A bed temperature prediction method suitable for fast load change condition of supercritical circulating fluidized bed boiler

Abstract

The application relates to a bed temperature prediction method suitable for a supercritical circulating fluidized bed boiler (SCFB) fast load change condition, which comprises the following steps: inputting historical working condition data in the operation process of the SCFB boiler into a state prediction network model to predict key system states required by an extended Kalman filter; inputting real-time working condition data in the operation process of the SCFB boiler and the key system states into the extended Kalman filter, adopting a double-layer online optimization strategy, and generating a long-time-domain bed temperature initial prediction value; and utilizing a Kalman residual error compensation network to perform residual error compensation on the long-time-domain bed temperature initial prediction value to obtain a final bed temperature prediction value. The application can realize high-precision prediction of the bed temperature under the fast load change condition.

Description

Technical Field

[0001] This invention relates to the field of intelligent environmental protection technology, and in particular to a method for predicting bed temperature under rapid load changes in supercritical circulating fluidized bed boilers. Background Technology

[0002] Against the backdrop of increasingly stringent global "dual carbon" goals and environmental policies, the power system urgently needs clean and efficient energy production equipment. Circulating fluidized bed (CFB) boilers, with their adaptability to low-quality fuels, high thermal efficiency, and the ability to achieve desulfurization through in-furnace technologies, are particularly valuable. , Its low emissions make it a core piece of equipment in the field of coal-fired power generation and biomass co-firing, which is of great significance for reducing the environmental impact of energy production.

[0003] However, with the rapid expansion of renewable energy sources (such as wind power and photovoltaics), their intermittent and fluctuating characteristics place higher demands on the frequency regulation and peak shaving capabilities of the power system. Traditional coal-fired power plants (including CFB boiler power plants) need to have a wider load operating range and faster response speed to maintain grid stability. Bed temperature, as a key parameter of the CFB boiler combustion system, directly affects the release of volatiles from the fuel, the coke reaction rate, and the pollutant formation process: excessively high bed temperatures may lead to slag coking, while excessively low temperatures will cause incomplete combustion, not only reducing combustion efficiency but also increasing unburned carbon emissions. This generates risks. Therefore, accurate bed temperature prediction under rapid load changes is a prerequisite for ensuring the efficient and clean operation of CFB boilers. Currently, bed temperature prediction technologies are mainly divided into two categories: Physically-informed models (PIMs) are built upon the principles of mass conservation, energy conservation, thermodynamic balance, and fluid mechanics. They offer advantages such as structural transparency and strong physical interpretability, and can characterize the dynamic changes in bed temperature through conservation laws. However, these models have significant drawbacks: CFB boiler bed temperature changes involve complex gas-solid two-phase flow and chemical reaction kinetics, which are difficult to fully characterize using mathematical equations; furthermore, under rapid load changes, model parameters (such as heat transfer coefficient and combustion rate) are prone to deviating from design values, leading to a significant decrease in prediction accuracy and failing to meet real-time environmental control requirements.

[0004] Data-driven models: These models do not rely on complex physical mechanisms and directly capture the nonlinear and dynamic characteristics of bed temperature from operational data using machine learning and deep learning techniques. While these models exhibit certain advantages under stable operating conditions, they suffer from two major problems: First, they lack physical consistency constraints, potentially leading to predictions that violate thermodynamic laws (e.g., predicted bed temperature values ​​exceeding a reasonable range), causing environmental control strategies to fail (e.g., excessive air supply). (1) Emissions); second, errors in long-term forecasts are prone to accumulate, especially under rapid load changes, making it difficult to adapt to sudden changes in system dynamics caused by load surges / drops, and thus difficult to support long-term environmental protection operation decisions.

[0005] To balance physical interpretability and prediction accuracy, existing technologies propose hybrid physics-based and data-driven models. However, these hybrid models still have shortcomings: First, they are not optimized for the core scenario of "rapid load changes"—when the load changes abruptly, the model's prediction of key system states lags, leading to bed temperature prediction errors; second, they lack effective residual compensation mechanisms, and system biases and noise can easily accumulate in long-term predictions, affecting the accuracy of pollutant control; third, the integration of physical constraints and data-driven approaches is insufficient, with some models simply combining the two methods without forming synergistic optimization, still failing to meet the environmental protection operation requirements of SCFB boilers.

[0006] In summary, existing bed temperature prediction technologies cannot simultaneously meet the requirements of "physical consistency, high accuracy, long time domain, and anti-interference" under rapid load change conditions, which may lead to problems such as decreased combustion efficiency and excessive pollutant emissions in SCFB boilers. There is an urgent need for a bed temperature prediction technology that is suitable for rapid load change scenarios and takes into account both environmental benefits and prediction performance. Summary of the Invention

[0007] Addressing the problems in the background technology—existing circulating fluidized bed boiler bed temperature prediction technologies suffer from weak physical models to handle dynamic conditions under rapid load changes, lack of physical consistency in data-driven models, long-term error accumulation in hybrid models, and insufficient environmental adaptability—this invention aims to provide a bed temperature prediction method suitable for supercritical circulating fluidized bed boilers (SCFB) under rapid load changes. By constructing a framework of "physical information fusion + data-driven collaborative optimization," it achieves high-precision, long-term time-domain bed temperature prediction, ensuring that SCFB boilers can maintain efficient combustion and low emissions even under sudden load changes. , Emissions meet the technical requirements of the environmental protection sector for the clean operation of industrial boilers.

[0008] To achieve the above objectives, the present invention provides the following solution: A method for predicting bed temperature under rapid load changes in supercritical circulating fluidized bed boilers includes: Historical operating data of the SCFB boiler during operation is input into the state prediction network model to predict the key system states required by the extended Kalman filter. The state prediction network model is built based on a CNN-BiLSTM hybrid network, and the extended Kalman filter is built based on a bed temperature thermodynamic model. The key system states include: coal feed rate, flue gas oxygen content, air intake rate, ash discharge rate, and working fluid temperature. The real-time operating data and key system status during the operation of the SCFB boiler are input into the extended Kalman filter, and a two-level online optimization strategy is adopted to generate the initial predicted value of the long-term time domain bed temperature. The Kalman residual compensation network is used to perform residual compensation on the initial predicted value of the long-term bed temperature to obtain the final predicted value of the bed temperature.

[0009] Optionally, the CNN part of the CNN-BiLSTM hybrid network uses 1D convolutional layers to extract local correlation features of multivariate time series, and combines max pooling operation to downsample to remove redundant information and retain key features. Then, a dropout layer is used to randomly discard some features to enhance the model's generalization ability. The BiLSTM part uses two independent LSTM networks, forward and backward, to process the feature sequences output by the CNN from the forward and backward time series respectively, to fully capture bidirectional time dependencies. Finally, a fully connected layer maps the concatenated bidirectional global features to the predicted values ​​of the key system states of the CFB boiler.

[0010] Optionally, a physical consistency loss function is introduced into the state prediction network model to predict the key system states required by the extended Kalman filter, thereby extending the time domain of bed temperature prediction. The physical consistency loss function includes: a prediction error term, a physical residual term, and a regularization term. The prediction error term is the squared difference between the predicted output of the state prediction network model and the actual critical system state. The physical residual term is the squared difference between the actual critical system state and the output of the bed temperature thermodynamic model. The regularization term is the L2 norm of the trainable parameters of the state prediction network model and the parameters of the bed temperature thermodynamic model; and is weighted by coefficients. The contribution of each of the three items is controlled separately.

[0011] Optionally, the bed temperature thermodynamic model is constructed based on the principles of mass conservation, energy conservation, and thermodynamic equilibrium, and the bed temperature thermodynamic model includes: The coal feed rate model is used to characterize the coal feeder transport delay; the carbon balance model is used to characterize the cyclic combustion process of unburned carbon; the oxygen concentration change model is used to characterize the relationship between combustion performance and oxygen consumption; and the bed temperature energy balance model is used.

[0012] Optionally, the two-level online optimization strategy includes: Layer 1: Construct an offline model library covering load increase, load stabilization, and load decrease conditions; when the SCFB boiler operating condition meets the triggering conditions, select a matching baseline model from the offline model library, optimize the objective function by minimizing the residual through recursive least squares, and update the parameters of the bed temperature thermodynamic model. Layer 2: Optimize and fine-tune the noise covariance matrix of the extended Kalman filter through particle swarm optimization to achieve short-term estimation of the SCFB boiler bed temperature.

[0013] Optionally, each particle in the particle swarm optimization algorithm represents a candidate parameter vector of the noise covariance matrix of the extended Kalman filter. The particle velocity and position are iteratively updated based on the optimal position of the particle and the optimal position of the swarm to minimize the prediction error of the extended Kalman filter for bed temperature. The inertia weight of the particle swarm optimization algorithm adopts a linear decreasing method.

[0014] Optionally, the triggering condition for the dual-level online optimization strategy is: the real-time load change rate exceeds a preset threshold, or the bed temperature prediction error of the extended Kalman filter exceeds a set tolerance.

[0015] Optionally, residual compensation for the initial predicted long-time-domain bed temperature using a Kalman residual compensation network includes: Input the preset windowed sequence into the Kalman residual compensation network to obtain the residual compensation term; The initial predicted value of the long-term bed temperature is added to the residual compensation term to obtain the final predicted value of the bed temperature.

[0016] Optionally, the preset windowed sequence includes: an extended Kalman filter residual sequence spliced ​​by time windows, a system state vector, and a control input vector; The Kalman residual compensation network adopts the same CNN-BiLSTM hybrid network architecture as the state prediction network model.

[0017] The beneficial effects of this invention are as follows: This invention enables high-precision prediction of bed temperature under rapid load changes (load increase, load stabilization, load decrease), with a mean absolute error (MAE) as low as 1.5364℃ and a root mean square error (RMSE) of 1.9738℃. It effectively suppresses the accumulation of long-term prediction errors, improves the robustness and adaptability of the model, and provides a reliable basis for boiler combustion optimization and safety control. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of a bed temperature prediction method applicable to rapid load change conditions in a supercritical circulating fluidized bed boiler, according to an embodiment of the present invention. Figure 2 This is a structural diagram of the bed temperature prediction system according to an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] like Figure 1 and Figure 2 As shown in the figure, this embodiment proposes a method for predicting bed temperature under rapid load change conditions in supercritical circulating fluidized bed boilers, including: Historical operating data of the SCFB boiler during operation is input into the state prediction network model to predict the key system states required by the extended Kalman filter. The state prediction network model is built based on a CNN-BiLSTM hybrid network, and the extended Kalman filter is built based on a bed temperature thermodynamic model. The key system states include: coal feed rate, flue gas oxygen content, air intake rate, ash discharge rate, and working fluid temperature. The real-time operating data and key system status during the operation of the SCFB boiler are input into the extended Kalman filter, and a two-level online optimization strategy is adopted to generate the initial predicted value of the long-term time domain bed temperature. The Kalman residual compensation network is used to perform residual compensation on the initial predicted value of the long-term bed temperature to obtain the final predicted value of the bed temperature.

[0023] Specifically, the technical concept of the method in this embodiment is as follows: An extended Kalman filter (EKF) is constructed based on a bed temperature thermodynamic model, and a two-level online optimization strategy is designed to achieve accurate short-term estimation of the bed temperature of SCFB boilers. A State Prediction Network (SAN) is constructed. The SAN adopts a CNN-BiLSTM hybrid architecture and introduces a physical consistency loss function. The SAN predicts the key system states required for the EKF to extend the time domain of bed temperature prediction. A Kalman residual compensation network (KRC-Net) is constructed. The KRC-Net takes the residual sequence of the EKF output and the real bed temperature, the system state vector and the control input vector as inputs, and learns the residual law to correct the model bias. Based on the historical operating data of the SCFB boiler, the initial parameters of the EKF, SAN and KRC-Net were trained offline to establish a model library containing baseline models for different operating conditions. During the operation of the SCFB boiler, the dual-level online optimization strategy is triggered based on real-time operating data to update the EKF. Combining the future critical system status predicted by the SAN with the residual compensation term output by the KRC-Net, online prediction of the SCFB boiler bed temperature under rapid load change conditions is achieved.

[0024] Furthermore, the bed temperature thermodynamic model is constructed based on the principles of mass conservation, energy conservation, and thermodynamic equilibrium, and includes: The coal feed rate model is used to characterize the coal feeding delay; the carbon balance model is used to characterize the cyclic combustion process of unburned carbon; the oxygen concentration change model is used to characterize the relationship between combustion performance and oxygen consumption; and the bed temperature energy balance model is used.

[0025] The two-level online optimization strategy includes: Layer 1: Construct an offline model library covering load increase, load stabilization, and load decrease conditions; when the SCFB boiler operating condition meets the triggering conditions, select a matching baseline model from the offline model library, minimize the residual optimization objective function through recursive least squares or gradient descent method, and update the parameters of the bed temperature thermodynamic model. Layer 2: Optimize and fine-tune the noise covariance matrix of the extended Kalman filter through particle swarm optimization algorithm to achieve short-term estimation of the SCFB boiler bed temperature.

[0026] Each particle in the particle swarm optimization algorithm represents a candidate parameter vector of the noise covariance matrix of the extended Kalman filter. The particle velocity and position are iteratively updated based on the optimal position of the particle and the optimal position of the swarm to minimize the prediction error of the extended Kalman filter for bed temperature. The inertia weight of the particle swarm optimization algorithm adopts a linear decreasing method.

[0027] The triggering conditions for the dual-level online optimization strategy are: the real-time load change rate exceeds a preset threshold, or the bed temperature prediction error of the extended Kalman filter exceeds a set tolerance.

[0028] Specifically, in this embodiment, the extended Kalman filter (EKF) based on the thermodynamic model and the two-level online optimization specifically include: Based on the thermodynamic mechanism of SCFB boiler bed temperature, a physical model is constructed that includes heat release from coal combustion, heat exchange of bed material, and heat transfer in gas-solid two-phase flow. The fuel carbon balance equation is derived based on the law of conservation of mass, and the differential equation for bed temperature change is established based on the law of conservation of energy. The continuous-time system is discretized into a nonlinear state-space model using the forward Euler method.

[0029] Based on this physical model, an EKF (Earth Temperature Flow Framework) is constructed to achieve accurate short-term bed temperature estimation; simultaneously, a two-level online optimization strategy is designed to adapt to rapid load changes. Layer 1 (Adaptive Update of Physical Model Parameters): Construct an offline model library covering load increase, load stabilization, and load decrease conditions; during online operation, if the real-time load change rate exceeds the preset threshold or the EKF prediction error exceeds the tolerance, select a matching baseline model from the model library, minimize the residual optimization objective function through recursive least squares (RLS), update the key parameters of the model, and ensure that the physical model matches the real-time conditions. Layer 2 (EKF Noise Parameter Optimization): Fine-tuning the noise covariance matrix Q (process noise) and R (measurement noise) of the EKF using Particle Swarm Optimization (PSO) – each particle represents a group The candidate parameters are updated iteratively to determine the particle's velocity and position: ; To minimize the bed temperature prediction error; where the inertial weight of PSO is linearly decreasing ( This ensures optimized efficiency and stability.

[0030] Furthermore, the CNN part of the CNN-BiLSTM hybrid network uses 1D convolutional layers to extract local features of multivariate time series, combines max pooling operations to reduce temporal resolution while retaining key information, and sets a dropout layer after Conv1D to enhance the model's generalization ability; the BiLSTM part processes the feature sequences output by the CNN from both forward and backward directions to capture bidirectional temporal dependencies, and finally maps the bidirectional hidden states to predicted values ​​of key system states through fully connected layers.

[0031] The state prediction network model introduces a physical consistency loss function to predict the key system states required by the extended Kalman filter, thereby extending the time domain of bed temperature prediction. The physical consistency loss function includes: a prediction error term, a physical residual term, and a regularization term. The prediction error term is the squared difference between the predicted output of the state prediction network model and the actual critical system state. The physical residual term is the squared difference between the actual critical system state and the output of the bed temperature thermodynamic model. The regularization term is the L2 norm of the trainable parameters of the state prediction network model and the parameters of the bed temperature thermodynamic model; and is weighted by coefficients. The contribution of each of the three items is controlled separately.

[0032] Specifically, in this embodiment, to address the problem of insufficient long-term prediction capability of EKF, a SAN using a CNN-BiLSTM hybrid architecture is constructed to predict the key system states (coal feed rate, flue gas oxygen content, air intake rate, slag discharge rate, and working fluid temperature) required for EKF prediction, expanding the prediction time domain: - CNN module: Employs 1D convolutional layers (Conv1D) to extract local features of multivariate time series (such as historical load, coal feed rate, and air flow rate), combines max pooling to reduce temporal resolution while retaining key information, and sets a dropout layer (dropout probability 0.5) to suppress overfitting; - BiLSTM module: Performs bidirectional processing on the feature sequence output by the CNN—forward LSTM captures future temporal dependencies, and backward LSTM captures historical temporal dependencies. Comprehensive temporal features are obtained by concatenating bidirectional hidden states, and finally mapped to the prediction of key system states through a fully connected layer; - Physical consistency constraint: Designs a composite loss function. ; in, The contribution of prediction error, physical residual (the difference between the actual state and the physical model output), and parameter regularization term are controlled separately to ensure that the SAN prediction results conform to thermodynamic laws.

[0033] Furthermore, residual compensation for the initial predicted long-term bed temperature using a Kalman residual compensation network includes: Input the preset windowed sequence into the Kalman residual compensation network to obtain the residual compensation term; The initial predicted value of the long-term bed temperature is added to the residual compensation term to obtain the final predicted value of the bed temperature.

[0034] The preset windowed sequence includes: an extended Kalman filter residual sequence spliced ​​by time windows, a system state vector, and a control input vector; The Kalman residual compensation network adopts the same CNN-BiLSTM hybrid network architecture as the state prediction network model.

[0035] Specifically, in this embodiment, the construction and training of the Kalman Residual Compensation Network (KRC-Net) aims to suppress error accumulation in long-term predictions. The KRC-Net learns the residual relationship between the EKF output and the actual bed temperature to achieve bias correction. Input design: The input is a windowed sequence, which consists of the EKF residual sequence (actual bed temperature - EKF predicted value), the system state vector (such as bed temperature and oxygen concentration), and the control input vector (such as coal feed instruction and secondary air volume), arranged in time windows (the window length can be set to...). It is composed of individual sampling steps; Architecture and Training: A CNN-BiLSTM architecture consistent with SAN is adopted to ensure feature extraction and temporal modeling capabilities; training is performed using a weighted mean squared error (LwMSE) loss function, where the weights... ( For true residuals, (For residual standard deviation), prioritize learning the large residual patterns; at the same time, adopt a sliding window normalization strategy to dynamically update the mean and variance of the residuals and stabilize the residual distribution under different working conditions.

[0036] Offline training and online prediction: Offline training: Real operating data from a 350 MW SCFB boiler at a power plant in Shandong Province was used as the training set. The data was not denoised or smoothed, and sensor noise (such as airflow disturbance and coal particle fluctuation) and load fluctuation characteristics were preserved. The sampling interval was 5 minutes, covering three typical operating conditions: load increase, load stabilization, and load decrease. Based on this data, the initial parameters of EKF were calibrated, and SAN and KRC-Net were trained to establish an initial model library. Online prediction: During real-time operation, a two-level optimization strategy is triggered by real-time operating data (such as load change rate and EKF prediction error) to update the EKF parameters; SAN predicts the key system states for the next 1-20 sampling steps and inputs them into the EKF to generate initial bed temperature prediction values; KRC-Net outputs compensation terms based on real-time residuals and state sequences to correct the initial prediction values, and finally obtains accurate bed temperature prediction results.

[0037] The following describes the method of this embodiment in detail using a 350 MW supercritical circulating fluidized bed (SCFB) boiler as an example. This boiler is located in a thermal power plant in Shandong Province, China. It is a supercritical once-through unit and adopts single air distribution plate, single furnace, M-type layout, balanced ventilation, single intermediate reheat and circulating fluidized bed combustion technology. It is equipped with a high-temperature cooled cyclone separator to achieve gas-solid separation. The furnace is a membrane water-cooled wall structure (without external heat exchanger), and the tail flue adopts a steam-cooled enclosure structure.

[0038] I. Implementation Prerequisites and Environmental Preparation: 1.1 Data Sources and Preprocessing: The experimental data are the actual operating data of the SCFB boiler in the power plant. The sampling interval is set to 5 minutes (consistent with the power plant's regular data acquisition frequency). The data is not denoised or smoothed, and the sensor noise (such as airflow disturbance, coal particle composition fluctuation, sensor drift) and load fluctuation characteristics are fully preserved to ensure that the experimental scenario is consistent with the actual industrial situation.

[0039] The data includes the following core variables (corresponding to the core variable definitions in the bed temperature mechanism model in the document): Controlled variable: Bed temperature (unit: Measurement range ); Input / Disturbance Variable: Coal Feeding Command (unit: ), motor speed (unit: ), primary air volume (unit: Secondary air volume (unit: ), slag discharge rate (unit: ); State variable: Oxygen concentration (Unit: %) Fuel Flow Rate (unit: ), carbon combustion rate (unit: ).

[0040] Data partitioning: A total of 10,000 data sets were collected, divided into... The dataset is divided into a training set (7000 sets), a validation set (1000 sets), and a test set (2000 sets). The test set includes data on load increases (load change rate). Stable load (load change rate) ), load reduction (load change rate) Three typical operating conditions cover the dynamic load scenarios that need to be verified in the document.

[0041] II. Specific implementation steps of the core module: 2.1 Bed temperature thermodynamic model and EKF construction and two-level optimization implementation: 2.1.1 Derivation of the bed temperature thermodynamic model: Based on the principles of mass conservation, energy conservation, and thermodynamic equilibrium, a dynamic bed temperature model is constructed, with the specific equations as follows: 1. Coal feed rate model: Considering the coal feeder transmission delay, the coal feed rate... The expression is: ; in, It is a constant (calibrated based on the boiler transmission path length). ), Motor speed (positively correlated with conveyor belt speed); transmission delay Approximate integer delay (That is, a delay of 1 sampling step, 5 minutes), therefore after discretization .

[0042] 2. Carbon Balance Equation: Describes the cyclic combustion process of unburned carbon. ; in, The fixed carbon content in coal (as defined by the experimental coal type) ), For solid waste emissions, This represents the total carbon content in fly ash.

[0043] 3. Oxygen concentration change equation: reflects the relationship between combustion performance and oxygen consumption. ; in, Total air volume The total parameters of the empirical set.

[0044] 4. Bed temperature energy balance equation: ; in: (Specific heat capacity of the bed material is set according to the characteristics of the quartz sand bed material). (Total mass of bed material, calculated using boiler design parameters); (The heat released during combustion, For carbon content, (Weighted average of volatile components) ; ; (The metal wall and the working fluid absorb heat,) The average heat transfer coefficient, The total heat transfer area on the flue gas side. (This refers to the working fluid temperature).

[0045] 2.1.2 EKF Discretization and Initial Parameter Settings: The continuous-time model is discretized using the forward Euler method to obtain a discrete-time nonlinear state-space model: ; Where: state vector Input vector ; System Matrix Input matrix Configure as follows: ; A 3rd order identity matrix; - Observation matrix (Only bed temperature was observed) ), measuring noise ,initial ; - Process noise covariance matrix Initially set as a diagonal matrix initial state value ( (This refers to the boiler's rated operating value).

[0046] 2.1.3 Two-level online optimization implementation: Layer 1: Adaptive update of physical model parameters (based on RLS); Offline model library construction: For three operating conditions—increase, stability, and decrease—three baseline models were trained respectively, and the matrices for each operating condition were stored. and nonlinear functions parameter; Online trigger condition: When the real-time load change rate ( (Rated power) or EKF prediction error When this happens, parameter updates are triggered; RLS update process: 1. Select a baseline model from the model library that matches the current operating conditions; 2. Define the residual optimization objective function The length of the sliding window. It is an adjustable parameter, that is (coefficients) 3. Minimize using the RLS algorithm The RLS forgetting factor is set to 0.95, and the gain matrix is... Iterative updates Until the residuals converge.

[0047] Layer 2: EKF noise parameter optimization (based on PSO): PSO parameter settings: particle number Maximum number of iterations Inertial weight The cognitive acceleration coefficient decreased linearly from 0.9 to 0.4. Social acceleration coefficient ; Particle encoding: Each particle (correspond The 3 diagonal elements and ), Parameter range: ; Fitness function: (Given the length of historical data, minimize the mean squared prediction error). Iterative optimization: Update particle velocity and position according to the following formula: ; ( Uniformly random numbers, For particles The optimal position, (For the optimal position of the group), output the optimal position after the iteration. ; 2.2 Construction and Training Implementation of State Prediction Network (SAN): 2.2.1 SAN Architecture Design: SAN employs a CNN-BiLSTM hybrid architecture, with the following specific structure (input is a multivariate time series, output is the key state variable required by EKF: fuel flow). Primary wind speed Secondary wind speed ): 1. Input layer: Length of the input sequence (i.e., data from the past 20 sampling steps, 100 minutes), feature number (Include The input shape is (None, 20, 6). 2. CNN module: 1D Convolutional Layer (Conv1D): 64 filters, 3 kernels, 1 stride, ReLU activation function, padding="same", output shape (None, 20, 64). Max pooling layer: pooling kernel size 2, stride 2, output shape (None, 10, 64). Dropout layer: Dropout probability 0.5 to prevent overfitting, and the output shape remains unchanged; 3. BiLSTM module: Hidden units: 128, Layers: 2, Return sequence False, output shape (None, 256) (bidirectional hidden state concatenation, ); Forward LSTM: by time step Processing sequences to capture future dependencies; Inverse LSTM: by time step Process sequences and capture historical dependencies; 4. Fully connected layer: Output dimension 3, activation function Linear, output shape.

[0048] 2.2.2 SAN Loss Function and Training Parameters: Composite loss function: ;in, (Prediction error weights) (Physical residual weights, (State values ​​output by the thermodynamic model) (Regular term weights) The number of samples in the training set; Training parameters: Optimizer Adam, learning rate Batch size Maximum number of training epochs The validation set early stopping strategy (stop training when the validation set loss does not decrease for 20 consecutive rounds).

[0049] 2.2.3 SAN Training Process: 1. Divide the training set into consecutive input windows and corresponding output labels according to the time series; 2. Initialize SAN parameters; 3. Input data in batches and perform forward propagation calculations. Calculate the loss ; 4. Backpropagation (using Adam optimizer) update (Trainable parameters of SAN) and (Thermodynamic model parameters); 5. After each training round, evaluate the loss using the validation set and save the model with the minimum loss as the optimal SAN model.

[0050] 2.3 Construction and Training of Kalman Residual Compensation Network (KRC-Net): 2.3.1 KRC-Net Architecture and Input Design; KRC-Net employs the same CNN-BiLSTM architecture as SAN (ensuring that feature extraction and temporal modeling capabilities are matched), with windowed sequences as input: Input sequence composition: EKF residual sequence (1-dimensional) System state sequence (3D) Control Input Sequence (3D); Window length 10, number of input features 7; 2.3 Construction and Training of Kalman Residual Compensation Network (KRC-Net): 2.3.1 KRC-Net Architecture and Input Design KRC-Net uses the same CNN-BiLSTM architecture as SAN, with windowed sequences as input: Input sequence composition: EKF residual sequence (System state sequence) Control input sequence ; The window length is 10, and the number of input features is 7. Output layer: Fully connected layer with output dimension 1 (corresponding to 1 residual compensation term) (used to correct the bed temperature prediction value of EKF).

[0051] 2.3.2 KRC-Net Loss Function and Training Parameters: Weighted mean squared error (LwMSE) loss: ; in, For true residuals The residuals predicted by KRC-Net (Standard deviation of training set residuals, calculated offline) It increases with the increase of the absolute value of the residual, and the learning of the law of large residuals is given priority; Training parameters: consistent with SAN (optimizer Adam, learning rate 0.001, batch size...) epochs ), and uses sliding window normalization.

[0052] 2.4 Offline Training Integration and Model Library Establishment: 1. First, train the thermodynamic model and EKF initial parameters: calibrate based on the training set data. Physical parameters are used to determine the initial EKF. ; 2. Training SAN: Using the state values ​​output by the thermodynamic model as physical constraints, optimize the SAN parameters to ensure that the predicted state conforms to physical laws; 3. Training KRC-Net: Using the prediction residuals of EKF on the training set as labels, train the residual compensation capability of KRC-Net; 4. Establish a working condition model library: Store the integrated model of "EKF + dual-level optimization + SAN + KRC-Net" according to three working conditions: rising, stable and falling. Record the optimal parameters (such as RLS forgetting factor, PSO particle optimal position, SAN / KRC-Net weights) under each working condition for online prediction.

[0053] III. The complete implementation process of online forecasting; Step 1: Real-time data acquisition and operating condition assessment; Real-time data is collected every 5 minutes: Calculate the load change rate to determine the load condition, and call the "load baseline model" from the model library.

[0054] Step 2: EKF two-level optimization and short-term forecasting; Layer 1 (RLS Parameter Update): Input historical data from the past 20 sampling steps and calculate the residuals. This triggers an RLS update; Layer 2 (PSO Noise Optimization): The optimal PSO is obtained after 50 iterations, taking the EKF prediction error from the past 100 sampling steps as input. ; EKF Short-Term Prediction: Calculates the state estimate for the current step based on the updated parameters. Output the predicted bed temperature values ​​for the next 1-5 steps.

[0055] Step 3: SAN Long-Term State Prediction: Input a multivariate sequence of the past 20 sampling steps; SAN outputs critical state predictions for the next 6-20 steps.

[0056] Step 4: EKF long-term initial prediction: Input the critical states predicted by SAN for the next 6-20 steps into EKF to generate initial long-term time-domain bed temperature predictions.

[0057] Step 5: KRC-Net Residual Compensation and Final Prediction: Input the residual sequence, state sequence, and input sequence of the past 10 sampling steps; KRC-Net outputs the residual compensation terms for the next 6-20 steps: Final bed temperature prediction EKF initial predictions + KRC-Net compensation terms.

[0058] The method in this embodiment includes the following steps: S1. Constructing an Extended Kalman Filter (EKF) based on a bed temperature thermodynamic model and designing a two-level online optimization strategy. Layer 1 adaptively updates the model parameters through Recursive Least Squares (RLS), while Layer 2 fine-tunes the EKF noise covariance matrix through Particle Swarm Optimization (PSO) to achieve accurate short-term bed temperature estimation; S2. Constructing a State Prediction Network (SAN), using a CNN-BiLSTM hybrid architecture and introducing a physical consistency loss function to predict key system states such as coal feed, flue gas oxygen content, air intake, ash discharge, and working fluid temperature required by the EKF, thus extending the time domain of bed temperature prediction; S3. Constructing a Kalman Residual Compensation Network (KRC-Net), using the EKF output residuals, system states, and input sequences as inputs, learning the residual patterns to correct model bias; S4. Based on historical operating data of a 350MW SCFB boiler in a power plant in Shandong, optimizing the EKF parameters, SAN, and KRC-Net. Offline training is performed to establish an initial prediction model library. In step S5, during boiler operation, a two-level optimization strategy is triggered based on real-time operating data to update the EKF. Combining the future state predicted by SAN with the residual compensation of KRC-Net, online bed temperature prediction under all operating conditions is achieved. This embodiment can achieve high-precision prediction of bed temperature under rapid load changes (load increase, load stabilization, load decrease), with a mean absolute error (MAE) as low as 1.5364℃ and a root mean square error (RMSE) of 1.9738℃, effectively suppressing the accumulation of long-term prediction errors, improving model robustness and adaptability to operating conditions, and providing a reliable basis for boiler combustion optimization and safety control.

[0059] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A bed temperature prediction method suitable for fast load change conditions of a supercritical circulating fluidized bed boiler, characterized in that, include: Historical operating data of the SCFB boiler during operation is input into the state prediction network model to predict the key system states required by the extended Kalman filter. The state prediction network model is built based on a CNN-BiLSTM hybrid network, and the extended Kalman filter is built based on a bed temperature thermodynamic model. The key system states include: coal feed rate, flue gas oxygen content, air intake rate, ash discharge rate, and working fluid temperature. The real-time operating data and key system status during the operation of the SCFB boiler are input into the extended Kalman filter, and a two-level online optimization strategy is adopted to generate the initial predicted value of the long-term time domain bed temperature. The two-level online optimization strategy includes: Layer 1: Construct an offline model library covering load increase, load stabilization, and load decrease conditions; when the SCFB boiler operating condition meets the triggering conditions, select a matching baseline model from the offline model library, minimize the residual optimization objective function through recursive least squares method, and update the parameters of the bed temperature thermodynamic model. Layer 2: Optimize and fine-tune the noise covariance matrix of the extended Kalman filter through particle swarm optimization algorithm to achieve short-term estimation of SCFB boiler bed temperature. The Kalman residual compensation network is used to perform residual compensation on the initial predicted value of the long-term bed temperature to obtain the final predicted value of the bed temperature.

2. The bed temperature prediction method for fast load change conditions of supercritical circulating fluidized bed boilers according to claim 1, characterized in that, The CNN-BiLSTM hybrid network employs 1D convolutional layers to extract local correlation features from multivariate time series, combined with max pooling downsampling to remove redundant information and retain key features. Subsequently, a dropout layer is used to randomly discard some features to enhance the model's generalization ability. The BiLSTM part uses two independent LSTM networks, forward and backward, to process the feature sequences output by the CNN from the forward and backward time series respectively, comprehensively capturing bidirectional time dependencies. Finally, a fully connected layer maps the concatenated bidirectional global features to predicted values ​​of the key system states of the CFB boiler.

3. The bed temperature prediction method for supercritical circulating fluidized bed boilers under rapid load change conditions according to claim 1, characterized in that, The state prediction network model introduces a physical consistency loss function to predict the key system states required by the extended Kalman filter, thereby extending the time domain of bed temperature prediction. The physical consistency loss function includes: a prediction error term, a physical residual term, and a regularization term. The prediction error term is the squared difference between the predicted output of the state prediction network model and the actual critical system state. The physical residual term is the squared difference between the actual critical system state and the output of the bed temperature thermodynamic model. The regularization term is the L2 norm of the trainable parameters of the state prediction network model and the parameters of the bed temperature thermodynamic model; and is weighted by coefficients. The contribution of each of the three items is controlled separately.

4. The bed temperature prediction method for supercritical circulating fluidized bed boilers under rapid load change conditions according to claim 1, characterized in that, The bed temperature thermodynamic model is constructed based on the principles of mass conservation, energy conservation, and thermodynamic equilibrium. The bed temperature thermodynamic model includes: The coal feed rate model is used to characterize the coal feeding delay; the carbon balance model is used to characterize the cyclic combustion process of unburned carbon; the oxygen concentration change model is used to characterize the relationship between combustion performance and oxygen consumption; and the bed temperature energy balance model is used.

5. The bed temperature prediction method for supercritical circulating fluidized bed boilers under rapid load change conditions according to claim 1, characterized in that, Each particle in the particle swarm optimization algorithm represents a candidate parameter vector of the noise covariance matrix of the extended Kalman filter. The particle velocity and position are iteratively updated based on the optimal position of the particle and the optimal position of the swarm to minimize the prediction error of the extended Kalman filter for bed temperature. The inertia weight of the particle swarm optimization algorithm adopts a linear decreasing method.

6. The bed temperature prediction method for supercritical circulating fluidized bed boilers under rapid load change conditions according to claim 1, characterized in that, The triggering conditions for the dual-level online optimization strategy are: the real-time load change rate exceeds a preset threshold, or the bed temperature prediction error of the extended Kalman filter exceeds a set tolerance.

7. The bed temperature prediction method for supercritical circulating fluidized bed boilers under rapid load change conditions according to claim 1, characterized in that, The residual compensation for the initial long-term bed temperature prediction using a Kalman residual compensation network includes: Input the preset windowed sequence into the Kalman residual compensation network to obtain the residual compensation term; The initial predicted value of the long-term bed temperature is added to the residual compensation term to obtain the final predicted value of the bed temperature.

8. The bed temperature prediction method for supercritical circulating fluidized bed boilers under rapid load change conditions according to claim 7, characterized in that, The preset windowed sequence includes: an extended Kalman filter residual sequence spliced ​​by time windows, a system state vector, and a control input vector; The Kalman residual compensation network adopts the same CNN-BiLSTM hybrid network architecture as the state prediction network model.

Images