Boiler soot blowing method, system, medium and equipment

Abstract

The invention discloses a boiler soot blowing method, system, medium and equipment, and relates to the technical field of coal-fired boiler heating surface cleaning. According to the boiler operation state data of the recent historical time period, the soot deposition change condition of the boiler in the subsequent time period is predicted, the optimal soot blowing strategy of each soot blower in the boiler is determined through the trained reinforcement learning model according to the prediction result and the boiler operation state data, traditional trigger type soot blowing control is avoided, and the soot blowing efficiency is improved. Preventive and continuous cleaning measures can be taken before the dust accumulation problem is serious, and the boiler is always kept in an efficient operation interval. Not only can consumption of soot blowing media be obviously reduced, but also unnecessary erosion corrosion to boiler pipelines can be reduced to the maximum extent, and the service life of equipment is prolonged. And meanwhile, the comprehensive influence of different soot blowing operation sequences on future long-term benefits of the system is evaluated based on state conversion of the whole system of the boiler, so that a soot blowing strategy most favorable for the overall efficiency and economical efficiency of a power plant can be found.

Description

Technical Field

[0001] This invention relates to the field of cleaning technology for the heating surfaces of coal-fired boilers, and particularly to a boiler soot blowing method, system, medium, and equipment. Background Technology

[0002] Currently, boilers are crucial equipment in the industrial sector, and their operating efficiency directly impacts a company's economic benefits. During boiler operation, ash particles generated from coal combustion inevitably adhere to heat exchange surfaces such as the furnace water-cooled walls, superheaters, reheaters, and economizers, forming ash and slag layers. The thermal conductivity of this ash layer is significantly lower than that of the metal tube wall material, typically hundreds of times lower; therefore, even slight ash accumulation can form an effective insulation layer. This phenomenon poses a serious threat to the safe and efficient operation of boilers, primarily manifested in a significant decrease in thermal efficiency, increased operational safety risks, and compromised operational stability.

[0003] To address these issues, boilers are equipped with soot blowing systems that use high-pressure steam, compressed air, or sound waves to remove ash and scale from the heated surfaces. However, existing soot blowing control technologies have several limitations:

[0004] Traditionally, sootblowers are started based on the experience and judgment of operators or fixed time procedures. This method is extremely inefficient and cannot adapt to dynamic operating conditions such as boiler load and coal type changes. The result is often either "over-purging" or "under-purging." Over-purging not only wastes a large amount of sootblowing media (which is itself a high-quality energy source) but also causes erosion and corrosion of the tube walls, shortening equipment lifespan; while under-purging fails to effectively control ash accumulation, leading to the continued existence of the aforementioned problems.

[0005] Some existing "intelligent" soot blowing systems assess the "cleanliness" of heated surfaces online by incorporating thermodynamic models and sensors (such as heat flux meters and boom weighers). When the cleanliness of a certain area falls below a preset threshold, the system automatically triggers the soot blower in that area. While these systems represent an improvement over traditional methods, they are essentially still a passive, reactive control logic. They only take action after ash accumulation has already significantly impacted thermal efficiency, exhibiting a clear lag. Summary of the Invention

[0006] Therefore, it is necessary to provide a boiler soot blowing method, system, medium, and equipment to address the aforementioned technical problems.

[0007] The present invention adopts the following technical solution: This invention provides a boiler soot blowing method, comprising: Obtain boiler operating status data for historical time periods; Based on historical boiler operating status data, predict the changes in boiler ash accumulation in subsequent target periods. The current boiler operating status data and ash accumulation changes are used as the current state space, and the combination of operating statuses of each sootblower in the boiler is used as the action space. The pre-trained reinforcement learning model determines the sootblowing action sequence of each sootblower in the boiler from the action space based on the current state space. The boiler is soot blown according to the soot blowing sequence of each soot blower.

[0008] Optionally, acquiring boiler operating status data within a historical time period specifically includes: Acquire power plant operation data and monitoring data on ash accumulation in boilers within historical time periods. Clean and normalize the power plant data and monitoring data, and use the resulting data as boiler operation status data within historical time periods. The power plant operation data represents the changes in the material and energy input entering the boiler system; the power plant operation data includes unit load, total coal feed, speed of each coal feeder, fuel analysis data, total primary and secondary air volume, oxygen content at the furnace outlet, carbon monoxide concentration, burner swivel angle, and steam inlet and outlet temperatures and flow rates of each heating surface area in the boiler. The monitoring data includes at least one of the following: heat flux density of the furnace water-cooled wall, temperature field distribution of multiple cross sections of the boiler flue gas path, and weight of the superheater and reheater tube bundle assemblies.

[0009] Optionally, predicting the ash accumulation change of the boiler in a subsequent target period based on boiler operating status data from historical periods specifically includes: Based on boiler operating status data from historical periods, a pre-trained time series model is used to predict changes in boiler ash accumulation during subsequent target periods. The time series model is a long short-term memory network model based on an attention mechanism, which is obtained by embedding an attention mechanism layer between the LSTM layer and the output layer of the long short-term memory network model.

[0010] Optionally, the operating status of each soot blower in the boiler includes: active status, blowing pressure, and blowing time; The combination of operating states of each sootblower includes the combination of operating states of sootblowers in each heated surface area of ​​the boiler.

[0011] Optionally, training the reinforcement learning model specifically includes: Based on the state space and action space at historical moments, determine the next state space to which the player will transition after executing the sootblowing action sequence selected from the action space. The reward for transitioning to the next state space after executing the sootblowing action sequence is determined by the following formula: ; in, As a reward for performing the dust-blowing action sequence, The thermal efficiency improvement value of the boiler performing soot blowing sequence. The cost of executing the soot blowing sequence, The penalty term corresponding to the blowing frequency and intensity of the blowing action sequence. , and These are the weighting coefficients; Based on the sequence of blowing dust, state transitions, and corresponding rewards, the network parameters of the reinforcement learning model are updated with the optimization objective of maximizing the long-term cumulative reward corresponding to the action selected by the reinforcement learning model from the action space.

[0012] This invention provides a boiler soot blowing system, comprising: The data acquisition module is used to acquire boiler operating status data within historical time periods; The intelligent control unit is used to predict the changes in ash accumulation in the boiler during the subsequent target period based on the boiler operating status data in the historical period. The soot blowing action determination module is used to determine the soot blowing action sequence of each soot blower in the boiler from the action space based on the current boiler operating status data and ash accumulation changes as the current state space and the combination of operating status of each soot blower in the boiler as the action space through a pre-trained reinforcement learning model. The control execution module is used to perform soot blowing on the boiler according to the soot blowing action sequence of each soot blower.

[0013] The present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described boiler soot blowing method.

[0014] The present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described boiler soot blowing method.

[0015] The above-mentioned at least one technical solution adopted in this invention can achieve the following beneficial effects: This invention predicts the changes in boiler ash accumulation in subsequent periods based on recent historical boiler operating status data. Then, based on the prediction results and boiler operating status data, a trained reinforcement learning model determines the optimal sootblowing strategy for each sootblower within the boiler. This avoids traditional trigger-based sootblowing control, enabling preventative and continuous cleaning measures before ash accumulation becomes severe, keeping the boiler in a consistently high-efficiency operating range. This significantly reduces the consumption of sootblowing media and minimizes unnecessary erosion and corrosion of boiler pipes, extending equipment lifespan. Furthermore, by evaluating the comprehensive impact of different sootblowing operation sequences on the long-term benefits of the entire boiler system through state transition assessments, it avoids suboptimal decisions caused by focusing only on the cleanliness of localized areas, thus finding the sootblowing strategy most beneficial to the overall efficiency and economy of the power plant. Attached Figure Description

[0016] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0017] Figure 1 This invention provides a schematic diagram of a boiler soot blowing method. Figure 2 A schematic diagram of a boiler soot blowing system architecture provided by the present invention; Figure 3 This is a schematic diagram of the division of the boiler heating surface area and the arrangement of sensors provided by the present invention; Figure 4 A schematic diagram of a long short-term memory network model structure based on an attention mechanism provided by the present invention; Figure 5 A schematic diagram of the online operation closed-loop process of the method of the present invention is provided for this invention; Figure 6 A schematic diagram of a computer device for implementing a boiler soot blowing method provided by the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0019] The fundamental flaw of existing technologies lies in their time limitation; they can only respond to current or recently passed conditions. They cannot answer a crucial question for optimal boiler operation: "Given that the unit is currently increasing its load and has just switched to a new batch of fuel, which superheater zone will reach a critical state of ash accumulation first in the next four hours? And, to prevent this from happening, what is the most cost-effective soot blowing sequence to be implemented now?" The ash accumulation process in a boiler is a dynamic process with a significant time delay; changes in fuel composition or combustion conditions gradually accumulate and become apparent over several hours.

[0020] Therefore, a more intelligent system must be able to model and utilize this temporal dependency. It needs to learn the complex mapping between a series of historical inputs (such as load, air volume, and coal quality) and a series of future outputs (ash accumulation rates in different regions). This directly reveals the necessity of introducing advanced time series forecasting models, such as Long Short-Term Memory (LSTM) networks, which are specifically designed to capture long-term dependencies in sequential data—something lacking in current technologies.

[0021] Based on this forward-thinking approach, this invention provides a machine learning-based intelligent soot blowing method and system for boilers. By introducing predictive models to anticipate the future, it improves upon passive response control.

[0022] The technical solutions provided by the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Figure 1 This is a schematic diagram of a boiler soot blowing method according to the present invention, which specifically includes the following steps: S101: Obtain boiler operating status data within a historical time period.

[0024] S102: Based on boiler operating status data from historical periods, predict the changes in boiler ash accumulation during subsequent target periods.

[0025] S103: Using the current boiler operating status data and ash accumulation changes as the current state space, and the combination of operating statuses of each sootblower in the boiler as the action space, the pre-trained reinforcement learning model determines the sootblowing action sequence of each sootblower in the boiler from the action space based on the current state space.

[0026] S104: Perform soot blowing on the boiler according to the soot blowing sequence of each soot blower.

[0027] For ease of explanation, the following description focuses solely on the server as the executing entity. The server mentioned in this invention can be a server set up on a business platform, or a device such as a desktop computer or laptop computer capable of executing the solution of this invention.

[0028] To address the problems of excessive energy consumption and pipeline erosion caused by unnecessary soot blowing operations, as well as the decline in boiler thermal efficiency due to untimely soot blowing, this invention provides a boiler intelligent soot blowing method with predictive, forward-looking, and global optimization capabilities. It replaces the traditional passive control logic with a forward-looking decision-making framework driven by artificial intelligence.

[0029] Specifically, the overall technical solution of this invention includes two stages. The first stage is the ash accumulation trend prediction stage. This invention utilizes a time series model, preferably an Attention-LSTM network model. This model learns the dynamic, nonlinear relationship between boiler operating status and ash accumulation rate on heated surfaces by offline training on historical multimodal operating data. During online operation, the model receives continuous real-time data streams and predicts the "ash accumulation index" of heated surfaces in various discrete areas of the boiler over a future period (e.g., the next 6 hours). This index is a quantitative indicator that can characterize the predicted ash accumulation thickness, thermal resistance, or cleanliness level.

[0030] The second stage is the soot blowing decision optimization stage. This invention utilizes a decision optimization model, preferably a deep reinforcement learning (DRL) agent. This agent takes the predicted ash accumulation index for all regions, the current real-time operating status of the boiler, and relevant economic parameters (such as steam costs and power generation revenue) as its comprehensive state input. Based on this input, the agent outputs an optimal soot blowing action sequence through its learned optimal strategy. This sequence explicitly indicates when, in which regions, and with what intensity (such as purging pressure and duration) to perform soot blowing operations. The goal of this strategy is to maximize a pre-designed long-term cumulative reward function. This reward function balances several conflicting objectives: the immediate thermal efficiency improvement brought by soot blowing (positive reward), the consumption cost of the soot blowing medium (negative reward), and the potential long-term wear and tear on the equipment (negative reward).

[0031] Figure 2 This is a schematic diagram of a boiler soot blowing system architecture according to the present invention. The intelligent boiler soot blowing system proposed in this invention mainly includes three parts: a data acquisition module (101), an intelligent control unit (102), and a control execution module (103).

[0032] The data acquisition module (101) is the system's information source. This module connects to the power plant's distributed control system (DCS) or real-time database via standard industrial communication protocols (such as OPC and Modbus) to acquire high-frequency time-series data of boiler operation. Simultaneously, this module is also responsible for integrating data from various dedicated sensors installed on the boiler body to form a comprehensive multimodal dataset reflecting the boiler's status.

[0033] The intelligent control unit (102) is the brain of this invention, typically deployed on a high-performance industrial control computer or server. It integrates three core software sub-modules:

[0034] Data preprocessing module (102a): Responsible for processing the raw data received from the data acquisition module, including data cleaning (such as handling outliers and missing values), data normalization or standardization, and feature engineering (such as constructing new meaningful features).

[0035] Ash accumulation prediction module (102b): This module embeds an offline-trained Attention-LSTM model, which is responsible for predicting the ash accumulation trend.

[0036] When making predictions, the Attention-LSTM model can dynamically assign different attention weights to different historical time steps in the input sequence, enabling it to "focus" on historical working conditions that have the greatest impact on future dust accumulation trends, thereby improving the accuracy and robustness of predictions.

[0037] Decision optimization module (102c): This module embeds a deep reinforcement learning (DRL) agent trained offline, which is responsible for generating the optimal dust-blowing strategy.

[0038] The control execution module (103) is the system's actuator. This module receives the optimal strategy from the decision optimization module and translates it into specific control instructions that the sootblower can understand (e.g., "Start sootblower number SB-SH2-L, set the purging pressure to 1.5MPa, and perform two rotational purging cycles"). These instructions are then sent to the sootblower's own programmable logic controller (PLC) system to drive the hardware to complete the specified purging actions.

[0039] To achieve accurate zone prediction and control, the complex structure of the boiler first needs to be conceptually divided. For example... Figure 3 As shown, Figure 3 This is a schematic diagram of the boiler heating surface area division and sensor arrangement according to the present invention. The boiler is divided into N monitoring zones (Zone 1, Zone 2, ..., Zone N) along the flue gas flow path. Each zone corresponds to one or a group of specific heating surfaces, such as the furnace water-cooled wall area, different levels of superheaters (SH1, SH2), reheater (RH1), and economizer (ECO), etc., and each zone is covered by one or more soot blowers.

[0040] The system's performance is highly dependent on the quality and dimensionality of the input data. By fusing macroscopic operational data from the DCS with microscopic physical measurement data from dedicated sensors, a rich, high-dimensional feature space can be constructed for the artificial intelligence model, enabling it to gain insights into the complex physicochemical processes inside the boiler.

[0041] Specifically, in one or more embodiments of the present invention, the server can obtain power plant operation data and monitoring data on ash accumulation in the boiler within a historical period, perform data cleaning and normalization on the power plant data and monitoring data, and use the obtained data as boiler operation status data within a historical period.

[0042] Among them, the power plant operation data characterizes the changes in the material and energy input entering the boiler system; the power plant operation data includes at least one of the following: unit load, total coal feed, speed of each coal feeder, fuel analysis data, total primary and secondary air volume, oxygen content at the furnace outlet, carbon monoxide concentration, burner sway angle, and steam inlet and outlet temperatures and flow rates in each heat transfer surface area of ​​the boiler; the monitoring data includes at least one of the following: heat flux density of the furnace water-cooled wall, temperature field distribution of multiple cross sections of the boiler flue gas path, and weight of the superheater and reheater tube bundle assemblies.

[0043] DCS data provides macroscopic background information on the combustion process. This data represents the material and energy input entering the boiler system and is the root cause affecting the ash accumulation process.

[0044] Dedicated sensor data provides direct physical evidence of ash accumulation. This data is a direct physical result of the combustion process and ash accumulation. Traditional systems often rely on only one type of data, such as building a thermodynamic model based on DCS data, or triggering an alarm solely based on heat flux density sensor readings. This invention fuses both types of data, enabling the artificial intelligence model to learn the complete causal chain from "cause" (DCS input) to "result" (sensor physical measurement). For example, the model can learn that a 1% change in a certain coal quality characteristic, combined with a specific air distribution method, will cause the reheater boom load cell reading to increase at a specific rate within the next 6 hours, while simultaneously causing a corresponding decrease in the heat flux density of the water-cooled wall. This profound, multi-dimensional understanding cannot be achieved with a single data modality.

[0045] To provide direct physical measurement of ash accumulation, the system integrates the following dedicated sensors: Heat flux meter: Directly installed on the tube panel of the furnace water-cooled wall, used to measure the heat flux density (unit: kW / m2) passing through the tube wall in real time. Since the ash layer is insulating, its thickness is inversely proportional to the heat flux density value. This data is the most direct indicator for judging the cleanliness of the water-cooled wall.

[0046] Acoustic pyrometers or infrared scanners: These devices can non-contactly measure the temperature field distribution along key sections of the flue gas path inside the boiler. By calculating the temperature drop before and after the flue gas flows through each area, the actual heat absorption in that area can be deduced. When the actual heat absorption deviates from the theoretical calculation based on the current operating conditions, it indicates that ash accumulation has occurred in that area.

[0047] Intelligent load cells (or weighing pins): These are mounted on the hangers of the suspended superheater and reheater tube bundles. These sensors accurately measure the total weight of the tube bundle assembly. As dust adheres and accumulates, the measured weight gradually increases, thus directly quantifying the mass of the accumulated dust.

[0048] Subsequently, the server can predict the changes in boiler ash accumulation in subsequent target periods based on boiler operating status data from historical time periods and a pre-trained time series model. In one or more embodiments of this invention, there is no limitation on the specific time series model used. For example, the time series model can be a long short-term memory network model based on an attention mechanism, obtained by embedding an attention mechanism layer between the LSTM layer and the output layer of the long short-term memory network model.

[0049] Traditional neural network models struggle to handle time-dependent sequential data. Long Short-Term Memory (LSTM) networks, a special type of recurrent neural network (RNN), utilize internal "memory units" and "gating mechanisms" to effectively capture and learn long-term dependencies in data sequences, making them ideal for simulating time-varying processes such as dust accumulation. Furthermore, attention mechanisms are introduced based on this foundation, such as... Figure 4 As shown, Figure 4 This is a schematic diagram of an Attention-LSTM network model structure based on an attention mechanism, as described in this invention. The attention mechanism further enhances the performance of the long short-term memory network model. It allows the model to dynamically assign different weights to different historical time steps in the input sequence during prediction, thereby enabling it to "focus" on the historical information most important to the current prediction, or even on the historical operating conditions that have the greatest impact on future dust accumulation trends. This improves the accuracy and robustness of predictions, which is particularly effective for handling long sequences and responding to sudden changes in operating conditions.

[0050] The model's input is a time-series matrix composed of feature vectors from the past period (e.g., the past 24 hours, sampled every minute). The model's output is a prediction vector containing the "ash accumulation index" for all N regions over a future period (e.g., the next 6 hours, one prediction per hour). This "ash accumulation index" can be a normalized cleanliness index (0 representing complete blockage, 1 representing complete cleanliness) or a predicted physical quantity, such as thermal resistance (unit: m). 2 ·K / W).

[0051] The model is trained using offline supervised learning with months or even years of historical power plant data. The "true label" during training can be calculated from historical data, such as historical cleanliness coefficients calculated based on thermodynamic equilibrium principles, or by directly using historical sensor measurements.

[0052] The dust-blowing problem is essentially a sequential decision-making problem in an uncertain environment, which perfectly aligns with the solution paradigm of reinforcement learning (RL). Deep reinforcement learning (DRL), by combining the powerful representational capabilities of deep learning with the decision-making capabilities of reinforcement learning, can learn a complex control policy that can directly map high-dimensional system states to optimal actions, thereby maximizing long-term cumulative rewards—something difficult to achieve using traditional programming methods or explicit rules.

[0053] The key components of the DRL framework are defined as follows: (1) Agent: that is, the DRL model itself.

[0054] (2) Environment: A digital twin model of the boiler system or the actual boiler system.

[0055] (3) State (S): The environmental information observed by the agent at each decision point. It is a vector containing all the information required for the current decision, such as: [Zone1 predicted ash accumulation index, ..., ZoneN predicted ash accumulation index, current unit load, fuel type indicator, ...].

[0056] (4) Action (A): The decision output by the agent based on the current state. It can be a discrete or continuous vector representing the operation command for each sootblower, indicating the subsequent operating state of each sootblower after responding to the operation command, including the activation state, purging pressure, and purging time. The combination of operating states of each sootblower includes the combination of operating states of sootblowers in different heating surface areas within the boiler. For example: [Action of sootblower 1, Action of sootblower 2, ...], where each action can be {no operation, low-pressure purging, high-pressure purging}, etc.

[0057] (5) Reward (R): This is the core signal guiding the agent's learning. The design of the reward function is crucial; it must accurately quantify the quality of each action and guide the agent to learn in a direction that aligns with the overall interests of the power plant.

[0058] In one or more embodiments of the present invention, when training a reinforcement learning model, the server can determine the next state space to which the model will transition after executing a sequence of blowing actions from each selected sootblower in the action space, based on the state space and action space at historical moments, and determine the reward for transitioning to the next state space by executing the sootblowing action sequence using the following formula: ; in, As a reward for performing the dust-blowing action sequence, The thermal efficiency improvement value for performing the soot blowing sequence on the boiler is a positive reward, representing the benefit brought by the operation; The cost of performing the soot blowing sequence (steam or other media) is a negative reward or penalty; This is a penalty term corresponding to the frequency and intensity of soot blowing in the soot blowing sequence. Too frequent or high-intensity blowing of the same area will increase this penalty to suppress the behavior of excessive flushing of the pipeline. , and These are the weighting coefficients.

[0059] By adjusting , and This allows for flexible balancing of the power plant's operational priorities at different times; for example, during peak electricity price periods, the focus can be appropriately increased. The weighting is used to pursue maximum power generation, and can be increased before equipment maintenance. The weight is used to protect the device.

[0060] Then, based on the sequence of blowing dust, state transitions, and corresponding rewards, the network parameters of the reinforcement learning model can be updated with the optimization objective of maximizing the long-term cumulative reward corresponding to the action selected by the reinforcement learning model from the action space.

[0061] DRL agents are typically trained in offline simulation environments built on historical data. This allows them to explore millions of different soot-blowing strategies without compromising actual production safety, learn the optimal control logic, and then deploy them to online systems for fine-tuning and application.

[0062] Specifically, the training process of a reinforcement learning model is an iterative optimization process. When an agent interacts with its environment, it generates behaviors such as... empirical data, among which This is the current state. For the action to be performed, The immediate reward calculated based on the above reward function. This is the state for the next time step. This empirical data is stored in the empirical replay pool. During training, batch data is sampled from the empirical replay pool, and deep reinforcement learning algorithms (such as Deep Q-Network, Proximal Policy Optimization (PPO), or Deep Deterministic Policy Gradient (DDPG)) are used to update the model's network parameters.

[0063] For example, if a value-based DQN algorithm is used, it minimizes the predicted Q-value and the target Q-value (based on immediate reward). The mean squared error loss function between the reward signal and the maximum Q-value of the next state is used to update the network, enabling the model to more accurately assess the long-term value of taking a specific blowing action in a given state. If a policy-based PPO algorithm is used, the policy network is updated directly along the gradient direction based on the reward signal to increase the probability of high-reward actions. By repeatedly performing the above "interaction-sampling-update" process, the model eventually converges to the optimal blowing policy.

[0064] The online operation of this invention follows a closed-loop "perception-prediction-decision-execution-feedback" process, such as... Figure 5 As shown, the specific steps include: (Step 201) Continuous data acquisition: The system continuously acquires real-time data from the DCS and various sensors.

[0065] (Step 202) Data preprocessing: Cleaning, normalizing and other operations are performed on the collected data.

[0066] (Step 203) Rolling prediction: Every preset time period T1 (e.g., 15 minutes), the ash accumulation prediction module (102b) uses the latest historical data to perform a rolling update prediction of the ash accumulation index of each area in the next T hours.

[0067] (Step 204) Decision making: Every preset time period T2 (e.g., 30 minutes), the decision optimization module (102c) obtains the latest predicted ash accumulation index and the real-time operating condition of the boiler as the current state input.

[0068] (Step 205) Strategy output and instruction issuance: The strategy network of the DRL agent outputs an optimal soot blowing action sequence based on the current state, and the control execution module (103) translates the action sequence into PLC instructions and issues them to the soot blowing system.

[0069] (Step 206) Perform soot blowing operation: The soot blower hardware performs physical blowing according to the instructions.

[0070] (Step 207) Feedback and Learning: After the system performs soot blowing, it will continuously monitor changes in the boiler status (such as heat flux density recovery and steam temperature changes) and calculate the actual reward value obtained for the action. This new data (state-action-reward-new state) will be used for online fine-tuning and continuous learning of the model, enabling the system to continuously adapt and improve its performance.

[0071] Working principle: This invention is a closed-loop, data-driven intelligent optimization process, which can be summarized as "perception-prediction-decision-execution-feedback".

[0072] First, in the sensing phase, the data acquisition module (101) acts as the system's "sensors," continuously collecting massive amounts of real-time operational data from the DCS and dedicated sensors to comprehensively capture the boiler's current state and dynamic changes. Next, the prediction phase begins. These raw data streams are sent to the data preprocessing module (102a) for cleaning and standardization.

[0073] The data is then input into the ash accumulation prediction module (102b). The Attention-LSTM model within this module, through its powerful time-series data analysis capabilities, analyzes past operating condition changes (such as load increases and coal quality fluctuations), and, like an experienced expert, predicts the potential development trend of ash accumulation in various heating surfaces of the boiler over the next few hours, quantifying it as an "ash accumulation index." This step represents a leap from "passive response" to "proactive prediction."

[0074] Next, the system enters the core decision-making phase. The decision optimization module (102c) receives future ash accumulation trend information from the prediction module and combines it with the boiler's current real-time operating conditions. The DRL agent within the module acts as the "optimal scheduler." In a virtual environment, it repeatedly weighs the pros and cons of each possible combination of soot blowing operations through its built-in reward function, that is, the balance between the thermal efficiency improvement (benefit) brought by soot blowing and the steam cost and equipment wear (cost). Ultimately, it outputs a soot blowing sequence that can obtain the maximum overall benefit in the future, which precisely answers the questions of "when to blow, where to blow, and how to blow."

[0075] Finally, there is the execution and feedback phase. The control execution module (103) translates the optimal strategy output by the DRL agent into specific, executable hardware instructions and sends them to the PLC system of the sootblower for execution. After the sootblowing action is completed, the data acquisition module (101) continues to monitor changes in the boiler status (e.g., whether the heat flux density has increased or the flue gas temperature has decreased). These changes serve as feedback information, verifying the effectiveness of the decision on the one hand, and serving as new data samples for the model's continuous online learning and self-optimization on the other, thereby enabling the system to continuously adapt to changes in boiler operating conditions and achieve continuous performance improvement. The entire process forms a complete intelligent closed-loop control system.

[0076] based on Figure 1 The boiler soot blowing method described herein can predict soot accumulation problems before they become severe, allowing for preventative and timely cleaning measures to maintain the boiler in a consistently efficient operating range. By accurately identifying areas requiring cleaning ("when and where"), this invention significantly reduces the consumption of soot blowing media (steam or air) and minimizes unnecessary scouring and corrosion of pipes in clean areas, extending equipment lifespan. The deep reinforcement learning agent can make decisions from the perspective of the entire boiler system, evaluating the comprehensive impact of different soot blowing operation sequences on the system's long-term future benefits. This avoids suboptimal decisions caused by focusing only on the cleanliness of local areas, thus finding the soot blowing strategy most beneficial to the overall efficiency and economy of the power plant. The machine learning model of this invention can continuously learn from online data to self-optimize and adjust, automatically adapting to changes in fuel characteristics, natural degradation of equipment performance, and changes in operating modes, without the need for complex and costly manual recalibration required by methods based on physical models.

[0077] When applying the boiler soot blowing method provided by this invention, it is not necessary to follow the... Figure 1 The steps shown are executed in sequence. The specific execution order of each step can be determined as needed, and this invention does not impose any restrictions on it.

[0078] Specific limitations regarding the boiler soot blowing system can be found in the limitations of the boiler soot blowing method described above, and will not be repeated here. Each module in the aforementioned boiler soot blowing system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0079] The present invention also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described... Figure 1 The provided boiler soot blowing method.

[0080] The present invention also provides Figure 6 The schematic diagram of the computer device shown is as follows: Figure 6 As shown, at the hardware level, this computer device includes a processor, internal bus, network interface, memory, and non-volatile memory, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then executes it to achieve the above. Figure 1 The provided boiler soot blowing method.

[0081] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0082] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this invention.

Claims

1. A method of boiler soot blowing, characterized in that, The method comprises the following steps: acquiring boiler operation state data in a historical period; predicting ash deposition variation of the boiler in a subsequent target period according to the boiler operation state data in the historical period; taking the current boiler operation state data and the ash deposition variation as a current state space, taking a running state combination of each soot blower in the boiler as an action space, and determining a soot blowing action sequence of each soot blower in the boiler from the action space according to the current state space through a pre-trained reinforcement learning model; blowing ash according to the soot blowing action sequence of each soot blower.

2. The method of claim 1, wherein, The method of acquiring the boiler operation state data in the historical period comprises the following steps: acquiring power plant operation data and monitoring data of ash deposition accumulation in the historical period, performing data cleaning and normalization on the power plant data and the monitoring data, and taking the obtained data as the boiler operation state data in the historical period; wherein the power plant operation data represents changes in the input of substances and energy into the boiler system; the power plant operation data comprises at least one of unit load, total coal supply, speed of each coal feeder, fuel analysis data, total primary air and secondary air, oxygen content at the furnace outlet, carbon monoxide concentration, burner swing angle, and steam inlet and outlet temperature and flow of each heating surface region in the boiler; the monitoring data comprises at least one of heat flux density of the furnace water-cooled wall, temperature field distribution of multiple cross sections of the boiler flue gas path, and weight of the superheater and reheater tube bundle assembly.

3. The method of claim 1, wherein, The method of predicting the ash deposition variation of the boiler in the subsequent target period according to the boiler operation state data in the historical period comprises the following steps: predicting the ash deposition variation of the boiler in the subsequent target period through a pre-trained time series model according to the boiler operation state data in the historical period; wherein the time series model is a long short-term memory network model based on an attention mechanism, which is obtained by embedding an attention mechanism layer between an LSTM layer and an output layer of the long short-term memory network model.

4. The method of claim 1, wherein, The running state of each soot blower in the boiler comprises an enabled state, a blowing pressure, and a blowing time; the running state combination of each soot blower comprises a running state combination of soot blowers in each heating surface region in the boiler.

5. The method of claim 1, wherein, The method of training the reinforcement learning model comprises the following steps: determining a next state space converted after executing a soot blowing action sequence of each soot blower selected from the action space according to the state space and the action space at a historical moment, and determining a reward of converting to the next state space after executing the soot blowing action sequence by the following formula: ； wherein, a reward for performing the sequence of soot blowing actions, a heat efficiency improvement value for the boiler for performing the sequence of soot blowing actions, a cost consumption for performing the sequence of soot blowing actions, a penalty term corresponding to a soot blowing frequency and intensity of the sequence of soot blowing actions, , and are weight coefficients; updating network parameters of the reinforcement learning model according to the executed soot blowing action sequence, state conversion, and corresponding reward, with the optimization goal of maximizing a long-term cumulative reward corresponding to an action selected from the action space by the reinforcement learning model.

6. A sootblowing system for a boiler, characterized in that The method comprises the following steps: a data acquisition module configured to acquire boiler operation state data in a historical period; The intelligent control unit is configured to predict the ash deposition variation of the boiler in a subsequent target period according to the boiler operation state data in a historical period; the soot blowing action determination module is configured to take the current boiler operation state data and the ash deposition variation as a current state space, take the operation state combination of each soot blower in the boiler as an action space, and determine the soot blowing action sequence of each soot blower in the boiler from the action space according to the current state space by using the pre-trained reinforcement learning model; The control execution module is configured to perform soot blowing on the boiler according to the soot blowing action sequence of each soot blower.

7. The sootblowing system of claim 6, wherein Further comprising: at least one of a heat flux density meter, an acoustic pyrometer or an infrared scanner, and a weighing sensor; The heat flux density meter is arranged on a tube panel of a furnace water wall and is configured to monitor the heat flux density of the furnace water wall in real time. The acoustic pyrometer or the infrared scanner is arranged on multiple sections of a boiler flue gas path in a non-contact manner and is configured to monitor the temperature field distribution of the multiple sections of the boiler flue gas path in real time. The weighing sensor is arranged on a hanger of a suspension type superheater and reheater tube bundle and is configured to monitor the weight of a superheater and reheater tube bundle assembly in real time.

8. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, and the computer program is executed by the processor to implement the method according to any one of claims 1-5.

9. A computer device, comprising: The computer program is stored in the memory and can be run on the processor, and the processor implements the method according to any one of claims 1-5 when executing the computer program.

Images