Boiler efficiency optimization method based on big data

Abstract

The invention discloses a boiler efficiency optimization method based on big data. The method comprises the following steps: S01, carrying out real-time multi-modal data acquisition and verification fusion; s02, performing adaptive data processing and feature engineering on the high-quality data stream, and dynamically adjusting verification depth and frequency; s03, performing dynamic prediction model management and adaptive re-calibration; s04, executing multi-target robust optimization based on dynamic preference, automatically generating preference setting by using a reinforcement learning agent, searching a Pareto optimal solution by using an evolutionary algorithm, and ensuring the security of reinforcement learning; and S05, executing optimization control and fine regulation and control of a specific subsystem. The invention aims to solve the problems of low data quality, poor model adaptability, single optimization target and the like in the prior art, and comprehensively improves the operation efficiency, economical efficiency and intelligent level of the boiler through systematic data processing, self-adaptive modeling, multi-target intelligent decision and fine control.

Description

Technical Field

[0001] The present invention relates to the technical field of boiler control and optimization, and in particular to a boiler efficiency optimization method based on big data. Background Art

[0002] As core equipment for energy conversion, the operating efficiency of boilers directly impacts energy consumption and operating costs. Traditional boiler control methods often rely on empirical models or fixed control logic, making them difficult to adapt to complex and changing operating conditions, such as fuel quality fluctuations, load changes, and equipment aging. In recent years, big data-based boiler efficiency optimization methods have begun to attract attention. These methods typically utilize historical and real-time operating data, combined with predictive control, intelligent control, and automatic optimization techniques, to achieve precise control of the boiler's air-coal ratio and supply air ratio, thereby maintaining optimal boiler efficiency.

[0003] However, existing big data-based boiler efficiency optimization methods still face many challenges in practical applications. First, the effectiveness of big data analysis is highly dependent on the quality of input data. Sensor noise, drift, failure or data loss will seriously affect the accuracy of the model and the control effect. Secondly, the boiler is a dynamic and complex nonlinear system. The pre-established intelligent model is difficult to fully adapt to its time-varying characteristics, resulting in a decrease in model prediction accuracy and a degradation of control performance. Furthermore, boiler optimization often involves multiple conflicting objectives, such as thermal efficiency, pollutant emissions, equipment life, etc. Simply pursuing the optimality of a single objective may lead to losing sight of one thing while focusing on another. In addition, complex prediction models and optimization algorithms may bring a large computational burden, affecting the real-time response capability of the control system. Finally, some key parameters that have a significant impact on the combustion process (such as real-time coal quality) are difficult to obtain quickly and accurately, which also restricts the accuracy of optimization control. Summary of the Invention

[0004] The purpose of the present invention is to provide a boiler efficiency optimization method, device and computer-readable storage medium based on big data, aiming to solve the problems in the existing technology such as low data quality, poor model adaptability, single optimization target, insufficient real-time performance and difficulty in obtaining key parameters, so as to improve the boiler operation efficiency in a more comprehensive, intelligent and robust manner.

[0005] To achieve the above objectives, the present invention provides a boiler efficiency optimization method based on big data, comprising the following steps:

[0006] S01. Real-time multimodal data collection and verification fusion, including: using sensors to collect multimodal data of boiler operation in real time, and performing real-time verification and cleaning on the collected multimodal data to generate high-quality data streams;

[0007] S02. Performing adaptive data processing and feature engineering on the high-quality data stream, including dynamically adjusting the depth and frequency of data verification based on the stability of the boiler's operating status and the sensitivity of key parameters, and prioritizing the verification quality of core data when computing resources are limited;

[0008] S03. Perform dynamic prediction model management and adaptive recalibration, including: online monitoring of the performance of the prediction model based on the high-quality data stream; when the performance degradation index of the prediction model exceeds a preset threshold, triggering online fine-tuning of the parameters of the prediction model or switching to an alternative model structure; and verifying the stability and effectiveness of the updated model through a shadow model and compliance assessment mechanism;

[0009] S04. Perform multi-objective robust optimization based on dynamic preferences, including: constructing the boiler optimization problem as a multi-objective optimization problem that includes multiple objectives such as thermal efficiency, emissions, and equipment losses; utilizing a reinforcement learning agent to automatically generate adaptive preference settings based on the current boiler state and external environmental factors; and employing an evolutionary algorithm to search for a Pareto optimal solution set based on the preference settings. Furthermore, digital twin rehearsals and safety envelope constraints are used to ensure the safety of the reinforcement learning agent's training and deployment.

[0010] S05. Execute optimization control and fine-tuning of specific subsystems, including: adjusting boiler operating parameters based on the optimization set point obtained by the multi-objective robust optimization, and dynamically optimizing and scheduling specific subsystems such as the sootblowing system.

[0011] As a preferred solution, in step S01, the real-time multimodal data acquisition and verification fusion may further include: performing dynamic threshold verification and change rate verification on single-point data through statistical process control principles, sensor redundancy information and expected values ​​based on simple physical models, performing consistency verification on multi-point related sensor data, and performing anomaly detection through model residual analysis; and / or using multivariate state estimation and hidden Markov models to perform online evaluation of the health status of the sensor and early fault warning, and using the warning information to guide the data verification strategy; and / or using multimodal process data fusion and machine learning models to perform real-time soft measurement of key coal quality parameters that are not directly measured to supplement the multimodal data.

[0012] As a preferred solution, in the single point data verification in step S01, the dynamic threshold L low,i ,L high,i Determined based on historical data distribution:

[0013] L low,i =μ i -k·σ i

[0014] Lhigh,i =μ i +k·σ i

[0015] Among them, x i (t) is the reading of sensor i at time t, μ i is the historical mean of sensor i’s readings, σ i is its historical standard deviation, and k is an adjustable parameter.

[0016] As a preferred solution, in step S02, the adaptive data processing and feature engineering may further include: dividing the sensor data and monitored parameters into different levels according to their criticality, and dynamically selecting the corresponding verification strategy combination and adjusting the verification intensity or frequency based on the boiler operation status evaluation results and calculated load.

[0017] As a preferred solution, in step S03, the dynamic prediction model management and adaptive recalibration may further include: online fine-tuning of the parameters of the current model using recursive least squares or gradient descent; and / or when parameter fine-tuning fails to effectively improve the performance degradation index, selecting the optimal model for replacement based on an information criterion (such as Akaike Information Criterion AIC) from a library of alternative models containing one or more different structures or complexities.

[0018] As a preferred solution, in step S04, the multi-objective robust optimization based on dynamic preference may further include: the state space of the reinforcement learning agent includes key boiler operating parameters, model performance indicators and external environmental factors; the action space is a multi-objective optimization weight vector or reference point; the reward function comprehensively reflects the overall effect of the multi-objective optimization, the control cost and the compliance with high-level constraints; and / or the evolutionary algorithm adopts a non-dominated sorting genetic algorithm NSGA-II or a decomposition-based multi-objective evolutionary algorithm MOEA / D.

[0019] As a preferred solution, in step S05, the fine control of the specific subsystem can further include: real-time monitoring and identification of the deterioration rate of the heat transfer performance of each heat exchange surface of the boiler, combining the sootblowing medium consumption cost, the potential impact of the sootblowing operation on the equipment life and the efficiency gain brought by cleaning, constructing a multi-objective optimization problem, and dynamically deciding the optimal sootblowing time, location and intensity.

[0020] As a preferred solution, the heat transfer coefficient K of the heat exchange surface area j is j (t) is calculated in real time using the following formula:

[0021]

[0022] in is the steam flow rate in region j at time t, hout,j (t) and h in,j (t) are the outlet and inlet steam enthalpies, A j is the heat transfer area, ΔT lm,j (t) is the log mean temperature difference.

[0023] The beneficial effects of the present invention are: compared with the prior art,

[0024] 1. Through multi-level data verification, sensor health warnings, and soft measurement of key parameters, the quality of the basic data for big data analysis and intelligent control is ensured, reducing decision-making errors caused by data problems. This is achieved through the comprehensive data processing strategy in step S01, which combines statistical methods, physical models, machine learning, and sensor health status assessment to systematically improve data credibility.

[0025] 2. Through online model performance monitoring, adaptive parameter correction, and secure model update and deployment mechanisms, the prediction and control model can better adapt to the time-varying characteristics of the boiler and maintain high accuracy over the long term. The dynamic model management in step S03, particularly the shadow model and progressive deployment, ensures a smooth and secure model update process, avoiding the risks associated with direct switching.

[0026] 3. By introducing a multi-objective optimization framework and adaptive generation of dynamic preferences based on reinforcement learning, intelligent trade-offs and collaborative optimization can be achieved across multiple objectives, such as boiler efficiency, emissions, and losses, more closely aligned with actual operational needs. The combination of multi-objective optimization and reinforcement learning in step S04 enables the system to autonomously learn and adjust optimization priorities, rather than relying on fixed weights or frequent manual intervention.

[0027] 4. Through dynamic optimization and scheduling of specific subsystems, such as sootblowing, and adaptive management of computing resources, deeper, finer-grained control and intelligent management of boiler operations are achieved. The dynamic allocation of computing resources in step S02 ensures real-time performance, while dynamic sootblowing optimization in step S05 embodies refined management of specific operations, further tapping into energy-saving potential.

[0028] 5. Combining these improvements, the present invention enables the boiler to operate stably and long-term at an optimal operating point, significantly improving energy efficiency, reducing pollutant emissions, and minimizing unnecessary equipment wear and tear, thereby delivering significant economic and environmental benefits. This is the result of the synergistic effect of steps S01 through S05, forming a closed-loop intelligent optimization system from data to decision-making to execution. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 Schematic diagram of the overall process of the boiler efficiency optimization method based on big data of the present invention;

[0030] Figure 2 Detailed flowchart of real-time multimodal data acquisition and verification fusion in step S01 of the method of the present invention;

[0031] Figure 3 Schematic diagram of the process of dynamic prediction model management and adaptive recalibration in step S03 of the method of the present invention;

[0032] Figure 4 This is a schematic diagram of the module structure of the boiler efficiency optimization device based on big data of the present invention;

[0033] Figure 5 Schematic diagram of the process of multi-objective robust optimization based on dynamic preference in step S04 of the method of the present invention;

[0034] Figure 6 Schematic diagram of the process of sensor health assessment and early fault warning in the method of the present invention;

[0035] Figure 7 Schematic diagram of the process of real-time soft measurement of coal quality parameters in the method of the present invention;

[0036] Figure 8 Schematic diagram of the process of dynamic intelligent sootblowing scheduling in the method of the present invention. DETAILED DESCRIPTION

[0037] The following further describes software embodiments of the present invention in conjunction with the accompanying drawings, but the present invention is not limited to the following embodiments. Unless otherwise specified, a "module" as used in this specification may refer to a hardware entity capable of performing a specific function, a software entity (e.g., a program code segment), or a combination of the two. A software entity may be an instruction stored on a computer-readable storage medium and executed by a processor.

[0038] Example 1:

[0039] Reference Figures 1 to 3 as well as Figures 5 to 8 The present invention provides a detailed description of the boiler efficiency optimization method based on big data. This method aims to achieve intelligent, adaptive, and multi-objective optimization of boiler operating efficiency through a series of collaborative steps.

[0040] Step S01: Real-time multi-modal data acquisition and verification fusion. Figure 2As shown in Figure 1, this step is the basis of the entire optimization method. First, the multimodal data of the boiler operation are collected in real time through various sensors equipped with the boiler (such as temperature, pressure, flow, flue gas analyzer, flame detector, acoustic sensor, etc.). These data constitute the original input of big data analysis. Then, the collected raw data are fully verified and cleaned in real time, which may include the following sub-steps: S01.1: Data verification and cleaning based on statistics and model residual analysis. For each single data point x collected by the sensor, i (t), set the dynamic threshold L low,i ,L high,i Check, where L low,i =μ i -k·σ i , L high,i =μ i +k·σ i (μ i is the historical mean, σ i is the historical standard deviation, and k is an adjustable parameter, such as 3. At the same time, check whether the rate of change of the data point exceeds the physically allowed range. For multiple sensors measuring the same physical quantity or strongly correlated physical quantities, calculate the difference in their readings. If the difference persists, mark it as suspicious. In addition, establish a simplified dynamic or static correlation model y between the key parameters of the boiler. pred,j (t)=f(x1(t),...,x m (t)), calculate the residual error e between the model prediction value and the actual sensor measurement value j (t) = y obs,j (t)-y pred,j (t), and apply statistical process control techniques (such as CUSUM charts) to the residual series to detect anomalies. Data identified as bad points or highly suspicious are replaced by interpolation, model estimates, or confirmed values ​​from redundant sensors or marked as unavailable.

[0041] S01.2: Sensor health assessment and early fault warning. Figure 6 As shown, this sub-step provides support for S01.1. By fusing information from multiple related sensors, a multivariate state estimation algorithm (such as a Kalman filter) is used to generate an optimal estimate of the measured physical quantity. The actual reading of each sensor z i,t The corresponding part of this optimal estimate The residual e i,t Used to extract statistical features F i,t These features are input into a pre-trained Hidden Markov Model (HMM), where the hidden states of the HMM represent different health levels of the sensor (healthy, slightly drifted, severely noisy, etc.). The forward algorithm calculates the probability of each hidden state under the current observation sequence to obtain the sensor health Hi (t). When the probability of a certain fault state is too high or the health level is lower than the threshold, an early fault warning is triggered. This warning information can adjust the weight or strategy of data verification in S01.1. HMM model parameter λ=(A HMM ,B HMM ,π0) includes the state transition probability matrix A HMM , observation probability distribution B HMM (usually a Gaussian mixture model) and an initial state distribution π0, trained using historical data and the Baum-Welch algorithm.

[0042] S01.3: Real-time soft measurement of coal quality parameters. Figure 7 As shown in the figure, in order to obtain key coal quality parameters that are difficult to measure directly (such as calorific value Q net,ar ), this sub-step utilizes multimodal data that is easy to measure (such as flame image features f image , furnace acoustic signal characteristics f acoustic , flue gas analysis data, time series characteristics of relevant process parameters f process ) for online estimation. Using deep learning models, such as CNN (processing images), RNN (processing time series) and attention mechanism (fusion of multimodal features f fused =Attention(f image ,f acoustic ,f process )) network structure, which maps the fused features to the estimated values ​​of the target coal quality parameters through the fully connected layer The model is trained using historical data (including multimodal process data and corresponding offline coal quality test results), and the loss function is usually the mean square error The real-time coal quality parameters obtained by soft sensing will be used as part of the high-quality data stream for subsequent model use.

[0043] Through the above sub-steps, a high-quality data stream that has been fully verified and integrates direct measurement and indirect estimation information is finally generated.

[0044] Step S02: Adaptive data processing and feature engineering. This step takes the high-quality data stream generated by S01 and further processes it to meet the needs of subsequent modeling, while taking into account the efficient use of computing resources. According to the boiler operating state S(t) (such as stable operation, rapid load changes, etc.), the key parameter volatility S(t) = f state (Var(P ms ), dL / dt, alarms) evaluation) and the criticality of each data parameter C i(divided into levels I, II, and III), dynamically adjust the depth and frequency of data verification in S01. For example, when the load changes rapidly, a comprehensive verification (strategy C) is performed on the level I parameters; during stable operation, only a quick verification (strategy A) is performed on the level III parameters. At the same time, the computing load L of the data processing unit is monitored. cpu (t), if it exceeds the threshold L cpu,max , then reduce the verification intensity or frequency according to the priority (first level III, then level II, then level I), and adjust the coefficient α i (t) = f adjust (C i ,S(t),L cpu (t)). This step ensures a dynamic balance between real-time data processing and complete verification under varying operating conditions and resource constraints. The processed data undergoes further feature engineering to extract features highly correlated with target variables such as boiler efficiency and emissions for use in subsequent models.

[0045] Step S03: Dynamic prediction model management and adaptive recalibration. Figure 3 As shown in Figure 2, this step is responsible for the maintenance and update of the boiler key process prediction model (such as the prediction of main steam temperature, flue gas composition, etc.) to ensure its long-term accuracy. S03.1: Online monitoring of model performance. For the core prediction model M k , continuously calculating its short-term forecast error And calculate the cumulative performance degradation indicators such as root mean square error based on the error in the rolling time window S03.2: Online model parameter fine-tuning and structure selection. RMSE k (t) Set the warning threshold T warn,k and action threshold T act,k When RMSE k (t)>T ac x ,k When , the model correction is triggered. First try to use recursive least squares (RLS) or gradient descent method For the current model M k The parameter θ k Perform online fine-tuning. If parameter fine-tuning is ineffective or the system undergoes significant structural changes, select an alternative model from a library containing more than one model with different structures or complexities. In the example, an optimal model M is selected based on the information criterion (such as Akaike information criterion AIC = 2p-2ln(L), where p is the number of parameters and L is the maximum likelihood value). k,best Replace the current model. S03.3: Shadow model and conformity assessment. The updated model (whether it is M′ obtained by fine-tuning the parameters) k Or the newly selected M new,k ) First, as the shadow model M sh adow,kWith the current master model M active,k By comparing the performance differences of the two in key performance indicators (KPIs) such as prediction accuracy, simulation control signal stability, and economic benefit estimation, ΔKPI j (t), and combined with stability check, compliance assessment is performed. valid Demonstrated sustained excellence and stability (e.g., ΔKPI j (t) are greater than a certain threshold, or at least one has improved significantly and the others have not deteriorated), the switch is approved. The switch can use a gradual control weight transfer u final (t) = λ(t)u sh adow,k (t)+(1-λ(t))u active,k (t), where λ(t) increases smoothly from 0 to 1 to ensure the smoothness and security of the model update.

[0046] Step S04: Execute multi-objective robust optimization based on dynamic preference. Figure 5 As shown in Figure 2, this step aims to balance multiple conflicting objectives in boiler operation, such as maximizing thermal efficiency η(x), reducing NOx emissions E NOx (x) is minimized, equipment wear W(x) is minimized, etc. S04.1: Multi-objective optimization problem construction and evolutionary algorithm solution. Boiler optimization is defined as a multi-objective optimization problem (MOOP), with objective functions f1(x),...,f m (x), where the decision variable x is the operating parameter such as the air-coal ratio. An evolutionary multi-objective optimization algorithm (EMOA) such as NSGA-II is used to search for the Pareto optimal frontier. S04.2: Dynamic preference adaptive generation based on reinforcement learning. To overcome the problem of manually set preferences (such as weight vector w i (t) or reference point Z * (t)) hysteresis and suboptimality, a reinforcement learning (RL) agent is introduced. The state space S of the RL agent is t Including boiler operating parameters, model performance, external environmental factors (electricity price, emission limits, etc.); action space A t The optimized weight w(t) or reference point Z for the output * (t); reward function R t+1 Designed to comprehensively reflect the overall effect of multi-objective optimization, such as R t+1 =αΔη(t)-βΔE NOx (t)-γControlCost(t). Use DDPG or PPO algorithms to train the RL agent so that it can automatically generate preference settings that adapt to the current working conditions. S04.3: Digital twin preview and safety envelope constraints. To ensure the safety of RL agent training and deployment, first use the high-fidelity boiler digital twin model M DTPerform sufficient offline pre-training on the RL agent. At the same time, define a dynamically updated safe action envelope SE(t) to generate the original preference A of the RL agent. raw (t) Perform real-time verification, if Then correct it to the nearest point within the safety envelope Or adopt preset conservative preferences. This ensures that RL-driven preference adjustments remain within a safe and controllable range. Evolutionary optimizers (such as NSGA-II) use the safety-verified preference settings generated by the RL agent to select one or a set of trade-off solutions from the Pareto frontier as the optimal operating set point for the boiler.

[0047] Step S05: Execute optimization control and fine-tune specific subsystems. This step applies the optimized set point obtained in S04 to the actual boiler control system, adjusting key operating parameters such as air-coal ratio and air damper opening. In addition, it also includes the refinement and dynamic optimization scheduling of specific subsystems, such as dynamic intelligent soot blowing. Figure 8 As shown, by real-time monitoring and identification of the heat transfer performance deterioration rate of each heat exchange surface of the boiler (such as superheater, reheater) (for example, by calculating the heat transfer coefficient Or fouling thermal resistance R f,j (t) changes), combined with the sootblowing media consumption costs, the potential impact of sootblowing operations on equipment life, and the efficiency gains from cleaning (Benefit B j (t), cost C total,j ), a multi-objective sootblowing decision optimization problem is constructed. The goal is to maximize the net benefit∑ j (B j -C totla,j ), while maintaining the fouling thermal resistance within an acceptable range. Using heuristic optimization algorithms or rule-based expert systems, dynamic decisions are made on the optimal sootblowing timing, location, and intensity. At the same time, based on the identified fouling thermal resistance growth rate dR f,j (t) / dt, which can predict the time T when the dirt reaches the critical value in the future crit,j (t) = t + (R f,max,j -R f,j (t)) / (dR f,j (t) / dt), and perform predictive sootblowing scheduling.

[0048] Through the collaborative work of the above steps S01 to S05, a closed-loop, adaptive, multi-objective collaborative boiler intelligent optimization control system is formed, thereby significantly improving the overall operating performance of the boiler.

[0049] Example 2:

[0050] Reference Figure 4The present invention also provides an embodiment of a big data-based boiler efficiency optimization device 100. This device 100 is designed to implement the aforementioned big data-based boiler efficiency optimization method. This device 100 can be an integrated control system or deployed in a distributed computing environment. It primarily includes a data processing module 110, a model management module 120, an optimization decision module 130, and a control execution module 140. These modules can communicate and exchange data via an internal bus or network.

[0051] Data processing module 110: Responsible for executing the functions described in steps S01 and S02 of the aforementioned method. Specifically, this module is connected to the boiler sensor network through an interface to collect multimodal operating data in real time. It integrates a data verification and cleaning submodule (implementing the functions of S01.1, including statistical verification, residual analysis, etc.), a sensor health assessment submodule (implementing the HMM logic of S01.2, such as Figure 6 Related processes shown), coal quality parameter soft measurement submodule (implementing the deep learning model logic of S01.3, such as Figure 7 The module outputs high-quality, feature-rich data streams to the model management module 120 and the optimization decision module 130.

[0052] Model management module 120: responsible for executing the functions described in step S03 of the above method (such as Figure 3 The module maintains prediction models for one or more key boiler processes. It contains an online model performance monitoring unit for real-time evaluation of the prediction accuracy of the currently active model (such as calculation of RMSE). When model performance degradation is detected, the model update mechanism is triggered. The model update mechanism includes an online parameter fine-tuning unit (such as based on the RLS algorithm) and a model structure selection unit (selected from the alternative model library based on criteria such as AIC). In addition, the module also includes a key shadow model verification and deployment unit, which is used for parallel operation, compliance assessment and progressive weight transfer before the updated model is officially put into use to ensure smooth and safe model switching.

[0053] Optimization decision module 130: responsible for executing the functions described in step S04 of the above method (such as Figure 5The core of this module is a multi-objective optimization engine (such as NSGA-II or MOEA / D), which is used to solve boiler optimization problems with multiple objectives such as efficiency, emissions, and losses. It also integrates a dynamic preference generation unit based on reinforcement learning, whose RL agent can autonomously learn and output adaptive multi-objective optimization weights or reference points based on real-time operating conditions and external factors (such as electricity prices). To ensure the safety of RL, the module also includes a digital twin interaction interface (for RL pre-training) and a safety envelope constraint verification unit to filter and correct the preferences output by the RL agent in real time. The optimization decision module 130 ultimately outputs a set of optimal boiler operation set points.

[0054] Control execution module 140: Responsible for executing the functions described in step S05 of the above method. This module receives the optimized set point from the optimization decision module 130 and converts it into specific control instructions for the boiler actuators (such as dampers, coal feeders, water pumps, etc.), realizing closed-loop control of the main operating parameters of the boiler. In addition, this module also contains the fine control logic of specific subsystems, such as the dynamic intelligent sootblowing scheduling unit (such as Figure 8 The unit dynamically determines the soot blowing operation based on the fouling of the heat exchange surface and the cost-benefit analysis, and sends the instruction to the soot blowing system for execution.

[0055] These modules work together to enable the device 100 to automatically, intelligently, and efficiently optimize boiler operation. The device can implement the functions of the above modules by executing computer program instructions stored in the memory through the processor.

[0056] The above is a further detailed description of the present invention in conjunction with specific preferred embodiments, and the specific implementation of the present invention should not be considered to be limited to these descriptions. For those skilled in the art of the present invention, without departing from the concept of the present invention, several simple deductions or substitutions can be made, which should be considered to fall within the scope of protection of the present invention.

Claims

1. A boiler efficiency optimization method based on big data, characterized in that: The steps include: S01. Real-time multimodal data collection and verification fusion, including: using sensors to collect multimodal data of boiler operation in real time, and performing real-time verification and cleaning on the collected multimodal data to generate a data stream; S02. performing adaptive data processing and feature engineering on the data stream, including dynamically adjusting the depth and frequency of data verification based on the stability of the boiler's operating status and the sensitivity of key parameters, and prioritizing the verification quality of core data when computing resources are limited; S03. Perform dynamic prediction model management and adaptive recalibration, including: online monitoring of the performance of the prediction model based on the high-quality data stream; when the performance degradation index of the prediction model exceeds a preset threshold, triggering online fine-tuning of the parameters of the prediction model or switching to an alternative model structure; and verifying the stability and effectiveness of the updated model through a shadow model and compliance assessment mechanism; S04. Perform multi-objective robust optimization based on dynamic preferences, including: constructing the boiler optimization problem as a multi-objective optimization problem that includes multiple objectives such as thermal efficiency, emissions, and equipment losses; utilizing a reinforcement learning agent to automatically generate adaptive preference settings based on the current boiler state and external environmental factors; and employing an evolutionary algorithm to search for a Pareto optimal solution set based on the preference settings. Furthermore, digital twin rehearsals and safety envelope constraints are used to ensure the safety of the reinforcement learning agent's training and deployment. S05. Execute optimization control and fine-tuning of specific subsystems, including: adjusting boiler operating parameters based on the optimization set point obtained by the multi-objective robust optimization, and dynamically optimizing and scheduling specific subsystems such as the sootblowing system.

2. The method according to claim 1, characterized in that The real-time multimodal data acquisition and verification fusion in step S01 further includes: Perform dynamic threshold and rate-of-change checks on single-point data, consistency checks on multi-point correlated sensor data, and anomaly detection through model residual analysis using statistical process control principles, sensor redundancy information, and expected values ​​based on simple physical models; and / or Use multivariate state estimation and hidden Markov models to perform online assessment of sensor health and early warning of faults, and use warning information to guide data verification strategies; and / or Multimodal process data fusion and machine learning models are used to perform real-time soft measurements of key coal quality parameters that are not directly measured to supplement the multimodal data.

3. The method according to claim 2, characterized in that The single point data verification in step S01, its dynamic threshold L low,i ,L high,i Determined based on historical data distribution: L low,i =μ i -k·σ i , L high,i =μ i +k·σ i , where x i (t) is the reading of sensor i at time t, μ i is the historical mean of sensor i’s readings, σ i is its historical standard deviation, and k is an adjustable parameter.

4. The method according to claim 1, wherein The adaptive data processing and feature engineering in step S02 further includes: dividing the sensor data and monitored parameters into different levels according to their criticality, and dynamically selecting the corresponding verification strategy combination and adjusting the verification intensity or frequency based on the boiler operation status evaluation results and calculated load.

5. The method according to claim 1, wherein The dynamic prediction model management and adaptive recalibration in step S03 further includes: Fine-tune the parameters of the current model online using recursive least squares or gradient descent; and / or When parameter fine-tuning fails to effectively improve the performance degradation index, the optimal model is selected for replacement based on the information criterion from a library of alternative models containing at least one model with different structures or complexities.

6. The method according to claim 1, wherein The multi-objective robust optimization based on dynamic preference in step S04 further includes: The state space of the reinforcement learning agent includes key boiler operating parameters, model performance indicators and external environmental factors; the action space is a multi-objective optimization weight vector or reference point; the reward function comprehensively reflects the overall effect of the multi-objective optimization, control cost and compliance with high-level constraints; and / or The evolutionary algorithm adopts a non-dominated sorting genetic algorithm NSGA-II or a decomposition-based multi-objective evolutionary algorithm MOEA / D.

7. The method according to claim 1, characterized in that The fine control of the specific subsystem in step S05 further includes: real-time monitoring and identification of the heat transfer performance deterioration rate of each heat exchange surface of the boiler, combining the sootblowing medium consumption cost, the potential impact of the sootblowing operation on the equipment life, and the efficiency gain brought by cleaning, constructing a multi-objective optimization problem, and dynamically deciding the optimal sootblowing timing, location, and intensity.

8. The method according to claim 7, characterized in that The heat transfer coefficient K of the heat exchange surface area j j (t) is calculated in real time using the following formula: in is the steam flow rate in area j at time t, h out,j (t) and h in,j (t) are the outlet and inlet steam enthalpies, A j is the heat transfer area, ΔT lm,j (t) is the log mean temperature difference.

Images