Power plant monitoring management system
By integrating data acquisition, fusion and intelligent analysis modules, combined with advanced machine learning algorithms, accurate monitoring and fault prediction of boiler coking are achieved, solving the problems of blind soot blowing and delayed monitoring in existing technologies, and improving the operating efficiency and safety of power plants.
Patent Information
- Application Number
- CN202510860642.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-25
- Publication Date
- 2025-09-16
AI Technical Summary
Existing technologies are unable to accurately monitor the location and severity of boiler coking, resulting in blind soot blowing, wasting resources and shortening equipment life. At the same time, the lag in fault monitoring leads to the escalation of accidents.
By adopting data acquisition module, data fusion module, intelligent analysis module and control and visualization module, combined with convolutional neural network, long short-term memory network and deep reinforcement learning, real-time diagnosis and fault prediction of boiler coking can be achieved, and the optimal sootblowing control strategy can be generated.
It improves boiler thermal efficiency, reduces fuel consumption and operation and maintenance costs, enhances equipment safety and reliability, and realizes predictive maintenance.
Smart Images
Figure CN120652934A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of industrial process control and monitoring, and in particular to a power plant monitoring and management system. Background Art
[0002] As the core component that converts fuel chemical energy into high-temperature, high-pressure steam, the operating status of a thermal power plant's boiler directly determines the safety, economy, and environmental performance of the entire plant. To effectively monitor this complex thermal equipment, modern power plants commonly deploy distributed control systems (DCS). However, as the operational requirements of units continue to increase, the limitations of existing technologies are becoming increasingly prominent.
[0003] First, existing technologies face significant challenges in operational efficiency management. Ash produced by the combustion of solid fuels (such as coal and biomass) is prone to coking and depositing on the boiler's heating surface, forming an adhesion layer with extremely poor thermal conductivity. This problem is the primary cause of reduced boiler thermal efficiency and increased fuel consumption. To address this issue, power plants rely on sootblowing systems. However, the currently commonly used sootblowing mode, which is based on fixed-time programs or the operator's rough experience, is unable to perceive the actual location and severity of coking, leading to a "blind blowing" phenomenon: not only does it cause unnecessary blowing of clean areas, wasting a large amount of valuable steam and reducing the net efficiency of power generation; at the same time, the high-intensity medium scouring also causes long-term wear on the pipelines, shortening the life of the equipment.
[0004] Secondly, existing monitoring technology exhibits significant lags in ensuring equipment safety and reliability. Leaks and bursts in the boiler's "four tubes" (water wall, superheater, reheater, and economizer) are the most significant threats to the safe operation of power plants. Coking, a key cause of tube bursts, is a major factor. The uneven local heat load caused by coking can cause the tube metal to overheat for extended periods, ultimately leading to failure due to creep. Existing DCS systems monitor these progressive failures with essentially a post-facto response mechanism. Their alarm logic is typically triggered only after a pipe rupture has occurred or macro-parameters such as steam pressure or boiler feedwater volume have experienced significant fluctuations. This model deprives power plants of valuable opportunities for early intervention in the incipient stages of a failure, easily leading to the escalation of the incident, unplanned downtime, and significant safety and economic losses. Summary of the Invention
[0005] In order to solve the above technical problems, the present invention provides a power plant monitoring and management system, which includes: a data acquisition module, a data fusion module, an intelligent analysis module and a control and visualization module.
[0006] The data acquisition module includes a thermal imaging acquisition unit, an acoustic signal acquisition unit and a process parameter acquisition unit, which are respectively used to collect thermal imaging image data, acoustic signal data and process parameter data of the power plant boiler; The data fusion module is connected to the data acquisition module and is used to time-align and pre-process the thermal imaging images, acoustic signals and process parameter data, and fuse them into a multivariate time series data set; The intelligent analysis module includes a coking diagnosis unit, an abnormality prediction unit and a soot blowing optimization unit; The coking diagnosis unit is used to process the thermal imaging image to quantitatively evaluate the coking state of the boiler heating surface in real time; The abnormality prediction unit is used to process a time series data including the coking state evaluation result, the acoustic signal and the process parameters to predict the future abnormal state of the boiler; The sootblowing optimization unit is used to generate an optimal sootblowing control strategy according to the coking state evaluation result and the future abnormal state prediction; The control and visualization module is connected to the intelligent analysis module and the distributed control system of the power plant, and is used to display the coking state assessment results and predictive alarms, and execute the soot blowing control instructions.
[0007] Preferably, the coking diagnosis unit adopts a convolutional neural network model, which is trained to perform semantic segmentation on the thermal imaging image, classify pixels in the image into different coking levels, and calculate a quantitative coking index of one or more regions based on the segmentation results.
[0008] Preferably, performing semantic segmentation on the thermal imaging image, classifying pixels in the image into different coking levels, and calculating a quantitative coking index of one or more regions based on the segmentation results, includes: Inputting the thermal imaging image as input data into a pre-trained convolutional neural network model; The convolutional neural network model performs semantic segmentation processing on the thermal imaging image to generate a coking segmentation map; in the coking segmentation map, each pixel or pixel region of the image is intelligently classified into a specific level in a predefined coking level set, wherein the coking level set includes: "clean", "light coking" and "severe coking"; Automatically calculating the quantitative coking index based on the output result of the coking segmentation map; The quantitative coking index is output.
[0009] Preferably, the automatic calculation of the quantitative coking index based on the output result of the coking segmentation map specifically includes: Statistics and identification: Counting the total number of pixels classified into different coking levels in the coking segmentation map, and converting them into the heating surface area or area ratio covered by the corresponding level; Weighted scoring: Multiply the area ratio of each coking level by a pre-set severity weight coefficient that is positively correlated with its degree of hazard, and then sum all the products to obtain a global severity score representing the overall coking condition of the boiler; Regional positioning: Based on the preset boiler structure partitions, the coking segmentation map is regionalized and analyzed, and the local coking index is calculated for each independent area.
[0010] Preferably, the abnormality prediction unit adopts an autoencoder model based on a long short-term memory network, which calculates the reconstruction error by learning the data pattern of the boiler under normal operating conditions, and generates an abnormality score and predictive alarm according to the size of the reconstruction error.
[0011] Preferably, the abnormality prediction unit adopts the autoencoder model based on the long short-term memory network to predict the future abnormal state of the boiler, specifically including: Constructing a multidimensional input sequence: Time-aligning and fusing the acoustic signal data and process parameter data collected by the acoustic signal acquisition unit and the process parameter acquisition unit with the quantitative coking index from the coking diagnosis unit to construct a multidimensional time series input vector; Using a pre-trained model: A pre-trained LSTM autoencoder model is applied; the model is trained using a data set of normal and healthy operating conditions of the boiler over various historical time periods and loads to learn "normal" data patterns; Perform data reconstruction: Continuously input the constructed real-time multi-dimensional time series input vector into the trained LSTM autoencoder model, and drive the model to reproduce the input vector according to its learned normal pattern, generating a reconstructed vector at the output of the model; Calculating the reconstruction error: Calculating the mean square error between the original input vector and the reconstructed vector generated by the model in real time as a reconstruction error value that measures the degree to which the current margin deviates from normality; Generate anomaly scores and alerts: The reconstruction error value is converted into a real-time anomaly score ranging from 0 to 100 through a normalization function. The real-time anomaly score is then continuously compared with the preset alert threshold. If the score continuously exceeds the threshold, the system automatically generates and issues a predictive alert.
[0012] Preferably, the sootblowing optimization unit uses a deep reinforcement learning model to generate an optimal sootblowing control strategy, specifically including: Defining the system state: constructing a system state vector in real time and periodically, the vector including: quantitative coking index data from the coking diagnosis unit, real-time anomaly scores from the anomaly prediction unit, and process parameters from the boiler distributed control system; Define the action space: predefine a set of discrete actions that can be selected by the reinforcement learning agent, including all available sootblowing strategies, where each "action" corresponds to a specific operation instruction that can be executed by the distributed control system; Define a reward function: Establish a reward function to quantify the benefits of each "action" under a specific "state." The goal of this function is to maximize the long-term net benefit of boiler operation. The positive reward term is the increased power generation revenue due to the improved heat exchange efficiency after sootblowing, and the negative penalty term is the cost of steam or compressed air consumed during the sootblowing operation. Strategy learning and execution: A reinforcement learning agent is deployed to learn an optimal control strategy by analyzing historical operating data. The sootblowing optimization unit automatically generates the optimal timing and sequence of sootblowing instructions based on the real-time recommendations of the strategy and sends them to the distributed control system for execution.
[0013] A power plant monitoring and management method comprises the following steps: S1. Synchronously collecting thermal imaging images, acoustic signals, and process parameters of the boiler through a data acquisition module; S2, fusing the thermal imaging images, acoustic signals, and process parameter data collected in step S1 into a unified multivariate time series data set; S3. Inputting the thermal imaging image into a coking diagnosis unit to generate a quantitative assessment of the coking state of the boiler heating surface; S4. Inputting the multivariate time series data, including the quantitative assessment of the coking state generated in step S3, into an abnormality prediction unit to generate a prediction of the future abnormal state of the boiler; S5 provides the outputs of step S3 and step S4 as state input to the sootblowing optimization unit to generate an optimal sootblowing control instruction; S6. Display the analysis results of steps S3 and S4 to the operator through the control and visualization module, and execute the control instructions generated in step S5.
[0014] Compared with the prior art, the advantages of the present invention are: Improved thermal efficiency and economy: Through the real-time and precise monitoring of coking by the coking diagnosis unit and the intelligent optimization of soot blowing by the soot blowing optimization unit, the present invention can always maintain the boiler heating surface at a high level of cleanliness, thereby maximizing heat transfer efficiency, optimizing the combustion process, and reducing fuel consumption and power generation costs.
[0015] Improved safety and reliability: The abnormality prediction unit can provide early predictions of critical faults such as pipeline leaks, providing early warning of faults. This allows power plants to convert costly unplanned outages into controllable planned maintenance, thereby avoiding secondary damage and improving unit operation reliability and personnel safety.
[0016] Reduced operation and maintenance costs: This invention realizes the transformation from "scheduled maintenance" or "post-fault maintenance" to "predictive maintenance". It can arrange maintenance activities according to the actual health status of the equipment, avoid unnecessary inspections and spare parts replacement, and thus minimize the total maintenance cost. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] Figure 1 This is a system architecture diagram of a power plant monitoring and management system proposed by the present invention. DETAILED DESCRIPTION
[0018] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present 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 used to explain the present invention and are not intended to limit the present invention.
[0019] Example 1: Reference Figure 1 This embodiment describes in detail the overall architecture of a power plant monitoring and management system proposed by the invention. The system includes: a data acquisition module, a data fusion module, an intelligent analysis module, and a control and visualization module.
[0020] Data acquisition module: This module is the data foundation of the entire system and is responsible for comprehensive and real-time sensing and acquisition of the boiler's operating status. It includes a thermal imaging acquisition unit, an acoustic signal acquisition unit, and a process parameter acquisition unit, which are used to collect thermal imaging image data, acoustic signal data, and process parameter data of the power plant boiler, respectively. Among them: The thermal imaging acquisition unit consists of one or more fixed endoscopic thermal imaging cameras installed on the boiler wall observation hole. The camera works in the mid-wavelength infrared (MWIR) spectral range, especially around 3.9 In this band, the flame and main combustion products (such as , ) has relatively weak radiation, which enables the camera to "penetrate" the flame and high-temperature flue gas and directly and clearly image the inner wall of the furnace and the surface of the heating surface tube array.
[0021] The acoustic signal acquisition unit consists of an array of acoustic sensors distributed across key boiler areas (such as the superheater and reheater). These sensors are designed to detect the high-frequency acoustic emission signals generated by high-pressure steam leaks. Their operating frequency range is from tens of hertz to over 100 kHz. In this embodiment, ultrasonic frequencies (e.g., 20-65 kHz) are preferred to effectively distinguish leakage signals from low-frequency background combustion noise.
[0022] The process parameter acquisition unit integrates conventional sensors from the power plant's existing DCS system to measure key process operating parameters. These include, but are not limited to, boiler exhaust oxygen content, main steam / reheat steam temperature and pressure, feedwater flow at each stage, fuel (pulverized coal) feed rate, and generator load (indirectly reflecting boiler output).
[0023] Data fusion module: This module is connected to the data acquisition module and is responsible for collecting all raw data from the data acquisition module and processing it. Its main functions include: All input data is time-stamped with high precision to ensure temporal alignment of data from different sources.
[0024] Perform data preprocessing, including signal amplification, filtering (for example, filtering out low-frequency combustion noise in acoustic signals), and normalization.
[0025] Fuse pre-processed, time-synchronized heterogeneous data streams (images, acoustics, process parameters) into a unified, high-dimensional time series data vector.
[0026] Intelligent analysis module: This module includes a coking diagnosis unit, an abnormality prediction unit, and a soot blowing optimization unit, realizing multiple functions from state perception to fault prediction to decision optimization. Among them: The coking diagnosis unit is based on a CNN model optimized for image semantic segmentation tasks, preferably U-Net or its improved version. The model is first pre-trained on a large-scale general image dataset and then fine-tuned using the power plant's accumulated, expert-annotated boiler wall thermal maps (marking areas such as clean, lightly coked, and heavily coked). When the system is running, the CNN model receives a sequence of radiometric thermal images from a medium-wave infrared thermal imaging camera and performs forward propagation calculations on each frame of the input thermal image to achieve pixel-level classification. In other words, each pixel in the image is classified into a predefined category such as "clean pipe wall," "lightly coked," or "heavy coked." The model ultimately outputs: A real-time coke distribution map: A pseudo-color visualization shows the severity and spatial distribution of coke in different areas of the furnace. This image can be directly superimposed on the user interface in the control room, providing operators with an intuitive understanding of the coke situation in the furnace.
[0027] A quantitative coking index: The system divides the entire heated surface into multiple regions and calculates the area ratio of coked pixels in each region based on the CNN segmentation results. It also assigns different weights based on the coking level (mild, heavy), and finally calculates a comprehensive, continuously changing coking index for each region.
[0028] The anomaly prediction unit uses an autoencoder based on a long short-term memory network (LSTM), namely the LSTM-AE model. This model uses historical data collected from the boiler under various normal operating conditions (including different loads and different coal types) as training data. Its encoder part learns to compress the complex multivariate time series of the input into a low-dimensional "latent representation". In the process, it learns and remembers the normal and complex spatiotemporal correlations between all input variables. Then, the decoder part attempts to reconstruct the original input sequence from this latent representation. When the system is running, when the input data stream deviates from the "normal mode" learned by the model during the training phase (that is, a sign of a latent fault), the model will not be able to accurately reconstruct this data, resulting in a significant difference between the original input and the reconstructed output. This difference is the "reconstruction error." Based on this "reconstruction error", the model outputs: Anomaly score: A real-time, quantitative score whose size is proportional to the reconstruction error and reflects the degree to which the current system state deviates from normal behavior.
[0029] Predictive alerts: When the anomaly score continuously exceeds a dynamically set threshold, the system triggers a predictive alert.
[0030] It should be noted that the inputs received by the abnormality prediction unit include: the real-time coking index sequence from the coking diagnosis unit, the processed acoustic energy or spectrum feature sequence from the acoustic signal acquisition unit, and the key operating parameter sequence from the process parameter acquisition unit (such as steam temperature, pressure, oxygen content, load, etc. at each point).
[0031] The sootblowing optimization unit uses a deep reinforcement learning algorithm. For example, a model-free, off-policy algorithm such as Proximal Policy Optimization (PPO). The state space of the deep reinforcement learning algorithm includes the coking index output by the coking diagnosis unit, the anomaly score output by the anomaly prediction unit, and real-time process parameters.
[0032] The action space of the deep reinforcement learning algorithm includes instructions for starting and stopping sootblowers in different areas of the boiler.
[0033] The reward function of the deep reinforcement learning model aims to maximize the long-term net benefit of boiler operation, where the positive reward term is the increase in power generation revenue brought about by the improvement in heat exchange efficiency after soot blowing, and the negative penalty term is the cost of steam or compressed air consumed to perform the soot blowing operation.
[0034] At each decision time step, the sootblowing optimization unit outputs an optimal action (i.e., deciding which area’s sootblower to activate, or keep silent) based on the current state and its learned optimal strategy.
[0035] Control and visualization module: This module is used to provide an interactive interface, output the intermediate results of the system, and is also responsible for interaction with the existing DCS. It includes: Graphical User Interface: Provides a user-friendly interactive interface for control room operators. It centrally displays key information output from the intelligent analysis module, including real-time coking distribution diagrams, abnormality prediction alarms, and recommended sootblowing operations.
[0036] Control interface: The generated optimized sootblowing instructions are sent to the power plant's DCS system via a standard industrial communication protocol (such as OPC or Modbus), and the DCS system ultimately executes the start and stop control of the sootblower.
[0037] Example 2: Based on the power plant monitoring and management system, a power plant monitoring and management method is also proposed, which specifically includes: Data acquisition: The system continuously collects infrared thermal imaging images, multi-point acoustic signals and various key process parameters of the boiler through the data acquisition module.
[0038] Data fusion: The data fusion module performs time synchronization, preprocessing and fusion on the collected raw data to construct a unified multivariate time series dataset.
[0039] Coking diagnosis: Real-time thermal imaging images are fed into a pre-trained CNN model. The CNN model performs forward calculations and outputs a quantitative coking index for each heated surface area.
[0040] Anomaly Prediction: The fused time series data, including the coking index, is fed into a pre-trained LSTM-AE model. The LSTM-AE model calculates the reconstruction error of the current state and converts it into a real-time anomaly score.
[0041] State construction: The system constructs the state vector required by the reinforcement learning model at the current time step based on the output of coking diagnosis and anomaly prediction (coking index and anomaly score) combined with the real-time process parameters (load, efficiency, etc.) obtained from the DCS.
[0042] Optimal action selection step: The reinforcement learning model receives the current state vector and selects the action that maximizes the long-term cumulative reward (i.e., the optimal soot-blowing decision) based on its internally learned and optimized control strategy.
[0043] Visualization and control execution: The control and visualization module displays all analysis results (coking diagrams, abnormal alarms) and decision recommendations (soot blowing instructions) on the control room user interface. It also sends control instructions to the DCS system for execution.
[0044] In a preferred embodiment, the system also includes a feedback and learning loop: After executing an action, the boiler's state changes (for example, the coking index in a certain area decreases after sootblowing, improving boiler efficiency). These changes are captured by the data acquisition module, which updates the state vector during the next decision cycle. This process forms a complete feedback loop, enabling the reinforcement learning model to continuously adjust and optimize its control strategy based on the actual results of its decisions, achieving continuous learning and self-improvement.
[0045] Throughout this specification, references to terms such as "one embodiment," "example," or "specific example" indicate that the specific features, structures, materials, or characteristics described in conjunction with that embodiment or example are included in at least one embodiment or example of the present invention. In this specification, schematic representations of these terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.
[0046] The preferred embodiments of the present invention disclosed above are intended only to help illustrate the present invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the present invention to the specific embodiments described. Obviously, many modifications and variations are possible based on the content of this specification. These embodiments are selected and described in detail in this specification to better explain the principles and practical applications of the present invention, thereby enabling those skilled in the art to better understand and utilize the present invention. The present invention is limited only by the claims and their full scope and equivalents.
Claims
1. A power plant monitoring and management system, characterized in that: Including data acquisition module, data fusion module, intelligent analysis module and control and visualization module; The data acquisition module includes a thermal imaging acquisition unit, an acoustic signal acquisition unit and a process parameter acquisition unit, which are respectively used to collect thermal imaging image data, acoustic signal data and process parameter data of the power plant boiler; The data fusion module is connected to the data acquisition module and is used to time-align and pre-process the thermal imaging images, acoustic signals and process parameter data, and fuse them into a multivariate time series data set; The intelligent analysis module includes a coking diagnosis unit, an abnormality prediction unit and a soot blowing optimization unit; The coking diagnosis unit is used to process the thermal imaging image to quantitatively evaluate the coking state of the boiler heating surface in real time; The abnormality prediction unit is used to process a time series data including the coking state evaluation result, the acoustic signal and the process parameters to predict the future abnormal state of the boiler; The sootblowing optimization unit is used to generate an optimal sootblowing control strategy according to the coking state evaluation result and the future abnormal state prediction; The control and visualization module is connected to the intelligent analysis module and the distributed control system of the power plant, and is used to display the coking state assessment results and predictive alarms, and execute the soot blowing control instructions.
2. A power plant monitoring and management system according to claim 1, characterized in that: The coking diagnosis unit uses a convolutional neural network model that is trained to perform semantic segmentation on the thermal imaging image, classify pixels in the image into different coking levels, and calculate a quantitative coking index for one or more regions based on the segmentation results.
3. A power plant monitoring and management system according to claim 2, characterized in that: The performing semantic segmentation on the thermal imaging image, classifying pixels in the image into different coking levels, and calculating a quantitative coking index of one or more regions based on the segmentation results, includes: Inputting the thermal imaging image as input data into a pre-trained convolutional neural network model; The convolutional neural network model performs semantic segmentation processing on the thermal imaging image to generate a coke segmentation map; in the coke segmentation map, each pixel or pixel region of the image is intelligently classified into a specific level in a predefined coke level set, wherein the coke level set includes: "clean", "light coke" and "severe coke"; Automatically calculating the quantitative coking index based on the output result of the coking segmentation map; The quantitative coking index is output.
4. A power plant monitoring and management system according to claim 3, characterized in that: The automatically calculating the quantitative coking index based on the output result of the coking segmentation map specifically includes: Statistics and identification: Counting the total number of pixels classified into different coking levels in the coking segmentation map, and converting them into the heating surface area or area ratio covered by the corresponding level; Weighted scoring: Multiply the area ratio of each coking level by a pre-set severity weight coefficient that is positively correlated with its degree of hazard, and then sum all the products to obtain a global severity score representing the overall coking condition of the boiler; Regional positioning: Based on the preset boiler structure partitions, the coking segmentation map is regionalized and analyzed, and the local coking index is calculated for each independent area.
5. A power plant monitoring and management system according to claim 1, characterized in that: The anomaly prediction unit adopts an autoencoder model based on a long short-term memory network, which calculates the reconstruction error by learning the data pattern of the boiler under normal operating conditions, and generates an anomaly score and predictive alarm according to the size of the reconstruction error.
6. A power plant monitoring and management system according to claim 1 or 5, characterized in that: The abnormality prediction unit uses the autoencoder model based on the long short-term memory network to predict the future abnormal state of the boiler, including: Constructing a multidimensional input sequence: Time-aligning and fusing the acoustic signal data and process parameter data collected by the acoustic signal acquisition unit and the process parameter acquisition unit with the coking index from the coking diagnosis unit to construct a multidimensional time series input vector; Using a pre-trained model: A pre-trained LSTM autoencoder model is applied. This model is trained using a data set of normal and healthy operating conditions of the boiler over various historical time periods and loads to learn "normal" data patterns. Perform data reconstruction: Continuously input the constructed real-time multi-dimensional time series input vector into the trained LSTM autoencoder model, and drive the model to reproduce the input vector according to its learned normal pattern, generating a reconstructed vector at the output of the model; Calculating the reconstruction error: Calculating the mean square error between the original input vector and the reconstructed vector generated by the model in real time as a reconstruction error value that measures the degree to which the current margin deviates from normality; Generate anomaly scores and alerts: The reconstruction error value is converted into a real-time anomaly score ranging from 0 to 100 through a normalization function. The real-time anomaly score is then continuously compared with the preset alert threshold. If the score continuously exceeds the threshold, the system automatically generates and issues a predictive alert.
7. A power plant monitoring and management system according to claim 1, characterized in that: The sootblowing optimization unit uses a deep reinforcement learning model to generate the optimal sootblowing control strategy, including: Defining the system state: constructing a system state vector in real time and periodically, the vector including: quantitative coking index data from the coking diagnosis unit, real-time anomaly scores from the anomaly prediction unit, and process parameters from the boiler distributed control system; Define the action space: predefine a set of discrete actions that can be selected by the reinforcement learning agent, including all available sootblowing strategies, where each "action" corresponds to a specific operation instruction that can be executed by the distributed control system; Define a reward function: Build a reward function to quantify the benefits of each action in a specific state. The goal of this function is to maximize the long-term net benefit of boiler operation. The positive reward term is the increased power generation revenue resulting from the improved heat exchange efficiency after sootblowing, and the negative penalty term is the cost of steam or compressed air consumed during sootblowing. Strategy learning and execution: A reinforcement learning agent is deployed to learn an optimal control strategy by analyzing historical operating data. The sootblowing optimization unit automatically generates the optimal timing and sequence of sootblowing instructions based on the real-time recommendations of the strategy and sends them to the distributed control system for execution.
8. A power plant monitoring and management method, characterized in that: include: S1. Synchronously collecting thermal imaging images, acoustic signals, and process parameters of the boiler through a data acquisition module; S2, fusing the thermal imaging images, acoustic signals, and process parameter data collected in step S1 into a unified multivariate time series data set; S3. Inputting the thermal imaging image into a coking diagnosis unit to generate a quantitative assessment of the coking state of the boiler heating surface; S4. Inputting the multivariate time series data, including the quantitative assessment of the coking state generated in step S3, into an abnormality prediction unit to generate a prediction of the future abnormal state of the boiler; S5 provides the outputs of step S3 and step S4 as state input to the sootblowing optimization unit to generate an optimal sootblowing control instruction; S6. Display the analysis results of steps S3 and S4 to the operator through the control and visualization module, and execute the control instructions generated in step S5.