CN-122021435-A - Bed temperature prediction method suitable for rapid load-changing working condition of supercritical circulating fluidized bed boiler

CN122021435ACN 122021435 ACN122021435 ACN 122021435ACN-122021435-A

Abstract

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

Inventors

QIAO JUNFEI
XU MINGYUE
CHEN GUANGWEI
MA WENHUI
GE QINGPENG
WANG ZIPENG
DING HAIXU

Assignees

北京工业大学

Dates

Publication Date: 20260512
Application Date: 20260131

Claims (9)

1. The bed temperature prediction method suitable for the rapid load-changing working condition of the supercritical circulating fluidized bed boiler is characterized by comprising the following steps of: The method comprises the steps of inputting historical working condition data in the operation process of an SCFB boiler into a state prediction network model to predict a key system state required by an extended Kalman filter, wherein the state prediction network model is constructed based on a CNN-BiLSTM hybrid network, the extended Kalman filter is constructed based on a bed thermodynamic model, and the key system state comprises coal feeding amount, flue gas oxygen content, air inlet amount, slag discharging amount and working medium temperature; Inputting real-time working condition data and the key system state in the operation process of the SCFB boiler into the extended Kalman filter, and generating a long-time domain bed temperature initial predicted value by adopting a double-layer online optimization strategy; and carrying out residual error compensation on the long-time-domain bed temperature initial predicted value by using a Kalman residual error compensation network to obtain a final bed temperature predicted value.
2. The bed temperature prediction method suitable for the rapid variable load working condition of the supercritical circulating fluidized bed boiler according to claim 1 is characterized in that a CNN part of the CNN-BiLSTM hybrid network adopts a 1D convolution layer to extract local correlation characteristics of a multi-variable time sequence, performs downsampling in combination with maximum pooling operation to remove redundant information and preserve key characteristics, then randomly discards part of characteristics through a dropout layer to enhance model generalization capability, and BiLSTM part processes characteristic sequences output by the CNN from time sequence forward and reverse respectively through a forward and reverse independent LSTM network to comprehensively capture a bidirectional time dependence, and finally maps spliced bidirectional global characteristics into predicted values of a CFB boiler key system state through a full connection layer.
3. The bed temperature prediction method suitable for the rapid load change working condition of the supercritical circulating fluidized bed boiler according to claim 1, wherein a physical consistency loss function is introduced into the state prediction network model, and the critical system state required by expanding a Kalman filter is predicted through the state prediction network model so as to expand a bed temperature prediction time domain; The physical consistency loss function comprises a prediction error term, a physical residual error term and a regular term, wherein the prediction error term is the square difference between the predicted output of the state prediction network model and the state of a real key system, the physical residual error term is the square difference between the state of the real key system and the output of a bed thermodynamic model, and the regular term is the L2 norm of the trainable parameter of the state prediction network model and the parameter of the bed thermodynamic model and is obtained by weighting coefficients And controlling the contribution degree of the three items respectively.
4. The bed temperature prediction method suitable for a rapid load change condition of a supercritical circulating fluidized bed boiler according to claim 1, wherein the bed thermodynamic model is constructed based on principles of mass conservation, energy conservation and thermodynamic equilibrium, and the bed thermodynamic model comprises: The system comprises a coal feeding rate model for representing coal feeding transmission delay, a carbon balance model for representing the circulating combustion process of unburned carbon, an oxygen concentration change model for representing the relation between combustion performance and oxygen consumption, and a bed temperature energy balance model.
5. The bed temperature prediction method suitable for the rapid load change condition of the supercritical circulating fluidized bed boiler according to claim 1, wherein the double-hierarchy online optimization strategy comprises: And when the SCFB boiler operation condition is detected to meet the triggering condition, selecting a matched baseline model from the offline model library, and updating parameters of the bed thermodynamic model by a recursive least square method and a residual optimization objective function, wherein Layer2 is used for optimizing and fine-adjusting a noise covariance matrix of the extended Kalman filter by a particle swarm algorithm so as to realize short-term estimation of the SCFB boiler bed temperature.
6. The bed temperature prediction method suitable for the rapid load change condition of the supercritical circulating fluidized bed boiler according to claim 5, wherein each particle in the particle swarm algorithm represents a candidate parameter vector of a noise covariance matrix of the extended kalman filter, the particle speed and the position are iteratively updated based on the optimal position and the group optimal position of the particle so as to minimize the prediction error of the extended kalman filter on the bed temperature, and the inertia weight of the particle swarm algorithm adopts a linear decreasing mode.
7. The method for predicting the bed temperature suitable for the rapid load change condition of the supercritical circulating fluidized bed boiler according to claim 5, wherein the triggering condition of the double-level online optimization strategy is that the real-time load change rate exceeds a preset threshold value or the bed temperature prediction error of the extended Kalman filter exceeds a set tolerance.
8. The bed temperature prediction method suitable for the rapid load change condition of the supercritical circulating fluidized bed boiler according to claim 1, wherein performing residual compensation on the long time domain bed temperature initial predicted value by using a kalman residual compensation network comprises: inputting a preset windowed sequence into a Kalman residual error compensation network to obtain a residual error compensation item; And adding the long-time-domain bed temperature initial predicted value and the residual error compensation term to obtain a final bed temperature predicted value.
9. The bed temperature prediction method suitable for the rapid variable load working condition of the supercritical circulating fluidized bed boiler according to claim 8, wherein the preset windowing sequence comprises an extended Kalman filter residual sequence, a system state vector and a control input vector which are spliced according to time windows; the Kalman residual error compensation network adopts a CNN-BiLSTM hybrid network architecture which is the same as the state prediction network model.

Description

Bed temperature prediction method suitable for rapid load-changing working condition of supercritical circulating fluidized bed boiler Technical Field The invention relates to the technical field of intelligent environmental protection, in particular to a bed temperature prediction method suitable for a supercritical circulating fluidized bed boiler under a rapid load-changing working condition. Background Under the background of the global 'double carbon' target and the stringent environmental protection policy, the power system has urgent demands on clean and efficient energy production equipment. Circulating Fluidized Bed (CFB) boiler is realized by virtue of adaptability to inferior fuel, high heat efficiency, and in-furnace desulfurization and other technologies、The advantage of low emission becomes the core equipment in the field of coal power generation and biomass mixed combustion, and has important significance for reducing the environmental impact of energy production. However, with the rapid expansion of renewable energy sources (such as wind power and photovoltaic), the intermittent and fluctuating characteristics place higher demands on the frequency modulation and peak regulation capabilities of the power system—conventional coal-fired power plants (including CFB boiler power plants) need to have a wider load operating range and a faster response speed to maintain the grid stable. The bed temperature is used as a key parameter of a CFB boiler combustion system to directly influence the volatile matter release, the coke reaction rate and the pollutant generation process of the fuel, wherein excessive bed temperature can lead to coking of coal slag, insufficient combustion can be caused if the bed temperature is too low, the combustion efficiency is reduced, and the unburned carbon emission and the pollutant generation process are increasedA risk is generated. Therefore, the accurate prediction of the bed temperature is realized under the working condition of rapid load change, and the method is a precondition for guaranteeing the high-efficiency and clean operation of the CFB boiler. At present, bed temperature prediction techniques are mainly divided into two categories: The physical model (Physics-informed models) is constructed based on principles of conservation of mass, conservation of energy, thermodynamic equilibrium and hydrodynamics, has the advantages of transparent structure and strong physical interpretability, and can describe the dynamic change law of bed temperature through conservation law. However, the model has obvious defects that the bed temperature change of the CFB boiler relates to complex gas-solid two-phase flow and chemical reaction dynamics, is difficult to completely describe through a mathematical equation, and in the rapid load-changing working condition, model parameters (such as heat transfer coefficient and combustion rate) are easy to deviate from design values, so that the prediction accuracy is greatly reduced, and the real-time environment-friendly control requirement cannot be met. The Data driven model (Data-driven models) captures the non-linear and dynamic characteristics of bed temperature from the running Data directly through machine learning and deep learning technology without relying on complex physical mechanisms. The model has certain advantages under stable working conditions, but has two major problems that firstly, the physical consistency constraint is lacking, the predicted result may violate the thermodynamic law (such as the bed temperature predicted value exceeds a reasonable interval), and the environmental protection control strategy is invalid (such as excessive air supply increase)And secondly, errors are easy to accumulate in long-term prediction, particularly under the working condition of rapid load change, the method cannot adapt to the dynamic abrupt change of the system caused by sudden load rise/sudden drop, and is difficult to support long-term environment-friendly operation decision. In order to achieve both physical interpretability and prediction accuracy, the prior art proposes Hybrid physics-based and data-driven models. However, the hybrid models still have the defects that firstly, the model is not optimized for a core scene of 'quick load change', namely, when the load suddenly changes, the prediction of the key system state is delayed, so that the bed temperature prediction deviation is caused, secondly, an effective residual error compensation mechanism is lacked, the system deviation and noise are easy to overlap in long-term prediction, the pollutant control precision is affected, thirdly, the fusion degree of physical constraint and data driving is insufficient, part of the models are simply spliced by two methods, the cooperative optimization is not formed, and the environmental protection operation requirement of the SCFB boiler still cannot be met. In summary, the existing bed temperature prediction technology