a high-fidelity mechanistic model and a data-driven model. The high-fidelity mechanistic model employs a boiler combustion mechanism model built using a mechanistic modeling language with 10,000 to 50,000 grid cells. The data-driven model uses a neural network surrogate model with 3 to 7 hidden layers. This digital twin system operates independently of the production system. In an offline environment, it performs closed-loop verification and iterative optimization of the control decisions output by the fuzzy decision system at a time ratio of 0.5:1 to 2:1, and then distributes the optimized control strategy to the actual control system through an industrial isolation gateway.

[0037] like Figure 2 As shown, the present invention also provides an autonomous decision-making control system for all operating conditions of a coal-fired power unit, comprising: The Unified Data Fusion Platform 101 is deployed in the production control area and includes a hyper-converged server cluster, a distributed time-series database, and a streaming data processing engine. The 102 autonomous decision-making intelligent control engine for all operating conditions integrates a large artificial intelligence model, a decision knowledge graph, and a dynamic feature decoupling network, and is deployed on an artificial intelligence inference server. The integrated control execution module 103 includes predictive control algorithms and adaptive control algorithms deployed in the DCS controller, as well as a communication interface connecting the decision layer and the control layer. The digital twin parallel verification system 104 operates independently of the production system and is built based on a hybrid model of high-fidelity mechanism model and data-driven model. The Intelligent Human-Computer Interaction Engine 105, deployed at the operator station, integrates a speech recognition module, a natural language understanding module, an active early warning module, and a video linkage module.

[0038] Example 2 Based on Example 1, this embodiment further illustrates another specific implementation of the autonomous decision-making and control method for all operating conditions of coal-fired power units described in this invention.

[0039] As described in Example 1, the method of the present invention includes core steps such as multi-source heterogeneous data fusion and blind spot processing, construction of an autonomous decision-making architecture under all operating conditions, multi-interference identification and suppression correction, large model-driven fuzzy decision-making and integrated control, and digital twin parallel verification and strategy iteration. This embodiment will focus on these steps and, in conjunction with specific system deployment and control logic, explain in detail the technical implementation principles.

[0040] In this embodiment, the unified data analysis platform is deployed in the production control area and adopts a distributed microservice architecture. The platform establishes bidirectional data channels with the distributed control system, video surveillance system, and unmanned inspection system through standardized communication protocols. To achieve real-time fusion of multi-source heterogeneous data, the platform has a built-in multimodal data alignment engine. This engine, based on timestamp mapping and event triggering mechanisms, converts raw data with different sampling frequencies and data formats into a unified time-series data structure.

[0041] Specifically, distributed control systems continuously generate analog measurement points and switch status data at millisecond intervals; video surveillance systems generate image stream data in frames, containing visualized information about equipment operating status; unmanned inspection systems generate structured or semi-structured data such as infrared thermal imaging, vibration spectrum, and partial discharge at minute or hour intervals. These data differ significantly in time granularity, data dimension, and physical meaning. To address this issue, the data alignment engine employs a sliding time window-based fusion strategy: for video and inspection data, the system automatically identifies key event tags (such as abnormal equipment temperature, excessive vibration, etc.), converts them into timestamped discrete event sequences, and correlates and matches them with contemporaneous time-series data in the distributed control system. Based on this, the system uses a weighted fusion algorithm to dynamically allocate fusion weights according to the confidence and timeliness of the data source, generating a unified feature vector for subsequent decision-making processes.

[0042] In terms of blind zone identification, this embodiment employs a gridded monitoring area method. Based on the layout diagram of key unit equipment and the sensor distribution diagram, the system divides the monitoring area into several logical grid units. For each grid unit, the system calculates in real-time the spatial distance and signal strength attenuation characteristics between it and adjacent sensors. If the signal coverage redundancy of a grid unit is lower than a preset threshold (i.e., the sensor signal strength deviation is too small and the straight-line distance is too large), the unit is determined to be a monitoring blind zone. For blind zone units, the system automatically triggers a data completion process: first, it selects a sensor with a spatially adjacent location and data quality that meets the validity conditions as a reference source; then, it estimates the data of the blind zone unit based on a spatial interpolation algorithm (such as Kriging interpolation or inverse distance weighted interpolation); finally, it verifies the physical rationality of the completed data through correlation analysis. If the correlation index between the completed data and the data of adjacent measuring points is lower than a preset low threshold, the system determines that there may be a physical anomaly (such as duct blockage or equipment failure) and proactively prompts operators to intervene and check.

[0043] The autonomous decision-making intelligent control system architecture constructed in this embodiment is divided into a perception layer, a decision layer, and an execution layer. The perception layer is responsible for the real-time acquisition, cleaning, fusion, and feature extraction of multi-source data; the decision layer carries the large-scale artificial intelligence model, dynamic feature decoupling network, and decision knowledge graph, and completes the identification of working conditions, state prediction, and control decision generation; the execution layer converts the control commands output by the decision layer into action signals of the actuators and feeds back the execution status.

[0044] One of the key innovations of this invention lies in decentralizing the core functional modules of traditional plant-level monitoring information systems to the decision-making level of the production control area. In traditional architectures, plant-level monitoring information systems are mainly used for historical data storage, offline performance analysis, and operational condition backtracking. Their data processing cycles are typically on the order of seconds to minutes, making them unable to directly participate in the real-time control link. This embodiment uses lightweight refactoring technology to compress and optimize the algorithm logic of functional modules such as online performance calculation, real-time operational condition identification, and trend prediction, transforming them from a high-latency batch processing mode to a low-latency stream processing mode. The refactored functional modules are deployed as microservices in the decision-making server of the production control area, interacting with the perception and execution layers via a high-speed memory data bus. This decentralization mechanism transforms the original post-event analysis function into a real-time decision-making tool, significantly improving the control system's response speed to changes in unit performance.

[0045] In the actual operating environment of coal-fired power units, sensor data is susceptible to interference from electromagnetic radiation and environmental factors. This embodiment designs a multi-level interference identification and adaptive correction mechanism to ensure the quality of data entering the decision-making process.

[0046] Electromagnetic interference mainly originates from frequency converters, high-power switching equipment, and high-frequency communication devices. The electromagnetic noise generated by these devices couples into the sensor signal loop, causing periodic fluctuations at specific frequencies in the data. This embodiment uses Fast Fourier Transform (FFT) to perform spectral analysis on the sensor signal, extracting the center frequency and bandwidth of the discrete spectral interference components. When the detected interference component intensity exceeds a preset threshold, the system dynamically constructs a digital bandstop filter bank. Each filter corresponds to one interference component, with its stopband center frequency aligned with the interference frequency, and the stopband width adaptively expanded according to the interference bandwidth. If the signal-to-noise ratio of the filtered data meets the requirements, it is considered qualified data; otherwise, the environmental interference correction process is further triggered.

[0047] Environmental interference mainly manifests as sensor zero-point drift or sensitivity changes caused by temperature and humidity variations. This embodiment establishes an environmental correction mapping set, constructed based on historical calibration data and sensor physical characteristics, recording the sensor's correction coefficients under different temperature and humidity conditions. When environmental parameters are detected to exceed the normal range, the system queries the correction coefficient based on the current temperature and humidity values ​​and gradually adjusts the sensor data correction amount using this coefficient as a step size. The correction process employs a closed-loop control strategy: the system continuously monitors the deviation between the corrected data and nearby reference measurement points. If the deviation gradually converges and eventually falls below a threshold, the correction is deemed effective; if the correction amount exceeds a preset maximum threshold and the deviation still does not converge, the environmental correction is deemed ineffective and an alarm is issued.

[0048] The core decision-making module of this embodiment consists of three parts: an artificial intelligence large model, a dynamic feature decoupling network, and a decision knowledge graph. The three parts work together to form a fuzzy decision-making system with expert-like reasoning capabilities.

[0049] The large-scale AI model is built based on the Transformer architecture, employing a two-stage strategy of pre-training and fine-tuning. The pre-training stage utilizes historical operating data from the unit (including normal operating conditions, variable load conditions, and fault conditions) for self-supervised learning, enabling the model to master a general representation of the dynamic characteristics of the coal-fired unit. The fine-tuning stage utilizes high-value operating data (such as start-up and shutdown processes, dry-wet state transition processes, and typical fault recovery processes) for supervised learning, making the model's decision boundaries for key operating conditions more accurate. To meet the low latency requirements of real-time control, this embodiment optimizes the large-scale model for inference acceleration, including model pruning, layer fusion, and quantization compression techniques, keeping the latency of a single inference iteration within an acceptable range.

[0050] Dynamic feature decoupling networks are used to address the "black box" problem of large models and improve the interpretability of decisions. This network treats the time-series variation curves of each key parameter as topological objects, extracting morphological features such as rising slope, falling slope, inflection point location, extreme point distribution, and oscillation frequency. Based on this, a graph neural network is constructed, where nodes represent the morphological features of different parameters, and edges represent the physical relationships and time delay weights between parameters. Through multi-layer message passing in the graph neural network, the network can automatically capture the dynamic coupling relationships between parameter features and output the degree of coupling as attention weights, forming an interpretable feature distillation module. For example, in the process of dry-wet conversion, there is a strong coupling relationship between the outlet temperature of the steam-water separator and the water level in the storage tank; this relationship is explicitly expressed through the weight changes of the corresponding edges in the graph neural network.

[0051] The decision-making knowledge graph is constructed based on the feature distillation module output by the dynamic feature decoupling network, combined with power plant operation procedures, expert experience rule base, and accident handling plans. The knowledge graph is stored in the form of a directed graph, with node types including operating condition nodes (e.g., "wet operation," "dry operation," "wet-dry transition"), state nodes (e.g., "high temperature," "large pressure fluctuation"), and operation nodes (e.g., "increase feedwater flow," "open the desuperheating valve"). Edge types include conditional relationships, causal relationships, and preventive relationships. During the decision-making process, the large model first outputs preliminary decision suggestions based on current data. Then, the dynamic feature decoupling network provides the feature coupling state under the current operating condition. Finally, the knowledge graph verifies and corrects the suggestions: if a suggestion conflicts with a rule in the knowledge graph (e.g., suggesting closing the water level regulating valve in a wet operating condition), the system automatically rejects the suggestion and triggers a rule-driven alternative decision.

[0052] The collaborative workflow of the three components is as follows: After processing, the data from the perception layer is simultaneously input into the large model and the dynamic feature decoupling network. The large model generates end-to-end decisions and outputs the probability distribution of control commands; the feature decoupling network extracts the feature coupling state under the current operating condition and outputs the influence relationship matrix between key parameters. The knowledge graph receives the outputs of the large model and the feature decoupling network, and performs consistency verification through graph query and rule-based reasoning. If consistent, the final control decision is output; if inconsistent, the system enters a fuzzy arbitration mode, correcting the decision according to preset priority rules (such as "safety over economy" and "stability over speed"). This mechanism retains the powerful representational capabilities of the large model while introducing constraints from physical rules and expert experience, achieving a balance between security and intelligence.

[0053] In this embodiment, intelligent control employs a strategy combining predictive and adaptive control. Predictive control, based on a state-space model, solves for the optimal control sequence within each control cycle by considering the current state and the trajectory of the setpoint at several future times. This algorithm can explicitly handle multivariate coupling and constraints, making it suitable for complex loops such as coordinated control systems. Adaptive control uses a model reference adaptive control strategy, identifying the parameters of the controlled object online through recursive least squares and adjusting the controller gain in real time to adapt to slow time-varying characteristics such as unit aging and coal quality changes.

[0054] Cruise mode is one of the key innovations of this invention. When the unit is operating in a steady state and the fluctuations of various parameters are within the allowable range, the system automatically switches to cruise mode. In this mode, the system no longer responds to minor fluctuations in automatic power generation control commands cycle by cycle, but instead maintains the unit at the optimal operating point obtained through offline optimization. The objective function for optimizing the optimal operating point comprehensively considers multiple dimensions such as coal consumption for power generation, nitrogen oxide emissions, and equipment wear, and is completed in the digital twin system using a multi-objective evolutionary algorithm. In cruise mode, the system only actively adjusts for disturbances exceeding the set dead zone, and adopts a passive tolerance strategy for small fluctuations within the dead zone, thereby reducing the frequent operation of actuators and extending equipment life.

[0055] The intelligent monitoring system is based on an intelligent human-computer interaction engine. This engine integrates a voice recognition module, a natural language understanding module, an active early warning module, and a video linkage module. The voice recognition module allows operators to query parameters, switch screens, or confirm alarms via natural language voice commands; the natural language understanding module, based on a pre-trained language model, can parse complex query intents, such as "Why is the current of the blower on side A 10% higher than that on side B?" and return the analysis results; the active early warning module, based on a time-series prediction model, issues early warnings before parameters exceed limits, such as "The main steam temperature is expected to reach the alarm value in 5 minutes"; and the video linkage module automatically retrieves camera footage from the corresponding area to the operator station when an alarm is triggered, achieving rapid location with "seeing the alarm immediately".

[0056] The high-fidelity digital twin parallel system constructed in this embodiment operates independently of the actual production system. The system employs a hybrid modeling approach: core thermal equipment (such as boilers and steam turbines) uses one-dimensional or quasi-three-dimensional mechanistic models based on the conservation of mass, energy, and momentum; for aspects difficult to model precisely (such as combustion processes and pollutant generation processes), neural network proxy models are used. The mechanistic model and the data-driven model exchange data bidirectionally through a coupling interface, forming a complete virtual image of the unit.

[0057] The digital twin system uses real-time data as boundary conditions to simulate operation at a speed faster than or equal to actual time. When the decision-making layer generates a new control strategy, the system first performs closed-loop verification in the digital twin environment: the strategy is applied to a group of virtual machines, and its dynamic response and safety boundaries are observed. If the verification passes, the strategy enters the iterative optimization phase, using Bayesian optimization or reinforcement learning algorithms to perform local optimization in the parameter space to further improve the control effect. The optimized strategy is then distributed to the actual control system through an industrial isolation gateway. For strategies that fail verification, the system automatically records the reasons for the failure and feeds them back to the decision-making layer for negative sample learning of the large model and rule updates of the knowledge graph.

[0058] This embodiment also includes a system fault self-healing mechanism. When a critical auxiliary machine (such as a blower or feedwater pump) trips, the system automatically invokes a pre-set self-healing strategy library. The self-healing strategy library uses an event-driven rule engine, with each rule containing trigger conditions, action sequences, and recovery conditions. Taking a blower trip as an example, the trigger condition is "the blower's operating signal is lost and the standby blower has not been interlocked to start," and the action sequence includes "rapidly reducing the load to below 50% of the rated load," "interlocking and starting the standby blower," and "adjusting the blade opening of other blowers to maintain negative pressure in the furnace," etc. The rapid load reduction strategy reduces the unit load at a set rate while coordinating coal feed, water feed, and air volume to prevent boiler flameout or a sudden drop in steam temperature. The standby equipment interlocking and starting strategy automatically switches the control loop after the standby equipment starts, removing the faulty equipment from the control logic. The entire self-healing process is completed within seconds to tens of seconds without manual intervention, significantly improving the unit's survivability under fault conditions.

[0059] The above embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for autonomous decision-making and control of a coal-fired power unit under all operating conditions, characterized in that, Includes the following steps: S1. Build a unified data analysis platform in the production control area to collect and intelligently integrate multi-source heterogeneous data from distributed control systems, video surveillance systems and unmanned inspection systems in real time, and determine whether there are monitoring blind spots. S2. Construct an intelligent control system architecture that covers all operating conditions, including unit start-up and shutdown, normal operation, load regulation and abnormal handling, and push the functions of the plant-level monitoring information system down to the production control area; determine whether to perform data multi-interference identification: obtain the environmental parameters of the monitoring unit, and if the environmental parameters do not meet the preset qualified conditions, perform data multi-interference identification; otherwise, jump to the dynamic calculation of the wind network. S3. In the data multi-interference identification, electromagnetic radiation interference characteristic parameters and environmental interference characteristic parameters are obtained. If the electromagnetic radiation interference judgment condition is met, electromagnetic radiation interference suppression is adopted. If the environmental interference judgment condition is met, environmental interference correction is adopted. Otherwise, the process jumps to the wind network dynamic solution. S4. Introduce a large-scale artificial intelligence model into the decision-making layer and integrate it with a dynamic feature decoupling network based on curve morphology topology and a decision knowledge graph to form a fuzzy decision-making system with expert-like capabilities, output control decisions, and execute intelligent control, cruise and monitoring integrated control strategies. S5. Construct a high-fidelity digital twin parallel system, perform closed-loop verification and strategy iteration optimization of the control decisions in an offline environment, and then send the optimized strategy to the actual control system.

2. The autonomous decision-making and control method for all operating conditions of a coal-fired power unit according to claim 1, characterized in that, The blind spot data completion in step S1 includes the following sub-steps: S101. Obtain a preset number of sensor monitoring data sequences that are closest to the blind zone unit and meet the blind zone data completion conditions as mine reference data. The blind zone data completion conditions indicate that the data acquisition success rate is greater than a preset data acquisition success threshold and the data fluctuation coefficient is less than a preset data fluctuation threshold. S102. Data fusion and completion are performed based on the Kriging interpolation method to obtain fused and completed data; S103. Based on the correlation analysis method, the fused and completed data is correlated with the monitoring data of the neighboring monitoring units to verify the correlation index. If the correlation index is greater than the preset high validity threshold, the completion is confirmed to be effective. If the correlation index is less than the preset low validity threshold, a ventilation abnormality prompt is sent and manual intervention is performed.

3. The autonomous decision-making and control method for all operating conditions of a coal-fired power unit according to claim 1, characterized in that, The full operating conditions in step S2 include at least the dry-wet transition point and the low-load operating condition where the unit load is less than 30% of the rated load; the system architecture is divided into a perception layer, a decision layer, and an execution layer. The perception layer is the unified data analysis platform, the decision layer is the decision engine driven by the artificial intelligence big model, and the execution layer is the control loop of the underlying execution mechanism; the function sinking specifically means: after the online performance calculation, real-time operating condition identification, and trend prediction function modules in the plant-level monitoring information system are refactored in a lightweight manner, they are deployed to the decision layer of the production control area.

4. The autonomous decision-making and control method for all operating conditions of a coal-fired power unit according to claim 1, characterized in that, The electromagnetic radiation interference suppression in step S3 includes the following sub-steps: S301. Obtain the mine signal spectrum distribution based on Fast Fourier Transform, extract discrete spectrum interference components, and generate an interference spectrum feature set; S302. Dynamically construct a digital band-stop filter bank based on the interference spectrum feature set, using the center frequency of the interference component as the center stopband frequency, and multiplying the bandwidth of the interference component by a preset bandwidth expansion coefficient as the stopband bandwidth. S303. Input the mine data sequence into the digital band-stop filter bank for filtering. After filtering, obtain the signal-to-noise ratio of the mine data. If the signal-to-noise ratio is greater than the preset signal-to-noise ratio threshold, mark it as qualified data and jump to the dynamic solution of the ventilation network. Otherwise, take environmental interference correction.

5. The autonomous decision-making and control method for all operating conditions of a coal-fired power unit according to claim 1, characterized in that, The environmental interference correction in step S3 includes the following sub-steps: S311. Input the environmental interference characteristic parameters and the humidity of the monitoring unit into the preset environmental correction mapping set, and query the environmental correction coefficient of the mine sensor. S312. Set the amplitude corresponding to the environmental correction coefficient of the mine sensor as the adjustment step size, gradually adjust the sensor data correction amount in the opposite direction of data drift, and superimpose the sensor data correction amount into the original monitoring data; S313. Continuously monitor the validity value of the data correction. If the validity value of the data correction is greater than the preset validity threshold, it is marked as qualified data and the process is redirected to the dynamic calculation of the wind network. If the validity value is still not greater than the preset validity threshold when the sensor data correction amount is greater than the preset maximum data correction threshold, an environmental correction processing failure prompt is sent.

6. The autonomous decision-making and control method for all operating conditions of a coal-fired power unit according to claim 1, characterized in that, The construction of the dynamic feature decoupling network in step S4 includes: treating the time-series change curve of each key parameter as a topological object, and extracting the rising slope, falling slope, inflection point position, extreme point distribution and oscillation frequency as morphological features; constructing a graph neural network, where the nodes of the graph neural network represent the features of different parameters, the edges of the graph neural network represent the physical association and delay time weight between parameters, and capturing the dynamic coupling relationship between the features of different parameters through message passing of the graph neural network to form an interpretable feature distillation module.

7. The autonomous decision-making and control method for all operating conditions of a coal-fired power unit according to claim 1, characterized in that, The intelligent control in step S4 uses a predictive control algorithm or an adaptive control algorithm to generate control commands for the actuators of the coal feeder, coal mill, fan, and water pump; the predictive control algorithm uses a predictive controller based on a state-space model; the adaptive control algorithm uses a model reference adaptive controller based on recursive least squares. The cruise mode is specifically defined as follows: when the unit is in steady-state operation, the system automatically switches to the cruise mode to automatically maintain the unit at the optimal operating point obtained in the pre-optimization process; the monitoring panel is implemented based on an intelligent human-machine interaction engine, which integrates a speech recognition module, a natural language understanding module based on BERT or GPT architecture, an active early warning module, and a video linkage module.

8. The autonomous decision-making and control method for coal-fired power units under all operating conditions according to claim 1, characterized in that, It also includes a system fault self-healing step: when a blower trips or a feedwater pump trips, the system automatically calls a preset self-healing strategy library to perform fault isolation and system reconstruction; the self-healing strategy library includes a rapid load reduction strategy and a standby equipment interlocking start strategy, the rapid load reduction strategy reduces the unit load to below 50% of the rated load at a rate of not less than 50% of the rated load / minute.

9. A fully autonomous decision-making control system for coal-fired power units under all operating conditions, characterized in that, The system employing the full-condition autonomous decision-making control method for coal-fired power units according to any one of claims 1 to 8 includes: A unified data fusion platform is deployed in the production control area, comprising a hyperconverged server cluster, a distributed time-series database, and a streaming data processing engine; The full-condition autonomous decision-making intelligent control engine integrates a large artificial intelligence model, a decision knowledge graph, and a dynamic feature decoupling network, and is deployed on an artificial intelligence inference server. The integrated control execution module includes predictive control and adaptive control algorithms deployed in the distributed control system controller, as well as a communication interface connecting the decision layer and the control layer; The digital twin parallel verification system operates independently of the production system and is built based on a hybrid model of high-fidelity mechanism and data-driven model. The intelligent human-computer interaction engine is deployed at the operator station and integrates a speech recognition module, a natural language understanding module, an active early warning module, and a video linkage module.

10. The autonomous decision-making control system for all operating conditions of a coal-fired power unit according to claim 9, characterized in that, The dynamic feature decoupling network is constructed by treating the time-series variation curves of each key parameter as topological objects and extracting morphological features, and by constructing a graph neural network to capture the dynamic coupling relationship between different parameter features; the decision knowledge graph is constructed based on the interpretable feature distillation module output by the dynamic feature decoupling network, combined with operating experience and control rules, and is used to verify and correct the output of the large model.

Images