CN-122015084-A - Circulating fluidized bed temperature prediction method based on dynamic correlation guided graph space-time learning

CN122015084ACN 122015084 ACN122015084 ACN 122015084ACN-122015084-A

Abstract

The invention discloses a circulating fluidized bed temperature prediction method based on dynamic correlation-guided graph space-time learning, and belongs to the technical field of industrial process control and energy power engineering. Aiming at the problem that the existing bed temperature prediction method fails to simultaneously consider the space coupling relation and time evolution characteristics between operation variables to cause insufficient prediction precision and robustness, the method adopts multi-strategy characteristic selection to screen key input from candidate variables, constructs a graph convolution network self-adaptive depicting variable topological structure guided by dynamic correlation and extracts space characteristics, fuses long-term and short-term memory network modeling time dynamic characteristics, designs multi-objective loss function optimization model training comprising fitting errors, trend consistency and output smoothness, and finally realizes high-precision stable bed temperature prediction of the circulating fluidized bed boiler under complex working conditions, thereby providing reliable basis for operation optimization and intelligent control.

Inventors

QIAO JUNFEI
MA WENHUI
CHEN GUANGWEI
GE QINGPENG
WANG ZIPENG
XU MINGYUE
LI ZHU
LI MAOJING

Assignees

北京工业大学

Dates

Publication Date: 20260512
Application Date: 20260131

Claims (8)

1. The circulating fluidized bed temperature prediction method based on dynamic correlation-guided graph space-time learning is characterized by comprising the following steps of: S1, screening out key variables related to bed temperature height from historical operation data of a circulating fluidized bed boiler by adopting a multi-strategy feature selection method, and inputting the key variables as a model; s2, constructing a dynamic weighted graph structure by taking the key variable as a node, and fusing and extracting spatial features and time features based on a graph convolution network and a long-term and short-term memory network; S3, training a network by using a multi-objective loss function, wherein the multi-objective loss function combines a fitting error, a trend consistency error and an output smoothness error; And S4, deploying the trained model to a circulating fluidized bed boiler control system, and outputting a bed temperature predicted value according to real-time operation data.
2. The circulating fluidized bed temperature prediction method based on dynamic correlation-guided graph space-time learning of claim 1, wherein in S1, the process of the multi-strategy feature selection method comprises: removing the characteristics with variance smaller than a preset threshold by using a variance threshold method; calculating a Pearson correlation coefficient of the reserved characteristic and the bed temperature, and filtering out the characteristic that the absolute value of the correlation coefficient is larger than a preset threshold value; And (3) carrying out importance scoring on the residual features by using a random forest regression model, selecting the features with the highest scores, and adding the historical bed temperature as input.
3. The circulating fluidized bed temperature prediction method based on dynamic correlation-guided graph space-time learning of claim 1, wherein in S2, constructing a dynamically weighted graph structure comprises: Defining a node set, wherein the node set comprises primary air quantity, secondary air quantity, main steam flow, coal supply quantity, generator power, economizer inlet flue gas temperature, hearth left side bed pressure, slag cooler inlet temperature and historical bed temperature; initializing an edge weight based on the Pearson correlation coefficient among the variables, setting the edge weight as the absolute value of the correlation coefficient if the absolute value of the correlation coefficient is larger than a set threshold value, otherwise, setting the edge weight as zero; In the training process, the edge weight matrix is set as a trainable parameter, and the graph topology is dynamically updated through back propagation.
4. The method for circulating fluidized bed temperature prediction based on dynamic correlation-guided graph space-time learning of claim 1, wherein in S2, extracting spatial features based on a graph convolution network comprises: information propagation and aggregation are carried out by using the symmetrical normalized adjacency matrix; node characteristics are extracted through at least two layers of graph roll lamination, and a ReLU activation function is adopted; And fusing the spatial characteristics output by the graph convolution network with the original input characteristics.
5. The circulating fluidized bed temperature prediction method based on dynamic correlation-guided graph space-time learning of claim 1, wherein in S2, fusing long-term and short-term memory network extraction time features comprises: inputting the sequence after the spatial features are fused into a long-short-period memory network; The long-term and short-term memory network is provided with at least two hidden layers, the dimension of each hidden layer is 64, and Dropout is used for preventing overfitting; outputting the hidden state of the last step of the long-term and short-term memory network, and mapping the hidden state into a bed temperature predicted value through a full-connection layer.
6. The method for circulating fluidized bed temperature prediction based on dynamic correlation-guided graph space-time learning of claim 1, wherein in S3, the multi-objective loss function comprises: Calculating the mean square error of the predicted value and the true value by fitting the error loss; Calculating the difference of the predicted value and the first-order difference sign of the true value according to the trend consistency loss; Outputting a sum of squares of second-order differences of the predicted value calculated by the smoothness loss; The total loss is a weighted sum of the three losses, and the weighting coefficients are determined by grid search.
7. The method for circulating fluidized bed temperature prediction based on dynamic correlation-guided graph space-time learning of claim 1, wherein in S4 deploying the trained model to a circulating fluidized bed boiler control system comprises: The model receives real-time operation parameter input, including primary air quantity, secondary air quantity, main steam flow, coal supply quantity, generator power, inlet flue gas temperature of the economizer, left side bed pressure of a hearth, inlet temperature of the slag cooler and historical bed temperature, and outputs bed temperature predicted values at future set moments for boiler control and optimization.
8. The circulating fluidized bed temperature prediction method of dynamic correlation-guided graph space-time learning of claim 1, wherein in S1, the preprocessing process of the historical operating data comprises: Removing abnormal values in the original data by setting a threshold value; Filling the missing value by adopting a linear interpolation method; Normalizing the processed data to a [0,1] interval; The normalized data set is divided into a training set, a validation set and a test set.

Description

Circulating fluidized bed temperature prediction method based on dynamic correlation guided graph space-time learning Technical Field The invention belongs to the technical field of industrial process control and energy power engineering, and particularly relates to a circulating fluidized bed temperature prediction method based on dynamic correlation-guided graph space-time learning. Background In the field of thermal power generation, as the installed capacity of renewable energy sources such as wind energy, photovoltaic power generation and the like continuously increases, the inherent intermittence and fluctuation of the renewable energy sources form challenges to the stability of a power grid, so that the thermal power generation still bears key functions such as basic load, peak regulation, frequency modulation, emergency standby and the like in a quite long period in the future. In order to realize more efficient and clean thermal power conversion, the circulating fluidized bed combustion technology has become a main flow technical route of medium and small-sized cogeneration, low-calorific-value fuel utilization and solid waste energy recovery due to the advantages of strong fuel adaptability, high combustion efficiency, low pollutant emission, flexible load adjustment and the like. In practical application, the excessive high bed temperature can cause coking and corrosion, and the incomplete combustion and efficiency reduction are easily caused, so that the accurate prediction of the bed temperature is realized, and the method has important significance for identifying abnormal working conditions in advance, optimizing a control strategy and ensuring the safe and stable operation of a unit. At present, the bed temperature modeling method mainly comprises two types of mechanism modeling and data driving modeling. The mechanism model is usually based on a thermal balance and reaction dynamics equation, and has physical interpretability, but the mechanism model has the problems of complex modeling, difficult parameter calibration, poor adaptability and the like in an actual industrial scene, and is difficult to cope with dynamic working conditions such as fuel type switching or severe load fluctuation. However, the existing data driving method mostly regards each operation variable such as primary air quantity, secondary air quantity, coal supply quantity and the like as independent input, cannot effectively describe the space coupling relation between the variables, and simultaneously lacks sufficient modeling of dynamic response and time lag characteristics of the system, which results in insufficient prediction precision and robustness under complex operation conditions. For example, in a load fluctuation or sensor noise interference scenario, the existing method is prone to generate phase deviation or non-physical jitter, and cannot provide stable and reliable multi-step prediction results, so that practical application of the method in real-time control and optimization is limited. Therefore, a bed temperature prediction method capable of capturing spatial correlation and time evolution characteristics between variables simultaneously is needed to solve the limitation of the prior art in practical engineering. Disclosure of Invention In order to solve the technical problems, the invention provides a circulating fluidized bed temperature prediction method based on dynamic correlation-guided graph space-time learning, which realizes high-precision and strong-robustness bed temperature prediction under complex scenes such as load fluctuation, fuel switching, sensor noise and the like and provides reliable technical support for abnormal early warning, combustion optimization and intelligent control of a circulating fluidized bed boiler. To achieve the above object, the present invention provides a circulating fluidized bed temperature prediction method based on dynamic correlation-guided graph space-time learning, comprising: S1, screening out key variables related to bed temperature height from historical operation data of a circulating fluidized bed boiler by adopting a multi-strategy feature selection method, and inputting the key variables as a model; s2, constructing a dynamic weighted graph structure by taking the key variable as a node, and fusing and extracting spatial features and time features based on a graph convolution network and a long-term and short-term memory network; S3, training a network by using a multi-objective loss function, wherein the multi-objective loss function combines a fitting error, a trend consistency error and an output smoothness error; And S4, deploying the trained model to a circulating fluidized bed boiler control system, and outputting a bed temperature predicted value according to real-time operation data. Optionally, in S1, the process of the multi-policy feature selection method includes: removing the characteristics with variance smaller than a preset thre