Circulating fluidized bed boiler bed temperature prediction method based on Koopman operator

Through the Koopman operator prediction network, combined with time series encoding and evolution operator module, the problems of low accuracy and response delay of circulating fluidized bed boiler bed temperature prediction are solved, and fast and accurate bed temperature prediction is achieved, supporting the stable operation of the boiler and environmental protection.

CN119026478APending Publication Date: 2024-11-26ELECTRIC POWER RES INST STATE GRID SHANXI ELECTRIC POWER +2
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411217164.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-02
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

The existing circulating fluidized bed boiler bed temperature prediction technology relies on traditional physical models and statistical methods, and has problems such as low accuracy, delay in response and poor adaptability, making it difficult to accurately predict bed temperature in a dynamic environment.

Method used

Using the circulating fluidized bed boiler bed temperature prediction method based on Koopman operator, the Koopman operator prediction network is constructed, combined with the time series encoding module, the autogenous evolution operator approximation module and the exogenous evolution operator approximation module, the bed temperature is achieved in a synergistic way. The latent state prediction and the latent variables are converted into the original state space through the decoder module, providing accurate bed temperature prediction.

Benefits of technology

It significantly improves the response speed and accuracy of bed temperature prediction, can quickly adapt to boiler operation changes, provide more accurate prediction results, and supports the stable operation and environmental protection of the boiler.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119026478A_ABST
    Figure CN119026478A_ABST
Patent Text Reader

Abstract

The invention discloses a circulating fluidized bed boiler bed temperature prediction method based on a Koopman operator, and the method comprises the steps: firstly obtaining the operation historical data of a circulating fluidized bed boiler, carrying out the preprocessing, carrying out the normalization of the preprocessed historical data, carrying out the segmentation according to a time window, and obtaining a historical time window sequence data set; constructing a Koopman operator prediction network, training the Koopman operator prediction network by using a historical time window sequence data set, and adjusting hyper-parameters in the Koopman operator prediction network to obtain a bed temperature prediction model of the circulating fluidized bed boiler; and S3, inputting the collected real-time data of the circulating fluidized bed boiler into the bed temperature prediction model obtained in the step S3 to obtain a predicted value of the bed temperature at the future moment. According to the method, the bed temperature of the circulating fluidized bed boiler can be effectively and accurately predicted, and the stability and safety of unit operation are improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of boiler operation control, and in particular to a circulating fluidized bed boiler bed temperature prediction method based on Koopman operator. Background Art

[0002] Circulating fluidized bed (CFB) boilers have become an important part of the modern energy industry due to their high combustion efficiency and excellent low emission characteristics. CFB technology achieves continuous combustion of fuel and efficient heat transfer through the circulation and fluidization of bed materials, effectively reducing the emission of nitrogen oxides and sulfides. In addition, CFB boilers can adapt to a variety of fuels, such as low-quality coal, biomass fuels and waste, so they have high flexibility and broad application prospects.

[0003] Bed temperature is a key parameter in the operation of CFB boilers, which has a direct impact on combustion efficiency, emission level and equipment safety. Therefore, accurate prediction and control of bed temperature is crucial to ensure stable operation of the boiler and environmental protection. However, real-time and accurate prediction of bed temperature has always been a technical challenge, because bed temperature is affected by many factors such as fuel type, load changes, air supply, etc., and exhibits highly nonlinear and dynamic characteristics.

[0004] Existing bed temperature prediction technologies mainly rely on traditional physical models and statistical methods. Although these methods are effective in some cases, they are often limited by the model's assumptions and high computational complexity. For example, physical models rely heavily on accurate initial and boundary conditions, which are difficult to obtain in actual operations; statistical methods rely on a large amount of historical data, have poor adaptability to new changes or abnormal conditions, and often have prediction delays when responding to sudden changes in operating conditions, resulting in reduced accuracy in actual applications. Summary of the invention

[0005] The purpose of the present invention is to provide a circulating fluidized bed boiler bed temperature prediction method based on the Koopman operator, which can quickly and accurately respond to the prediction scheme of boiler operation changes and significantly improve the response speed and accuracy of the prediction.

[0006] To achieve the above object, the technical solution adopted by the present invention is:

[0007] A circulating fluidized bed boiler bed temperature prediction method based on Koopman operator comprises the following steps:

[0008] S1, obtaining historical data of the operation of the circulating fluidized bed boiler and preprocessing it, normalizing the preprocessed historical data, and segmenting it by time window to obtain a historical time window sequence data set, wherein the historical data includes: bed temperature, total coal volume, total air volume, and outlet flue gas oxygen content;

[0009] S2, constructing a Koopman operator prediction network, wherein the Koopman operator prediction network includes a time series encoding module, a self-evolution operator approximation module, an exogenous evolution operator approximation module and a decoder module;

[0010] The time series encoding module is used to encode the historical sequence information of bed temperature and the exogenous variable information of coal supply, air supply and oxygen supply into the potential invariant subspace of Koopman;

[0011] The self-evolution operator approximation module and the exogenous evolution operator approximation module are used to respectively evolve the encoded potential information of the bed temperature and the potential information of the exogenous variable, and the potential state prediction of the next step of the bed temperature in the potential invariant subspace is realized through the synergistic effect of the self-evolution operator approximation module and the exogenous evolution operator approximation module;

[0012] The decoder module is used to convert the bed temperature potential state prediction in the latent variable invariant subspace into the original nonlinear state space to obtain the prediction result of the bed temperature of the circulating fluidized bed boiler at a future moment;

[0013] S3, using the historical time window series data set obtained in step S1 to train the Koopman operator prediction network, and adjusting the hyperparameters in the Koopman operator prediction network to obtain a bed temperature prediction model for a circulating fluidized bed boiler;

[0014] S4, using the collected real-time data of the circulating fluidized bed boiler, inputting it into the bed temperature prediction model obtained in step S3 to obtain the predicted value of the bed temperature at a future moment.

[0015] As a preferred embodiment of the present invention, in step S1, the data preprocessing method includes removing abnormal values ​​of historical data and filling missing values, and the method of normalizing historical data is as follows:

[0016]

[0017] Among them, x max and x min are the maximum and minimum values ​​in the data, respectively, and x′ i represents the normalized data, x i Represents the data before processing.

[0018] As a preferred embodiment of the present invention, in step S2, when constructing the time series encoding module, the state information of the time series is first extracted using the GRU network, and then mapped to the Koopman potential subspace through the fully connected network. The calculation formula of the GRU network and the fully connected network of the time series encoding module is:

[0019] z t=σ(W z ·[h t-1 ,x t ]+b z )

[0020] r t =σ(W r ·[h t-1 ,x t ]+b r )

[0021]

[0022] h k =MLP(h t )

[0023] Among them, W represents the weight, b represents the bias term, and σ is the sigmoid activation function, which is expressed as The expression of tanh activation function is x t represents the current input of the network, h t-1 Represents the output of the previous moment in the network cycle, z t is the update gate that controls how much past information is passed to the future, r t To reset the gate, it determines whether to combine the new input information with the previous memory. The hidden state h at the current moment t The hidden state of the previous moment after selective forgetting (1-z t )⊙h t-1 and selective memory of the candidate hidden transition at the current moment Composition, h k represents the Koopman potential subspace state, h k =MLP(h t ) represents mapping the hidden state into the latent subspace through a fully connected network.

[0024] As a preferred implementation of the present invention, the hidden layer state value of the GRU network is 8, and the number of stacked layers is 2.

[0025] As a preferred embodiment of the present invention, in step S2, the self-evolution operator approximation module A and the exogenous evolution operator approximation module B are constructed, both of which are implemented by a single-layer fully connected network without bias; the Koopman operator evolution process is evolved in the form of a fourth-order Runge-Kutta iteration, and the calculation formula is as follows:

[0026] k1=(Ah k +Bh uk )·△At

[0027]

[0028] k4=(A(h k +k3)+Bh uk )·△t

[0029]

[0030] Among them, A is the approximate matrix of the self-evolution operator, B is the approximate matrix of the exogenous evolution operator, and h k is the potential initial state obtained by the time series encoder from the historical series data of bed temperature at time k, h uk is the potential exogenous state of the historical series data of the historical exogenous variable at time k through the time series encoder, and Δt is the Runge-Kutta iteration time of 1.

[0031] As a preferred implementation manner of the present invention, the dimensions of the approximate matrix A of the self-generated evolution operator and the approximate matrix B of the exogenous evolution operator are 64 dimensions.

[0032] As a preferred embodiment of the present invention, in step S2, when constructing the decoder module, the decoder network is implemented by a deep feedforward fully connected network, and its calculation formula is:

[0033] y k =f(W k y k-1 +b k )

[0034] Among them, W is the weight matrix, b is the bias term, and y k is the output of the kth layer of the deep feedforward fully connected network, y k-1 is the output of the k-1th layer of the deep feedforward fully connected network, and f indicates that the linear activation function is the ReLU function;

[0035] The loss function used in the Koopman operator learning network training is composed of:

[0036]

[0037] Among them, L pred represents the long-term prediction loss from time k+1 to time k+H, L ae Represents the prediction loss of the encoder and decoder. Since the present invention adopts a time series encoder, L ae represents the encoding-decoding loss of the initial state, L linear represents the linear evolution loss of the Koopman invariant subspace, x k and represent the true value and predicted value of the bed temperature information at time k, respectively, and h k and They represent the true coded value and predicted value of the Koopman invariant subspace latent variable at time k respectively;

[0038] All loss terms are weighted together to obtain the total loss function of the Koopman operator learning network, which is expressed as follows:

[0039] Loss = λ pred L pred +λ ae L ae +λ linear L linear

[0040] Among them, λ pred , ae , linear is the weight used to control each loss function.

[0041] As a preferred implementation manner of the present invention, the number of hidden layers of the decoder deep feedforward fully connected network is 4, and the number of nodes in the hidden layer is 128.

[0042] As a preferred embodiment of the present invention, the pred , ae , linear Take 1, 0.1, and 0.1 respectively.

[0043] As a preferred implementation of the present invention, in step S3, the historical time window series data set obtained in step S1 is divided into a training set and a validation set, the trained Koopman operator prediction network is verified with the validation set, and an ablation experiment is performed to adjust the network hyperparameters until the predetermined conditions are met, thereby obtaining a trained CFB boiler bed temperature prediction model.

[0044] The present invention proposes a circulating fluidized bed boiler bed temperature prediction method based on the Koopman operator. The method simultaneously considers the bed temperature information of the circulating fluidized bed and the exogenous variable information of the coal supply rate, air supply rate, and outlet flue gas oxygen content. Through the joint action of the autogenous evolution operator and the exogenous evolution operator, the accurate prediction of the CFB boiler bed temperature is achieved.

[0045] The prediction method proposed in the present invention breaks through the limitations of traditional prediction technology in dynamic environments, provides a prediction scheme that quickly and accurately responds to changes in boiler operation, significantly improves the response speed and accuracy of the prediction, provides strong support for boiler operation control and optimization, and is of great significance for promoting clean, efficient and safe operation in the field of thermal power generation. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1The overall flow chart of the circulating fluidized bed boiler bed temperature prediction method based on Koopman operator proposed in the present invention;

[0047] Figure 2 A schematic diagram of the Koopman operator evolution transformation proposed in the present invention;

[0048] Figure 3 This is a framework diagram of the circulating fluidized bed boiler bed temperature prediction based on the Koopman operator proposed in the present invention. DETAILED DESCRIPTION

[0049] The basic concept of the present invention is to establish a bed temperature prediction model based on the Koopman operator theory, which can capture and predict the rapid changes of bed temperature in real time and adapt to the dynamic characteristics under different fuel and load conditions. In addition, the model also takes into account the influence of exogenous variable information such as coal supply, air supply, and outlet flue gas oxygen content on the bed temperature. Through the joint action of the self-evolution operator (responsible for processing the evolution of the internal state of the system) and the exogenous evolution operator (responsible for processing the influence of exogenous variables on the state), accurate prediction of the bed temperature is achieved.

[0050] The method of the present invention not only takes into account the specific operating parameters of the CFB boiler, but also utilizes the high adaptability and predictive ability of the Koopman operator to effectively address the challenge of accurate bed temperature prediction faced by CFB boilers in complex industrial applications.

[0051] Among them, the core advantage of the Koopman operator theory is that it can handle nonlinear complex dynamic systems through linearization means, allowing the dynamic characteristics of the system to be extracted directly from real-time measurement data. This is especially important for systems such as CFB boilers where operating conditions change frequently, because the Koopman operator can reflect the dynamic response of the bed temperature to changes in operating conditions in real time without relying on pre-set physical models or large amounts of historical data.

[0052] like Figure 1 , Figure 2 , Figure 3 As shown, this embodiment provides a circulating fluidized bed boiler bed temperature prediction method based on Koopman operator, including the following steps S1 to S4.

[0053] Step S1, obtaining historical data of CFB boiler operation for preprocessing, normalizing the preprocessed historical data, and segmenting the data by time window to obtain a historical time window series data set.

[0054] The historical data of CFB boiler operation obtained in this step include: bed temperature, total coal volume, total air volume, and outlet flue gas oxygen content.

[0055] The data preprocessing method includes outlier removal, missing value filling, etc. In order to ensure the continuity of historical data, the outlier removal and missing value filling are both replaced by the average value of adjacent samples.

[0056] For example, consider the following historical data sample:

[0057] Bed temperature (℃): 853.9, 854.0, 999.0, 854.0, 854.0, 853.9

[0058] Total coal volume (t / h): 170.76, 171.37, 168.72, 171.24, 171.28, 172.95

[0059] Total air volume (kNm 3 / h): 531.29, 528.39, 529.29, 535.08, 533.32, 535.48

[0060] Outlet flue gas oxygen content (%): 2.83, 2.83, 2.83, 2.82, null, 2.82

[0061] Specifically, the preprocessing of the above historical data is as follows:

[0062] The 999.0 in the bed temperature is obviously an abnormal value, which is replaced by the average of the two adjacent values, that is, 854.0; the null in the outlet flue gas oxygen content is a missing value, which is filled with the average of the adjacent values, that is, 2.82.

[0063] The method for normalizing the preprocessed historical data is as follows:

[0064]

[0065] Among them, x max and x min are the maximum and minimum values ​​in the data, respectively, and x′ i represents the normalized data, x i Indicates the data before processing. The maximum and minimum values ​​mentioned above are set by the operation expert in engineering application.

[0066] The maximum values ​​of the historical data are: bed temperature 952.9℃, total coal volume 305.58t / h, total air volume 993.08kNm 3 / h, outlet flue gas oxygen content 7.38%; minimum value: bed temperature 844.3℃, total coal volume 127.54t / h, total air volume 452.75e kNm 3 / h, the oxygen content of the outlet flue gas is 0.25%, so the normalized data is:

[0067] Bed temperature (normalized): 0.0884, 0.0893, 0.0893, 0.0893, 0.0893, 0.0884

[0068] Total coal volume (normalized): 0.2427, 0.2461, 0.2313, 0.2455, 0.2457, 0.2551

[0069] Total air volume (normalized): 0.1454, 0.1399, 0.1417, 0.1524, 0.1491, 0.1531

[0070] Outlet flue gas oxygen content (normalized): 0.3619, 0.3619, 0.3619, 0.3604, 0.3604, 0.3604

[0071] Normalized data can be used more effectively by the Koopman operator prediction network because it eliminates the impact of different dimensions and numerical ranges, allowing the network to focus more on the inherent dynamic changes of the data.

[0072] Step S2, constructing a Koopman operator prediction network, wherein the Koopman operator prediction network includes a time series encoding module, a self-evolved operator approximation module, an exogenous evolution operator approximation module and a decoder module.

[0073] The time series encoding module is used to encode the historical sequence information of the bed temperature and the exogenous variable information of the coal supply, air supply and oxygen supply into the Koopman's latent invariant subspace. The self-evolution operator approximation module and the exogenous evolution operator approximation module are used to evolve the encoded latent information of the bed temperature and the latent information of the exogenous variables respectively, and through the synergistic effect of the self-evolution operator approximation module and the exogenous evolution operator approximation module, the latent state prediction of the bed temperature in the latent invariant subspace is realized. The decoder module is used to convert the latent state prediction of the bed temperature in the latent variable invariant subspace into the original nonlinear state space to obtain the prediction result of the CFB boiler bed temperature at the future moment.

[0074] In a relatively specific embodiment of the present invention, constructing a Koopman operator prediction network includes steps S2.1 to S2.3.

[0075] S2.1 constructs the time series encoding module.

[0076] In order to fully extract the information of the historical sequence of states, the time series encoder module first uses a GRU network to extract the state information of the time series, and then maps it to the Koopman latent subspace through a fully connected network.

[0077] Specifically, the calculation formulas of the GRU network and the fully connected network of the time series encoding module are as follows:

[0078] z t =σ(W z ·[h t-1 , x t ]+b z )

[0079] rx=σ(W r ·[h t-1 ,x t ]+b r )

[0080]

[0081] h k =MLP(h t )

[0082] Among them, W represents the weight, b represents the bias term, and σ is the sigmoid activation function, which is expressed as The expression of tanh activation function is x t represents the current input of the network, h t-1 Represents the output of the previous moment in the network cycle, z t is the update gate that controls how much past information is passed to the future, r t To reset the gate, it determines whether to combine the new input information with the previous memory. The hidden state h at the current moment t The hidden state of the previous moment after selective forgetting (1-z t )⊙h t-1 and selective memory of the candidate hidden transition at the current moment Composition, h k represents the Koopman potential subspace state, h k =MLP(h t ) means mapping the hidden state to the potential subspace through a fully connected network; the hidden layer state of the GRU network is 8, and the number of stacked layers is 2.

[0083] S2.2 constructs the self-evolution operator approximation module and the exogenous evolution operator approximation module.

[0084] For the convenience of calculation, both the self-evolution operator approximation module and the exogenous evolution operator approximation module are implemented using a single-layer fully connected network without bias.

[0085] Specifically, in order to improve the evolution accuracy of the Koopman operator, the operator evolution process is evolved through the fourth-order Runge-Kutta iteration form, and its calculation formula is:

[0086] k1=(Ah k +Bh uk )·Δt

[0087]

[0088] k4=(A(h k +k3)+Bh uk )·Δt

[0089]

[0090] Among them, A is the approximate matrix of the self-evolution operator, B is the approximate matrix of the exogenous evolution operator, and h k is the potential initial state obtained by the time series encoder from the historical series data of bed temperature at time k, h uk is the potential exogenous state of the historical series data of the historical exogenous variables at time k through the time series encoder, and Δt is the Runge-Kutta iteration time; among them, the dimensions of matrices A and B are 64, and the value of Δt is 1.

[0091] S2.3 constructs the decoder module.

[0092] For ease of implementation, the decoder network is implemented through a deep feedforward fully connected network, and its calculation formula is:

[0093] y k =f(W k y k-1 +b k )

[0094] Among them, W is the weight matrix, b is the bias term, and y k is the output of the kth layer of the deep feedforward fully connected network, y k-1 is the output of the k-1th layer of the deep feedforward fully connected network, and f is the nonlinear activation function; the number of hidden layers of the deep feedforward fully connected network is 4, the number of nodes in the hidden layer is recommended to be 128, and the nonlinear activation function is the Relu function.

[0095] Finally, the present invention designs the following loss function form for the Koopman operator prediction network. The total loss function contains three loss terms, which are as follows:

[0096]

[0097] Among them, L pred represents the long-term prediction loss from time k+1 to time k+H, L ae Represents the prediction loss of the encoder and decoder. Since the present invention adopts a time series encoder, L aerepresents the encoding-decoding loss of the initial state, L linear represents the linear evolution loss of the Koopman invariant subspace, x k and represent the true value and predicted value of the bed temperature information at time k, respectively, and h k and They represent the true coded value and predicted value of the Koopman invariant subspace latent variable at time k respectively.

[0098] All loss functions are weighted together to obtain the total loss function used in Koopman operator prediction network training. The formula is as follows:

[0099] Loss = λ pred L pred +λ ae L ae +λ linear L linear

[0100] Among them, λ pred , ae , linear To control the weights of various loss functions, these three weight parameters can be optimized according to actual needs. In the present invention, they are finally optimized to λ pred , ae , linear They are 1, 0.1, and 0.1 respectively.

[0101] Step S3, using the historical time window sequence data set in S1 to train the Koopman operator prediction network, such as Figure 3 As shown, the hyperparameters in the Koopman operator prediction network are adjusted to obtain a bed temperature prediction model for a CFB boiler. Specifically, the prediction model is:

[0102]

[0103] Among them, T b is the bed temperature, w c is the coal supply, u Air is the total air volume, c O2 is the oxygen content of the outlet flue gas, is the predicted value of bed temperature, (t+1:t+H) represents the data of H future moments, (tL:t) represents the data of L historical moments, and f is the bed temperature prediction model of the CFB boiler. Specifically, L is 500, and H is 30.

[0104] In an embodiment of the present invention, the historical time window series data set can be divided into a training set and a validation set. The trained Koopman operator prediction network is verified using the validation set, and an ablation experiment is performed to adjust the network hyperparameters until predetermined conditions are met, thereby obtaining a trained CFB boiler bed temperature prediction model.

[0105] Specifically, the normalized historical data was divided into 80% training set and 20% validation set. The training set was used to train the Koopman operator prediction network, while the validation set was used to verify the accuracy and generalization ability of the model. During the training process, hyperparameters such as learning rate, batch size, and training rounds were carefully adjusted. After 50 rounds of training, the model achieved a prediction accuracy of 90% on the validation set, showing excellent bed temperature prediction capabilities. In addition, through further parameter optimization, the prediction error of the model was successfully reduced, and its robustness under different operating conditions was enhanced.

[0106] Step S4, using the collected real-time data of the CFB boiler, inputting it into the bed temperature prediction model to obtain the predicted value of the bed temperature at a future moment.

[0107] It should be noted that the real-time data input into the model also needs to be preprocessed accordingly in the form of training data samples. The prediction of the model can be a single-step prediction or a multi-step prediction.

[0108] Specifically, based on the real-time data sample, the historical bed temperature data of L=500 and the historical total coal volume, total air volume, and outlet flue gas oxygen volume are used to predict the bed temperature of the next H=30 steps. The bed temperature data after the 15th step (1 minute) in the future is obtained as follows:

[0109] Predicted values: 857.5276, 857.3882, 857.1380, 856.8527, 856.4286, 856.1042

[0110] Actual values: 858.1, 858.1, 858.1, 858.0, 857.9, 857.7

[0111] By comparing the predicted values ​​with the actual observed values, it can be seen that the model has a strong predictive ability for bed temperature changes in the future.

[0112] In addition, in order to verify the superiority of the circulating fluidized bed boiler bed temperature prediction method based on the Koopman operator in the present invention, several other time series prediction methods in the prior art are used for comparison, and the specific methods are as follows:

[0113] LSTM (Long Short-Term Memory): The LSTM model is a variant of a recurrent neural network that captures long-term dependencies in sequence data through a gating mechanism. LSTM is particularly suitable for processing and predicting time series data with long intervals between important events. Its structure includes three main gates: input gate, forget gate, and output gate, which control the inflow, retention, and outflow of information.

[0114] GRU (Gated Recurrent Unit): GRU is a simplified version of LSTM, which merges the forget gate and input gate in LSTM into a single update gate, and merges the cell state and hidden state, simplifying the structure of the model while maintaining similar performance. GRU manages the flow of information through two gate control mechanisms: update gate and reset gate, which makes it more efficient than LSTM in some tasks.

[0115]

[0116] This example compares the prediction accuracy of LSTM, GRU and the method of the present invention in predicting the bed temperature for the next 30 steps (2 minutes). Compared with the LSTM and GRU models, the prediction accuracy of the method of the present invention is significantly improved. Taking the prediction of bed temperature after 1 minute as an example, the MAE and RMSE of the method of the present invention are improved by 51.58% and 51.13% respectively compared with LSTM, and by 42.2% and 41.68% respectively compared with GRU.

[0117] Finally, it should be noted that the above is only used to illustrate the technical solution of the present invention, and is not used to limit the present invention. For those skilled in the art, the present invention may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A circulating fluidized bed boiler bed temperature prediction method based on Koopman operator, characterized in that: The steps include: S1, obtaining historical data of the operation of the circulating fluidized bed boiler and preprocessing it, normalizing the preprocessed historical data, and segmenting it by time window to obtain a historical time window sequence data set, wherein the historical data includes: bed temperature, total coal volume, total air volume, and outlet flue gas oxygen content; S2, constructing a Koopman operator prediction network, wherein the Koopman operator prediction network includes a time series encoding module, a self-evolution operator approximation module, an exogenous evolution operator approximation module and a decoder module; The time series encoding module is used to encode the historical sequence information of bed temperature and the exogenous variable information of coal supply, air supply and oxygen supply into the potential invariant subspace of Koopman; The self-evolution operator approximation module and the exogenous evolution operator approximation module are used to respectively evolve the encoded potential information of the bed temperature and the potential information of the exogenous variable, and the potential state prediction of the next step of the bed temperature in the potential invariant subspace is realized through the synergistic effect of the self-evolution operator approximation module and the exogenous evolution operator approximation module; The decoder module is used to convert the bed temperature potential state prediction in the latent variable invariant subspace into the original nonlinear state space to obtain the prediction result of the bed temperature of the circulating fluidized bed boiler at a future moment; S3, using the historical time window series data set obtained in step S1 to train the Koopman operator prediction network, and adjusting the hyperparameters in the Koopman operator prediction network to obtain a bed temperature prediction model for a circulating fluidized bed boiler; S4, using the collected real-time data of the circulating fluidized bed boiler, inputting it into the bed temperature prediction model obtained in step S3 to obtain the predicted value of the bed temperature at a future moment.

2. The circulating fluidized bed boiler bed temperature prediction method based on Koopman operator according to claim 1 is characterized in that: In step S1, the data preprocessing method includes removing abnormal values ​​from historical data and filling missing values. The method for normalizing historical data is as follows: Among them, x max and x min are the maximum and minimum values ​​in the data, respectively, and x′ i represents the normalized data, x i Represents the data before processing.

3. The method for predicting bed temperature of a circulating fluidized bed boiler based on Koopman operator according to claim 1 or 2, characterized in that: In step S2, when constructing the time series encoding module, the state information of the time series is first extracted using the GRU network, and then mapped to the Koopman potential subspace through the fully connected network. The calculation formula of the GRU network and the fully connected network of the time series encoding module is: z t =σ(W z ·[h t-1 ,x t ]+b z } r t =σ(W r ·[h t-1 ,x t ]+b r ) h k =MLP(h t ) Among them, W represents the weight, b represents the bias term, and σ is the sigmoid activation function, which is expressed as The expression of tanh activation function is x t represents the current input of the network, h t-1 Represents the output of the previous moment in the network cycle, z t is the update gate that controls how much past information is passed to the future, r t To reset the gate, it determines whether to combine the new input information with the previous memory. The hidden state h at the current moment t The hidden state of the previous moment after selective forgetting (1-z t )⊙h t-1 and selective memory of the candidate hidden transition at the current moment Composition, h k represents the Koopman potential subspace state, h k =MLP(h t ) represents mapping the hidden state into the latent subspace through a fully connected network.

4. The circulating fluidized bed boiler bed temperature prediction method based on Koopman operator according to claim 3 is characterized in that: The hidden layer state value of the GRU network is 8, and the number of stacked layers is 2.

5. The method for predicting bed temperature of a circulating fluidized bed boiler based on Koopman operator according to claim 1 or 4, characterized in that: In step S2, the self-evolution operator approximation module A and the exogenous evolution operator approximation module B are constructed, both of which are implemented by a single-layer fully connected network without bias; the Koopman operator evolution process is evolved through a fourth-order Runge-Kutta iteration form, and the calculation formula is as follows: k1=(Ah k +Bh uk )·Δt k4=(A(h k +k3)+Bh uk )·Δt Among them, A is the approximate matrix of the self-evolution operator, B is the approximate matrix of the exogenous evolution operator, and h k is the potential initial state obtained by the time series encoder from the historical series data of bed temperature at time k, h uk is the potential exogenous state of the historical series data of the historical exogenous variable at time k through the time series encoder, and Δt is the Runge-Kutta iteration time of 1.

6. The circulating fluidized bed boiler bed temperature prediction method based on Koopman operator according to claim 5, characterized in that: The dimensions of the approximate matrix A of the self-evolution operator and the approximate matrix B of the exogenous evolution operator are 64 dimensions.

7. The method for predicting bed temperature of a circulating fluidized bed boiler based on Koopman operator according to claim 1 or 6, characterized in that: In step S2, when constructing the decoder module, the decoder network is implemented by a deep feedforward fully connected network, and its calculation formula is: y k =f(W k y k-1 +b k ) Among them, W is the weight matrix, b is the bias term, and y k is the output of the kth layer of the deep feedforward fully connected network, y k-1 is the output of the k-1th layer of the deep feedforward fully connected network, and f indicates that the linear activation function is the ReLU function; The loss function used in the Koopman operator learning network training is composed of: Among them, L pred represents the long-term prediction loss from time k+1 to time k+H, L ae Represents the prediction loss of the encoder and decoder. Since the present invention adopts a time series encoder, L ae represents the encoding-decoding loss of the initial state, L linear represents the linear evolution loss of the Koopman invariant subspace, x k and represent the true value and predicted value of the bed temperature information at time k, respectively, and h k and They represent the true coded value and predicted value of the Koopman invariant subspace latent variable at time k respectively; All loss terms are weighted together to obtain the total loss function of the Koopman operator learning network, which is expressed as follows: Loss=λ pred L pred +λ ae L ae +λ linear L linear Among them, λ pred , ae , linear is the weight used to control each loss function.

8. The circulating fluidized bed boiler bed temperature prediction method based on Koopman operator according to claim 7, characterized in that: The number of hidden layers of the decoder deep feedforward fully connected network is 4, and the number of nodes in the hidden layer is 128.

9. The method for predicting bed temperature of a circulating fluidized bed boiler based on Koopman operator according to claim 7 or 8, characterized in that: The lambda pred , ae , linear Take 1, 0.1, and 0.1 respectively.

10. The circulating fluidized bed boiler bed temperature prediction method based on Koopman operator according to claim 9, characterized in that: In step S3, the historical time window sequence data set obtained in step S1 is divided into a training set and a validation set. The trained Koopman operator prediction network is verified with the validation set, and an ablation experiment is performed to adjust the network hyperparameters until the predetermined conditions are met, thereby obtaining a trained CFB boiler bed temperature prediction model.