Tin smelting hearth oxygen content delay compensation control method based on TCN

Abstract

The invention discloses a tin smelting hearth oxygen content delay compensation control method based on a TCN, and belongs to the technical field of automation. Performing time sequence decomposition on the cleaned data by using a seasonal trend decomposition method to obtain a trend component, a seasonal component and a residual component; constructing a dynamic prediction model of the oxygen content in the chamber based on a time convolution network (TCN), and evaluating a training result through a mean square error; and replacing the transfer function model in the reference adaptive control framework with the trained time convolutional network TCN model, constructing a time delay compensation control framework based on the time convolutional network TCN to calculate a control law for compensating time lag, and completing model construction. According to the method, the control performance of a delay time-varying system is improved, and a curve has a higher convergence speed.

Description

Technical Field

[0001] The present invention relates to the field of automation technology, and particularly relates to a method for delaying oxygen content compensation control in a tin smelting furnace hearth based on TCN. Background Art

[0002] The top-blown furnace is an important link in the tin smelting process, and high-efficient recovery of valuable metals such as tin, germanium, and copper is achieved through high-temperature smelting in the furnace. The high-temperature smelting in the furnace hearth depends on carbon combustion for heat supply. By introducing oxygen into the furnace hearth to control the combustion of raw material carbon, the temperature in the furnace hearth is controlled within a reasonable range to ensure the efficiency and stability of the recovery process. Therefore, precise control of the oxygen content in the furnace hearth is crucial. Existing methods mainly use PID control, model reference adaptive control (MRAC), etc.

[0003] However, the temperature control system of the top-blown furnace has extremely strong nonlinearity, large time delay, and time-varying characteristics, specifically manifested as follows: 1. Nonlinear characteristics: The combustion process in the furnace hearth involves complex chemical reactions and heat transfer, resulting in strong nonlinearity in the entire smelting process. 2. Large time delay characteristics: After the raw materials enter the furnace hearth, there is a metal heat transfer process, resulting in a significant delay in the system response, and the data measured by the sensors in the insulation layer is also affected by the time delay. 3. Time-varying characteristics: The system delay time will change with the changes in reaction temperature, raw material composition, and process conditions, further increasing the difficulty of control. Existing technologies rely on the establishment of transfer function models, and other control methods such as model predictive control and robust control rely more on the establishment of system state space equations. Whether it is a transfer function or a state space equation, it is difficult to accurately describe the nonlinear and large time delay characteristics of chemical reactions and thermal effects in the smelting process, and the parameter identification step in the modeling process also increases a huge workload. Traditional control methods are difficult to achieve fast and stable oxygen content control, resulting in fluctuations in the production process, affecting product quality and production safety. Summary of the Invention

[0004] To solve the above problems, the present invention provides a method for delaying oxygen content compensation control in a tin smelting furnace hearth based on TCN.

[0005] To achieve the above technology, the specific steps are as follows:

[0006] S1. Collect furnace hearth data and perform data cleaning. Use the Seasonal and Trend decomposition using Loess (STL) method to decompose the cleaned data into three parts: trend component, seasonal component, and residual component. The steps are as follows:

[0007] S1.1. Collect furnace hearth data;

[0008] Furnace data, including: oxygen content in the furnace, oxygen input, temperature, and raw material composition, where the oxygen content data in the furnace is a time series;

[0009] The acquisition method is: export real-time production data from the Distributed Control System (DCS);

[0010] S1.2. Clean the collected furnace data, and the steps are as follows:

[0011] S1.2.1. Set a moving window with a slidable fixed step size in the exported real-time production data to preprocess the data within the window size;

[0012] S1.2.2. Preprocess the data within the window;

[0013] The preprocessing includes: mean removal processing and outlier detection and removal processing;

[0014] The mean removal processing is: calculate the mean of the data within the window and subtract the mean from each data point to eliminate the local trend;

[0015] The outlier detection and removal processing is: after detecting jump values and missing values caused by sensor failures or measurement errors within the window, remove them, and use the median filling method to complete the removed outliers to ensure the continuity of the data;

[0016] S1.3. Perform STL decomposition on the cleaned oxygen content, oxygen input, and temperature in the furnace, and sequentially extract the trend component, seasonal component, and residual component;

[0017] The trend component reflects the long-term change law; the seasonal component reflects the periodic change; perform statistical analysis on the residual component to remove noise and interference data.

[0018] S2. Construct a dynamic prediction model for the oxygen content in the furnace based on the Temporal Convolutional Network (TCN);

[0019] The Temporal Convolutional Network TCN is set with 3 residual blocks, namely: the first residual block, the second residual block, and the third residual block;

[0020] Each residual block includes two convolutional operations and one residual calculation; the convolutional operation is as follows: after the data input passes through the basic convolutional layer, weight normalization is performed, and the output of the data is obtained after passing through the Relu activation function; in order to prevent overfitting of the data, a Dropout layer is also provided after the Relu activation function; the basic convolutional layer uses dilated convolution operation and causal convolution operation, and the dilation coefficient of the dilated convolution is set to twice the number of the residual block where it is located.

[0021] The data is input from the first residual block and output from the third residual block after passing through the second residual block; the output channels of each residual block are different, and by setting different output channels, it is used to capture the implicit relationship between features.

[0022] The construction steps are as follows:

[0023] S2.1. Input the data into the first residual block, perform residual calculation on the result obtained after two convolutional operations and the original features, and then output.

[0024] The residual calculation is a 1×1 residual connection operation.

[0025] The data is input into the first residual block and the first convolutional operation and the second convolutional operation are performed successively. The first convolutional operation sequentially performs causal convolution operation, dilated convolution operation, weight normalization operation, Relu activation function operation, and Dropout operation; among them, the convolutional kernel sizes of the dilated convolution and the causal convolution are 3, and the dilation coefficient of the dilated convolution is 2; Dropout is set to 0.3.

[0026] The second convolutional operation is the same as the first convolutional operation.

[0027] The output channel is set to 5.

[0028] S2.2. Input the output of the first residual block into the second residual block.

[0029] The operations performed by the second residual block are similar to those of the first residual block, except that the dilation coefficient of the dilated convolution in the second residual block is 4, and the output channel is set to 12.

[0030] S2.3. Input the output of the second residual block into the third residual block.

[0031] The operations performed by the third residual block are similar to those of the first residual block and the second residual block, except that the dilation coefficient of the dilated convolution in the third residual block is 6, and the output channel is set to 1.

[0032] S3. Group the data obtained by STL decomposition and the raw material components, divide each group into a training set, a validation set, and a test set according to a ratio, and input them into the dynamic prediction model of the oxygen content in the furnace based on the temporal convolutional network for training to obtain the trained model, and evaluate the training results using the mean square error. The steps are as follows:

[0033] S3.1. Group the data obtained by STL decomposition and the raw material components, and input them into the dynamic prediction model of the oxygen content in the furnace based on the temporal convolutional network for training;

[0034] The grouping situation is as follows:

[0035] The trend component of the oxygen content in the furnace, the trend component of the oxygen input amount, the trend component of the temperature, and the raw material components;

[0036] The seasonal component of the oxygen content in the furnace, the seasonal component of the oxygen input amount, the seasonal component of the temperature trend, and the raw material components;

[0037] The residual component of the oxygen content in the furnace, the residual component of the oxygen input amount, the residual component of the temperature, and the raw material components;

[0038] S3.2. Set the window of the TCN model and divide the training set, validation set, and test set according to a ratio;

[0039] The setting method is as follows: the input channel is 4, and the sliding window length is 20, obtaining an input matrix dimension of 20×4;

[0040] The ratio setting is: 6:2:2;

[0041] S3.3. Initialize the TCN model;

[0042] S3.4. Input the training set into the model for training respectively;

[0043] The training set data can be used multiple times. The role of the training set is to adjust the model parameters and prevent overfitting;

[0044] After all the data in the training set have been trained for a complete training epoch, use the validation set to evaluate the model performance;

[0045] Select the mean square error MSE to evaluate the model performance. MSE can measure the difference between the predicted value and the true value. The smaller the MSE value, the better. The MSE expression is:

[0046]

[0047] In the formula, y i is the actual value of the oxygen content at the i-th moment, is the predicted value of the oxygen content at the i-th moment, and n represents the length of the oxygen content time series;

[0048] During the evaluation process, hyperparameters are adjusted. The hyperparameters of the TCN include: the Dropout ratio and the number of channels of the model;

[0049] Update the hyperparameters of the model with the best evaluation results on the validation set;

[0050] S3.7. Evaluate the generalization performance of the model using the test set. The evaluation metric is also selected as MSE. Different from the training set, the test set data is used after the training is completed and can only be used once. The reason for choosing MSE as the metric is to ensure that the test set and the training set are under the same standard and make the evaluation results consistent.

[0051] S4. Replace the transfer function model in the reference adaptive control framework (MRAC) with the trained TCN model, construct a time-delay compensation control framework based on TCN to calculate the control law for compensating the time delay, and complete the model construction;

[0052] The time-delay compensation control framework based on TCN includes: an outer loop, an inner loop, and a TCN model; among them, the outer loop is a conventional PID controller G c (s), and the inner loop includes: an ideal operating condition model G p (s) and an ideal operating condition model G p (s) under the condition of using a Smith predictor;

[0053] The output of the conventional PID controller G c (s) is denoted as U(s);

[0054] The output of the inner loop is denoted as K τ '(s);

[0055] G c (s) satisfies the following expression:

[0056] G c (s) = Y(s) - R(s)

[0057] In the formula, Y(s) is the output under the environmental disturbance; R(s) is the preset target value; in the present invention, Y(s) is set as the oxygen content in the time-delay link under the environmental disturbance of the TCN model; R(s) is the target oxygen content;

[0058] Add the output of the conventional PID controller G c (s) and the output K τ '(s) of the inner loop to obtain the control signal M τ (s), and the expression is as follows:

[0059] M τ (s) = U(s) + K τ '(s);

[0060] Control signal M τ (s) are respectively used as the signal inputs of the TCN model, the ideal operating condition model G p (s) and the ideal operating condition model G p (s) under the condition of using the Smith predictor; the delay link of the TCN model is obtained respectively Ideal operating condition model G p (s) delay link e -τs and the ideal operating condition model G p (s) delay link

[0061] Through the delay link of the TCN model and the delay link e of the ideal operating condition model G p (s), the error E -τs (s) is obtained, and the expression is as follows: τ (s), the expression is as follows:

[0062]

[0063] In the formula, T1 and T2 are the first time constant and the second time constant respectively; is the time-varying delay factor; τ is the delay factor; s is the complex variable in the complex frequency domain; K0 is the ratio coefficient;

[0064] Through the error E τ (s) and the output of the Smith predictor, the error E τ ’(s) after eliminating the delay is obtained;

[0065] The output expression of the Smith predictor is:

[0066]

[0067] In the formula, is the delay link of the ideal operating condition model G p (s) under the condition of using the Smith predictor; among them, the time constant τ max The value rule is as follows:

[0068]

[0069] The error E τ ’(s) after eliminating the delay is:

[0070] E τ ’(s) = output of the Smith predictor + E τ (s);

[0071] The error E τ’(s) After the integration operation The output K of the inner loop is obtained τ '(s), and the expression is as follows:

[0072]

[0073] In the formula, T τ is the time constant of the integration link, where β is the step size of descent;

[0074] S5. Verify the model effect;

[0075] The verification method is as follows: The TCN-based delay compensation control framework, PID, and MRAC are simultaneously used for the control of the oxygen content in the furnace under the same conditions.

[0076] Advantages of the present invention:

[0077] The present invention improves the control performance of the time-delay time-varying system. It is reflected in the following aspects: 1. By introducing the TCN network, this method can effectively capture the long-distance dependence relationship in the oxygen content data of the furnace, and combined with its parallel computing advantages, it realizes the efficient modeling and real-time control of complex dynamic systems. 2. The STL data preprocessing technology is used to decompose the oxygen content data into trend, seasonal, and residual components, removing noise and periodic fluctuations, improving the quality of the data and the robustness of the model, and enabling it to adaptively process non-stationary time series. 3. An adaptive compensation framework based on TCN is designed, and the control parameters are adjusted according to the prediction error to offset the error caused by the time-varying delay. Through optimizing the objective function and the principle of the steepest descent, the rapid convergence of the model tracking error is achieved. Compared with traditional methods, the present invention makes the oxygen content curve have a faster convergence speed. In addition, the present invention has intelligent parameter adjustment and real-time adaptability, can automatically optimize the control parameters and dynamically respond to the changes in working conditions, is not only applicable to the control of oxygen content, but also can be extended to other industrial process control scenarios with time-varying delay, and has a wide range of application prospects. Description of the Drawings

[0078] Figure 1 is the step flow chart of the present invention;

[0079] Figure 2 is the structural diagram of the time convolutional network of the present invention;

[0080] Figure 3 is the schematic diagram of sample collection in the embodiment of the present invention;

[0081] Figure 4 is the decomposition diagram of the seasonal trend decomposition method of the oxygen content data in the furnace in the embodiment of the present invention;

[0082] Figure 5This is the decomposition diagram of the seasonal trend decomposition method for the temperature data in the implementation of the present invention;

[0083] Figure 6 This is the decomposition diagram of the seasonal trend decomposition method for the oxygen input data in the implementation of the present invention;

[0084] Figure 7 This is the delay compensation control framework diagram of the present invention;

[0085] Figure 8 This is the result comparison diagram between the present invention and the prior art when the oxygen content is 50%;

[0086] Figure 9 This is the result comparison diagram between the present invention and the prior art when the oxygen content is 25%. Detailed implementation manners

[0087] The present invention will be further described in detail below in conjunction with specific embodiments.

[0088] As Figure 1 shown, a delay compensation control method for the oxygen content in the tin smelting furnace based on TCN comprises the following steps:

[0089] S1. Collect furnace data and perform data cleaning, and use the Seasonal and Trend decomposition using Loess (STL) method to decompose the cleaned data into three parts: trend component, seasonal component and residual component. The steps are as follows:

[0090] S1.1. Collect furnace data;

[0091] The furnace data includes: oxygen content in the furnace, oxygen input, temperature and raw material composition, wherein the oxygen content data in the furnace is a time series;

[0092] The collection method is: export real-time production data from the Distributed Control System (DCS);

[0093] S1.2. Perform data cleaning on the collected furnace data. The steps are as follows:

[0094] S1.2.1. Set a sliding fixed-step moving window in the exported real-time production data for preprocessing the data within the window size;

[0095] As Figure 3 shown, in this embodiment, the length of the moving window is set to 20; the fixed step is set to 15; the time length is 254 minutes;

[0096] S1.2.2. Preprocess the data within the window;

[0097] The preprocessing includes: mean removal processing and outlier detection and removal processing;

[0098] The mean removal processing is: calculate the mean of the data within the window, and subtract the mean from each data point to eliminate the local trend;

[0099] The outlier detection and removal processing is: after detecting jump values and missing values caused by sensor failures or measurement errors within the window, remove them, and use the median filling method to complete the removed outliers to ensure the continuity of the data;

[0100] S1.3. Perform STL decomposition on the oxygen content, oxygen input amount, and temperature in the furnace after cleaning, and sequentially extract the trend component, seasonal component, and residual component;

[0101] The trend component reflects the long-term change law; the seasonal component reflects the periodic change; perform statistical analysis on the residual component to remove noise and interference data.

[0102] S2. As Figure 3 shown, construct a dynamic prediction model for the oxygen content in the furnace based on the Temporal Convolutional Network (TCN);

[0103] The Temporal Convolutional Network TCN has 3 residual blocks, namely: the first residual block, the second residual block, and the third residual block;

[0104] Each residual block includes two convolutional operations and one residual calculation; the convolutional operation is: after the data input passes through the basic convolutional layer, perform weight normalization operation, and obtain the output of the data after passing through the Relu activation function; in order to prevent overfitting of the data, a Dropout layer is also provided after the Relu activation function; the basic convolutional layer uses dilated convolution operation and causal convolution operation, and the dilation coefficient of the dilated convolution is set to twice the number of the residual block where it is located;

[0105] The data is input from the first residual block and output from the third residual block after passing through the second residual block; the output channels of each residual block are different, and by setting different output channels, it is used to capture the implicit relationship between features;

[0106] The construction steps are as follows:

[0107] S2.1. Input the data into the first residual block, perform residual calculation on the result obtained after two convolutional operations and the original features, and then output;

[0108] The residual calculation is a 1×1 residual connection operation;

[0109] The data input first residual block performs the first convolution operation and the second convolution operation successively. The first convolution sequentially performs causal convolution operation, dilated convolution operation, weight normalization operation, Relu activation function operation, and Dropout operation. Among them, the convolution kernels of the dilated convolution and the causal convolution are of size 3, and the dilation coefficient of the dilated convolution is 2; Dropout is set to 0.3.

[0110] The second convolution operation is the same as the first convolution operation.

[0111] The output channels are set to 5.

[0112] S2.2. Input the output of the first residual block into the second residual block.

[0113] The operations performed by the second residual block are similar to those of the first residual block, except that the dilation coefficient of the dilated convolution in the second residual block is 4, and the output channels are set to 12.

[0114] S2.3. Input the output of the second residual block into the third residual block.

[0115] The operations performed by the third residual block are similar to those of the first residual block and the second residual block, except that the dilation coefficient of the dilated convolution in the third residual block is 6, and the output channels are set to 1.

[0116] S3. Group the data after STL decomposition and the raw material components, divide each group into a training set, a validation set, and a test set according to a ratio, and input them into the dynamic prediction model of the oxygen content in the furnace based on the temporal convolutional network for training to obtain the trained model, and evaluate the training results through the mean square error.

[0117] The steps are as follows:

[0118] S3.1. As Figure 4 . Figure 5 and Figure 6 shown, group the data after STL decomposition and the raw material components, and input them into the dynamic prediction model of the oxygen content in the furnace based on the temporal convolutional network for training respectively.

[0119] The grouping situation is as follows:

[0120] The trend component of the oxygen content in the furnace, the trend component of the oxygen input amount, the trend component of the temperature, and the raw material components;

[0121] The seasonal component of the oxygen content in the furnace, the seasonal component of the oxygen input amount, the seasonal component of the temperature trend, and the raw material components;

[0122] The residual component of the oxygen content in the furnace, the residual component of the oxygen input amount, the residual component of the temperature, and the raw material components;

[0123] S3.2. Set the window of the TCN model and divide the training set, validation set, and test set proportionally;

[0124] The setting method is as follows: the input channel is 4, and the sliding window length is 20, resulting in an input matrix dimension of 20×4;

[0125] The proportion is set to 6:2:2;

[0126] S3.3. Initialize the TCN model;

[0127] S3.4. Input the training set into the model for training respectively;

[0128] The training set data can be used multiple times. The role of the training set is to adjust the model parameters to prevent overfitting;

[0129] After all the data in the training set has been trained for a complete epoch, use the validation set to evaluate the model performance;

[0130] Select the mean squared error MSE to evaluate the model performance. MSE can measure the difference between the predicted value and the true value. The smaller the MSE value, the better. The MSE expression is:

[0131]

[0132] In the formula, y i is the actual value of the oxygen content at the i-th moment, is the predicted value of the oxygen content at the i-th moment, and n represents the length of the oxygen content time series;

[0133] Adjust the hyperparameters during the evaluation process. The hyperparameters of the TCN include: Dropout ratio, number of channels of the model;

[0134] When the model is overfitting, increase the Dropout ratio;

[0135] If additional variables are added to the input matrix, the number of input channels needs to be consistent with the dimension of the input variables. The adjustment of these hyperparameters needs to be manually adjusted according to actual needs;

[0136] Update the hyperparameters of the model with the best evaluation result on the validation set;

[0137] S3.7. Use the test set to evaluate the generalization performance of the model. The evaluation index also selects MSE. Different from the training set, the test set data is used after the training is completed and can only be used once. Selecting MSE as the index is to ensure that the test set and the training set are under the same standard and make the evaluation results consistent.

[0138] S4. Such as Figure 7As shown, the trained TCN model is used to replace the transfer function model in the reference model reference adaptive control (MRAC) framework, and a time-delay compensation control framework based on the TCN is constructed to calculate the control law for compensating the time delay, thus completing the model construction.

[0139] The time-delay compensation control framework based on the TCN includes: an outer loop, an inner loop, and a TCN model; where the outer loop is a conventional PID controller G c (s), and the inner loop includes: an ideal operating condition model G p (s) and an ideal operating condition model G p (s) under the condition of using a Smith predictor;

[0140] The output of the conventional PID controller G c (s) is denoted as U(s);

[0141] The output of the inner loop is denoted as K τ '(s);

[0142] G c (s) satisfies the following expression:

[0143] G c (s) = Y(s) - R(s)

[0144] In the formula, Y(s) is the output under environmental disturbance; R(s) is the preset target value; in the present invention, Y(s) is set to the oxygen content in the time-delay link under environmental disturbance of the TCN model; R(s) is the target oxygen content;

[0145] Using the expression of G c (s), a target function J is designed by the integral square error (ISE) index to calculate the time-varying time-delay adaptive compensation control law, and the expression is as follows:

[0146]

[0147] In the formula, e τ (t) represents the error between the predicted value and the ideal value, that is, e τ (t) = G c (s);

[0148] Adding the output of the conventional PID controller G c (s) and the output K τ '(s) of the inner loop to obtain the control signal M τ (s), and the expression is as follows:

[0149] M τ (s) = U(s) + K τ '(s);

[0150] The control signal M τ(s) are respectively used as the signal inputs of the TCN model and the ideal operating condition model G p (s) and the ideal operating condition model G p (s) under the condition of using the Smith predictor; the delay link of the TCN model is obtained respectively The delay link e of the ideal operating condition model G p (s); -τs and the delay link of the ideal operating condition model G p (s) under the condition of using the Smith predictor

[0151] Through the delay link of the TCN model and the delay link e of the ideal operating condition model G p (s), the error E -τs (s) is obtained, and the expression is as follows: τ (s):

[0152]

[0153] In the formula, T1 and T2 are the first time constant and the second time constant respectively. After calculation by the formula, T1 = 66.6 and T2 = 25.46; is the time-varying delay factor; τ is the delay factor; s is the complex variable in the complex frequency domain; K0 is the ratio coefficient, K0 = 3.1, and the expression is as follows:

[0154]

[0155] By differentiating the objective function J with respect to the time-varying delay factor , the rate of change of the objective function can be obtained, and the expression of the rate of change of the objective function is as follows:

[0156]

[0157] By differentiating the error E with respect to the time-varying delay factor τ (s), the rate of change of the error can be obtained, and the expression of the rate of change of the error is as follows:

[0158]

[0159] In the formula, s' is the differential operation;

[0160] Converting the rate of change of the error into a time-domain equation, the following expression can be obtained:

[0161]

[0162] In the formula, y(t) is the time-domain expression of Y(s);

[0163] The error change rate is converted into a time-domain equation and substituted into the objective function J to obtain the time-domain equation of the objective function error change rate, and the expression is as follows:

[0164]

[0165] Through the error E τ (s) and the output of the Smith predictor, the error E τ ’(s) after eliminating the delay is obtained;

[0166] The output expression of the Smith predictor is:

[0167]

[0168] In the formula, is the ideal working condition model G p (s) delay link; among them, the time constant τ max The value rule is as follows:

[0169]

[0170] The error E τ ’(s) expression after eliminating the delay is:

[0171] E τ ’(s) = the output of the Smith predictor + E τ (s);

[0172] The error E τ ’(s) after integral operation obtains the output K τ '(s) of the inner loop, and the expression is as follows:

[0173]

[0174] In the formula, T τ is the integral link time constant, where β is the descent step size, set to β = 0.01;

[0175] According to the principle of the steepest descent, the change law of the output K τ '(s) of the inner loop with time is as follows:

[0176]

[0177] The output K τ '(s) of the inner loop obtains the time-domain equation of the change law with time by substituting the time-domain equation of the objective function error change rate into the output K τThe conversion formula can be obtained from the variation law of (s) with time, and the expression of the conversion formula is as follows:

[0178]

[0179] Substitute the conversion formula into the output K of the inner loop τ The output K of the inner loop is obtained from the variation law of (s) with time τ The time-domain equation of the variation law of (s) with time, the expression is as follows:

[0180]

[0181] At this time, let where P τ (s) is p τ (t) of the time-domain expression, the control signal M τ (s) can be rewritten as:

[0182]

[0183] S5. Verify the model effect;

[0184] The verification method is: the TCN-based delay compensation control framework, PID and MRAC are simultaneously used for the control of the oxygen content in the furnace under the same conditions.

[0185] The present invention compares the comparison results when the oxygen content standard needs to be controlled at 50% and 25%. The green curve is the reference curve, the blue curve is the PID control curve, and the control effect shows large fluctuations and slow convergence speed; the purple curve is the curve of the unimproved model reference adaptive control (MRAC). Due to the large delay characteristics in the smelting process, the system response lags, and the oxygen content cannot be adjusted in time; the red curve is the delay compensation control curve based on TCN. It can be seen that this control strategy can not only quickly adjust the change of oxygen content, but also reduce the oscillation amplitude of the system, and the improvement effect is obvious.

Claims

1. A delayed compensation control method for oxygen content in a tin smelting furnace based on TCN, characterized in that: The following steps are involved: S1. Collect furnace data and clean the data. Use seasonal trend decomposition method to decompose the cleaned data into three parts: trend component, seasonal component and residual component. S2. Construct a dynamic prediction model of oxygen content in the chamber based on the temporal convolutional network TCN; S3, grouping the data and raw material components after seasonal trend decomposition method STL, dividing each group into training set, validation set and test set in proportion, inputting them into the dynamic prediction model of oxygen content in the furnace based on time convolution network for training, obtaining the trained model, and evaluating the training results by mean square error; S4. Use the trained time convolution network TCN model to replace the transfer function model in the reference adaptive control framework, build a time delay compensation control framework based on the time convolution network TCN to calculate the control law for compensating the time delay, and complete the model construction; S5. Verify the model effect.

2. The method for delay compensation control of oxygen content in a tin smelting furnace based on TCN according to claim 1 is characterized in that: The steps of collecting furnace data and cleaning the data, and using the seasonal trend decomposition method to decompose the cleaned data into three parts: trend component, seasonal component and residual component are as follows: S1.

1. Collect furnace data; Furnace data, including: oxygen content in the furnace, oxygen input, temperature and raw material composition, among which the oxygen content data in the furnace is a time series; The collection method is: exporting real-time production data from the distributed control system; S1.

2. Clean the collected furnace data. The steps are as follows: S1.2.

1. A sliding window with a fixed step size is set in the exported real-time production data to pre-process the data within the window size; S1.2.2, preprocess the data in the window; Preprocessing includes: removing the mean value and detecting and removing outliers; S1.

3. Perform STL decomposition on the oxygen content, oxygen flow rate and temperature in the furnace after cleaning, and extract the trend component, seasonal component and residual component in turn.

3. The method for delay compensation control of oxygen content in a tin smelting furnace based on TCN according to claim 1 is characterized in that: The dynamic prediction model of the oxygen content in the chamber based on the temporal convolutional network TCN is constructed as follows: the temporal convolutional network TCN is provided with three residual blocks, namely: a first residual block, a second residual block and a third residual block; Each residual block includes two convolution operations and one residual calculation. The convolution operation is as follows: the data input is weighted normalized after passing through the basic convolution layer, and the data output is obtained after passing through the Relu activation function. In order to prevent data overfitting, a Dropout layer is also set after the Relu activation function. The basic convolution layer uses dilated convolution operations and causal convolution operations, and the dilation coefficient of the dilated convolution is set to twice the number of residual blocks. The data is input from the first residual block, passed through the second residual block, and output from the third residual block; the output channels of each residual block are different, and different output channels are set to capture the implicit relationship between features.

4. The method for delay compensation control of oxygen content in a tin smelting furnace based on TCN according to claim 1 is characterized in that: The data and raw material components decomposed by the seasonal trend decomposition method STL are grouped, each group is divided into a training set, a validation set and a test set in proportion, and are respectively input into a dynamic prediction model of oxygen content in a furnace based on a time convolutional network for training, to obtain a trained model, and the steps of evaluating the training results by the mean square error are as follows: S3.1, grouping the data and raw material components after STL decomposition, and inputting them into the dynamic prediction model of oxygen content in the chamber based on the time convolution network for training; The groups are: The trending components of oxygen content, oxygen input, temperature and raw material composition in the furnace; The seasonal composition of oxygen content in the furnace, the seasonal composition of oxygen input, the seasonal composition of temperature trend and the composition of raw materials; Residual components of oxygen content in the furnace, residual components of oxygen input, residual components of temperature and raw material composition; S3.2, set the window of the TCN model and divide the training set, validation set and test set in proportion; S3.3, initialize the TCN model; S3.4, input the training sets into the model for training; S3.

5. After a complete training epoch for all the training set data, use the validation set to evaluate the model performance; To evaluate the model performance, we choose mean square error MSE; S3.

7. Use the test set to evaluate the generalization performance of the model. MSE is also selected as the evaluation indicator. Unlike the training set, the test set data is used after the training is completed and can only be used once. MSE is selected as the indicator to ensure that the test set and the training set are under the same standard and the evaluation results are consistent.

5. The method for delay compensation control of oxygen content in a tin smelting furnace based on TCN according to claim 1 is characterized in that: The trained time convolution network TCN model is used to replace the transfer function model in the reference adaptive control framework, and a delay compensation control framework based on the time convolution network TCN is constructed to calculate the control law for compensating the time lag. The delay compensation control framework based on TCN in the model construction includes: an outer loop, an inner loop and a TCN model; wherein the outer loop is a conventional PID controller G c (s), the inner loop includes: ideal working condition model G p (s) and the ideal operating model G using the Smith predictor p (s); Conventional PID controller G c The output of (s) is denoted as U(s); The output of the inner loop is denoted by K τ '(s); G c (s) satisfies the following expression: G c (s)=Y(s)-R(s) In the formula, Y(s) is the output under environmental disturbance; R(s) is the preset target value; The conventional PID controller G c The output of (s) and the output of the inner loop K τ '(s) add to get the control signal M τ (s), the expression is as follows: M τ (s)=U(s)+K τ '(s); Control signal M τ (s) are respectively used as TCN model and ideal working condition model G p (s) and the ideal operating model G using the Smith predictor p (s) signal input; respectively get the delay link of the TCN model Ideal working condition model G p (s) delay link e -τs and the ideal operating model G using the Smith predictor p (s) delay link Through the delay link of the TCN model and the ideal operating model G p (s) delay link e -τs Get the error E τ (s), the expression is as follows: Where T1 and T2 are the first time constant and the second time constant respectively; is the time-varying delay factor; τ is the delay factor; s is the complex variable in the complex frequency domain; K0 is the ratio coefficient; Through the error E τ (s) and the output of the Smith predictor to obtain the error E after eliminating the delay τ '(s); The output expression of the Smith predictor is: In the formula, G is the ideal operating model when using Smith predictor p (s) delay link; where the time constant τ max The value selection rules are as follows: The error after eliminating the delay E τ '(s) expression is: E τ '(s) = output of Smith predictor + E τ (s); The error after eliminating the delay E τ '(s) After integration operation After that, the output K of the inner loop is obtained τ '(s), the expression is as follows: Where, T τ is the time constant of the integration link, Among them, β is the descending step length;

Images