Combustion optimization method and system of coal-fired boiler, electronic equipment and storage medium

Abstract

The invention provides a combustion optimization method and system for a coal-fired boiler, electronic equipment and a storage medium. The combustion optimization method for the coal-fired boiler comprises the steps that historical operation data and operation condition data of the boiler are obtained; determining the combustion efficiency of the boiler according to the historical operation data of the boiler; working condition division is carried out according to the boiler combustion efficiency and the operation working condition data, and information of all working condition intervals is obtained; constructing a combustion efficiency prediction model driven by time series data according to the information of each working condition interval; the combustion efficiency prediction model is embedded into a deep reinforcement learning method framework for training, and an optimization model is generated; and inputting real-time operation data of the boiler into the optimization model to generate optimization decision data.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of coal-fired power plant boiler control, and in particular to a coal-fired boiler combustion optimization method and system, an electronic device, and a storage medium. BACKGROUND

[0002] As the core energy pillar supporting the operation of modern social economy, coal-fired power generation plays an irreplaceable fundamental role in ensuring the safe and stable operation of the power system. Although the installed capacity of new energy power generation represented by wind power and photovoltaic power continues to grow rapidly, there is significant instability and intermittency in its output characteristics. This strong random power characteristic brings serious challenges to real-time balancing and precise scheduling of the power grid. It is worth noting that in the process of frequent participation in deep peak shaving, the mismatch of combustion dynamics parameters and the attenuation of heat transfer system efficiency caused by variable load conditions result in a loss of thermal efficiency of a typical 350MW subcritical unit of up to 12%. The waste of standard coal caused by a single unit per year is as high as 18,200 tons, which is equivalent to the emission of 47,600 tons of carbon dioxide.

[0003] The existing coal-fired boiler combustion optimization method has significant noise in the boiler operating parameters collected by the DCS system (Distributed Control System) and missing data in some measurement points, resulting in large output error fluctuations (±1.8%) of the data-driven combustion prediction model, which affects the stability of the optimization control. SUMMARY

[0004] The embodiments of the present application provide a coal-fired boiler combustion optimization method, system, electronic device, and storage medium to solve the problems of related technologies. The technical solutions are as follows: In a first aspect, the embodiments of the present application provide a coal-fired boiler combustion optimization method, comprising: obtaining historical operating data and operating condition data of a boiler; determining a boiler combustion efficiency according to the historical operating data of the boiler; dividing operating conditions according to the boiler combustion efficiency and the operating condition data to obtain each operating condition interval information; constructing a time series data-driven combustion efficiency prediction model according to the each operating condition interval information; embedding the combustion efficiency prediction model into a deep reinforcement learning method framework for training to generate an optimization model; inputting real-time operating data of the boiler into the optimization model to generate optimization decision data.

[0005] In a second aspect, the embodiments of the present application provide a coal-fired boiler combustion optimization system, comprising: The acquisition module is configured to acquire historical operation data and operation condition data of the boiler. The determination module is configured to determine a boiler combustion efficiency according to the historical operation data of the boiler. The obtaining module is configured to divide conditions according to the boiler combustion efficiency and the operation condition data to obtain condition interval information. The construction module is configured to construct a time-series data-driven combustion efficiency prediction model according to the condition interval information. The generation module is configured to embed the combustion efficiency prediction model into a deep reinforcement learning method framework to train an optimization model. The input module is configured to input real-time operation data of the boiler into the optimization model to generate optimization decision data.

[0006] In a third aspect, an electronic device is provided, which includes at least one processor and a memory communicatively connected to the at least one processor. The memory stores instructions executable by the at least one processor, so that the at least one processor can execute the combustion optimization method of the coal-fired boiler.

[0007] In a fourth aspect, a computer-readable storage medium is provided, which stores computer instructions. When the computer instructions are executed on a computer, the method in any of the embodiments is executed.

[0008] The advantages or beneficial effects of the above technical solutions include at least the following: In the combustion optimization method, the historical operation data and operation condition data of the boiler are used to obtain condition interval information, so that the state intervals of stable conditions and unstable conditions in the conditions can be identified. The combustion efficiency prediction model is constructed according to the identified condition interval information. The deep reinforcement learning method framework is trained offline in the virtual environment of the combustion prediction model to generate an optimization model. The optimization model is deployed in the real-time data stream processing system of the boiler. Real-time operation data is input into the optimization model to generate optimization decision data. The boiler is optimized by the optimization decision data. The coal consumption is effectively reduced and the combustion thermal efficiency is improved. The online dynamic optimization control of the coal-fired boiler is realized. The thermal efficiency of the unit is improved, and the coal consumption is significantly reduced. The variables in the boiler operation process are analyzed comprehensively. The problem that the output error of the data-driven combustion prediction model fluctuates greatly due to the significant noise in the boiler operation parameters and the missing data of some measurement points, which affects the stability of the optimization control, is effectively solved.

[0009] The above summary is intended to illustrate only and is not intended to be limiting in any way. Further aspects, implementations, and features of the present application will be apparent to those skilled in the art from the description of the illustrative aspects, implementations, and features described above in conjunction with the accompanying drawings and the detailed description below. BRIEF DESCRIPTION OF DRAWINGS

[0010] In the drawings, like reference numerals refer to same or similar components throughout the several views. The drawings are not necessarily to scale. It should be understood that the drawings only depict some embodiments in accordance with the disclosure and should not be considered limiting of the scope of the disclosure.

[0011] Figure 1 Flow chart of a method for combustion optimization of a coal-fired boiler according to an embodiment of the present application.

[0012] Figure 2 Block diagram of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0013] In the following, only some exemplary embodiments are described in brief. As will be appreciated by those skilled in the art, the described embodiments can be modified in various different ways without departing from the spirit or scope of the present application. Therefore, the drawings and description are to be regarded as being illustrative in nature rather than restrictive.

[0014] In the related art, coal-fired power generation plays an irreplaceable basic role in ensuring the safe and stable operation of the power system. Although the installed capacity of new energy power generation represented by wind power and photovoltaic continues to grow rapidly, there is significant instability and intermittency in its output characteristics. This strong random power supply characteristic brings serious challenges to real-time balance and accurate scheduling of the power grid. According to the 2023 Annual Report on the Electric Power Industry, coal-fired power generation still accounts for an absolute proportion of 65.2% of the total power generation in China. Not only does it continue to play a key role in basic load power in the power supply structure, but also provides more than 70% of the peak shaving capacity for the power grid due to its rapid response capability. It is worth noting that in the process of frequent participation in deep peak shaving, the mismatch of combustion dynamics parameters and the attenuation of heat transfer system efficiency caused by variable load conditions result in a loss of thermal efficiency of a typical 350 MW subcritical unit of up to 12%-15%. The waste of standard coal caused by a single unit per year is as high as 18.25 million tons, equivalent to the emission of 47.66 million tons of carbon dioxide.

[0015] When the unit heat value of coal decreases, more coal is needed to maintain the same boiler load, and the primary air volume required for conveying the pulverized coal increases, leading to increased power consumption of the coal mill and increased resistance of the coal pulverizing system. This in turn requires the fan to provide a higher pressure head, increasing the power consumption of the fan and ultimately pushing up the plant power consumption rate. At the same time, the decrease in coal heat value will lower the theoretical combustion temperature and the furnace temperature level, causing the coal powder airflow to delay ignition and the combustion stability to decrease, affecting the burnout rate; in addition, it will also lead to an increase in the flue gas temperature and an increase in the flue gas heat loss.

[0016] As the initial link in the energy conversion chain of coal-fired power plants, the thermal efficiency of the boiler directly determines the overall system energy efficiency. According to statistics, the typical coal-fired boiler has a comprehensive thermal efficiency loss of 6-12 percentage points due to factors such as flue gas heat loss (accounting for about 5%-8%), solid unburned carbon loss (accounting for about 1%-3%), and heat loss (accounting for about 0.5%-1%). Through combustion optimization control technology (such as dynamic adjustment of air-coal ratio, oxygen content control of economizer flue gas, etc.), coal consumption can be reduced by 1%-5% (corresponding to about 30-40 thousand tons of coal per year for a 350MW unit), and the boiler thermal efficiency can be improved by 2.5-4.2 percentage points.

[0017] When calculating the combustion efficiency based on the mechanism model of thermodynamic laws, the mass / energy conservation equations, combustion reaction kinetics equations, and heat transfer differential equations need to be solved simultaneously, involving multi-physical field coupling solution, resulting in a long time of up to 23 minutes for a single efficiency evaluation, which is difficult to meet the real-time optimization demand of minutes. More seriously, the mechanism model has inherent defects in modeling dynamic processes: the deviation of dynamic heat transfer coefficient caused by the load regulation rate in the actual operation of the coal-fired unit can be as high as 12.18%, resulting in an efficiency prediction error of more than 23 percentage points.

[0018] Secondly, at the data-driven optimization level, the existing combustion optimization method faces a double data dilemma: 1. Data quality defects: Due to the significant noise in the boiler operation parameters collected by the DCS system and the missing data of some measurement points, the output error of the data-driven combustion prediction model fluctuates greatly (±1.8%), affecting the stability of the optimization control; 2. Insufficient coverage of working conditions: In the historical operation data, the proportion of deep peak shaving working conditions with a load rate of less than 40% is only 5.2%, and the lack of samples leads to the instability of the control command of the optimization strategy under extreme working conditions (such as the fluctuation amplitude of the coal supply amount command exceeding ±15%).

[0019] Therefore, the technical personnel of the present application are committed to developing a method system suitable for combustion optimization of different capacity coal-fired boilers, using the data of the boiler operation of the coal-fired power plant to realize an intelligent combustion optimization system from predicting the combustion efficiency to optimizing the combustion process.

[0020] Figure 1 A flow chart of a combustion optimization method of a coal-fired boiler according to an embodiment of the present application is shown. As shown in the figure, a combustion optimization method of a coal-fired boiler includes: Figure 1 S110: obtaining historical operation data and operation condition data of the boiler; S120: determining a combustion efficiency of the boiler according to the historical operation data of the boiler; S130: performing condition division according to the combustion efficiency of the boiler and the operation condition data to obtain each condition interval information; S140: constructing a time-series data-driven combustion efficiency prediction model according to the each condition interval information; S150: embedding the combustion efficiency prediction model into a deep reinforcement learning method framework to perform training and generate an optimization model; S160: inputting real-time operation data of the boiler into the optimization model to generate optimization decision data.

[0021] In the combustion optimization method of the present embodiment, the historical operation data and operation condition data of the boiler are used to obtain each condition interval information, so that the state intervals of stable conditions and unstable conditions in the conditions can be identified, and then a combustion efficiency prediction model is constructed using the identified each condition interval information. The deep reinforcement learning method framework is trained offline in the virtual environment of the combustion prediction model to generate an optimization model. The optimization model is deployed in the real-time data stream processing system of the boiler, and real-time operation data is input into the optimization model to generate optimization decision data. The boiler is optimized by using the optimization decision data, which can effectively reduce the coal consumption and improve the combustion thermal efficiency, thereby achieving online dynamic optimization control of the coal-fired boiler, improving the thermal efficiency of the unit while significantly reducing the coal consumption, and enabling comprehensive analysis of each variable in the operation process of the boiler, effectively solving the problem that the output error of the data-driven combustion prediction model fluctuates greatly due to significant noise in the operation parameters of the boiler and missing data in some measurement points, thereby affecting the stability of the optimization control.

[0022] In step S110, the historical operation data and operation condition data of the boiler are obtained. In the embodiments of the present application, temperature, pressure, gas composition, and flow sensors are installed in key areas of the boiler combustion (such as burners, flues, air pipes, water pipes, etc.), and the corresponding temperature, pressure, gas composition, and flow data collected by the sensors are input into the distributed control system (DCS) for monitoring combustion state parameters (such as furnace temperature distribution, flue gas oxygen content, air-coal ratio, etc.), thereby obtaining the historical operation data and operation condition data of the boiler.

[0023] ​According to actual demand and coal bunker fuel consumption, the data acquisition frequency can be set to 10 seconds on the distributed control system DCS to ensure the timeliness and accuracy of the data.

[0024] The historical operation data of the boiler monitored by the distributed control system DCS is used to calculate the boiler thermal efficiency, and the real data that does not exist directly is obtained by recording and differentiating, decomposing, correlation analyzing and conventional arithmetic operations according to the thermal efficiency and other characteristics.

[0025] The operation condition data includes combustion heat value, unit load and cooling medium temperature and other related parameter data, which can be directly obtained on the distributed control system DCS or calculated by the data measured on the distributed control system DCS.

[0026] In step S120, the boiler combustion efficiency is determined according to the historical operation data of the boiler. In the embodiment of the present application, the historical operation data of the boiler monitored by the distributed control system DCS is used to calculate each heat loss (flue gas, incomplete combustion of gas, incomplete combustion of solid, heat dissipation) by using the counterbalance method, and then the combustion efficiency is calculated.

[0027] In the process of boiler operation, the higher the combustion efficiency is, the higher the fuel utilization rate is, and the smaller the energy waste is.

[0028] The DCS (Distributed Control System) will collect and store the boiler operation parameters in real time, such as fuel consumption, flue gas composition, boiler steam output, temperature at each place, etc. By using these historical data, the combustion efficiency can be calculated by the counterbalance method (inverse algorithm).

[0029] The counterbalance method is to calculate each heat loss of the boiler, and then subtract the total proportion of these losses from 100% to obtain the combustion efficiency, so it is called "counterbalance" (not directly calculating the effective utilization, but inversely calculating from the loss).

[0030] At the same time, the historical thermal efficiency values of each period and their corresponding main parameters, such as coal heat value and unit load, are recorded and saved.

[0031] In step S130, the operation condition is divided in combination with the start-stop state of the coal mill, the unit load and the boiler combustion efficiency and other multi-dimensional operation characteristics to accurately obtain the information of each operation condition interval. In this embodiment, the operating condition data covers key parameter data such as coal mill start-stop state, unit load, combustion heat value, cooling medium temperature, etc. Based on the combination mode of coal mill start-stop, unit load level, and in cooperation with factors such as coal heat value and combustion efficiency, the operating condition division work is carried out. Through this multi-parameter cooperative division method, the operating condition interval can be carefully divided into stable operating condition interval and non-stable operating condition interval. The information of each operating condition interval can be characterized and embodied by various historical data generated during the operation of the boiler in the corresponding operating condition interval.

[0032] Change the past single reliance on "load" to divide the rough mode of boiler operation state, realize the fine, dynamic and automatic identification of the boiler operating condition by fusing the coal characteristics and key operating parameters, so as to provide decision basis for the precise control and optimized operation of the unit.

[0033] The offline modeling stage uses historical operation data to complete the construction of the model and the definition of the operating condition. In the online identification stage, the real-time collected data is used to automatically match the current unit operating state to the defined operating condition.

[0034] A quantitative mapping model of coal industry analysis parameters and combustion efficiency is established; a mathematical model is created, which can accurately predict the combustion efficiency of the boiler according to the input coal quality parameters (heat value, volatile matter, ash melting characteristics, etc.). Combustion efficiency is a core indicator, and combustion efficiency is one of the most critical performance indicators for evaluating the good or bad operation of a boiler, but it is usually difficult to measure directly, in real time and accurately online. Coal quality is a key disturbance factor, and the fuel coal of a coal-fired power plant has varying characteristics (heat value, flammability, and easy coking, etc.), which is the most important external disturbance affecting combustion efficiency. By collecting a large amount of historical data of coal industry analysis parameters, each data contains a complete set of coal quality test reports (heat value, volatile matter, ash content, moisture content, ash melting point, etc.) and the corresponding actual boiler combustion efficiency (usually calculated by heat balance).

[0035] A variety of machine learning or statistical methods can be used to establish this mapping relationship, such as: multiple linear regression (Multiple Linear Regression), support vector machine (Support Vector Machines, SVM), neural network (Neural Networks) or gradient boosting tree (Gradient Boosting Trees) etc. Use the above-mentioned large amount of historical data of coal industry analysis parameters, and use independent test data set to verify the accuracy of the model. The final obtained model can be formalized as: combustion efficiency = f(heat value, volatile matter, ash melting characteristics,...); generate a trained combustion efficiency prediction model.

[0036] Constructing the working condition characteristic vector; the real-time coal calorific value is the core of energy input, directly determines the amount of fuel and air required to produce the same amount of heat, and is the fundamental influence on the combustion process. The unit load is the demand of energy output, which determines the overall operation rhythm and parameter setting range of the boiler. The cooling medium temperature (such as the circulating water inlet temperature) is an important boundary condition, which affects the vacuum degree of the steam turbine, and then affects the efficiency of the whole thermal cycle. Under the same power generation load, different cooling water temperatures require different boiler evaporation capacities.

[0037] Vector construction: at any time t, the characteristic vector V(t) of the working condition can be represented as: V(t) = [real-time coal calorific value (t), unit load (t), cooling medium temperature (t)]; The real-time coal calorific value comes from the online coal quality analyzer, or is estimated by the weighted average of the coal entering the furnace. The load and temperature are directly obtained from the DCS (Distributed Control System), which can generate a multi-dimensional data set that comprehensively and quantitatively describes the real-time operating state of the boiler. With the real-time coal calorific value as the primary division basis, the working condition characteristic vector is constructed in coordination with the unit load and cooling medium temperature.

[0038] Fine segmentation of working conditions is achieved using an improved density clustering algorithm; using unsupervised learning algorithms, the inherent clustering patterns in the above-mentioned massive historical feature vector data are automatically and objectively identified, and these clustered "data clouds" are recognized as a stable and repeatedly appearing typical working condition interval.

[0039] Traditional manual division of working conditions (such as high, medium, and low load) is subjective and has a very coarse granularity. Clustering algorithms can divide based on the distribution characteristics of the data itself, which is more scientific and fine. The improved density clustering algorithm may discover some specific working condition combinations that are ignored by humans but actually appear frequently (for example, "medium load + poor quality coal + high summer temperature" may be an independent working condition that needs special attention).

[0040] The improved density clustering algorithm can be used to select the clustering algorithm. The improved density clustering algorithm refers to DBSCAN (Density-Based Spatial Clustering of Applications with Noise) and its variants.

[0041] In this embodiment, the improved density clustering algorithm can be used without pre-setting the number of categories. Unlike algorithms such as K-Means, DBSCAN does not need to specify in advance how many working conditions to divide into, and the improved density clustering algorithm can automatically find the number of working conditions according to the data density. The improved density clustering algorithm can identify clusters of any shape. The real working condition may not be circular or spherical in the feature space, and DBSCAN can well handle irregularly shaped working condition distribution. The improved density clustering algorithm can handle noise / abnormal points. Those transition states or abnormal data points that do not belong to any stable working condition will be identified as noise, which is very consistent with the actual industry. The improved density clustering algorithm optimizes the problem of sensitivity to parameters of DBSCAN, such as adaptive parameter selection algorithm, or HDBSCAN that can handle multi-density data sets.

[0042] The obtained large number of historical feature vectors V(t) are input into the improved density clustering algorithm. The improved density clustering algorithm outputs the category (cluster) label to which each data point belongs.

[0043] Each identified cluster represents a "refined working condition" defined by data driving. For example: working condition A: [high calorific value coal, high load, winter low temperature]; working condition B: [high calorific value coal, high load, summer high temperature]; working condition C: [low calorific value coal, medium load,...]; and finally generate working condition interval information.

[0044] In step S140, based on the working condition interval information, a time series data driven combustion efficiency prediction model is constructed using time series modeling techniques such as long short-term memory network (LSTM).

[0045] In the embodiments of the present application, first, for each divided working condition interval, the historical time series data of the boiler operation in the interval is extracted, covering the time series of key process parameters such as fuel supply amount, air distribution, furnace temperature, flue gas composition. Then, these time series data are input into the LSTM network, and the model is trained to learn the dynamic change law of each operating parameter with time and the correlation between the operating parameter and the combustion efficiency under different working conditions. The LSTM network can effectively mine the time series characteristics of parameter changes in the boiler combustion process due to its ability to capture long-time sequence dependencies, thereby achieving accurate prediction of the combustion efficiency and providing data support and prediction basis for subsequent combustion optimization control.

[0046] The boiler operation history data is divided into several intervals (such as high load, low load, different coal quality combinations, etc.) according to different working conditions. Each working condition interval has corresponding historical operation data, including: process variables (state variables): real-time collected temperature, pressure, Concentration, coal fineness, primary air volume, secondary air volume, etc. (25-dimensional state variables); and decision variables (control variables): fuel distribution, damper opening, superheater water injection, online number of coal mills, etc. (11-dimensional decision variables) Based on the time sequence dependence between these process variables and decision variables, an initial model is constructed to predict the combustion efficiency.

[0047] The current combustion efficiency is not only affected by the current state and operation, but also by the control actions and state evolution of the previous few minutes (or tens of seconds). Training the model with static samples alone ignores this process inertia and dynamic response lag, resulting in poor prediction results.

[0048] Establishing a time series dataset not only records the state and control variables at a single time point, but also records the change sequence within a time window (e.g. past 60 seconds, 120 seconds).

[0049] LSTM (Long Short-Term Memory) is a specially designed recurrent neural network (RNN) architecture that can remember important information over a long time span and ignore irrelevant noise: its long-term memory can perceive the cumulative impact of coal quality changes or combustion state changes on future minutes; its short-term response can perceive the efficiency changes of instantaneous adjustments of boiler operation parameters; LSTM selectively updates information through a gating mechanism (input gate, forget gate, output gate), making it very suitable for industrial time series data such as boilers, which are nonlinear, strongly lagging, and have many interference factors.

[0050] Specifically, within each divided operating condition interval, historical boiler operation data in each operating condition interval information is used to construct a time series dataset containing state variable and control variable sequences, LSTM is used to learn the time sequence dependence between process variables and decision variables, and a nonlinear dynamic mapping relationship between input and combustion efficiency is fitted to obtain preliminary combustion efficiency prediction values for each operating condition interval, providing a mean value benchmark for subsequent uncertainty modeling and optimization control strategies.

[0051] The preliminary combustion efficiency prediction value does not indicate uncertainty, i.e. how much fluctuation range and risk distribution the prediction may have; in industrial practice, there are unpredictable factors such as coal quality fluctuations, measurement noise, and environmental changes that can affect efficiency; if a single prediction value is used directly to guide the reinforcement learning optimization strategy, once the prediction error is large, the strategy may produce unsafe or suboptimal decisions.

[0052] The diffusion model is essentially a probabilistic generative model, which has the advantages of: learning the complete distribution of the data (not just the mean), being able to generate any quantile or sample set to characterize the uncertainty of the model output, and being more flexible and robust in modeling complex multi-modal error distributions than the Gaussian assumption of Bayesian methods. When training the preliminary LSTM model, the difference between the preliminary combustion efficiency prediction value and the historical actual combustion efficiency value, i.e. the prediction residual. These residuals have different distribution characteristics under different operating conditions, some have small fluctuations in the interval, and some have large fluctuations in the interval.

[0053] The historical residual data is classified according to the operating condition interval and sent to the diffusion model for training. The diffusion model gradually learns the change pattern and distribution form of the residual, such as the upper and lower boundaries of the residual and which conditions are prone to large deviations. When new real-time boiler state and control decisions are input, first use LSTM to get the preliminary combustion efficiency prediction value. Input the preliminary combustion efficiency prediction value into the diffusion model corresponding to the operating condition interval, and let it sample the possible residual multiple times to generate a batch of potential true values with slight differences. By counting this batch of potential true values, the mean update prediction center, maximum and minimum value, and specific quantile can be obtained. For example, the middle 90% of the data can be taken as the confidence interval of the prediction, and different quantile points can be taken to represent the range of different confidence levels. Each operating condition interval has its own diffusion model, so when predicting, the corresponding residual distribution characteristics can be selected according to the operating condition to get a more accurate confidence interval. Finally, a combustion efficiency prediction model for each operating condition is constructed, and the output of each combustion efficiency prediction model is not only a number, but also a reliable range with upper and lower boundaries.

[0054] Based on the preliminary prediction value of LSTM, a diffusion model is connected to each operating condition interval to generate a large number of efficiency prediction samples, extract quantiles, mean and variance, and then establish a probability prediction model that fully reflects the dynamic characteristics of the boiler, providing a solid mathematical foundation for intelligent optimization and risk management. By introducing a diffusion generation model to model and sample the prediction residual based on the preliminary efficiency prediction model in each operating condition interval, the prediction value confidence interval containing the mean, variance and each quantile is output, thereby expanding the original deterministic prediction into a sub-condition probability prediction model that can reflect the prediction uncertainty, providing a more robust reward evaluation basis for subsequent reinforcement learning optimization control.

[0055] To maintain the prediction accuracy, set a rolling update of the combustion efficiency prediction model based on the newly added historical boiler operation data every 6 hours, so that the combustion efficiency prediction model can be optimized and updated, improving the accuracy of the combustion efficiency prediction model.

[0056] In step S150, the combustion efficiency prediction model is embedded into a deep reinforcement learning method framework, a virtual simulation environment is constructed to carry out offline training, and an optimization model is generated.

[0057] In this embodiment, firstly, a state transition model and a composite reward function are obtained. The state transition model is used to depict the evolution rule of the subsequent state under the action of the current state (such as fuel quantity, air quantity and the like) and the control action (such as coal supply quantity adjustment, air door opening degree adjustment and the like) in the process of boiler operation; the composite reward function comprehensively considers the multi-objective such as combustion efficiency improvement and operation stability, and provides feedback basis for reinforcement learning training. Subsequently, in the virtual environment driven by the combustion efficiency prediction model, the deep reinforcement learning method framework algorithm is trained offline according to the state transition model and the composite reward function. This offline training process does not need to rely on the online operation of the real boiler, and through simulating a large number of boiler operation scenes in the virtual environment, the reinforcement learning intelligent agent continuously explores and learns the optimal combustion control strategy, and finally generates an optimization model capable of guiding the actual combustion optimization.

[0058] By modeling the combustion efficiency based on the historical boiler operation data, a working condition probability prediction model (LSTM model + diffusion model) is obtained. Based on each working condition probability prediction model, a virtual environment is constructed as a simulation environment, which can predict the future state according to the given control decision and give the combustion efficiency and confidence interval. A deep reinforcement learning (DRL) algorithm (such as DDPG) is used to train an optimization control strategy offline in the virtual environment, and finally an online optimization model is obtained to guide the real boiler operation. The composite reward function is a numerical score used to evaluate the good or bad of an operation in reinforcement learning. There is not only one reward indicator, but multiple target indicators are integrated together: the core target is to improve the combustion efficiency; the constraint / auxiliary targets are to control the pollutant emission (NOx,

[0059] According to the historical operation data of the boiler, 25-dimensional process variables are extracted as state observation values and 11-dimensional decision variables are defined as action space; according to the state observation values, the action space and the long short-term memory network, a state transition model is determined; An action change penalty term, a first weight coefficient and a second weight coefficient are obtained; the preliminary efficiency prediction value, the combustion efficiency prediction model, the action change penalty term, the first weight coefficient and the second weight coefficient generate a composite reward function.

[0060] The boiler combustion process is modeled as a Markov decision process (MDP), which specifically includes extracting 25-dimensional process variables as state observation values from the historical operation data of the boiler operation , and defining 11-dimensional decision variables as action space from the historical operation data​ , using LSTM neural network Constructing state transition model: .

[0061] MDP (Markov Decision Process) is the standard modeling framework of reinforcement learning, which describes the relationship between "state - decision - state transition - reward". Abstracting the boiler combustion process into MDP means that we regard the operation of the boiler as a decision-making process that evolves at all times. The current state determines the next action, and the action affects the subsequent state and reward. State definition (State, S) is derived from the historical operation data of the boiler. Correlation analysis and feature selection are performed first. Finally, 25-dimensional process variables (such as furnace temperature distribution, coal supply, primary air / secondary air flow, steam pressure, oxygen content, etc.) are selected as the observed state of the boiler at a certain time. The state represents both the "snapshot" of the current operation of the boiler and the basis for decision-making. Action definition (Action, A) defines 11-dimensional decision variables, representing adjustable control means such as coal mill switch, secondary air damper opening, air distribution ratio, coal conveying speed, etc.

[0062] State transition model (LSTM implementation) uses LSTM (Long Short-Term Memory Network) to fit the state transition law of the boiler, i.e., to predict the next state given the current state and the action taken.

[0063] LSTM is suitable for handling the dependence of operating variables on time, and can capture the dynamic characteristics and time series correlation of boiler combustion. The resulting model is the state transition model, which is one of the core components of the subsequent virtual simulation environment.

[0064] Compound reward function: where is the predicted value of the combustion efficiency at time t, , is the corresponding combustion efficiency prediction model for each operating condition, is the action change penalty term, , is the weight coefficient .

[0065] Core index: t time combustion efficiency prediction value, the main positive target of the reward function is to improve the combustion efficiency of the boiler. Call the combustion efficiency prediction model (probability prediction based on LSTM+diffusion model) of each working condition to reduce the uncertainty risk. Select the corresponding prediction model according to the current working condition interval, and give the efficiency prediction value in this state. The penalty term is the action change penalty; in actual operation, large adjustment of control quantity will cause boiler fluctuation and even safety risk, so the penalty term of control change amplitude must be added in the reward function. The penalty term is used to suppress frequent large adjustment or excessive operation, and to encourage smooth optimization.

[0066] The weight coefficient can be calibrated or optimized by historical data to balance multiple objectives such as efficiency, stability and environmental protection. The generated composite reward function can provide optimization direction and constraints for reinforcement learning.

[0067] Through the combustion efficiency prediction model to calculate the efficiency change after control action, the state transition model to calculate the boiler state evolution after control action, and the composite reward function to judge the operation, the virtual environment is constructed: the state transition model and the combustion efficiency prediction model are combined to form an interactive simulation environment; the reward function is embedded in the environment feedback system, and the deep reinforcement learning method framework is trained offline: in the virtual environment, the agent of the deep reinforcement learning method framework is repeatedly tested; through the process of "trial→state change→reward feedback", the optimal control strategy is learned, and after training, an optimization model is output; the optimization model can be deployed to the real boiler operation, and the optimal control suggestion is given in real time according to the state.

[0068] Based on the combustion efficiency prediction model and the state transition model, a virtual environment for boiler operation is constructed, and a composite reward function considering efficiency, emission and stability is combined. In this environment, deep reinforcement learning (such as DDPG) method is used for offline training, and finally an optimization model capable of outputting optimal control instructions is obtained, realizing energy-saving, stable and efficient and risk-controllable boiler operation.

[0069] In step S160, the real-time operation data of the boiler is input into the optimization model to generate optimization decision data.

[0070] By deploying the trained optimization model to the real-time control system, and inputting the operation data of the boiler into the optimization model, the optimal decision (optimization decision data) is obtained, realizing online dynamic optimization control of coal-fired boiler, which significantly reduces coal consumption while improving the thermal efficiency of the unit.

[0071] Real-time operation data of the boiler is acquired in real time from a DCS system of the boiler, which can be process variables (i.e., state variables) of the boiler, for example, temperatures of various points in a furnace, coal supply amount, primary air / secondary air volume, main steam pressure, temperature, discharge concentration, and the like. The dimensions of these variables can be consistent with the state space defined when the MDP is trained (for example, 25 dimensions). The real-time state data collected at present is packaged and input into the optimization model. The optimization model calculates, based on the previously trained strategy, an action value of obtaining the maximum long-term comprehensive benefit in the current state, i.e., optimization decision data.

[0072] The output is a control suggestion defined by the action space, for example, values of 11 continuous control variables: damper opening, coal belt speed, burner switching signal, air preheater adjustment, and the like. The control variables can be directly issued as boiler control instructions to an actuator, or can be first transmitted to an artificial operator for review and confirmation.

[0073] The optimization model can enable the system to automatically adjust the control strategy according to the operation state. The optimization model can adjust the decision according to different coal quality, load change, and environmental conditions, instead of a fixed strategy. The control strategy takes into account the control strategies of ensuring stable combustion (safety, low fluctuation), improving thermal efficiency (energy saving), and reducing emissions (environmental protection), is safer and more reliable, has higher combustion efficiency, and has better energy saving and environmental protection effects. At the same time, the artificial dependence is reduced, the optimization decision data is continuously output, the delay is low, and the control measures can be updated at a second or minute level.

[0074] The optimization model trained by the historical data can quickly output the corresponding optimal control variable by taking the real-time collected boiler operation state as input after deployment, so as to realize real-time, dynamic, and closed-loop combustion process optimization control.

[0075] The combustion optimization method of the embodiment can effectively solve the problem that the boiler operation parameters collected by the DCS system have significant noise and part of the data of the measuring points is missing, the output error of the combustion prediction model based on data driving fluctuates greatly (±1.8%), and the stability of the optimization control is affected. At the same time, the technical problem that the proportion of deep peak shaving working conditions with a load rate lower than 40% in the historical operation data is only 5.2%, and the optimization strategy is unstable in extreme working conditions (for example, the fluctuation range of the coal supply amount instruction is more than ±15%) due to insufficient samples can also be solved.

[0076] In an embodiment of the present application, the combustion efficiency of the boiler is determined according to historical operation data of the boiler. The heat loss of each operation item is determined by using a counterbalance method to calculate the historical operation data of the boiler. The combustion efficiency of the boiler is determined according to the heat loss of each operation item.

[0077] In the embodiments of the present application, the historical operation data of the boiler monitored by the distributed control system DCS is used to calculate various heat losses (flue gas, incomplete combustion of gas, incomplete combustion of solid, heat dissipation) by using the counterbalance method, and then the combustion efficiency is calculated.

[0078] During the operation of the boiler, the higher the combustion efficiency is, the higher the fuel utilization rate is, and the smaller the energy waste is.

[0079] The distributed control system will collect and store the boiler operation parameters in real time, such as fuel consumption, flue gas composition, boiler steam output, temperature at each location, etc. Using these historical data, the combustion efficiency can be calculated by the counterbalance method (back calculation method).

[0080] The counterbalance method is to calculate the various heat losses of the boiler, and then subtract the total proportion of these losses from 100% to obtain the combustion efficiency, so it is called counterbalance (not directly calculating the effective utilization, but inversely calculating from the loss). The total heat input (fuel heat value) = useful heat (heat for converting water into steam) + various heat losses. The combustion efficiency = 1 - the sum of various heat loss rates.

[0081] Common heat loss items include: flue gas heat loss The temperature of the discharged flue gas is higher than the ambient temperature, and a large amount of heat is taken away; chemical incomplete combustion heat loss The flammable gas (such as CO) in the flue gas is not completely combusted, and the chemical energy is taken away; mechanical incomplete combustion heat loss The residual combustible carbon in fly ash and slag is not completely combusted; heat dissipation loss of the furnace wall The heat is lost to the environment through conduction / radiation of the furnace wall, pipe, etc.; other special losses such as unutilized air preheating loss caused by air leakage, heat loss during soot blowing, etc.

[0082] According to the historical data and the boiler thermal working condition formula, the proportion of each loss to the fuel input heat is calculated, for example: flue gas heat loss rate = (flue gas volume x specific heat capacity x (flue gas temperature - ambient temperature)) / fuel input heat; chemical incomplete combustion heat loss rate = (CO volume x combustion heat value) / fuel input heat; mechanical incomplete combustion heat loss rate = (unburned carbon content in fly ash x carbon combustion heat value + unburned carbon content in slag x carbon combustion heat value) / fuel input heat; heat dissipation loss rate = (boiler heat dissipation area x (furnace wall average temperature - ambient temperature) x heat dissipation coefficient) / fuel input heat; other heat loss rate (such as air leakage heat loss, soot blowing heat loss, etc.) = (heat of cold air brought in by air leakage - heat absorbed by corresponding air heating + heat consumed by soot blowing medium) / fuel input heat. Counterbalance method calculation: combustion efficiency That is, efficiency = 100% - sum of all heat loss rates.

[0083] At the same time, record and save the historical thermal efficiency value of each period and its corresponding main parameters, such as the calorific value of coal and the unit load, etc.

[0084] In an embodiment of the present application, the operating condition data includes the combustion heat value, the unit load, and the cooling medium temperature, and the operating condition division according to the boiler combustion efficiency and the operating condition data obtains the operating condition interval information, which includes: establishing a quantitative mapping model of the coal industry analysis parameters and the boiler combustion efficiency; generating a historical operating condition feature vector set according to the quantitative mapping model, the combustion heat value, the unit load, and the cooling medium temperature; obtaining the operating condition interval information according to the historical operating condition feature vector set and an improved density clustering algorithm.

[0085] In the embodiment, the operating condition data includes the combustion heat value, the unit load, and the cooling medium temperature, and the operating condition division according to the unit load, the cooling medium temperature, the combustion heat value, and the combustion efficiency divides the operating condition interval into stable operating condition intervals and unstable operating condition intervals. The operating condition interval information can be obtained through various historical operating data and operating public data of the boiler in various operating condition intervals.

[0086] The coarse mode of past single reliance on "load" to divide the boiler operating state is changed, the boiler operating condition is finely, dynamically, and automatically identified by fusing the coal characteristics and the key operating parameters, thereby providing a decision basis for the precise control and optimized operation of the unit.

[0087] The offline modeling stage utilizes the historical operating data to complete the model construction and the operating condition definition.

[0088] The online identification stage utilizes the real-time collected data to automatically match the operating state of the current unit to the defined operating condition.

[0089] Establish a quantitative mapping model of coal industrial analysis parameters and combustion efficiency; create a mathematical model that can accurately predict the combustion efficiency of a boiler based on input coal quality parameters (calorific value, volatile matter, ash melting characteristics, etc.). Combustion efficiency is a core indicator, and it is one of the most critical performance indicators for evaluating the operation of a boiler. However, it is usually difficult to measure directly, in real-time, and accurately online. Coal quality is a key disturbance factor, and the characteristics of the fuel coal in a coal-fired power plant (calorific value, flammability, ease of coking, etc.) vary greatly, which is the main external disturbance affecting combustion efficiency. By collecting a large amount of historical data on coal industrial analysis parameters, each data set contains a complete set of coal quality test reports (calorific value, volatile matter, ash content, moisture content, ash melting point, etc.) and the corresponding actual boiler combustion efficiency (usually calculated by heat balance).

[0090] A variety of machine learning or statistical methods can be used to establish this mapping relationship, such as: Multiple Linear Regression, Support Vector Machines (SVM), Neural Networks, or Gradient Boosting Trees, etc. Use the above-mentioned large amount of historical data on coal industrial analysis parameters to verify the accuracy of the model using an independent test data set. The final model can be formalized as: Combustion Efficiency = f(Calorific Value, Volatile Matter, Ash Melting Characteristics,...); generate a trained combustion efficiency prediction model.

[0091] Construct a working condition feature vector; real-time coal calorific value is the core of energy input, directly determining the amount of fuel and air required to produce the same amount of heat, and is the fundamental factor affecting the combustion process. Unit load is the demand for energy output, determining the overall operation rhythm and parameter setting range of the boiler. Cooling medium temperature (such as circulating water inlet temperature) is an important boundary condition, which affects the vacuum degree of the steam turbine, and thus the efficiency of the entire thermal cycle. Under the same power generation load, different cooling water temperatures require different boiler evaporation capacities.

[0092] The feature vector V(t) of this working condition at any time t can be represented as: V(t) = [Real-time coal calorific value(t), Unit load(t), Cooling medium temperature(t)].

[0093] Real-time coal calorific value comes from an online coal quality analyzer or is estimated by weighted average of the coal entering the furnace. Load and temperature are directly obtained from the DCS (Distributed Control System), which can generate a multi-dimensional data set that comprehensively and quantitatively describes the real-time operating state of the boiler. With real-time coal calorific value as the primary basis, the working condition feature vector is constructed in conjunction with unit load and cooling medium temperature.

[0094] The improved density clustering algorithm is used to realize fine segmentation of working conditions. In the constructed massive historical feature vector data, an unsupervised learning algorithm is used to automatically and objectively find the inherent clustering mode of the data, and identify these clustered "data cloud clusters" as stable and repeatedly appearing typical working condition intervals.

[0095] In this embodiment, the improved density clustering algorithm can be used without pre-setting the number of categories. Unlike algorithms such as K-Means, DBSCAN does not need to specify the number of working conditions to be divided in advance. The improved density clustering algorithm can automatically find the number of working conditions according to the data density.

[0096] The above obtained massive historical feature vectors V(t) are input into the improved density clustering algorithm. The improved density clustering algorithm outputs the category (cluster) label to which each data point belongs. Each identified cluster represents a "fine working condition" defined by data driving. For example: working condition A: [high calorific value coal, high load, winter low temperature]; working condition B: [high calorific value coal, high load, summer high temperature]; working condition C: [low calorific value coal, medium load,...]; and finally generate working condition interval information.

[0097] In an embodiment of the present application, the construction of the time series data-driven combustion efficiency prediction model according to the working condition interval information comprises: establishing a time series data set based on the working condition interval information; generating a preliminary efficiency prediction value according to the time series data set in a long short-term memory network processing process variable and decision variable time series correlation; obtaining a historical actual combustion efficiency value; determining a prediction residual according to the preliminary efficiency prediction value and the historical actual combustion efficiency value; diffusion model generates prediction residual samples based on the distribution of the prediction residual; extracting quantiles, mean and variance from the prediction residual samples to obtain a confidence interval of the prediction value; generating a time series data-driven combustion efficiency prediction model according to the confidence interval of the prediction value.

[0098] In this embodiment, in the divided working condition intervals, the historical boiler operation data in the working condition interval information is used to construct a time series data set containing state variable and control variable sequences. The long short-term memory network (LSTM) is used to learn the time series dependence relationship between process variables and decision variables, fit the nonlinear dynamic mapping relationship between input and combustion efficiency, and obtain the preliminary combustion efficiency prediction value of each working condition interval. The mean value benchmark is provided for subsequent uncertainty modeling and optimization control strategy.

[0099] When training the preliminary LSTM model, the difference between the preliminary combustion efficiency prediction value and the historical actual combustion efficiency value, i.e., the prediction residual, is obtained. These residuals have different distribution characteristics under different working conditions, some of which have small fluctuations in the interval, and some of which have large fluctuations in the interval.

[0100] The historical residual data is classified according to the working condition interval and sent to the diffusion model for training. The diffusion model learns the overall probability distribution of the residual data through the iterative process of forward noise addition and reverse noise removal. The working condition characteristics can be additionally input as a condition to obtain the residual distribution under different working conditions. Under the condition of the current working condition and the preliminary prediction value, the diffusion model starts the "reverse sampling" process from pure noise, generates new sample data that conforms to the previously learned residual distribution through multiple denoising iterations. A set of values obtained from the prediction residual sample set (possibly hundreds or thousands) reflects the possible error range of the model under this working condition.

[0101] The diffusion model gradually learns the change pattern and distribution form of the residual, such as the upper and lower boundaries of the residual and the conditions that are prone to large deviations. When the real-time boiler state and control decisions are input, the preliminary combustion efficiency prediction value is obtained using the LSTM. The preliminary combustion efficiency prediction value is input into the diffusion model corresponding to the working condition interval, and the model generates a batch of potential true values with slight differences by sampling the possible residual multiple times. By statistical analysis of this batch of potential true values, the mean update prediction center, maximum and minimum values, and specific quantiles can be obtained. For example, the middle 90% of the data can be taken as the confidence interval of the prediction, and different quantile points can represent different confidence levels.

[0102] Each working condition interval has its own diffusion model, so the corresponding residual distribution characteristics can be selected according to the working condition during prediction to obtain a more accurate confidence interval. Ultimately, a combustion efficiency prediction model is constructed for each working condition interval, and the model outputs not only a number but also a reliable range with upper and lower boundaries. Based on the preliminary prediction value of the LSTM, a diffusion model is connected to each working condition interval to generate a large number of efficiency prediction samples, extract quantiles, mean, and variance, and establish a probability prediction model that fully reflects the dynamic characteristics of the boiler, providing a solid mathematical foundation for intelligent optimization and risk management.

[0103] By introducing a diffusion generation model to model and sample the prediction residual based on the preliminary efficiency prediction model in each working condition interval, the prediction value confidence interval containing the mean, variance, and each quantile is output, thereby expanding the original deterministic prediction into a working condition probability prediction model that can reflect the prediction uncertainty, providing a more robust reward evaluation basis for subsequent reinforcement learning optimization control.

[0104] To maintain the prediction accuracy, the combustion efficiency prediction model is updated based on the newly added historical boiler operation data every 6 hours, so that the combustion efficiency prediction model can be optimized and updated, thereby improving the accuracy of the combustion efficiency prediction model.

[0105] In an embodiment of the present application, the establishment of the time series data set based on the working condition interval information comprises: obtaining the boiler operation process variables in the working condition interval; performing correlation analysis on the boiler operation process variables in the working condition interval to determine that the variables most related to the combustion efficiency are the core decision variables; taking the coal supply amount, the air-coal ratio, and the flue gas oxygen content as the core decision variables of the feature set, and establishing the time series data set.

[0106] In the present embodiment, the boiler operation historical data is divided into several intervals (such as high load, low load, different coal quality combinations, etc.) according to different working conditions. Each working condition interval has corresponding historical operation data, including: process variables (state variables): real-time collected temperature, pressure, concentration, coal fineness, primary air volume, secondary air volume, etc. (that is, 25-dimensional state variables); and decision variables (control variables): fuel distribution, damper opening, superheater water injection amount, online number of coal mills, etc. (11-dimensional decision variables) Based on the time series dependence relationship between the process variables and the decision variables, an initial model capable of predicting the combustion efficiency is constructed. The establishment of the time series data set not only records the state and control amount at a single time point, but also records the change sequence within a time window (for example, 60 seconds, 120 seconds).

[0107] In each divided working condition interval, the historical boiler operation data in the working condition interval information is used to construct a time series data set containing state variable and control variable sequences. The long short-term memory network (LSTM) is used to learn the time series dependence relationship between the process variables and the decision variables, fit the nonlinear dynamic mapping relationship between the input and the combustion efficiency, and obtain the preliminary combustion efficiency prediction value of each working condition interval, providing a mean value benchmark for subsequent uncertainty modeling and optimization control strategy.

[0108] The correlation indicators of all variables and the combustion efficiency (target variable) are calculated, such as the Pearson correlation coefficient, the Spearman rank correlation, or the feature importance ranking. Variables with a correlation coefficient below a certain threshold are removed. The coal supply amount (directly affecting the heat input), the air-coal ratio (affecting the combustion completeness and flue gas loss), and the flue gas oxygen content (at the economizer outlet) (reflecting the combustion air excess or deficiency) are not only highly correlated features, but also controllable variables in future reinforcement learning optimization.

[0109] For each operating condition interval, the feature set data of the corresponding time period is extracted. Each sample is arranged in chronological order, maintaining the dynamic dependency between variables. Contains: historical lag values (such as variable values in the past 5 minutes or 10 sampling points), possible derivative features (moving average, difference value, etc.), and each operating condition interval can generate an independent time series data set.

[0110] In an embodiment of the present application, the combustion efficiency prediction model is embedded into a deep reinforcement learning method framework for training to generate an optimization model, comprising: Obtaining a state transition model and a composite reward function; In the combustion efficiency prediction model driven virtual environment, the deep reinforcement learning method framework algorithm is trained offline according to the state transition model and the composite reward function to generate an optimization model.

[0111] Model the combustion efficiency by the boiler historical operation data to obtain a sub-working condition probability prediction model (LSTM model + diffusion model).

[0112] Based on each sub-working condition probability prediction model, a virtual environment is constructed as a simulation environment, which can predict the future state according to the given control decision and give the combustion efficiency and confidence interval.

[0113] An optimization control strategy is trained offline in the virtual environment using a deep reinforcement learning (DRL) algorithm (such as DDPG), and finally an online optimization model is obtained to guide the real boiler operation.

[0114] In an embodiment of the present application, the obtaining of the boiler combustion variable data and the composite reward function comprises: According to the historical operation data of the boiler, 25-dimensional process variables are extracted as state observation values and 11-dimensional decision variables are defined as action space; According to the state observation value, the action space and the long short-term memory network, a state transition model is determined; Obtaining an action change penalty term, a first weight coefficient and a second weight coefficient; The preliminary efficiency prediction value, the combustion efficiency prediction model, the action change penalty term, the first weight coefficient and the second weight coefficient generate a composite reward function.

[0115] The composite reward function is a numerical score used to evaluate the goodness of an operation in reinforcement learning. There is not only one reward indicator, but multiple goals are combined together: the core goal is to improve the combustion efficiency; The auxiliary goal is to control the pollutant emission (NOx, ), maintain the steam temperature stable, reduce the boiler wear and tear, etc.

[0116] According to historical operation data of the boiler, 25-dimensional process variables are extracted as state observations, and 11-dimensional decision variables are defined as action space; and according to the state observations, the action space and a long short-term memory network, a state transition model is determined.

[0117] An action change penalty term, a first weight coefficient and a second weight coefficient are obtained; and the preliminary efficiency prediction value, the combustion efficiency prediction model, the action change penalty term, the first weight coefficient and the second weight coefficient generate a composite reward function.

[0118] The boiler combustion process is modeled as a Markov decision process (MDP), specifically including extracting 25-dimensional process variables as state observations from historical operation data of the boiler , defining 11-dimensional decision variables as action space from historical operation data , using an LSTM neural network to construct a state transition model: .

[0119] MDP (Markov Decision Process) is a standard modeling framework for reinforcement learning. State definition (State, S) is from the historical operation data of the boiler, and correlation analysis and feature selection are performed first. Finally, 25-dimensional process variables (such as furnace temperature distribution, coal supply, primary air / secondary air flow, steam pressure, oxygen content, etc.) are selected as the observed state of the boiler at a certain time. The state represents both the "snapshot" of the current operation of the boiler and the basis for decision-making. Action definition (Action, A) defines 11-dimensional decision variables, representing adjustable control means such as coal mill switch, secondary air damper opening, air distribution ratio, coal conveying speed, etc.

[0120] The state transition model (LSTM implementation) uses LSTM (Long Short-Term Memory Network) to fit the state transition law of the boiler, i.e. to predict the next state given the current state and the action taken. LSTM is suitable for handling the dependence of process variables on time, and can capture the dynamic characteristics and time series correlation of boiler combustion. The resulting model is a state transition model, which is one of the core components of the subsequent virtual simulation environment.

[0121] Composite reward function: where is the predicted value of the combustion thermal efficiency at time t, , is the corresponding combustion efficiency prediction model for each working condition, is the action change penalty term, , is the weight coefficient .

[0122] The core index is a predicted value of the combustion heat efficiency at time t, and the main positive target of the reward function is to improve the combustion heat efficiency of the boiler. The combustion efficiency prediction model (probability prediction based on LSTM+diffusion model) of each working condition is called to reduce the uncertainty risk. The corresponding prediction model is selected according to the current working condition interval, and the efficiency prediction value in this state is given. The penalty term is the action change penalty; in actual operation, a large adjustment of the control quantity will cause the boiler to fluctuate or even pose a safety risk, so the penalty term of the control change amplitude must be added to the reward function. The penalty term is used to suppress frequent large adjustments or excessive operations and encourage smooth optimization.

[0123] The weight coefficient can be calibrated or optimized through historical data to balance multiple targets such as efficiency, stability and environmental protection, and the generated composite reward function can provide an optimization direction and constraint for reinforcement learning.

[0124] The combustion efficiency prediction model is used to calculate the efficiency change after the control action, the state transition model is used to calculate the boiler state evolution after the control action, and the composite reward function is used to judge the operation. The virtual environment is constructed by combining the state transition model and the combustion efficiency prediction model to form an interactive simulation environment; the reward function is embedded in the environment feedback system, and the deep reinforcement learning method framework is trained offline: in the virtual environment, the agent of the deep reinforcement learning method framework is repeatedly tested; through the continuous process of “trial→state change→reward feedback”, the optimal control strategy is learned, and after the training is completed, a strategy network (optimization model) is output. The optimization model can be deployed in real boiler operation to give optimal control suggestions in real time according to the state.

[0125] Based on the combustion efficiency prediction model and the state transition model, a virtual environment for boiler operation is constructed, a composite reward function considering efficiency, emission and stability is combined, and a deep reinforcement learning (such as DDPG) method is used for offline training in the environment, so as to finally obtain an optimization model capable of outputting optimal control instructions, thereby realizing energy-saving, stable efficiency and controllable risk of boiler operation.

[0126] In a second aspect, the embodiments of the present application provide a combustion optimization system of a coal-fired boiler, including: An acquisition module is configured to acquire historical operation data and operation condition data of the boiler; A determination module is configured to determine a combustion efficiency of the boiler according to the historical operation data of the boiler; A obtaining module is configured to divide working conditions according to the combustion efficiency of the boiler and the operation condition data, and obtain working condition interval information; A construction module is configured to construct a time series data-driven combustion efficiency prediction model according to the working condition interval information. The generating module is configured to embed the combustion efficiency prediction model into a deep reinforcement learning method framework to generate an optimization model; The input module is configured to input real-time operation data of the boiler into the optimization model to generate optimization decision data.

[0127] In the combustion optimization system, the stable and unstable state intervals in the working condition are identified through the historical operation data and the operation condition data of the boiler, and the combustion efficiency prediction model is constructed based on the identified working condition interval information. The deep reinforcement learning method framework is trained offline in the virtual environment of the combustion prediction model to generate an optimization model. The optimization model is deployed in the real-time data stream processing system of the boiler. Real-time operation data is input into the optimization model to generate optimization decision data. The boiler is optimized through the optimization decision data, which can effectively reduce the coal consumption and improve the combustion heat efficiency, realize online dynamic optimization control of the coal-fired boiler, significantly reduce the coal consumption while improving the unit heat efficiency, and comprehensively analyze each variable in the boiler operation process, effectively solve the problem of significant noise in the boiler operation parameters and missing data in some measurement points, which leads to large output error fluctuation of the data-driven combustion prediction model and affects the stability of the optimization control.

[0128] In an embodiment of the present application, the combustion efficiency of the boiler is determined according to the historical operation data of the boiler, including: The heat loss of each operation item is determined by using the counterbalance method to calculate the historical operation data of the boiler. The combustion efficiency of the boiler is determined according to the heat loss of each operation item.

[0129] In an embodiment of the present application, the operation condition data includes the combustion heat value, the unit load, and the cooling medium temperature. The working condition interval information is obtained by dividing the working condition according to the combustion efficiency of the boiler and the operation condition data, including: A quantitative mapping model of the coal industry analysis parameters and the combustion efficiency of the boiler is established. A historical working condition feature vector set is generated according to the quantitative mapping model, the combustion heat value, the unit load, and the cooling medium temperature. The working condition interval information is obtained according to the historical working condition feature vector set and the improved density clustering algorithm.

[0130] In an embodiment of the present application, the time series data-driven combustion efficiency prediction model is constructed according to the working condition interval information, including: A time series data set is established based on the working condition interval information. generate a preliminary efficiency prediction value according to the time series data set and a long short-term memory network processing time series correlation between the process variable and the decision variable; obtain a historical actual combustion efficiency value; determine a prediction residual according to the preliminary efficiency prediction value and the historical actual combustion efficiency value; generate a prediction residual sample based on a distribution of the prediction residual according to the diffusion model; extract quantiles, mean and variance from the prediction residual sample to obtain a confidence interval of the prediction value; generate a time series data-driven combustion efficiency prediction model according to the confidence interval of the prediction value.

[0131] In an embodiment of the present application, the time series data set is established based on the working condition interval information, which includes: obtain the boiler running process variable in the working condition interval; perform correlation analysis on the boiler running process variable in the working condition interval to determine that the variable most related to the combustion efficiency is the core decision variable; establish a time series data set with the coal supply amount, the air-coal ratio and the flue gas oxygen content as the feature set of the core decision variable.

[0132] In an embodiment of the present application, the combustion efficiency prediction model is embedded into a deep reinforcement learning method framework for training to generate an optimization model, which includes: obtain a state transition model and a composite reward function; offline train the deep reinforcement learning method framework algorithm in a virtual environment driven by the combustion efficiency prediction model according to the state transition model and the composite reward function to generate an optimization model.

[0133] In an embodiment of the present application, the boiler combustion variable data and the composite reward function are obtained, which includes: extract 25-dimensional process variables as state observation values and define 11-dimensional decision variables as action space according to the historical operation data of the boiler; determine a state transition model according to the state observation values, the action space and a long short-term memory network; obtain an action change penalty term, a first weight coefficient and a second weight coefficient; generate a composite reward function according to the preliminary efficiency prediction value, the combustion efficiency prediction model, the action change penalty term, the first weight coefficient and the second weight coefficient.

[0134] The functions of each module in each device in the embodiments of the present application can be referred to the corresponding description in the above method, which will not be repeated here.

[0135] Figure 2A structural block diagram of an electronic device according to an embodiment of this application is shown. Figure 2 As shown, the electronic device includes a memory 410 and a processor 420. The memory 410 stores instructions that can be executed on the processor 420. When the processor 420 executes the instructions, it implements the combustion optimization method for the coal-fired boiler in the above embodiments. The number of memories 410 and processors 420 can be one or more. This electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workbenches, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present application described and / or claimed herein.

[0136] The electronic device may also include a communication interface 430 for communicating with external devices and exchanging data. The devices are interconnected using different buses and can be mounted on a common motherboard or otherwise as needed. The processor 420 can process instructions executed within the electronic device, including instructions stored in or on memory to display graphical information of a GUI on an external input / output device (such as a display device coupled to the interface). In other embodiments, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple electronic devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). The bus can be divided into address buses, data buses, control buses, etc. For ease of illustration, Figure 2 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0137] Optionally, in a specific implementation, if the memory 410, processor 420 and communication interface 430 are integrated on a single chip, the memory 410, processor 420 and communication interface 430 can communicate with each other through an internal interface.

[0138] It is understood that the above-described processor can be a Central Processing Unit (CPU), and can also be other general-purpose processors, Digital Signal Processors (DSPs), Application-Specific Integrated Circuits (ASICs), Field Programmable Gate Arrays (FPGAs) or other programmable logic devices, discrete gates or transistor logic, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc. It is worth noting that the processor can be a processor supporting Advanced RISC Machines (ARM) architecture.

[0139] The embodiments of the present application provide a computer readable storage medium (such as the memory 410 described above) which stores computer instructions, and the program is executed by a processor to implement the method provided in the embodiments of the present application.

[0140] Optionally, the memory 410 can include a program storage area and a data storage area, wherein the program storage area can store an operating system and at least one application required by a function; and the data storage area can store data created according to the use of the electronic device, etc. In addition, the memory 410 can include a high-speed random access memory, and can also include a non-transitory memory, such as at least one magnetic disk storage device, a flash memory device, or other non-transitory solid-state memory device. In some embodiments, the memory 410 can optionally include a memory disposed remotely with respect to the processor 420, and these remote memories can be connected to the electronic device through a network. Examples of the above-mentioned network include but are not limited to the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof.

[0141] Any process or method descriptions or any other descriptions herein can be understood as representing any quantity of instructions, code, or statements that, when executed by a processing circuit, cause the processing circuit to perform a function described. The embodiments of the present application are preferably implemented using or including computer executable instructions in the form of program code, machine code, firmware, software, or any combination thereof, which is executed by a processing circuit. The program code includes program modules, routines, or functions.

[0142] The logic and / or steps represented in the flow diagrams and / or otherwise described herein, for example, can be embodied in non-transitory computer-readable media, executed by an instruction execution system, apparatus, or device, such as a computer-based system, processor-containing system, or other system that can fetch the instructions from the instruction execution system, apparatus, or device and execute the instructions, or in conjunction with which the instructions can be executed.

[0143] It should be understood that each part of the present application can be realized by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, a plurality of steps or methods can be realized by software or firmware stored in a memory and executed by a suitable instruction execution system. All or part of the steps of the above-mentioned embodiment methods can be instructed by a program to complete the relevant hardware, and the program can be stored in a computer readable storage medium, and when the program is executed, it includes one of the steps of the method embodiment or a combination thereof.

[0144] In addition, each functional unit in each embodiment of the present application can be integrated in one processing module, or each unit can be physically present separately, or two or more units can be integrated in one module. The above-mentioned integrated module can be realized in the form of hardware or in the form of software functional module. The above-mentioned integrated module, if realized in the form of software functional module and sold or used as an independent product, can also be stored in a computer readable storage medium. The storage medium can be a read-only memory, a magnetic disk or an optical disk, etc.

[0145] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of various changes or replacements within the technical range disclosed in the present application, and these should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A combustion optimization method for a coal-fired boiler, characterized in that, include: Acquire historical operating data and operating condition data of the boiler; The boiler combustion efficiency is determined based on the boiler's historical operating data. Based on the boiler combustion efficiency and the operating condition data, the operating conditions are divided to obtain information for each operating condition interval. Based on the information of each operating condition range, a time-series data-driven combustion efficiency prediction model is constructed. The combustion efficiency prediction model is trained by embedding a deep reinforcement learning framework to generate an optimized model. The real-time operating data of the boiler is input into the optimization model to generate optimization decision data.

2. The method according to claim 1, characterized in that, Determining the boiler combustion efficiency based on the boiler's historical operating data includes: The heat loss of each operating item was determined by using the inverse balance method to calculate the historical operating data of the boiler. The boiler combustion efficiency is determined based on the heat loss of each operating item.

3. The method according to claim 2, characterized in that, The operating condition data includes: combustion calorific value, unit load, and cooling medium temperature. The process of dividing the operating conditions based on the boiler combustion efficiency and the operating condition data to obtain information for each operating condition interval includes: Establish a quantitative mapping model between coal-fired industrial analysis parameters and the combustion efficiency of the boiler; Based on the quantitative mapping model, the combustion calorific value, the unit load, and the cooling medium temperature, a set of historical operating condition feature vectors is generated. Information for each operating condition interval is obtained based on the set of historical operating condition feature vectors and the improved density clustering algorithm.

4. The method according to claim 3, characterized in that, The step of constructing a time-series data-driven combustion efficiency prediction model based on the information from each operating condition interval includes: A time-series dataset is established based on the information of each operating condition interval; Based on the time-series dataset, the temporal correlation between process variables and decision variables is processed using a long short-term memory network to generate preliminary efficiency prediction values; Obtain historical actual combustion efficiency values; The prediction residual is determined based on the preliminary efficiency prediction and the historical actual combustion efficiency. The diffusion model generates predicted residual samples based on the distribution of the predicted residuals; The quantiles, mean, and variance are extracted from the predicted residual samples to obtain the confidence interval of the predicted values. Based on the confidence interval of the predicted values, a time-series data-driven combustion efficiency prediction model is generated.

5. The method according to claim 4, characterized in that, The time-series dataset established based on the information of each operating condition interval includes: Obtain the boiler operation process variables within the operating condition range; A correlation analysis was conducted on the boiler operation process variables within the aforementioned operating condition range to identify the variables most relevant to combustion efficiency as the core decision variables. A time-series dataset is constructed based on a feature set with coal feed rate, air-coal ratio, and flue gas oxygen content as the core decision variables.

6. The method according to claim 5, characterized in that, The optimized model generated by embedding the combustion efficiency prediction model into a deep reinforcement learning framework for training includes: Obtain the state transition model and the compound reward function; In a virtual environment driven by the combustion efficiency prediction model, an optimized model is generated by offline training of the deep reinforcement learning framework algorithm based on the state transition model and the composite reward function.

7. The method according to claim 6, characterized in that, The state transition model and composite reward function include: Based on the historical operating data of the boiler, 25-dimensional process variables are extracted as state observations, and 11-dimensional decision variables are defined as the action space. The state transition model is determined based on the state observations, the action space, and the long short-term memory network. Obtain the action change penalty, the first weight coefficient, and the second weight coefficient; The preliminary efficiency prediction value, the combustion efficiency prediction model, the action change penalty term, the first weight coefficient, and the second weight coefficient are used to generate a composite reward function.

8. A combustion optimization system for a coal-fired boiler, characterized in that, include: The acquisition module is used to acquire historical operating data and operating condition data of the boiler; The determination module is used to determine the boiler combustion efficiency based on the historical operating data of the boiler. The module is used to divide the operating conditions according to the boiler combustion efficiency and the operating condition data, and obtain information on each operating condition interval. The construction module is used to construct a time-series data-driven combustion efficiency prediction model based on the information of each operating condition interval. The generation module is used to embed the combustion efficiency prediction model into a deep reinforcement learning framework for training, and generate an optimized model. The input module inputs the boiler's real-time operating data into the optimization model to generate optimization decision data.

9. An electronic device, characterized in that, include: At least one processor; as well as A memory that is communicatively connected to the at least one processor; in, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-7.

10. A computer-readable storage medium storing computer instructions that, when executed by a processor, implement the method as described in any one of claims 1-7.

Images