All-working-condition online advanced circulating fluidized bed temperature prediction system based on S4 model

Abstract

The invention discloses an all-condition online advanced circulating fluidized bed temperature prediction system based on an S4 model. The system comprises an offline stage and an online stage. The off-line stage is responsible for collecting on-site historical data, selecting and preprocessing the historical data, carrying out interpretable working condition division on the historical data, and training an S4 optimal intelligent model which adapts to the whole working condition and has long-term dependence. The online stage is responsible for collecting and storing real-time data of the system, selecting online data for preprocessing, then achieving online advanced prediction of the bed temperature through the S4 optimal model, and finally conducting online updating and fine adjustment on the S4 model by continuously calculating accumulated errors so as to ensure that the model operates stably and can rapidly adapt to working condition changes. Therefore, safe and stable operation and rapid load change of the circulating fluidized bed are realized, and a foundation is laid for further control system optimization.

Description

Technical Field

[0001] The invention relates to the technical field of bed temperature prediction of a circulating fluidized bed boiler unit, and in particular to a full-operating-condition online advanced circulating fluidized bed bed temperature prediction system based on an S4 model. Background Art

[0002] As renewable energy sources continue to be connected to the power grid, their intermittency and volatility have brought great challenges to the safe and stable operation of thermal power units. Circulating fluidized bed boilers are widely used due to their wide fuel adaptability, high combustion efficiency, low stable combustion load and low pollutant control cost. The key to their stable operation is to maintain the stability of the bed temperature. This is because the bed temperature directly affects the boiler combustion rate, pollutant emissions, and the reliability and stability of the overall system. Therefore, it is crucial to establish a model that accurately predicts the bed temperature of a circulating fluidized bed.

[0003] However, the mechanism model of the combustion side of the circulating fluidized bed unit is a complex system with multiple factors and multiple physical processes coupled with each other. It has significant nonlinearity, large delay and thermal inertia. In the modeling process, a certain degree of simplification is usually required, which may affect the modeling accuracy and make it difficult to establish a highly accurate mechanism model. In addition, the dynamic characteristics of the circulating fluidized bed unit are complex. The changes in its internal parameters are not only affected by the current operating conditions, but also by the changes in operating conditions in the past period of time, and there is a long-term dependency problem. The mechanism model is usually limited by specific assumptions and boundary conditions, and it is difficult to accurately describe all its dynamic characteristics. At the same time, data-driven models based on input and output data also have challenges. The data collection cycle of this type of model is long and the experimental data cannot cover the full operating conditions of the unit, which will cause the model performance to deteriorate. The data-driven model based on historical data is difficult to adapt to the current operating conditions of the unit, so it cannot accurately reflect the new operating characteristics of the unit, making it difficult to accurately reflect the current operating status of the system.

[0004] In addition, the operating performance and efficiency of the circulating fluidized bed unit will be affected by a variety of external factors. It is necessary to introduce exogenous variables to provide additional information outside the system, so that it can improve its ability to adapt to changes in operating conditions, thereby improving the prediction accuracy of the model under different conditions. Most existing methods for the division of different operating conditions lack sufficient interpretability and effectiveness, and there are highly coupled spatiotemporal interactions between various process variables, which further hinders the understanding and effective interpretation of the operating condition division. Therefore, it is also necessary to adopt an interpretable operating condition division method based on global space and local time to fully understand the operating status of the system and its changing laws. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide an online advance circulating fluidized bed temperature prediction system for all operating conditions based on the S4 model, which accurately predicts the bed temperature of the circulating fluidized bed unit and provides a reliable basis for further optimization control of the system.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: a full-condition online advanced circulating fluidized bed temperature prediction system based on the S4 model, including an offline stage and an online stage;

[0007] The offline stage includes a DCS system and database, an offline stage data selection and preprocessing module, a working condition division module and an S4 intelligent offline model module;

[0008] The DCS system and database collect historical data of industrial sites; the offline data selection and preprocessing module selects exogenous variables of the system from historical data and performs preprocessing; the working condition division module divides the selected historical data using an interpretable working condition division method combining global space and local time characteristics; the S4 intelligent offline model module trains the S4 intelligent offline model through historical data representing different working conditions to obtain the S4 optimal intelligent model that adapts to all working conditions;

[0009] The online stage includes a real-time data receiving module, an online stage data selection and preprocessing module and a model correction module;

[0010] The real-time data receiving module receives and stores the real-time online data; the online stage data selection and preprocessing module standardizes the obtained real-time online data, selects the real-time online data and the historical data of the adjacent leading time as the model input variables and transmits them to the S4 optimal intelligent model to obtain the predicted value of the bed temperature at the leading time; the model correction module is bidirectionally connected with the S4 intelligent offline model module, calculates the error between the predicted value and the true value, and corrects the S4 intelligent offline model.

[0011] As a preferred implementation, in the offline stage, the historical data include the total boiler air volume, primary air volume, secondary air volume, oxygen content, unit load, main steam temperature, total coal volume and bed temperature.

[0012] As a preferred implementation, in the working condition division module, an interpretable working condition division method combining global space and local time features is used to divide the selected historical data, including the following steps:

[0013] Step 1: Establish the similarity coefficient between node i and other neighboring nodes j, as follows:

[0014] e ij =a([Wx i ,Wx j]),j∈N i

[0015] Step 2: After obtaining the similarity coefficient, normalize the correlation coefficient to obtain the attention coefficient:

[0016]

[0017] Among them, e i,j represents the importance of node i to node j; W represents the learnable weight matrix, which has the same dimension as the input variable; a also represents the learnable parameter vector; x i and x j Represent the feature vectors corresponding to node i and node j respectively;

[0018] Step 3: Aggregate the information of node i and its adjacent nodes weighted by the attention coefficient. The new feature of node i is as follows:

[0019]

[0020] Get the global spatial feature x s ;

[0021] Step 4: Perform a linear transformation on the input variables to obtain the query Q, key K, and value V. The attention score is calculated by taking the dot product of the query vector and all key vectors as follows:

[0022]

[0023] Where d represents the input dimension. Then, the new feature after extracting the time information is expressed as x t = att(x(t)). Aggregate the extracted spatial features and temporal features h = [x s ,x t ];

[0024] Step 5: Use the Kohonen network to divide the working conditions of these latent variables with spatiotemporal characteristics; initialize the weight vector of the Kohonen network neuron, randomly select a group of nodes from the latent variables, and calculate the distance between the node and the weight vector of other nodes, as shown in the following formula:

[0025]

[0026] Among them, h is the input sample, w i is the weight vector of the neuron, and the node with the minimum distance is found, which is the winning node, as follows:

[0027] w c =argmin i d(h,w i )

[0028] The weight of the winning node is adjusted as follows:

[0029] w c (t+1)=w c (t)+η(t)(h(t)-w c (t))

[0030] Among them, η(t) is the learning rate at time t;

[0031] Step 6: Adjust the weights of the neighboring nodes around the winning node. The nodes in the neighboring range are determined by the community radius of the winning node. The community radius usually decreases over time, so that the update range gradually shrinks from global to local. The community radius is as follows:

[0032]

[0033] Among them, d cj represents the distance between the winning neuron and its jth neighbor; σ(t) represents the field width, which usually decreases over time; then, the weight adjustment method of the adjacent nodes is as follows:

[0034] w j (t+1)=w j (t)+η(t)·h cj (t)·(h(t)-w j (t))

[0035] The clustering results of different working conditions are obtained (h1, h2, h3…).

[0036] As a preferred embodiment, the S4 intelligent offline model is to use the S4 model to establish a bed temperature intelligent model, including a state equation and an output equation, and use data representing different working conditions to train the model to achieve an optimal model that adapts to all working conditions. The S4 model structure is as follows:

[0037] h′(t)=Ah(t)+Bx(t)

[0038] y(t)=Ch(t)

[0039] Among them, x(t) is the input signal; h(t) is the state signal; y(t) is the output signal; A, B, and C are the parameters learned by the network through gradient descent.

[0040] As a preferred implementation, the state space equation is discretized as:

[0041]

[0042] in, and represents the discretized parameter matrix; f A and f B represents the discretization rule, which here represents the bilinear transformation; its specific expressions are as follows:

[0043]

[0044] The state space equation is derived into a convolutional form, in order to make it have the ability to scale linearly or nearly linearly in the sequence length, specifically expressed as:

[0045]

[0046] As a preferred implementation, the A matrix in the continuous state space model is converted into a HIPPO matrix to compress all current input signals into a coefficient vector, as shown in the following formula:

[0047]

[0048] As a preferred embodiment, the real-time online data includes the total air volume of the boiler, the primary air volume, the secondary air volume, the oxygen content, the unit load, the main steam temperature, the total coal volume and the bed temperature variable.

[0049] As a preferred implementation, the online data selection and preprocessing module selects real-time online data at 10 moments and historical data at the adjacent previous 10 moments as model input variables and inputs them into the S4 optimal intelligent model to achieve online prediction of the circulating fluidized bed temperature 10 moments in advance.

[0050] As a preferred implementation, the model correction module updates and fine-tunes the S4 model online by comparing the MSE error values ​​between the bed temperature prediction values ​​at 10 moments obtained in the last online stage and the actual values ​​at 10 moments corresponding to the current real-time system to adapt to the real-time operation changes of the circulating fluidized bed unit.

[0051] As a preferred implementation, when the online stage enters the next operation, the predicted value of the bed temperature 10 steps ahead calculated in the previous online stage and the actual value of the corresponding real-time system at 10 moments are obtained, and the error accumulation between the predicted value and the actual value at 10 moments is calculated as follows:

[0052]

[0053] When the error accumulation exceeds the threshold, it is considered that the operating conditions have undergone new changes, and the trained model needs to be back-propagated to fine-tune the model parameters; when the error accumulation does not exceed the threshold, it is considered that the operating conditions have not changed significantly, and the full-operating condition model trained offline can adapt well to the online operation status of the system.

[0054] The present invention establishes an S4 online model applicable to all working conditions, and uses a model correction method to update and fine-tune the model online, thereby realizing online advance prediction of the bed temperature of a circulating fluidized bed. In the offline stage, historical data is divided into working conditions by a working condition division method with spatiotemporal interpretability, thereby improving the interpretability and accuracy of the working condition division; different working condition information is used as input to the S4 model training to obtain the optimal intelligent model, thereby solving the long-term dependency problem and improving the generalization and applicability of the model. In the online stage, online bed temperature prediction of real-time data on site is realized, and operating parameters can be adjusted in time to reduce bed temperature fluctuations and ensure the smooth operation of the system. The online model correction and update enables the S4 model to have a faster re-learning ability. When production conditions change, the bed temperature of the circulating fluidized bed unit can be accurately predicted while ensuring production safety, providing a reliable basis for further optimization and control of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] The drawings herein are used to provide further explanation of the present invention and constitute a part of this application. The illustrative embodiments of the present invention and their description are used to explain the present invention and do not constitute improper limitations on the present invention.

[0056] Figure 1 This is a flow chart of the full-condition online advanced circulating fluidized bed temperature prediction system based on the S4 model proposed by the present invention;

[0057] Figure 2 This is a framework diagram of the full-condition online advanced circulating fluidized bed temperature prediction system based on the S4 model proposed by the present invention;

[0058] Figure 3 Schematic diagram of the model correction for calculating the accumulated error in the online stage. DETAILED DESCRIPTION

[0059] refer to Figure 1-2 In order to solve the problem of low accuracy of bed temperature prediction model under variable working conditions, a typical embodiment of the present invention provides an online advanced circulating fluidized bed temperature prediction system for all working conditions based on the S4 model. The basic concept is to adopt a working condition division method combining global space and local time characteristics, establish an S4 optimal model suitable for all working conditions, and adopt a model correction method to update and fine-tune the model online to achieve online advanced prediction of circulating fluidized bed temperature. This embodiment fully considers the long-term dependence of the unit under different working conditions and the influence of external conditions on the unit, which is of great significance to the safe and stable operation of the unit, rapid load change and control system optimization.

[0060] The system provided in this embodiment includes an offline phase and an online phase. The offline phase includes a DCS system and a database, an offline phase data selection and preprocessing module, a working condition division module, and an S4 intelligent offline model module. The online phase includes a real-time data receiving module, an online phase data selection and preprocessing module, and a model correction module.

[0061] The DCS system is connected to the database, the offline data selection and preprocessing module, the working condition division module, and the S4 intelligent offline model module in sequence. The real-time data receiving module is connected to the online data selection and preprocessing module, and then connected to the S4 intelligent offline model module. The model correction module is bidirectionally connected to the S4 intelligent offline model module.

[0062] In the offline stage, the industrial site historical data is obtained through sensors through the DCS system and database. The offline stage data selection and preprocessing module is responsible for selecting the system's exogenous variables from the field historical data, and cleaning and standardizing the historical data. The working condition division module uses an interpretable working condition division method that combines global space and local time characteristics to divide the selected historical data. The S4 intelligent model module is responsible for training different working condition information to achieve the optimal model that adapts to all working conditions. The main task of the offline stage is to train the S4 optimal model that adapts to all working conditions.

[0063] The historical data include the total air volume of the boiler, the primary air volume, the secondary air volume, the oxygen content, the unit load, the main steam temperature, the total coal volume and the bed temperature. Among them, the total air volume of the boiler, the primary air volume, the secondary air volume, the oxygen content, the unit load, the main steam temperature and the total coal volume are exogenous variables; the bed temperature is the target variable.

[0064] In the offline data selection and preprocessing module, the above exogenous variables are selected as training data. Data preprocessing is to smooth the missing values ​​and abnormal values ​​in the data and then standardize them.

[0065] The working condition division module comprehensively considers the global spatial distribution and local temporal changes, and accurately divides different working conditions of historical data in a highly interpretable way.

[0066] The interaction between different equipment will lead to changes in variables. Changes in the speed of the coal feeder and the blower will affect the fuel supply and combustion rate in the combustion chamber, thereby affecting the amount of oxygen in the boiler combustion process; changes in load will lead to changes in temperature and pressure in the combustion chamber, and the main steam temperature measures the boiler combustion efficiency. These factors have a comprehensive impact on the changes in bed temperature variables.

[0067] Therefore, a graph attention network mechanism is utilized to infer the global spatial relationship between different variables, so that the structured relationship between different variables can be represented, where each variable represents a different node.

[0068] First, establish the similarity coefficient between node i and other neighboring nodes j as follows:

[0069] e ij =a([Wx i ,Wx j ]),j∈N i (1)

[0070] After obtaining the similarity coefficient, the correlation coefficient is normalized to obtain the attention coefficient:

[0071]

[0072] Among them, e i,j represents the importance of node i to node j; W represents the learnable weight matrix, which has the same dimension as the input variable; a also represents the learnable parameter vector; x i and x j Represent the feature vectors corresponding to node i and node j respectively.

[0073] Next, the information of node i and its adjacent nodes is weighted and aggregated according to the attention coefficient, and the new feature of node i is as follows:

[0074]

[0075] The attention coefficient is calculated through the graph attention mechanism, and spatial connections are established for different nodes. The influence of each node on other neighboring nodes is obtained by weighted aggregation, so as to obtain the global spatial feature x s .

[0076] Industrial processes usually involve a large amount of time series data, which may contain a lot of noise and redundant information. In order to accurately obtain the key time features in industrial data, the self-attention mechanism is used to extract the potential time information in the variables to help the model identify the moment that has the greatest impact on bed temperature changes.

[0077] First, the input variables are linearly transformed to obtain the query Q, key K, and value V. The attention score is calculated by the dot product of the query vector and all key vectors, as shown below:

[0078]

[0079] Where d represents the input dimension. Then, the new feature after extracting the time information is expressed as x t = att(x(t));

[0080] Furthermore, the extracted spatial features and temporal features are aggregated h = [x s ,x t], the working conditions of these latent variables with spatiotemporal characteristics are divided through the Kohonen network.

[0081] First, initialize the weight vector of the Kohonen network neuron; second, randomly select a set of features (nodes) from the latent variables (which can be regarded as new nodes) and calculate the distance between the weight vectors of the node and other nodes, as follows:

[0082]

[0083] Among them, h is the input sample, w i is the weight vector of the neuron. Find the node with the minimum distance, that is, the winning node, as follows:

[0084] w c =argmin i d(h,w i ) (6)

[0085] The weight of the winning node is adjusted as follows:

[0086] w c (t+1)=w c (t)+η(t)(h(t)-w c (t)) (7)

[0087] Among them, η(t) is the learning rate at time t.

[0088] After that, the weights of the neighboring nodes around the winning node should also be adjusted. The nodes in the neighboring range are determined by the community radius of the winning node. The neighborhood radius usually decreases over time, so that the update range gradually shrinks from global to local. Common community radii are as follows:

[0089]

[0090] Among them, d cj represents the distance between the winning neuron and its jth neighbor; σ(t) represents the domain width, which usually decreases over time. Then, the weight adjustment method of the adjacent nodes is as follows:

[0091] w j (t+1)=w j (t)+η(t)·h cj (t)·(h(t)-w j (t)) (9)

[0092] The clustering results (h1, h2, h3…) of different working conditions can be obtained.

[0093] The S4 intelligent model module trains the S4 model based on the data of the above different working conditions, and obtains the long-term dependence in each working condition data through the model, thereby improving the accuracy of model prediction. The model mainly includes state equations and output equations, as shown below:

[0094] h′(t)=Ah(t)+Bx(t) (10)

[0095] y(t)=Ch(t) (11)

[0096] Among them, x(t) is the input signal; h(t) is the state signal; y(t) is the output signal; A, B, and C are the parameters learned by the network through gradient descent.

[0097] The state space equation is discretized as:

[0098]

[0099] in, and represents the discretized parameter matrix; f A and f B represents the discretization rule, which here represents the bilinear transformation; its specific expressions are as follows:

[0100]

[0101] The state space equation is derived into a convolutional form, in order to make it have the ability to scale linearly or nearly linearly in the sequence length, specifically expressed as:

[0102]

[0103]

[0104] At the same time, in order to effectively solve the long-distance dependency problem of sequence modeling in a limited storage space, the A matrix in the continuous state space model, that is, the A matrix in formula (10), is a HIPPO matrix that compresses all current input signals into a coefficient vector, as shown in the following formula:

[0105]

[0106] In order to avoid excessive HIPPO matrix calculation, matrix decomposition is used to express it using a low-rank matrix, as shown below:

[0107] A=VΛV * -PQ=V(Λ-(V * P)(V * Q) * )V * (20)

[0108] After the offline model is updated based on historical data training, the S4 intelligent optimal model applicable to all working conditions is obtained.

[0109] In order to ensure the stability and efficiency of the combustion process of the circulating fluidized bed unit, it is necessary to monitor and predict the change trend of the bed temperature in real time, so as to adjust the operating parameters in advance, optimize the operating strategy, and improve the overall performance of the system. Therefore, it is necessary to establish an online model of the unit to make advance predictions on the real-time data on site.

[0110] In the online stage, the real-time system of the circulating fluidized bed unit provides real-time online data. The real-time data receiving module connects to the functional sensors to receive and store data. The online stage data selection and preprocessing module performs structural processing on the real-time online data, and then uses the S4 optimal model to realize online bed temperature advance prediction. At the same time, the model correction module updates and fine-tunes the S4 model online, which can prevent the model from error accumulation, effectively respond to changes in operating conditions, and ensure the safe and stable operation of the system.

[0111] The data selection and preprocessing module is connected with the S4 intelligent offline model module to realize the prediction of dynamic bed temperature value, and the real-time online data and adjacent historical data are used as new input variables, and the optimal offline S4 model is used for online prediction to obtain the advance prediction value of bed temperature.

[0112] The online data includes the total air volume of the boiler, the primary air volume, the secondary air volume, the oxygen content, the unit load, the main steam temperature, the total coal volume and the bed temperature variable. In this embodiment, the real-time data receiving module stores and updates the real-time online data at ten moments.

[0113] The current data and the adjacent historical data have a certain short-term time correlation. The data selection and preprocessing module standardizes the online data and selects the real-time online data of ten moments and the historical data of the adjacent previous 10 moments as the model input variables. Furthermore, the input variables are transmitted to the S4 optimal intelligent model to realize the predicted value of the bed temperature 10 moments ahead, so as to adjust the combustion control parameters such as fuel supply and air volume in advance to ensure that the combustion process always runs in the best state.

[0114] The model correction module is designed to solve the problem that the model cannot adapt to new working conditions during online prediction, resulting in a decrease in prediction performance. The model correction module is bidirectionally connected with the S4 intelligent offline model module to calculate the error between the predicted value and the true value to correct the model.

[0115] When the unit's operating parameters, environmental conditions, load changes, and control systems change, the operating conditions will change accordingly. The new operating conditions will contain new patterns and dynamic characteristics that did not appear in the training data, resulting in a decrease in the accuracy of online model predictions. In actual operation, the model correction module will continuously update and fine-tune model parameters in response to large cumulative errors to adapt to changes in system characteristics and fluctuations in external conditions.

[0116] When the online stage enters the next operation, the predicted value of the bed temperature 10 steps ahead calculated in the previous online stage and the actual value of the corresponding real-time system at 10 moments will be obtained, and the error accumulation between the predicted value and the actual value at 10 moments will be calculated as follows:

[0117]

[0118] When the error accumulation exceeds the threshold, it is considered that the operating conditions have undergone new changes, and the trained model needs to be back-propagated to fine-tune the model parameters; when the error accumulation does not exceed the threshold, it is considered that the operating conditions have not changed significantly, and the full-operating condition model trained offline can adapt well to the online operation status of the system.

[0119] like Figure 3 The figure shows the model correction module composed of the core technology of the present invention. In this embodiment, the online real-time data of 10 moments and the adjacent 10 historical data are used as input variables, and the bed temperature prediction value 10 moments ahead is realized through the S4 optimal intelligent model. Considering that the on-site operating conditions may change, by comparing the error between the bed temperature prediction value obtained in the last online stage and the actual value corresponding to the current real-time system, the S4 model is updated and fine-tuned online to adapt to the real-time operation changes of the circulating fluidized bed unit.

[0120] The scope of protection claimed by the present invention is not limited to the above specific embodiments. For those skilled in the art, the present invention may have various deformations and changes. Any modifications, improvements and equivalent substitutions made within the concept and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A full-condition online advanced circulating fluidized bed temperature prediction system based on the S4 model, characterized in that: Includes offline and online stages; The offline stage includes a DCS system and database, an offline stage data selection and preprocessing module, a working condition division module and an S4 intelligent offline model module; The DCS system and database collect historical data of industrial sites; the offline data selection and preprocessing module selects exogenous variables of the system from historical data and performs preprocessing; the working condition division module divides the selected historical data using an interpretable working condition division method combining global space and local time characteristics; the S4 intelligent offline model module trains the S4 intelligent offline model through historical data representing different working conditions to obtain the S4 optimal intelligent model that adapts to all working conditions; The online stage includes a real-time data receiving module, an online stage data selection and preprocessing module and a model correction module; The real-time data receiving module receives and stores the real-time online data; the online stage data selection and preprocessing module standardizes the obtained real-time online data, selects the real-time online data and the historical data of the adjacent leading time as the model input variables and transmits them to the S4 optimal intelligent model to obtain the predicted value of the bed temperature at the leading time; the model correction module is bidirectionally connected with the S4 intelligent offline model module, calculates the error between the predicted value and the true value, and corrects the S4 intelligent offline model.

2. The full-condition online advanced circulating fluidized bed temperature prediction system based on the S4 model according to claim 1 is characterized by: During the offline stage, historical data include total boiler air volume, primary air volume, secondary air volume, oxygen content, unit load, main steam temperature, total coal volume and bed temperature.

3. The full-condition online advanced circulating fluidized bed temperature prediction system based on the S4 model according to claim 1 or 2, characterized in that: In the working condition division module, the selected historical data is divided by using an interpretable working condition division method that combines global space and local time characteristics, including the following steps: Step 1: Establish the similarity coefficient between node i and other neighboring nodes j, as follows: e ij =a([Wx i ,Wx j ]),j∈N i Step 2: After obtaining the similarity coefficient, normalize the correlation coefficient to obtain the attention coefficient: Among them, e i,j represents the importance of node i to node j; W represents the learnable weight matrix, which has the same dimension as the input variable; a also represents the learnable parameter vector; x i and x j Represent the feature vectors corresponding to node i and node j respectively; Step 3: Aggregate the information of node i and its adjacent nodes weighted by the attention coefficient. The new feature of node i is as follows: Get the global spatial feature x s ; Step 4: Perform a linear transformation on the input variables to obtain the query Q, key K, and value V. The attention score is calculated by taking the dot product of the query vector and all key vectors as follows: Where d represents the input dimension, then the new feature after extracting the time information is expressed as x t = att(x(t)) aggregates the extracted spatial and temporal features h = [x s ,x t ]; Step 5: Use the Kohonen network to divide the working conditions of these latent variables with spatiotemporal characteristics; initialize the weight vector of the Kohonen network neuron, randomly select a group of nodes from the latent variables, and calculate the distance between the node and the weight vector of other nodes, as shown in the following formula: Among them, h is the input sample, w i is the weight vector of the neuron, and the node with the minimum distance is found, which is the winning node, as follows: w c =argmin i d(h,w i ) The weight of the winning node is adjusted as follows: w c (t+1)=w c (t)+η(t)(h(t)-w c (t)) Among them, η(t) is the learning rate at time t; Step 6: Adjust the weights of the neighboring nodes around the winning node. The nodes in the neighboring range are determined by the community radius of the winning node. The community radius usually decreases over time, so that the update range gradually shrinks from global to local. The community radius is as follows: Among them, d cj represents the distance between the winning neuron and its jth neighbor; σ(t) represents the field width, which usually decreases over time; then, the weight adjustment method of the adjacent nodes is as follows: w j (t+1)=w j (t)+η(t)·h cj (t)·(h-w j (t)) The clustering results of different working conditions are obtained (h1, h2, h3…).

4. The full-condition online advanced circulating fluidized bed temperature prediction system based on the S4 model according to claim 3 is characterized by: The S4 intelligent offline model is to establish a bed temperature intelligent model using the S4 model, including a state equation and an output equation. The model is trained using data representing different working conditions to achieve an optimal model that adapts to all working conditions. The S4 model structure is as follows: h′(t)=Ah(t)+Bx(t) y(t)=Ch(t) Among them, x(t) is the input signal; h(t) is the state signal; y(t) is the output signal; A, B, and C are the parameters learned by the network through gradient descent.

5. The full-condition online advanced circulating fluidized bed temperature prediction system based on the S4 model according to claim 4 is characterized in that: The state space equation is discretized as: in, and represents the discretized parameter matrix; f A and f B represents the discretization rule, which here represents the bilinear transformation; its specific expressions are as follows: The state space equation is derived into a convolutional form, in order to make it have the ability to scale linearly or nearly linearly in the sequence length, specifically expressed as:

6. The full-condition online advanced circulating fluidized bed temperature prediction system based on the S4 model according to claim 5 is characterized by: The A matrix in the continuous state space model is converted into a HIPPO matrix to compress all current input signals into a coefficient vector, as shown below:

7. The full-condition online advanced circulating fluidized bed temperature prediction system based on the S4 model according to claim 1 is characterized by: The real-time online data include total boiler air volume, primary air volume, secondary air volume, oxygen volume, unit load, main steam temperature, total coal volume and bed temperature variables.

8. The full-condition online advanced circulating fluidized bed temperature prediction system based on the S4 model according to claim 1 or 7, characterized in that: The online data selection and preprocessing module selects the real-time online data of 10 moments and the historical data of the previous 10 moments as the model input variables and inputs them into the S4 optimal intelligent model to realize the online prediction of the circulating fluidized bed temperature 10 moments in advance.

9. The full-condition online advanced circulating fluidized bed temperature prediction system based on the S4 model according to claim 8 is characterized by: The model correction module updates and fine-tunes the S4 model online by comparing the MSE error values ​​between the bed temperature prediction values ​​at 10 moments obtained in the last online stage and the actual values ​​at 10 moments corresponding to the current real-time system to adapt to the real-time operation changes of the circulating fluidized bed unit.

10. The full-operating-condition online advanced circulating fluidized bed temperature prediction system based on the S4 model according to claim 9 is characterized in that: When the online stage enters the next operation, the predicted value of the bed temperature 10 steps ahead calculated in the previous online stage and the actual value of the corresponding real-time system at 10 moments will be obtained, and the error accumulation between the predicted value and the actual value at 10 moments will be calculated as follows: When the error accumulation exceeds the threshold, it is considered that the operating conditions have undergone new changes, and the trained model needs to be back-propagated to fine-tune the model parameters; when the error accumulation does not exceed the threshold, it is considered that the operating conditions have not changed significantly, and the full-operating condition model trained offline can adapt well to the online operation status of the system.

Images