A Real-Time Prediction Method for Furnace Temperature Field Based on Convolutional Long Short-Term Memory Networks

Through the real-time prediction method of temperature field based on convolutional long short-term memory network, the problems of instability in combustion of industrial heating furnaces and large amount of computational fluid mechanics are solved, and fast and accurate temperature field prediction is achieved, which significantly improves efficiency and accuracy .

CN110909508BActive Publication Date: 2024-11-26CHINA PETROLEUM & CHEMICAL CORP
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN201911036206.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2019-10-29
Publication Date
2024-11-26
Estimated Expiration
2039-10-29

AI Technical Summary

Technical Problem

The unstable combustion process of industrial heating furnaces leads to local overtemperature, making it difficult to optimize the temperature field in real time. The traditional computational fluid mechanics method has a large amount of calculation and is difficult to meet the real-time requirements.

Method used

The real-time prediction method of temperature field based on the convolutional long and short-term memory network is adopted. By building a 4-layer convolutional long and short-term memory network, data normalization and three-dimensional expansion are carried out to train the network to realize real-time prediction of the temperature field of the industrial heating furnace.

Benefits of technology

Fast and accurate temperature field prediction is achieved, and the calculation time is shortened from several days to several seconds compared to traditional methods, with small errors, meeting the needs of real-time and efficientness.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN110909508B_ABST
    Figure CN110909508B_ABST
Patent Text Reader

Abstract

A method for real-time prediction of the temperature field of a heating furnace based on a convolutional long short-term memory network includes: building a real-time prediction network for the temperature field; training the real-time prediction network for the temperature field; normalizing each data of the industrial heating furnace respectively; three-dimensionally expanding each data of the industrial heating furnace; sending the expanded data to the trained real-time prediction network for the temperature field to obtain a three-dimensional temperature field prediction result; and restoring the obtained three-dimensional temperature field prediction result to the pipe wall temperature field by retaining the spatial coordinates of the pipe wall temperature field corresponding to each point of the three-dimensional temperature field. The present invention can quickly obtain a predicted temperature field. Compared with the computational fluid dynamics method, which takes several days to calculate a temperature field once, the method of the present invention only takes a few seconds to obtain a predicted temperature field, which significantly improves efficiency and has a very small prediction error.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a method for predicting the temperature field of an industrial heating furnace, and more particularly to a method for real-time prediction of the temperature field of a heating furnace based on a convolutional long short-term memory network. Background Art

[0002] Industrial heating furnaces are important equipment in industrial refining plants. Whether they operate safely directly affects the service life, production capacity and economic benefits of the equipment. The combustion process of industrial heating furnaces is unstable. During operation, local overtemperature may occur at random locations. The furnace tubes of industrial heating furnaces are mostly composed of process media that are prone to coking. If a certain local position of the heating furnace runs in an overtemperature state for a long time, it will cause loss and damage to the furnace tubes. Therefore, measures must be taken to optimize the combustion conditions of the heating furnace temperature field. However, the heating furnace equipment is large and the environment is harsh, so it is difficult to measure the relevant physical quantity parameters online, resulting in a lack of reliable basis for combustion adjustment and difficulty in achieving combustion optimization operation. Therefore, only soft measurement solutions can be adopted to obtain the temperature field of industrial heating furnaces.

[0003] Soft measurement refers to the process of mathematically inferring variables that are difficult to measure or cannot be measured in industrial production by combining the measured quantities obtained with computer and other application technologies, thereby obtaining the values ​​of important variables that cannot be directly measured. Soft measurement has the advantage of rapid dynamic response, and can continuously give the values ​​of important variables in industrial processes.

[0004] Soft sensor models can be divided into principle-based soft sensor models (Model-Driven Models, MDM for short) and data-based soft sensor models (Data-Driven Models, DDM for short)[4], also known as data-driven soft sensor models.

[0005] MDM is based on the prior knowledge of the operating mechanism of industrial processes and the related theoretical knowledge of complex physical chemistry involved in industrial processes. The principle model usually has high numerical calculation accuracy and strong interpretability. However, traditional mechanism-based models are often composed of a large number of algebraic equations, with large calculation volume and slow convergence, which makes it difficult to meet the real-time requirements of soft measurement.

[0006] CFD is the abbreviation of Computational Fluid Dynamics, and CFD is also a type of MDM. CFD uses numerical analysis methods to solve problems involving fluid flow. The operating data of industrial heating furnaces such as temperature and flow can be used as the basis for CFD calculations, so that the temperature calculation values ​​of a series of points inside the industrial heating furnace can be obtained. CFD can calculate very detailed and rich three-dimensional temperature field data of industrial heating furnaces. However, CFD cannot get rid of the large amount of calculation of MDM, and it is difficult to obtain real-time temperature fields.

[0007] The soft sensor modeling method based on deep learning belongs to DDM, which uses deep neural network (DNN). Compared with traditional neural networks, DNN has a higher network layer depth and more advanced convergence methods, so it has stronger approximation and convergence capabilities. For complex and highly nonlinear processes in industrial processes, DNN shows good performance.

[0008] And in the calculation

[0009] Recurrent Neural Network (RNN) and its improved version Long Short Term Memory (LSTM) network, these two networks were proposed in the field of machine learning as early as the last century and have been widely used. In actual industrial processes, the current system state and output are closely related to one or more previous states. Therefore, RNN is very suitable for processing time series problems in industrial production processes. LSTM contains memory and forgetting structures. Compared with traditional RNN, LSTM can solve the problems of gradient disappearance and gradient explosion in the original RNN. LSTM has been applied to fields such as natural language processing and speech recognition, and has also been applied in the field of soft measurement. It has excellent performance in processing time series problems in industrial processes. Moreover, after training, the method has a short calculation time and can meet the real-time requirements. Summary of the invention

[0010] The technical problem to be solved by the present invention is to provide a real-time prediction method for the temperature field of an industrial heating furnace based on a convolutional long short-term memory network, which can realize soft measurement of the temperature field of the industrial heating furnace.

[0011] The technical solution adopted by the present invention is: a real-time prediction method of the temperature field of a heating furnace based on a convolutional long short-term memory network, characterized in that it comprises the following steps:

[0012] 1) Build a real-time prediction network for temperature field;

[0013] 2) Training the temperature field real-time prediction network;

[0014] 3) Normalize the data of industrial heating furnaces respectively;

[0015] 4) Three-dimensional expansion of industrial heating furnace data;

[0016] 5) Sending the expanded data to the trained temperature field real-time prediction network to obtain the three-dimensional temperature field prediction results;

[0017] 6) The obtained three-dimensional temperature field prediction result is restored to the pipe wall temperature field through the spatial coordinates of the pipe wall temperature field corresponding to each point of the three-dimensional temperature field retained in step 2).

[0018] The temperature field real-time prediction network described in step 1) has a total of 4 layers of convolutional long short-term memory networks (ConvLSTM), the number of input data channels is 18, the number of output data channels of the first layer of long short-term memory network (ConvLSTM1) is 24, the number of output data channels of the second layer of long short-term memory network (ConvLSTM2) is 48, and the number of output data channels of the third layer of long short-term memory network (ConvLSTM3) is 24. The output data of the first layer of long short-term memory network (ConvLSTM1) and the third layer of long short-term memory network (ConvLSTM3) are spliced ​​to obtain data with a channel number of 48, and sent to the fourth layer of long short-term memory network (ConvLSTM4), and finally a 24-channel temperature field output (output) is obtained;

[0019] At the same time, the input and output time series lengths of the first-layer long short-term memory network (ConvLSTM1), the second-layer long short-term memory network (ConvLSTM2), and the third-layer long short-term memory network (ConvLSTM3) are all 10, and the input time series length of the fourth-layer long short-term memory network (ConvLSTM4) is 10, and the output time series length is 3.

[0020] Step 2) includes:

[0021] (1) The data of industrial heating furnaces are normalized using the following formula:

[0022]

[0023] Among them, x is the data to be normalized, mean, max, and min are the mean, maximum, and minimum values ​​within the training sample range of the working condition data respectively;

[0024] (2) Three-dimensional expansion of industrial heating furnace data;

[0025] (3) Convert the pipe wall temperature field into a three-dimensional temperature field

[0026] (4) The temperature field real-time prediction network was trained without freezing the network parameters during training. The initial learning rate during training was 0.01, the batch size during training was 1, the L1 loss function was selected as the loss function, and the Adam algorithm was used to optimize the loss function. A total of 600 iterations were performed, and the learning rate was multiplied by 0.8 for decay every 40 iterations. The network achieved the best effect at the 480th iteration, and the optimal model was obtained. The mean absolute error of the optimal model output was 28.3 Kelvin.

[0027] The data in step 2) include: the values ​​of various temperatures, pressures, oxygen contents, and gas flow rates of the distributed control system in the industrial heating furnace measured at set times;

[0028] Step 2) The expansion method of step (2) is the same as that of step 3), which is to expand one operating condition data into a 32*652 matrix, and each element in the matrix is ​​filled with the operating condition data; and for the temperature data of the 8 thermocouples on the pipe wall, when expanding, only 50 elements near the spatial position corresponding to the thermocouple are filled with the temperature data measured by the thermocouple, and the remaining elements are set to zero, and finally 18*32*652 input data is obtained.

[0029] Step 2) Step (3) is to first divide the pipeline into 24 pipelines along the z-axis, and then cut the pipe wall along the axis and unfold it into a rectangle. The unfolding method is to select 32 points in the axial direction of the pipeline in sequence, and arrange the 32 points in the z-axis direction to obtain the first column of data; then select 32 points in the axial direction of the pipeline and continue to arrange them, repeat 652 times to obtain a 24*32*652 matrix. At the same time, the temperature values ​​of the temperature field are divided by 1000 so that the value range is between 0 and 1, and the spatial coordinates of the pipe wall temperature field corresponding to each point in the three-dimensional temperature field are retained.

[0030] The normalization process described in step 3) is performed using the following formula:

[0031]

[0032] Among them, x is the data to be normalized, mean, max, and min are the mean, maximum, and minimum values ​​within the training sample range of the working condition data, respectively.

[0033] The real-time prediction method of the heating furnace temperature field based on the convolutional long short-term memory network of the present invention is completely data-driven, and uses ConvLSTM to dynamically extract the internal relationship of the time series to obtain the predicted temperature field, thereby realizing the soft measurement of the industrial heating furnace temperature field, and the average absolute error MAE of the predicted temperature field is 31.7 K. The method of the present invention can quickly obtain the predicted temperature field. Compared with the computational fluid dynamics method, which takes several days to calculate a temperature field once, the method of the present invention only takes a few seconds to obtain the predicted temperature field, which significantly improves the efficiency and has a very small prediction error. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a schematic diagram of the pipeline temperature field of the present invention;

[0035] Figure 2 It is a schematic diagram of the temperature field real-time prediction network structure of the present invention;

[0036] Figure 3This is the trend graph of thermocouple No. 1;

[0037] Figure 4 This is the trend graph of thermocouple No. 2;

[0038] Figure 5 This is the trend graph of thermocouple No. 3;

[0039] Figure 6 This is the trend graph of thermocouple No. 4;

[0040] Figure 7 This is the trend graph of thermocouple No. 5;

[0041] Figure 8 This is the trend graph of thermocouple No. 6;

[0042] Fig. 9 This is the trend graph of thermocouple No. 7;

[0043] Fig.10 This is the trend graph of thermocouple No. 8;

[0044] Fig.11 It is the modeling diagram of the calculated temperature field;

[0045] Fig.12 It is the predicted temperature field modeling diagram. DETAILED DESCRIPTION

[0046] The real-time prediction method of the heating furnace temperature field based on the convolutional long short-term memory network of the present invention is described in detail below in conjunction with the embodiments and drawings.

[0047] The real-time prediction method of the heating furnace temperature field based on the convolutional long short-term memory network of the present invention comprises the following steps:

[0048] 1) Build a real-time prediction network for temperature field;

[0049] like Figure 2 As shown, the temperature field real-time prediction network has a total of 4 layers of convolutional long short-term memory networks (ConvLSTM), the number of input data channels is 18, the number of output data channels of the first layer of long short-term memory network (ConvLSTM1) is 24, the number of output data channels of the second layer of long short-term memory network (ConvLSTM2) is 48, and the number of output data channels of the third layer of long short-term memory network (ConvLSTM3) is 24. The output data of the first layer of long short-term memory network (ConvLSTM1) and the third layer of long short-term memory network (ConvLSTM3) are spliced ​​to obtain data with a channel number of 48, and sent to the fourth layer of long short-term memory network (ConvLSTM4), and finally a 24-channel temperature field output (output) is obtained;

[0050] At the same time, the input and output time series lengths of the first-layer long short-term memory network (ConvLSTM1), the second-layer long short-term memory network (ConvLSTM2), and the third-layer long short-term memory network (ConvLSTM3) are all 10, and the input time series length of the fourth-layer long short-term memory network (ConvLSTM4) is 10, and the output time series length is 3.

[0051] 2) Training the temperature field real-time prediction network; including:

[0052] (1) The data of industrial heating furnaces are normalized using the following formula:

[0053]

[0054] Among them, x is the data to be normalized, mean, max, and min are the mean, maximum, and minimum values ​​within the training sample range of the working condition data respectively;

[0055] The actual working condition in (2) is only a 1-dimensional vector, which does not match the target network output 3-dimensional temperature field. Therefore, the present invention expands the 1-dimensional working condition into 3 dimensions. It is necessary to perform three-dimensional expansion on the data of the industrial heating furnace. The data described here include: the values ​​of various temperatures, pressures, oxygen content, and gas flow of the distributed control system in the industrial heating furnace measured at the set time; the expansion method described here is the same, which is to expand one working condition data into a 32*652 matrix, and each element in the matrix is ​​filled with the working condition data; and for the temperature data of the 8 thermocouples on the pipe wall, when expanding, only 50 elements near the spatial position corresponding to the thermocouple are filled with the temperature data measured by the thermocouple, and the remaining elements are set to zero, and finally 18*32*652 input data is obtained.

[0056] (3) Convert the pipe wall temperature field into a three-dimensional temperature field, because the three-dimensional temperature field also contains the structural information of the pipe wall, but such a temperature field cannot be used as a network output. Therefore, the present invention rearranges the calculation points on the pipe wall temperature field into a three-dimensional temperature field.

[0057] First, the pipeline is divided into 24 pipelines along the z-axis, and then the pipe wall is cut and expanded into a rectangle along the axis. The expansion method is to select 32 points in the axial direction of the pipeline in sequence, and arrange the 32 points in the z-axis direction to obtain the first column of data; then select 32 points in the axial direction of the pipeline and continue to arrange them, repeat 652 times to obtain a 24*32*652 matrix. At the same time, the temperature values ​​of the temperature field are divided by 1000 so that the value range is between 0 and 1, and the spatial coordinates of the pipe wall temperature field corresponding to each point in the three-dimensional temperature field are retained.

[0058] (4) The temperature field real-time prediction network was trained without freezing the network parameters during training. The initial learning rate during training was 0.01, the batch size during training was 1, the L1 loss function was selected as the loss function, and the Adam algorithm was used to optimize the loss function. A total of 600 iterations were performed, and the learning rate was multiplied by 0.8 for decay every 40 iterations. The network achieved the best effect at the 480th iteration, and the optimal model was obtained. The mean absolute error of the optimal model output was 28.3 Kelvin.

[0059] 3) Each data of the industrial heating furnace is normalized separately; the data described here include: the values ​​of each temperature, pressure, oxygen content, and gas flow of the distributed control system in the industrial heating furnace measured at the set time; the expansion method described here is the same, which is to expand one working condition data into a 32*652 matrix, and each element in the matrix is ​​filled with the working condition data; and for the temperature data of the 8 thermocouples on the pipe wall, when expanding, only 50 elements near the spatial position corresponding to the thermocouple are filled with the temperature data measured by the thermocouple, and the remaining elements are set to zero, and finally 18*32*652 input data is obtained. The normalization processing described here is performed using the following formula:

[0060]

[0061] Among them, x is the data to be normalized, mean, max, and min are the mean, maximum, and minimum values ​​within the training sample range of the working condition data respectively;

[0062] 4) Perform three-dimensional expansion on the data of the industrial heating furnace; the expansion method is to expand one working condition data into a 32*652 matrix, and each element in the matrix is ​​filled with the working condition data; and for the temperature data of the 8 thermocouples on the pipe wall, when expanding, only 50 elements near the spatial position corresponding to the thermocouple are filled with the temperature data measured by the thermocouple, and the remaining elements are set to zero, and finally 18*32*652 input data is obtained.

[0063] 5) Sending the expanded data to the trained temperature field real-time prediction network to obtain the three-dimensional temperature field prediction results;

[0064] 6) The three-dimensional temperature field prediction result obtained is restored to the pipe wall temperature field through the spatial coordinates of the pipe wall temperature field corresponding to each point of the three-dimensional temperature field retained in step 2). The temperature field output by the temperature field real-time prediction network is only a three-dimensional matrix, and does not have the shape of the pipe wall. Therefore, after obtaining the prediction result, it is also necessary to restore it to the pipe wall temperature field.

[0065] Here are some examples

[0066] Since the temperature field calculated by CFD only has data at one moment, and network training requires a large amount of data to train the network, in order to ensure the smooth progress of this example, a series of simulated temperature fields are constructed according to actual conditions to be used as data for network training.

[0067] The example of the present invention selects 30 working conditions corresponding to the working condition attachments of each calculated temperature field to construct a simulated temperature field. Taking the temperature of a single pipe wall thermocouple in the working condition as the standard, the calculated temperature field is adjusted to obtain a simulated temperature field. For example, for the calculated temperature field of working condition No. 15, if you want to obtain the simulated temperature field of working condition No. 10, you should first calculate the growth percentage of the 8 thermocouple data in working condition No. 10 relative to the 8 thermocouples in working condition No. 15. Then adjust the temperature of the pipeline corresponding to the calculated temperature field of working condition No. 15 according to the 8 growth percentages, and for the pipeline without a thermocouple in the middle, take a gradual treatment according to the growth percentage of the two nearest thermocouples.

[0068] The temperature field prediction in this example is actually a regression task. For this problem, this example will use the mean absolute error (MAE) to evaluate the algorithm performance. The calculation formula of MAE is as shown in Formula 1.

[0069]

[0070] Where m is the number of points in the three-dimensional temperature field, y i is the actual temperature value of the i-th point in the three-dimensional temperature field, is the predicted temperature value of the i-th point in the three-dimensional temperature field,

[0071] The MAE of the best training result is 0.0283, and the corresponding test result is 0.0317, which is converted to Kelvin temperature, that is, 28.3K and 31.7K. Relative to the average temperature of the temperature field of 773.6K, the error is 3.6% and 4.1%.

[0072] It can be seen that the overall error of the prediction results is within an acceptable range. However, it is also necessary to measure the local error to better reflect the performance of the algorithm. As shown in Table 1, the predicted temperature field for working condition No. 4805 and the temperature difference of the corresponding 8 thermocouples, it can be seen that the network prediction results can also well match the real data in terms of local details.

[0073] Table 1 Comparison of LSTM predicted temperatures at eight locations

[0074]

[0075] In addition to requiring the predicted temperature field to conform to the actual working conditions, it is also hoped that the predicted results can conform to the temperature change trend. Figures 3 to 10 The actual and predicted values ​​of the 8 thermocouples shown are in condition 4805 and for the next 20 hours.

[0076] In the data preprocessing stage, this example rearranges the temperature field into a 24*32*652 matrix for easy network input and output. In the last step, the matrix needs to be restored to the pipe wall temperature field, so that the experimental results can be more intuitive. Fig.11 and Fig.12 The figure shows the three-dimensional scattered point modeling results of the calculated temperature field and the predicted temperature field in Matlab for working condition 4805. The gap in the pipeline is due to the uneven distribution of calculation points, which leads to the lack of corresponding calculation points in the restoration process.

Claims

1. A real-time prediction method for heating furnace temperature field based on convolutional long short-term memory network, characterized in that: The steps include: 1) Build a real-time prediction network for temperature field; 2) Training the temperature field real-time prediction network; including: (1) Each data of the industrial heating furnace is normalized; the data of the industrial heating furnace includes: each temperature, pressure, oxygen content, and gas flow value of the distributed control system in the industrial heating furnace measured at a set time; the normalization is performed using the following formula: Among them, x is the data to be normalized, mean, max, and min are the mean, maximum, and minimum values ​​within the training sample range of the corresponding working condition data respectively; (2) Three-dimensional expansion of industrial heating furnace data; The actual working condition is only a 1-dimensional vector, which does not match the 3-dimensional temperature field output by the target network. Therefore, to expand the 1-dimensional working condition into 3-dimensional, it is necessary to perform 3-dimensional expansion on the data of the industrial heating furnace. The values ​​of various temperatures, pressures, oxygen content, and gas flow of the distributed control system in the industrial heating furnace measured at the set time are expanded by expanding 1 working condition data into a 32*652 matrix, and each element in the matrix is ​​filled with the working condition data; and for the temperature data of the 8 thermocouples on the pipe wall, when expanding, only 50 elements near the spatial position corresponding to the thermocouple are filled with the temperature data measured by the thermocouple, and the remaining elements are set to zero, and finally 18*32*652 input data is obtained; (3) Convert the pipe wall temperature field into a three-dimensional temperature field The three-dimensional temperature field also contains the structural information of the pipe wall, but such a temperature field cannot be used as a network output, so the calculation points on the pipe wall temperature field must be rearranged into a three-dimensional temperature field; First, the pipeline is divided into 24 pipelines along the z-axis, and then the pipe wall is cut and unfolded into a rectangle along the axis. The unfolding method is to select 32 points in the direction of the pipeline axis in sequence, and arrange the 32 points in the direction of the z-axis to obtain the first column of data; then select 32 points in the direction of the pipeline axis and continue to arrange them, repeat 652 times to obtain a 24*32*652 matrix. At the same time, the temperature values ​​of the temperature field are divided by 1000 so that the value range is between 0 and 1, and the spatial coordinates of the pipe wall temperature field corresponding to each point in the three-dimensional temperature field are retained; (4) The temperature field real-time prediction network was trained without freezing the network parameters during training. The initial learning rate during training was 0.01, the batch size during training was 1, the L1 loss function was selected as the loss function, and the Adam algorithm was used to optimize the loss function. A total of 600 iterations were performed, and the learning rate was multiplied by 0.8 for decay every 40 iterations. The network achieved the best effect at the 480th iteration, and the optimal model was obtained. The mean absolute error of the optimal model output was 28.3 Kelvin. 3) Normalize the data of industrial heating furnaces respectively; 4) Three-dimensional expansion of industrial heating furnace data; The expansion method is to expand one working condition data into a 32*652 matrix, and each element in the matrix is ​​filled with the working condition data; and for the temperature data of the 8 thermocouples on the pipe wall, when expanding, only 50 elements near the spatial position corresponding to the thermocouple are filled with the temperature data measured by the thermocouple, and the remaining elements are set to zero, and finally 18*32*652 input data is obtained; 5) Sending the expanded data to the trained temperature field real-time prediction network to obtain the three-dimensional temperature field prediction results; 6) The obtained three-dimensional temperature field prediction result is restored to the pipe wall temperature field through the spatial coordinates of the pipe wall temperature field corresponding to each point of the three-dimensional temperature field retained in step 2).

2. The method for real-time prediction of heating furnace temperature field based on convolutional long short-term memory network according to claim 1 is characterized in that: The temperature field real-time prediction network described in step 1) has a total of 4 layers of convolutional long short-term memory networks (ConvLSTM), the number of input data channels is 18, the number of output data channels of the first layer of long short-term memory network (ConvLSTM1) is 24, the number of output data channels of the second layer of long short-term memory network (ConvLSTM2) is 48, and the number of output data channels of the third layer of long short-term memory network (ConvLSTM3) is 24. The output data of the first layer of long short-term memory network (ConvLSTM1) and the third layer of long short-term memory network (ConvLSTM3) are spliced ​​to obtain data with a channel number of 48, and sent to the fourth layer of long short-term memory network (ConvLSTM4), and finally a 24-channel temperature field output (output) is obtained; At the same time, the input and output time series lengths of the first-layer long short-term memory network (ConvLSTM1), the second-layer long short-term memory network (ConvLSTM2), and the third-layer long short-term memory network (ConvLSTM3) are all 10, and the input time series length of the fourth-layer long short-term memory network (ConvLSTM4) is 10, and the output time series length is 3.

3. The real-time prediction method for the heating furnace temperature field based on the convolutional long short-term memory network according to claim 1 is characterized in that: The data of the industrial heating furnace in step 3) include: the values ​​of the temperature, pressure, oxygen content, and gas flow of the distributed control system in the industrial heating furnace measured at the set time; the normalization process is performed using the following formula: Among them, x is the data to be normalized, mean, max, and min are the mean, maximum, and minimum values ​​within the training sample range of the corresponding working condition data, respectively.