A soft measurement method and system for flue gas components at the outlet of a circulating fluidized bed boiler furnace
By arranging measuring points in different fluidization zones of the circulating fluidized bed furnace and combining a time-series prediction model with a bidirectional long short-term memory network based on an attention mechanism, the real-time performance and stability issues of flue gas component detection in circulating fluidized bed boilers were resolved, achieving efficient continuous estimation of flue gas components.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SOUTHEAST UNIV
- Filing Date
- 2026-02-10
- Publication Date
- 2026-06-02
AI Technical Summary
Existing methods for detecting flue gas components in circulating fluidized bed boilers rely on expensive instruments and are difficult to meet the requirements for real-time and continuous operation. Traditional soft measurement methods have poor stability and reliability under complex operating conditions and are difficult to reflect the dynamic correlation of the combustion process.
Measurement points were arranged along the height direction in different fluidization zones of the circulating fluidized bed furnace to collect temperature and pressure parameters. Combined with the time series prediction model, time series data of internal state parameters of the furnace at multiple locations were established, and a bidirectional long short-term memory network model based on the attention mechanism was used for prediction.
It enables continuous estimation of flue gas components without the need for high-cost online analyzers, improving real-time performance and continuity, alleviating the limitations of traditional methods, and enhancing the accuracy and stability of operational monitoring.
Smart Images
Figure CN122133467A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the detection and control of circulating fluidized bed combustion processes, and more particularly to a soft measurement method and system for the components of flue gas at the furnace outlet of a circulating fluidized bed boiler. Background Technology
[0002] With the widespread application of circulating fluidized bed (CFB) boilers in the power and industrial heating sectors, higher demands are placed on the stable control of the combustion process in the furnace and the precise regulation of pollutant emissions. The concentrations of components such as carbon monoxide, nitrogen oxides, and oxygen in the flue gas at the furnace outlet are important parameters characterizing the combustion state and emission levels, and their changes are closely related to the combustion reaction process within the furnace. However, direct online detection of flue gas components typically relies on expensive analytical instruments and suffers from problems such as long sampling cycles, high maintenance costs, and response lags, making it difficult to meet the real-time and continuous operation monitoring requirements of CFB boilers. To reduce measurement costs and improve system real-time performance, existing technologies have attempted to employ soft measurement methods, which use readily available process variables within the furnace to infer and calculate flue gas components that are difficult or inconvenient to measure directly. Existing soft measurement methods are mostly based on empirical equations, SVR, BP neural networks, etc., and typically use parameters such as temperature and pressure at a single point or a small number of measurement points as input variables. However, the circulating fluidized bed combustion process exhibits strong nonlinearity, strong coupling, and significant dynamic time delay characteristics. Different height regions within the furnace correspond to different fluidization states and combustion reaction stages, making it difficult to comprehensively reflect the evolution characteristics of the combustion process along the height direction using a single location or static approach. Traditional soft sensor methods tend to overlook the dynamic correlation between combustion states within the furnace and time changes, and are highly sensitive to load changes, fuel switching, and operating condition drift. This results in insufficient model generalization ability, and the soft sensor results suffer from poor stability and reliability under complex operating conditions, thus affecting the accuracy of boiler operation regulation and emission control. Summary of the Invention
[0003] Purpose of the invention: The purpose of this invention is to provide a soft measurement method for flue gas components at the furnace outlet of a circulating fluidized bed boiler, which has the ability to comprehensively characterize furnace state parameters in multiple fluidized zones, depicts the dynamic relationship between furnace state and flue gas components based on a time-series prediction model, and is applicable to the online operation of circulating fluidized bed boilers; on the other hand, it provides a soft measurement system for flue gas components at the furnace outlet of a circulating fluidized bed boiler.
[0004] Technical solution: The soft measurement method for flue gas components at the furnace outlet of a circulating fluidized bed boiler according to the present invention includes the following steps:
[0005] Temperature and pressure parameters are collected at at least two different locations along the height of the furnace wall of a circulating fluidized bed boiler in different fluidization zones, as well as the outlet temperature and pressure parameters at the furnace outlet flue, and historical detection data of the target components of the flue gas at the furnace outlet; wherein the temperature and pressure parameters are time-series data that change over time.
[0006] The time series data are aligned and combined in chronological order to construct time series data of multi-location furnace internal state parameters for characterizing changes in combustion state inside the furnace.
[0007] Based on the correspondence between the time-series data of the internal state parameters of the multi-location furnace and the historical detection data in the time dimension, a time-series prediction model is established that can characterize the relationship between the change of the internal combustion state of the furnace over time and the change of the target components of the flue gas at the furnace outlet.
[0008] During the online operation phase, the time-series data of the internal state parameters of the furnace at multiple locations, which are collected in real time and constructed in the manner described above, are input into the time-series prediction model, and the soft measurement results of the target component concentration of the flue gas at the furnace outlet at the current time or a preset future time are output.
[0009] Preferably, the temperature and pressure parameters at each measuring point are collected synchronously with the same sampling period.
[0010] Preferably, the target components of the flue gas at the furnace outlet include carbon monoxide (C2O3). ),carbon dioxide( ),oxygen( ), nitrogen oxides ( One or more of the following, and may also be sulfur dioxide ( ), hydrogen sulfide ( ),ammonia( One or more of the following.
[0011] Preferably, the different fluidization zones include at least two different height zones among the lower dense phase zone of the furnace, the middle transition zone of the furnace, and the upper dilute phase zone of the furnace, to reflect the state differences of different combustion zones in the circulating fluidized bed furnace.
[0012] Preferably, the dense phase region is used to characterize the fuel pyrolysis and initial combustion state under conditions of high bed material concentration and strong gas-solid contact; the transition region is used to characterize the secondary air mixing, coke burnout, and nitrogen-containing intermediate product conversion process; the dilute phase region is used to characterize the particle residence time and the influence of subsequent oxidation or reduction reactions on the composition of flue gas at the furnace outlet; by arranging measuring points in different fluidization regions, the time-series prediction model can be provided with staged input information reflecting the evolution characteristics of the combustion reaction along the furnace height direction.
[0013] Preferably, the time series data is aligned and combined in chronological order, including time alignment processing of multi-source time series data; data repair and quality improvement processing of defective segments in the time series data; and the construction of time series data of internal state parameters of multi-location furnace that is consistent in time and complete and usable.
[0014] Preferably, the data repair and quality improvement processing includes performing at least one denoising process, outlier handling, and missing value completion process on the time-series data; wherein, the denoising process includes at least one of moving average filtering, median filtering, Savitzky-Golay filtering, or wavelet denoising; the outlier handling includes outlier identification and correction based on threshold rules, 3σ criteria, box plot criteria, or isolated forest algorithm; and the missing value completion includes at least one of linear interpolation, spline interpolation, K-nearest neighbor interpolation, or Kalman filter interpolation.
[0015] Preferably, the time-series prediction model includes a recurrent neural network model based on an attention mechanism.
[0016] Preferably, the attention-based recurrent neural network model is a bidirectional long short-term memory network model, wherein the model takes the time-series data of the internal state parameters of the multi-location furnace in the form of a sequence in the time dimension as the model input, and uses the attention mechanism to weight the features of different time steps in the time series, thereby enhancing the model's ability to focus on features of key time periods.
[0017] Preferably, the time-series prediction model includes at least one bidirectional long short-term memory network layer and a regression output layer connected to its output.
[0018] Preferably, during the model training phase, a sliding time window method is used to extract a continuous time period of a preset length from the time-series data of the internal state parameters of the multi-location furnace, and the data within the continuous time period is used as a single input to the model.
[0019] Preferably, the length of the sliding time window is determined by comparing the prediction error of the model under different window lengths, so as to minimize the prediction error of the model on the validation dataset; the window length that minimizes the prediction error is selected as the model parameter through this comparison process.
[0020] Preferably, the method further includes an online calibration step: during the online operation phase, when real-time or intermittent measured data of the target components of the flue gas are obtained, based on the difference between the measured data and the soft measurement results at the corresponding time, incremental updates, transfer learning updates, or periodic retraining of the time-series prediction model are triggered.
[0021] Preferably, the triggering method for the online calibration includes at least one of event triggering, performance triggering, or periodic triggering; wherein, event triggering includes fuel switching, significant load changes, or restarting of equipment after maintenance; performance triggering includes the deviation between soft measurement results and intermittent measured results exceeding a preset threshold and lasting for a preset time; and periodic triggering includes updating the model according to a preset period.
[0022] Preferably, the collected time-series data also includes auxiliary time-series variables directly related to the combustion conditions; in the step of establishing the time-series prediction model, the auxiliary time-series variables and the time-series data of the internal state parameters of the multi-location furnace are concatenated in the feature dimension and used together as the model input; wherein, the auxiliary time-series variables include the boiler's main steam flow rate, power generation data, and feed rate data reflecting the type or ratio of fuel.
[0023] Preferably, the auxiliary time-series variables further include boiler operating load, load change rate, and fuel condition identifier; the fuel condition identifier includes fuel type, fuel switching status, and multi-fuel ratio information.
[0024] A soft measurement system for flue gas components at the furnace outlet of a circulating fluidized bed boiler includes:
[0025] The data acquisition module is used to collect temperature and pressure parameters at at least two different locations along the height direction in different fluidization zones inside the furnace of a circulating fluidized bed boiler, outlet temperature and outlet pressure parameters at the furnace outlet flue, and historical detection data of the target components of the flue gas at the furnace outlet.
[0026] The data processing module is used to perform time alignment and combination of the collected temperature and pressure parameters to form time-series data of multi-location furnace internal state parameters that characterize the changes in combustion state inside the furnace.
[0027] The model building and inference module is used to establish a time-series prediction model based on the correspondence between the time-series data of the internal state parameters of the multi-location furnace and the historical detection data in the time dimension, and to output the soft measurement results of the target components of the flue gas at the furnace outlet based on the real-time collected time-series data of the internal state parameters of the furnace during the online operation phase.
[0028] Preferably, the data acquisition module includes a composite sensor in which a temperature sensor and a pressure sensor are installed in the same location.
[0029] Preferably, the model building and inference module maintains the same normalization parameters during the model update phase as during the training phase to ensure input consistency during the online phase.
[0030] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: By arranging measuring points along the height direction in different fluidization zones of the circulating fluidized bed furnace, time-series data of temperature and pressure at multiple locations are obtained. Combined with a time-series prediction model, the relationship between the combustion state inside the furnace and the changes in flue gas components at the furnace outlet is modeled, forming a soft measurement method based on multi-region, multi-parameter time-series characteristics. This achieves continuous estimation of flue gas components without the need for high-cost online analyzers, which not only improves the real-time and continuous nature of flue gas component measurement, but also alleviates the problem of insufficient adaptability caused by neglecting the dynamic characteristics of the combustion process in existing soft measurement methods, significantly improving the practicality of circulating fluidized bed boiler operation monitoring. Attached Figure Description
[0031] Figure 1 This is a schematic diagram of the overall process of the method of the present invention;
[0032] Figure 2 This is a schematic diagram of the arrangement of measuring points in the circulating fluidized bed furnace of the present invention;
[0033] Figure 3 This is a schematic diagram illustrating the data preprocessing and sample construction of the present invention;
[0034] Figure 4 This is a schematic diagram of the Bi-LSTM soft measurement model training / prediction of the present invention. Detailed Implementation
[0035] The technical solution of the present invention will be further described below with reference to the accompanying drawings.
[0036] like Figure 1-4 As shown, this embodiment provides a soft measurement method for the components of flue gas at the furnace outlet of a circulating fluidized bed boiler, including the following steps:
[0037] Step 1: Measurement point layout and data acquisition.
[0038] like Figure 2The diagram shown illustrates the arrangement of measuring points inside the furnace and the location of the outlet in this embodiment. The furnace, along its height, includes a dense phase zone at the bottom, a transition zone in the middle (not shown in the diagram), and a dilute phase zone above it. The lower part of the dense phase zone connects to the wind chamber, and the upper part gradually transitions to the transition zone and the dilute phase zone. The top of the dilute phase zone connects to the furnace outlet. Multiple measuring points are set along the height of the furnace sidewall to obtain furnace operating status parameters at different heights within the furnace. Measuring points 1 to 4 are distributed along the furnace height, corresponding to different locations within the dense phase zone, transition zone, and dilute phase zone, respectively. Measuring point 5 is located at the furnace outlet. Each measuring point is used to collect furnace operating status parameters at its corresponding location, mainly easily obtainable time-series variables such as temperature and pressure. The data collected at each measuring point is used to characterize the combustion and flow characteristics of different regions within the furnace and serves as input data for subsequent dynamic prediction of the outlet flue gas composition. This allows for the calculation of the concentration of target components in the outlet flue gas through a model, thereby achieving soft measurement of the outlet flue gas composition without the need for continuous reliance on high-cost gas analyzers in subsequent use. It should be noted that the furnace structure, the division of the dilute and dense phase regions, and related auxiliary structures shown in the figure are merely illustrative examples used to illustrate the arrangement of measuring points and data collection locations, and do not limit the specific furnace structure or internal flow boundaries.
[0039] In practical implementation, the measurement points can be arranged according to the zoning characteristics of gas-solid flow and combustion reaction within the circulating fluidized bed furnace. For example, K-1 temperature and pressure measurement points can be arranged along the height of the furnace wall according to a preset rule. In this embodiment, K=5 is used as an example, that is, four furnace wall measurement points are arranged to cover at least two of the following regions: the dense phase region in the lower part of the furnace, the transition region in the middle part of the furnace, and the dilute phase region in the upper part of the furnace, so as to comprehensively characterize the dynamic changes in bed flow and combustion reaction. At the same time, one outlet temperature and pressure measurement point is arranged at the furnace outlet flue, that is, at the flue gas passage between the furnace outlet and the inlet of the downstream first functional component. All measurement points can use temperature and pressure composite sensors, or the temperature sensor and pressure sensor can be installed in the same position.
[0040] During the data acquisition phase, a uniform sampling period is set. Its range is between 0.1 seconds and 300 seconds; in this example, we take... =1 second. Temperature time-series data (denoted as ) are collected synchronously at each measuring point using this sampling period. , , ..., , ;in, The temperature time series data for measuring point 1, Temperature time series data for measuring point 2, Temperature time series data for measuring point K-1, The data consists of temperature time series data and pressure time series data (denoted as temperature time series data) at the measuring point located at the furnace outlet flue. , , ..., , ;in, The pressure time series data for measuring point 1, For the pressure time series data of measuring point 2, For the pressure time series data of measuring point K-1, This refers to the pressure time-series data at the measuring point located at the furnace outlet flue. During the model training phase, a gas analyzer (such as non-dispersive infrared, ultraviolet differential absorption, electrochemical sensing, gas chromatography, or extraction / in-situ flue gas analyzer) is simultaneously used to collect one or more target components (such as...) in the furnace outlet flue gas. It can also be expanded to , , , , , Reference concentration time series data for (etc.). The above data is collected within a continuous running period (e.g., 30 days) and used for the construction and training of subsequent prediction models.
[0041] To significantly improve the model's generalization ability under different load, fuel switching, and blending ratio variations, auxiliary time-series data directly related to combustion conditions, such as load time-series data and fuel condition indicator time-series data, can also be collected simultaneously. The load time-series data is obtained from the boiler distributed control system (DCS) and can be selected from unit load (typically represented by power generation), load command, or load change rate (…). One or more of the following ( / dt) are used as continuous variables reflecting the overall energy output and dynamic process of the boiler; the fuel condition identification time series data is obtained through the feeding metering device or DCS recording, and is composed of the feed amount of multiple fuels (such as the feed amount of coal, biomass and solid waste) to form a ratio vector, or the normalized ratio vector is directly used as a continuous feature input to accurately characterize the fuel type and real-time blending ratio.
[0042] Step 2: Data preprocessing and feature construction.
[0043] To construct high-quality samples suitable for model training, the collected raw time-series data needs to be preprocessed. First, time alignment is performed: due to potential slight delays in the sampling response or transmission path of temperature and pressure sensors and flue gas analyzers, interpolation methods are used to calibrate the data from each channel to a unified timestamp, ensuring strict synchronization of all variables on the time axis. Second, data cleaning and repair are performed to improve data quality: this includes suppressing signal noise using methods such as moving averages, median filtering, or wavelet denoising; identifying and correcting outliers using the 3σ criterion, box plots, or isolated forest-based algorithms; and completing missing data segments using linear interpolation, spline interpolation, or K-nearest neighbor interpolation to obtain a complete and reliable time-series sequence. Finally, normalization and feature combination are performed: all time-series data, including cleaned temperature, pressure, load, and fuel identification data, are normalized (e.g., using min-max normalization or Z-score standardization). Subsequently, all features (temperature T, pressure P, load, fuel identifier Fuel) corresponding to each sampling time are directly concatenated along the feature dimension to form an extended feature vector. The extended feature vectors from all time steps are combined sequentially to form the final time-series dataset of multi-position furnace internal state parameters. The parameters upon which normalization depends are entirely statistically derived from the training data and are strictly maintained in the subsequent online prediction stage to ensure the consistency of the model input.
[0044] Step 3: Construction and training of the time series prediction model.
[0045] like Figure 4 The diagram illustrates the training and prediction process of a soft-sensing model based on a Bidirectional Long Short-Term Memory (Bi-LSTM) network with an attention mechanism. During model training, the soft-sensing results are compared with the measured concentration data of the target components in the flue gas at the furnace outlet to update the model parameters. During online operation, the soft-sensing results serve as the predicted output of the target components in the flue gas at the furnace outlet. This step is the core of the soft-sensing method, aiming to establish a model capable of characterizing the complex time-delay relationship between the internal dynamics of the furnace and the components at the outlet. First, a sliding time window is used to construct training samples. The time-series data of the internal state parameters of the furnace at multiple locations for L consecutive sampling times (L is an integer greater than 1, for example, determined to be 60 through optimization) is used as an input sample X, with its corresponding label y being the reference value of the concentration of one or more target components in the flue gas at the current time (Δt=0) or a future time Δt. A large number of training sample pairs (X, y) are generated by traversing the entire time series through a sliding window.
[0046] Secondly, a time-series prediction model is constructed and trained. The model employs a bidirectional long short-term memory network based on an attention mechanism. The model structure includes an input layer, at least one bidirectional LSTM network layer, an attention mechanism module, and a fully connected regression output layer. The attention mechanism enables the model to adaptively weight features at different time steps in the input sequence, thus focusing more on information in key time periods. When predicting multiple components, two strategies can be adopted: one is to construct a multi-output Bi-LSTM model with the same number of neurons in its output layer as the target component types, simultaneously outputting the concentrations of all components; the other is to train an independent single-output Bi-LSTM model for each component. During model training, the sample set is divided into training, validation, and test sets (e.g., 70%, 15%, 15%), and the Adam or SGD optimizer is used, with MSE or MAE as the loss function for training. By using the concatenated extended feature sequence [T, P, Load, Fuel] as input for each time step, the Bi-LSTM model can learn the temporal conditional mapping relationship between furnace internal temperature and pressure dynamics and outlet flue gas components under specific load and fuel condition constraints. This effectively reduces feature aliasing between different operating conditions and improves the model's prediction stability and generalization ability under operating condition switching and long-term operation. The sliding window length L can be determined by comparing the prediction errors of the model on the validation set under different L values using optimization methods such as grid search, with the goal of minimizing the prediction error.
[0047] The aforementioned time-series prediction model does not rely on a specific neural network structure to achieve technical effects. Instead, it introduces multi-location time-series parameters that reflect the combustion state in different fluidization zones of the furnace and combines them with a time window approach to characterize the dynamic evolution of the internal state of the furnace along the height direction, thereby achieving stable soft measurement of the flue gas components at the furnace outlet.
[0048] Step 4: Online soft measurement and model calibration.
[0049] After the model training is complete, it is put into online operation. During the online operation phase, the system only needs to collect the temperature at each measuring point in real time. ~ ) and pressure ( ~ Data (and optionally, load and fuel data) undergoes a preprocessing process and sliding window construction consistent with the training phase to generate a real-time input sequence, which is then fed into the deployed Bi-LSTM time-series prediction model. This model can then output soft-measured values of the concentration of one or more target components in the flue gas at the furnace outlet in real time. .
[0050] To address operating condition drift caused by fuel changes and load fluctuations, this method integrates an online correction step. Correction can be triggered by events (such as fuel switching or significant load changes), performance (the deviation |e| between the online soft measurement value and the intermittently collected reference concentration exceeds a threshold and persists for N sampling windows), or periodically (such as weekly or monthly updates). When the triggering condition is met, the system updates the deployed time-series prediction model using newly acquired measured data. Update methods can include incremental updates (using new data to continue training for several rounds with a small learning rate), transfer learning updates (freezing some network layers of the model and only fine-tuning the output layer or the last layer), or periodic retraining (retraining based on recent data). This closed-loop correction mechanism ensures the long-term adaptability and accuracy of the soft measurement system. This correction mechanism can work in conjunction with schemes that introduce operating condition identifiers to jointly address complex operating condition changes.
[0051] This embodiment also provides a soft measurement system for the flue gas components at the furnace outlet of a circulating fluidized bed boiler, used to implement the above method. The system includes:
[0052] The data acquisition module is used to collect temperature and pressure parameters at at least two different locations along the height direction in different fluidization zones within the furnace of a circulating fluidized bed boiler, as well as the outlet temperature and pressure parameters at the furnace outlet flue gas location, and to acquire historical detection data of the target components of the furnace outlet flue gas. The data acquisition module is also used to acquire load time-series data and fuel condition identification time-series data from the boiler control system. The data processing module is used to perform time alignment, data repair, and combination processing on the acquired multi-source temperature and pressure time-series data to construct time-series data of the internal state parameters of the furnace at multiple locations. The model building and inference module is used to establish a time-series prediction model based on the time-series data of the internal state parameters of the furnace at multiple locations and the historical detection data, and to output soft measurement results of the target components of the furnace outlet flue gas during the online operation phase. The data acquisition module includes a composite sensor with temperature and pressure sensors installed in the same location. The model building and inference module maintains the same normalized parameters during the model update phase as during the training phase to ensure input consistency during the online phase.
Claims
1. A soft measurement method for the components of flue gas at the furnace outlet of a circulating fluidized bed boiler, characterized in that, Includes the following steps: Temperature and pressure parameters are collected at at least two different locations along the height of the furnace wall of a circulating fluidized bed boiler in different fluidization zones, as well as the outlet temperature and pressure parameters at the furnace outlet flue, and historical detection data of the target components of the flue gas at the furnace outlet; wherein the temperature and pressure parameters are time-series data that change over time. The time series data are aligned and combined in chronological order to construct time series data of multi-location furnace internal state parameters for characterizing changes in combustion state inside the furnace. Based on the correspondence between the time-series data of the internal state parameters of the multi-location furnace and the historical detection data in the time dimension, a time-series prediction model is established that can characterize the relationship between the change of the internal combustion state of the furnace over time and the change of the target components of the flue gas at the furnace outlet. During the online operation phase, the time-series data of the internal state parameters of the furnace at multiple locations, which are collected in real time and constructed in the manner described above, are input into the time-series prediction model, and the soft measurement results of the target component concentration of the flue gas at the furnace outlet at the current time or a preset future time are output.
2. The method according to claim 1, characterized in that, The different fluidization zones include at least two different height zones among the lower dense phase zone of the furnace, the middle transition zone of the furnace, and the upper dilute phase zone of the furnace, which are used to reflect the state differences of different combustion zones in the circulating fluidized bed furnace.
3. The method according to claim 1, characterized in that, The time series data is aligned and combined in chronological order, including time alignment of multi-source time series data; and data repair and quality improvement of defective segments in the time series data. And to construct time-series data of internal state parameters of the furnace in multiple locations that are consistent in timing and fully usable.
4. The method according to claim 1, characterized in that, The time-series prediction model includes a recurrent neural network model based on an attention mechanism.
5. The method according to claim 4, characterized in that, The attention-based recurrent neural network model is a bidirectional long short-term memory network model. The model takes the time-series data of the internal state parameters of the multi-position furnace in the form of a sequence in the time dimension as the model input, and uses the attention mechanism to weight the features of different time steps in the time series.
6. The method according to claim 5, characterized in that, During the model training phase, a sliding time window method is used to extract a continuous time period of preset length from the time series data of the internal state parameters of the multi-location furnace, and the data within the continuous time period is used as the single input of the model.
7. The method according to claim 6, characterized in that, The length of the sliding time window is determined by comparing the prediction error of the model under different window lengths, in order to minimize the prediction error of the model on the validation dataset.
8. The method according to claim 1, characterized in that, The method further includes an online calibration step: during the online operation phase, when real-time or intermittent measured data of the target components of the flue gas are obtained, based on the difference between the measured data and the soft measurement results at the corresponding time, incremental updates, transfer learning updates, or periodic retraining of the time series prediction model are triggered.
9. The method according to claim 1, characterized in that, The collected time-series data also includes auxiliary time-series variables directly related to combustion conditions; in the step of establishing the time-series prediction model, the auxiliary time-series variables and the time-series data of the internal state parameters of the multi-location furnace are concatenated in the feature dimension and used together as model input; wherein, the auxiliary time-series variables include the boiler's main steam flow rate, power generation data, and feed rate data reflecting fuel type or ratio.
10. A soft measurement system for flue gas components at the furnace outlet of a circulating fluidized bed boiler, characterized in that, include: The data acquisition module is used to collect temperature and pressure parameters at at least two different locations along the height direction in different fluidization zones inside the furnace of a circulating fluidized bed boiler, outlet temperature and outlet pressure parameters at the furnace outlet flue, and historical detection data of the target components of the flue gas at the furnace outlet. The data processing module is used to perform time alignment and combination of the collected temperature and pressure parameters to form time-series data of multi-location furnace internal state parameters that characterize the changes in combustion state inside the furnace. The model building and inference module is used to establish a time-series prediction model based on the correspondence between the time-series data of the internal state parameters of the multi-location furnace and the historical detection data in the time dimension, and to output the soft measurement results of the target components of the flue gas at the furnace outlet based on the real-time collected time-series data of the internal state parameters of the furnace during the online operation phase.