Industrial heating furnace temperature prediction method and device based on deep learning and storage medium

Abstract

The invention provides an industrial heating furnace temperature prediction method and device based on deep learning, and a storage medium, and belongs to the technical field of temperature prediction, and the method comprises the steps: collecting the operation data of an industrial heating furnace, and carrying out the preprocessing; constructing a temperature prediction network model, and standardizing input data; constructing an input projection layer extension feature number; two layers of gating circulation units are constructed, and two stages of GRU networks are arranged and are connected through residual errors; the first-stage GRU network learns a time dependency relationship of the input data, and outputs a hidden state sequence to a second-stage GRU network to extract key time sequence context information; an output layer is constructed, final temperature prediction and normalization are carried out, an output temperature increment is subjected to destandardization, and an actual temperature value is obtained; performing hyper-parameter training on the temperature prediction network architecture to obtain a temperature prediction network model with a temperature change prediction function; and predicting the temperature prediction network model by using actual data to obtain a temperature prediction result.

Description

Technical Field

[0001] This invention belongs to the field of temperature prediction technology, specifically relating to the field of temperature prediction technology for industrial heating furnaces, and is used for accurate prediction of temperature changes in industrial heating furnaces over future periods. Background Technology

[0002] Industrial heating furnaces are widely used equipment in modern manufacturing, especially in industries such as metallurgy, chemicals, and glass manufacturing. The main function of a heating furnace is to raise the temperature of raw materials to the required process temperature to ensure product quality and production efficiency. Therefore, temperature control during the heating process is crucial for product quality, production stability, and efficient energy utilization.

[0003] Industrial heating furnaces are complex thermal systems characterized by large hysteresis, strong nonlinearity, and strong coupling. Their temperature is influenced by multiple factors, exhibiting nonlinearity, multiple time delays, and disturbance uncertainties. Traditional PID control often struggles to achieve high-precision temperature regulation in dynamic environments. Excessive temperature fluctuations not only lead to energy waste and reduced heating efficiency but can also affect workpiece quality, product performance, and yield. Furthermore, high-temperature operation of the furnace significantly impacts equipment lifespan and safe operation; insufficient prediction can easily result in excessive energy consumption and safety hazards.

[0004] Using deep learning to solve problems in intelligent manufacturing is becoming a growing trend. By predicting furnace temperature changes in advance, feedforward compensation and optimization decisions can be provided for the control system, enhancing its intelligence and adaptability. Summary of the Invention

[0005] The purpose of this invention is to solve the technical problem of accurately predicting the future temperature sequence of industrial heating furnaces.

[0006] This invention is achieved through the following technical solution: This invention proposes a method, device, and storage medium for predicting the temperature of an industrial heating furnace based on deep learning; This invention proposes a deep learning-based method for predicting the temperature of industrial heating furnaces, the method comprising: S1. Collect industrial heating furnace operation data, preprocess the heating furnace operation data to obtain training set and test set; the heating furnace operation data includes historical power and temperature sequences; S2. Construct a temperature prediction network model, including: The historical power and temperature sequences are standardized and then used as input data into the model. An input projection layer is constructed based on the input data; the input projection layer is set as a fully connected layer, and a linear layer is used to map from the original feature space to the hidden space to expand the number of features; A two-layer gated recurrent unit is constructed, which is set as a two-level GRU network and connected by residuals. The first-level GRU network learns the temporal dependency of the input data and outputs a hidden state sequence to the second-level GRU network. After deep feature learning and weighting, key temporal context information is extracted. Construct an output layer, set the output layer as a fully connected layer, and add the hidden state of the last time step of the second-level GRU network output to perform final temperature prediction and normalization, and output the temperature increment; The temperature increment is denormalized to obtain the actual temperature value; S3. The temperature prediction network architecture is trained using the training set, and a temperature prediction network model with temperature change prediction function is obtained by training hyperparameters. S4. Use actual data to make predictions on the temperature prediction network model to obtain temperature prediction results.

[0007] Furthermore, the industrial heating furnace described in S1 is made of copper, has a heating power range of 0-5.5kw, a target heating temperature range of 30-400℃, and a precision control range of ±1℃.

[0008] Furthermore, the number of extended features described in S2 is expanded from 2 to 32.

[0009] Furthermore, the hidden space described in S2 serves as a hidden unit, with a random discard rate of 10%.

[0010] Furthermore, the training hyperparameters mentioned in S3 include: a batch size of 505 in the training set, a total training batch size of 184, a learning rate of 0.001, a maximum number of iterations of 100, and an optimizer of Adam.

[0011] Furthermore, the training described in S3 employs an early stopping mechanism: if the validation loss based on the training set is monitored and it is determined that the validation loss has not decreased in 30 consecutive training rounds, then training is stopped.

[0012] Furthermore, the loss function is defined as follows: , In the formula, N is the number of training data. For the i-th training data, the predicted value is... Let be the true value of the i-th training data.

[0013] Furthermore, the prediction method described in S4 includes: Manually specify the prediction start point; Read the historical power and temperature sequence 1000 seconds before the predicted starting point; Input the power for the next 200 seconds; The temperature prediction network model is used to perform autoregressive prediction. The temperature increment predicted at the current time step is used as the input of the model at the next time step. The process is iterated until the temperature prediction sequence corresponding to the next 200 seconds is generated, which is then used as the final prediction result.

[0014] This invention also proposes a deep learning-based industrial heating furnace temperature prediction device, the device comprising: Data acquisition and preprocessing module: used to acquire industrial heating furnace operation data, preprocess the heating furnace operation data to obtain training set and test set; the heating furnace operation data includes historical power and temperature sequences; Model building module: Constructs a temperature prediction network model, including: standardizing the historical power and temperature sequences and then inputting them into the model as input data; An input projection layer is constructed based on the input data; the input projection layer is set as a fully connected layer, and a linear layer is used to map from the original feature space to the hidden space to expand the number of features; A two-layer gated recurrent unit is constructed, which is set as a two-level GRU network and connected by residuals. The first-level GRU network learns the temporal dependency of the input data and outputs a hidden state sequence to the second-level GRU network. After deep feature learning and weighting, key temporal context information is extracted. Construct an output layer, set the output layer as a fully connected layer, and add the hidden state of the last time step of the second-level GRU network output to perform final temperature prediction and normalization, and output the temperature increment; The temperature increment is denormalized to obtain the actual temperature value; Model training module: used to train the temperature prediction network architecture using the training set, and obtain a temperature prediction network model with temperature change prediction function by training hyperparameters; The results output module is used to use actual data to predict the temperature prediction network model and obtain the temperature prediction results.

[0015] The present invention also proposes a storage medium storing a computer program, which, when run, implements the method described in any one of the present invention.

[0016] The beneficial effects of this invention are: This invention innovatively proposes a deep learning-based temperature prediction network model, SimpleTempPredictor, for industrial heating furnaces. It employs a two-layer GRU with residual connections, and trains the model using industrial heating furnace operating data. The loss function is the mean squared error (MSELoss) to measure the difference between the predicted result and the true value. Whenever the loss decreases, the current optimal model state is saved. When making temperature predictions, the starting point is input, and the model and data can be loaded. Historical temperature and power data are used to predict future temperature increments, ultimately outputting the temperature over a future period and a visual image. Traditional temperature prediction methods struggle to adapt to the large hysteresis and strong nonlinearity of industrial heating furnaces, resulting in significant deficiencies in prediction accuracy and speed. In contrast, the temperature prediction network model described in this invention can quickly and accurately predict the future temperature of the heating furnace, exhibiting high accuracy and strong real-time performance.

[0017] The SimpleTempPredictor temperature prediction network model described in this invention predicts the temperature of industrial heating furnaces. Its inputs are the temperature and power of the industrial heating furnace in the previous 1000 seconds and the power in the next 200 seconds. The output is the temperature prediction result, including the temperature prediction for the next 200 seconds and the temperature change curve. This provides a guarantee for further temperature control of industrial heating furnaces and has certain methodological advantages.

[0018] The architecture of the temperature prediction network model is the SimpleTempPredictor deep learning network architecture designed from the perspective of time series modeling. This architecture maps the power-temperature sequence to the latent space through the input projection layer, and combines a two-layer GRU network with residual connections to fully capture long-term and short-term dynamic features. On this basis, an attention mechanism is introduced to perform weighted modeling of key time segments, and finally, the temperature increment of the next time step is predicted through a fully connected output layer, thereby completing the prediction of the temperature of the industrial heating furnace.

[0019] The industrial furnace operation data described in this invention encompasses samples of industrial furnace operation states, used in the SimpleTempPredictor network architecture for temperature prediction. Specifically, the industrial furnace operation state sample portion generates a large amount of temperature data under the heating state of the industrial furnace, forming a training sample library and a test sample library. The SimpleTempPredictor network architecture provides the network design structure. The temperature prediction portion trains and tests the industrial furnace operation state sample library based on SimpleTempPredictor to achieve temperature prediction functionality. Using the SimpleTempPredictor network model trained with the training sample library, in the scenario of furnace operation condition monitoring, it predicts and estimates the time-series data of temperature changes caused by fluctuations in operating conditions, estimating the temperature increment and temperature evolution trend at future time steps. This provides a guarantee for subsequent furnace operation control and energy efficiency optimization, meeting the actual needs of industrial applications and facilitating deployment and implementation. Furthermore, the industrial furnace operation data is obtained through multi-sensor power and temperature signal acquisition experiments of industrial furnaces, acquiring the required power and temperature information. The large amount of data obtained ensures the large sample requirements of deep learning methods and guarantees the smooth application of deep learning algorithms.

[0020] This invention utilizes deep learning technology to better meet the temperature prediction needs of industrial heating furnaces with large hysteresis and strong nonlinearity. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0022] Figure 1 This is a flowchart of a deep learning-based method for predicting the temperature of an industrial heating furnace, as described in this invention.

[0023] Figure 2 This is a diagram of the temperature prediction deep network model described in this invention.

[0024] Figure 3 This is a flowchart of the overall temperature prediction network model described in this invention.

[0025] Figure 4 This is a diagram showing the training results of the temperature prediction network model described in this invention.

[0026] Figure 5 This is a flowchart illustrating the verification process of the temperature prediction network model described in this invention.

[0027] Figure 6 This is a diagram showing the prediction results of the temperature prediction network model described in this invention. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] Specific implementation method one, the industrial heating furnace temperature prediction method based on deep learning described in this invention, such as... Figure 1 As shown, the method includes: S1. Collect industrial heating furnace operation data, preprocess the heating furnace operation data to obtain training set and test set; the heating furnace operation data includes historical power and temperature sequences; S2. Construct a temperature prediction network model and use the historical power and temperature sequences as model input data; The architecture of the temperature prediction network model consists of an input projection layer, two gated recurrent units, an intermediate linear transformation layer, an attention weighting layer, and an output layer; wherein, The input projection layer expands the number of features in the input data; The two-layer gated loop unit fuses temporal features from the expanded features; The attention-weighted layer extracts key temporal context information; The output layer predicts the temperature increment at the next time step; S3. The temperature prediction network architecture is trained using the training set, and a temperature prediction network model with temperature change prediction function is obtained by training hyperparameters. S4. Use the test set to make predictions on the temperature prediction network model to obtain temperature prediction results.

[0030] In this embodiment, the method begins with step one, in which industrial heating furnace operation data is collected, and the heating furnace operation data is preprocessed to obtain training set and test set; the heating furnace operation data includes historical power and temperature sequences; the industrial heating furnace is made of copper, the heating power range is 0-5.5kw, the target heating temperature range is 30-400℃, and the accuracy control range is ±1℃.

[0031] Specifically, based on the multi-sensor power and temperature signal acquisition experiment of the industrial heating furnace, the operating status samples of the industrial heating furnace were obtained, and the operating status samples of the industrial heating furnace were grouped into a training sample library and a test sample library. The SimpleTempPredictor architecture is constructed based on time series modeling. The SimpleTempPredictor architecture includes an input projection layer, two gated recurrent unit (GRU) modules, an intermediate linear transformation layer, an attention weighting layer, and an output layer. The SimpleTempPredictor architecture was trained using industrial furnace operation data as training samples to train hyperparameters and obtain a SimpleTempPredictor network model with temperature change prediction capabilities. The SimpleTempPredictor network model was then tested using industrial furnace operation data as test samples to predict the temperature of the industrial furnace under different power levels during the heating process.

[0032] In this embodiment, the industrial heating furnace operation data sample can be the operating status of the industrial heating furnace collected by sensors during the heating process; the industrial heating furnace operation data sample is the power and temperature during the heating process, which is input into the SimpleTempPredictor architecture to obtain the predicted temperature at the output layer.

[0033] Furthermore, the industrial heating furnace is made of copper, with a maximum heating power of 5.5 kW, an adjustable target heating temperature of 30-400℃ with an accuracy of ±1℃, and a heating method of resistance wire heating.

[0034] Then proceed to step two, in which, as follows: Figure 2 As shown, a temperature prediction network model is constructed, including: The historical power and temperature sequences are standardized and then used as input data into the model. An input projection layer is constructed based on the input data; the input projection layer is set as a fully connected layer, and a linear layer is used to map from the original feature space to the hidden space to expand the number of features; A two-layer gated recurrent unit is constructed, which is set as a two-level GRU network and connected by residuals. The first-level GRU network learns the temporal dependency of the input data and outputs a hidden state sequence to the second-level GRU network. After deep feature learning and weighting, key temporal context information is extracted. Construct an output layer, set the output layer as a fully connected layer, and add the hidden state of the last time step of the second-level GRU network output to perform final temperature prediction and normalization, and output the temperature increment; The temperature increment is denormalized to obtain the actual temperature value; The number of extended features is increased from 2 to 32.

[0035] The hidden space serves as a hidden unit, with a random discard rate of 10%.

[0036] Specifically, such as Figure 3 As shown, the SimpleTempPredictor architecture takes historical power and temperature sequences of length seq_len as input, expands the number of input features from 2 (temperature and power) to 32, and sequentially passes them through a two-stage GRU network and uses residual connections to fuse temporal features. During GRU training, 10% of hidden units are randomly dropped to prevent overfitting. Then, an attention mechanism is used to extract key temporal context information, and finally, the temperature increment of the next time step is predicted through a fully connected output layer. The model as a whole uses the ReLU activation function and Dropout regularization, which can balance the ability to model temporal dependencies and generalization performance.

[0037] Then, step three is performed. In step three, the temperature prediction network architecture is trained using the training set, and a temperature prediction network model with temperature change prediction function is obtained by training hyperparameters.

[0038] The training hyperparameters include: batch size of 505 in the training set, total training batches of 184, learning rate of 0.001, maximum number of iterations of 100, and the optimizer is Adam.

[0039] Furthermore, the training described in S3 employs an early stopping mechanism: if the validation loss based on the training set is monitored and it is determined that the validation loss has not decreased in 30 consecutive training rounds, then training is stopped.

[0040] Furthermore, the loss function is defined as follows: , In the formula, N is the number of training data. For the i-th training data, the predicted value is... Let be the true value of the i-th training data.

[0041] Specifically, the hyperparameters are set as follows: 505 training samples per batch, 184 batches in total, learning rate of 0.001, maximum iteration of 100 rounds, and the optimization method is Adam.

[0042] like Figure 4As shown, the SimpleTempPredictor architecture selects 80% of all samples from the industrial heating furnace operation data as training samples and inputs them into SimpleTempPredictor for training. The training hyperparameters are: 505 training trees per batch, a total of 184 batches, a learning rate of 0.001, a maximum of 100 iterations, and the optimization method is Adam. The trained SimpleTempPredictor network model is obtained. Furthermore, during model training, an early stopping mechanism is used. By monitoring changes in the validation loss, if the validation loss does not improve within 30 consecutive iterations, early stopping is triggered, and training is stopped.

[0043] Then, step four is performed, in which actual data is used to predict the temperature prediction network model and obtain the temperature prediction result.

[0044] The prediction method includes: Manually specify the prediction start point; Read the historical power and temperature sequence 1000 seconds before the predicted starting point; Input the power for the next 200 seconds; The temperature prediction network model is used to perform autoregressive prediction. The temperature increment predicted at the current time step is used as the input of the model at the next time step. The process is iterated until the temperature prediction sequence corresponding to the next 200 seconds is generated, which is then used as the final prediction result.

[0045] Specifically, such as Figure 5 As shown, when testing the SimpleTempPredictor network model, the starting point for prediction is manually entered. The model will read the historical data of the previous 1,000 points, including the power and temperature of the previous 1,000 points, as well as the power for the next 200 seconds, and predict the temperature for the next 200 seconds.

[0046] Furthermore, such as Figure 6 As shown, NumPy numerical calculations were used to calculate the point-by-point difference between the predicted and actual temperatures, yielding evaluation indicators such as mean absolute error (MAE), root mean square error (RMSE), and maximum error. Subsequently, based on Matplotlib plotting tools, the temperature-power curve, prediction interval curve, error distribution histogram, and predicted-actual value scatter plot of the industrial heating furnace were used for visualization analysis, thereby comprehensively evaluating the prediction accuracy and error characteristics of the SimpleTempPredictor model.

[0047] Specific implementation method two, the following describes the process of predicting the temperature of an industrial heating furnace using a deep learning-based industrial heating furnace temperature prediction method according to the present invention: An input layer is constructed to standardize the data before input. The input temperature and power data are passed through a linear layer to map them from the original feature space to the hidden space. This operation expands the number of input features from 2 (temperature and power) to 32. 10% of the hidden units are randomly dropped during GRU training to prevent overfitting. Each training sample contains 50 historical time steps.

[0048] A two-layer GRU is constructed using residual connections, feeding the processed input data into the first-layer GRU network. In the first layer, the GRU learns the temporal dependencies of the input data and outputs a sequence of hidden states. The output of the first-layer GRU serves as the input to the second-layer GRU for further data processing. The second-layer GRU continues to learn deeper features from the time-series data. The GRU output at each time step is weighted, focusing on more important time steps to improve prediction accuracy. The output layer is constructed by feeding the GRU's output (usually the hidden state at the last time step) into a fully connected layer for final temperature prediction. This layer maps the GRU's output to the target output, i.e., the temperature increment. The model output is a normalized temperature increment, which is then denormalized to convert the increment into the actual temperature value. The industrial heating furnace is connected to a server to collect data on heating power and temperature. After data collection, the dataset is divided into a training set and a validation set. Configure logging for easy tracking and debugging. Logs should include information such as the training process, data loading status, and validation results. Load temperature and power data from a file. Use a sliding window technique to predict the next temperature increment from the temperature and power sequence. Predict one temperature increment at a time and use it as the training target. Generate training samples for each file and return the features and target. Temperature and power data are standardized to ensure all feature values ​​are within a similar range. MinMaxScaler is used to standardize the data. The standardization process handles training and testing data separately, ensuring the same standardization parameters are used for both training and testing. The Adam optimizer is used to train the model, and the loss function chosen is MSELoss (mean squared error) to measure the difference between the predicted results and the true values. ReduceLROnPlateau is used to dynamically adjust the learning rate; the learning rate is reduced when the validation loss no longer decreases. The training progress is displayed using the trange and tqdm libraries. During each epoch of training, the training loss is calculated and the model weights are updated. After each epoch, the model is evaluated using a validation set, and the validation loss is calculated. By monitoring changes in the validation loss, if the validation loss does not improve within 30 consecutive rounds, early stop is triggered to stop training. Whenever the validation loss decreases, the current optimal model state is saved. The saved information includes the model weights, optimizer state, training loss, validation loss, etc. After training, load the best-performing model and evaluate it using a test set. Calculate the loss on the test set and output the model's performance. Output a summary of each stage of the training process, print metrics such as training loss, validation loss, and test loss, and record the save path and related information of the best model; Manually input the prediction starting point (i.e., the starting point of each prediction interval). Multiple custom starting points are allowed, and the model will start predictions from each of these starting points. Load the model and data; the data is a validation set from experimental sampling data of the heating system, used to test the effectiveness of temperature prediction. The read data will be standardized by a normalizer to ensure that the model input meets expectations. Extract historical temperature and power data from the raw data; For each interval, the model uses historical temperature and power data to predict future temperature increments. This process is autoregressive, meaning it predicts one temperature increment at a time and uses that increment as input for the next time step. Depending on the given number of prediction intervals and step size, the function selects multiple starting points in the data for prediction. For each interval, the function uses historical data for prediction and progressively updates the historical data window for the next time step. Calculate model errors, including mean absolute error and root mean square error. Create visualizations to compare predictions with actual data. Display the prediction error metrics in the lower right corner of the charts to help users understand the model's performance. Once the prediction is complete, the results are stored and a chart is generated. The chart is saved as an image file for users to view.

[0049] This temperature prediction model has achieved relatively ideal results. Based on the temperature and power of the first 1000 data points and the power for the next 200 seconds, it can predict the temperature for the next 200 seconds. The average absolute error of the temperature prediction can be controlled within 1℃, and the maximum error does not exceed 1℃.

[0050] Specific implementation method three: A deep learning-based industrial heating furnace temperature prediction device, the device comprising: Data acquisition and preprocessing module: used to acquire industrial heating furnace operation data, preprocess the heating furnace operation data to obtain training set and test set; the heating furnace operation data includes historical power and temperature sequences; The model building module is used to construct a temperature prediction network model and also to use the historical power and temperature sequences as model input data. The architecture of the temperature prediction network model consists of an input projection layer, two layers of gated recurrent units, an intermediate linear transformation layer, an attention weighting layer, and an output layer. The input projection layer expands the number of features in the input data; The two-layer gated loop unit fuses temporal features from the expanded features; The attention-weighted layer extracts key temporal context information; The output layer predicts the temperature increment at the next time step.

[0051] Model training module: used to train the temperature prediction network architecture using the training set, and obtain a temperature prediction network model with temperature change prediction function by training hyperparameters; The result output module is used to use the test set to make predictions on the temperature prediction network model and obtain temperature prediction results.

[0052] Specific implementation method four: The storage medium described in this embodiment stores a computer program, which implements the method described in any one of the present invention when running.

Claims

1. A method for predicting the temperature of an industrial heating furnace based on deep learning, characterized in that, The method includes: S1. Collect industrial heating furnace operation data, preprocess the heating furnace operation data to obtain training set and test set; the heating furnace operation data includes historical power and temperature sequences; S2. Construct a temperature prediction network model, including: The historical power and temperature sequences are standardized and then used as input data into the model. An input projection layer is constructed based on the input data; the input projection layer is set as a fully connected layer, and a linear layer is used to map from the original feature space to the hidden space to expand the number of features; A two-layer gated recurrent unit is constructed, which is set as a two-level GRU network and connected by residuals. The first-level GRU network learns the temporal dependency of the input data and outputs a hidden state sequence to the second-level GRU network. After deep feature learning and weighting, key temporal context information is extracted. Construct an output layer, set the output layer as a fully connected layer, and add the hidden state of the last time step of the second-level GRU network output to perform final temperature prediction and normalization, and output the temperature increment; The temperature increment is denormalized to obtain the actual temperature value; S3. The temperature prediction network architecture is trained using the training set, and a temperature prediction network model with temperature change prediction function is obtained by training hyperparameters. S4. Use actual data to make predictions on the temperature prediction network model to obtain temperature prediction results.

2. The method for predicting the temperature of an industrial heating furnace based on deep learning according to claim 1, characterized in that, The industrial heating furnace described in S1 is made of copper, with a heating power range of 0-5.5kw, a target heating temperature range of 30-400℃, and a precision control range of ±1℃.

3. The method for predicting the temperature of an industrial heating furnace based on deep learning according to claim 1, characterized in that, The number of extended features described in S2 is expanded from 2 to 32.

4. The method for predicting the temperature of an industrial heating furnace based on deep learning according to claim 1, characterized in that, The hidden space described in S2 is a hidden unit with a random discard rate of 10%.

5. The method for predicting the temperature of an industrial heating furnace based on deep learning according to claim 1, characterized in that, The training hyperparameters mentioned in S3 include: batch size of 505 in the training set, total training batches of 184, learning rate of 0.001, maximum number of iterations of 100, and the optimizer being Adam.

6. The method for predicting the temperature of an industrial heating furnace based on deep learning according to claim 1, characterized in that, The training described in S3 employs an early stopping mechanism: if the validation loss based on the training set is monitored and it is determined that the validation loss has not decreased in 30 consecutive training rounds, then training is stopped.

7. The method for predicting the temperature of an industrial heating furnace based on deep learning according to claim 6, characterized in that, The loss function is defined as follows: ， In the formula, N The number of training data, For the i-th training data, the predicted value is... Let be the true value of the i-th training data.

8. The method for predicting the temperature of an industrial heating furnace based on deep learning according to claim 1, characterized in that, The prediction methods described in S4 include: Manually specify the prediction start point; Read the historical power and temperature sequence 1000 seconds before the predicted starting point; Input the power for the next 200 seconds; The temperature prediction network model is used to perform autoregressive prediction. The temperature increment predicted at the current time step is used as the input of the model at the next time step. The process is iterated until the temperature prediction sequence corresponding to the next 200 seconds is generated, which is then used as the final prediction result.

9. A deep learning-based industrial heating furnace temperature prediction device, characterized in that, The device includes: Data acquisition and preprocessing module: used to acquire industrial heating furnace operation data, preprocess the heating furnace operation data to obtain training set and test set; the heating furnace operation data includes historical power and temperature sequences; Model building module: Constructs a temperature prediction network model, including: standardizing the historical power and temperature sequences and then inputting them into the model as input data; An input projection layer is constructed based on the input data; the input projection layer is set as a fully connected layer, and a linear layer is used to map from the original feature space to the hidden space to expand the number of features; A two-layer gated recurrent unit is constructed, which is set as a two-level GRU network and connected by residuals. The first-level GRU network learns the temporal dependency of the input data and outputs a hidden state sequence to the second-level GRU network. After deep feature learning and weighting, key temporal context information is extracted. Construct an output layer, set the output layer as a fully connected layer, and add the hidden state of the last time step of the second-level GRU network output to perform final temperature prediction and normalization, and output the temperature increment; The temperature increment is denormalized to obtain the actual temperature value; Model training module: used to train the temperature prediction network architecture using the training set, and obtain a temperature prediction network model with temperature change prediction function by training hyperparameters; The results output module is used to use actual data to predict the temperature prediction network model and obtain the temperature prediction results.

10. A storage medium, characterized in that, The storage medium stores a computer program that, when executed, implements the method described in any one of claims 1-8.

Images