Dry quenching furnace pre-storage chamber pressure control method based on artificial intelligence

Abstract

The invention belongs to the technical field of dry quenching pre-storage chamber pressure control, and particularly relates to a dry quenching furnace pre-storage chamber pressure control method based on artificial intelligence, which comprises the following steps: firstly, carrying out original data acquisition and data preprocessing, then carrying out dry quenching furnace pre-storage chamber process control model switching, then carrying out pre-storage chamber pressure disturbance prediction, and finally, carrying out pre-storage chamber pressure disturbance prediction. And finally, the intelligent control system adjusts the blow-off valve in advance according to the coke loading state signal and the disturbance prediction result. Full-automatic control can be achieved through multi-controller coordinated control, and manual intervention is not needed in the coke loading period; through pressure prediction of the pre-storage chamber, adjustment is performed in advance, and pressure fluctuation of the pre-storage chamber is reduced; and through dynamic switching of process models, the fire and smoke phenomena caused by mismatching of the models during coke loading are prevented.

Description

Technical Field

[0001] The present invention belongs to the technical field of pressure control of the pre-storage chamber of coke dry quenching, and specifically relates to a pressure control method for the pre-storage chamber of a coke dry quenching furnace based on artificial intelligence. Background Art

[0002] The coke dry quenching technology is widely used in the iron and steel industry for cooling coke and recovering heat energy. The pre-storage chamber is a key part of the coke dry quenching furnace, and its pressure control is crucial for the stable operation, safety, and environmental protection of the entire coke dry quenching furnace. However, due to the periodic operation of the process, the process model changes continuously. The PID control adopted by some plants has poor control effects and large pressure fluctuations, so most plants still adopt manual control. The existing pressure control methods for the pre-storage chamber of coke dry quenching furnaces have the following problems:

[0003] 1. The system model is not fixed and the pressure fluctuates greatly: When the coke dry quenching furnace is operating, it is necessary to periodically open the furnace lid to charge coke, resulting in changes in the process model of the system. Simple PID control cannot adjust in time, and most systems can only be adjusted manually. The manual control method is difficult to accurately predict and adjust the pressure of the pre-storage chamber, cannot correct disturbances in real time, resulting in frequent pressure fluctuations, affecting the stability of the coke dry quenching furnace system. Moreover, generally, a furnace of coke needs to be charged every 7 - 9 minutes, the operation is very frequent, and the labor intensity is large.

[0004] 2. It is not environmentally friendly and has potential safety hazards: Positive pressure often occurs when opening the furnace lid to charge coke, resulting in the phenomenon of fire and smoke during the charging process, causing environmental pollution and increasing the operation risk of the system. Summary of the Invention

[0005] In view of the various deficiencies of the prior art, the inventor has studied and designed a pressure control method for the pre-storage chamber of a coke dry quenching furnace based on artificial intelligence through long-term practice, realizing stable and safe operation of the coke dry quenching furnace through efficient pressure prediction and intelligent regulation.

[0006] A pressure control method for the pre-storage chamber of a coke dry quenching furnace based on artificial intelligence according to the present invention includes the following steps:

[0007] Step 1: Acquisition and preprocessing of original data.

[0008] Step 2: Switching of the process control model for the pre-storage chamber of the coke dry quenching furnace. During the coke charging process of the coke dry quenching furnace, independent process models are established for the two working conditions of opening the furnace lid to charge coke and closing the furnace lid without charging coke. When the furnace lid is open, the negative pressure of the pre-storage chamber is maintained by adjusting the opening degree of the relief valve. When the furnace lid is closed and no coke is charged, the pressure of the pre-storage chamber is controlled by adjusting the circulating gas relief valve to maintain the negative pressure of the pre-storage chamber.

[0009] Step 3: Prediction of the pressure disturbance in the pre-storage chamber. An LSTM neural network is used to construct a prediction model. By predicting the future pressure changes in advance, feedforward information is provided for the controller. The prediction of the pressure disturbance in the pre-storage chamber specifically includes the following sub-steps:

[0010] Sub-step 3.1 Data preprocessing; Data is collected once every 5 seconds, and a total of 60 data points are collected within 5 minutes. The length of the input sequence is 60. For each input sequence, the data of the past 5 minutes is used as the input, and the corresponding output is the pressure in the pre-storage chamber 1 minute later.

[0011] Set the time series data as X = {x_1, x_2,…, x_T}, where each x_t contains two feature quantities, namely the circulating air volume and the air guide volume, and the corresponding pre-storage chamber pressure sequence is Y = {y_1, y_2, …, y_T}.

[0012] Sub-step 3.2 Design of the LSTM prediction model.

[0013] Sub-step 3.3 Model training: The training process of the LSTM prediction model includes the following sub-steps:

[0014] Sub-step 3.3.1 Data preprocessing and sample construction: Normalize the circulating air volume and the air guide volume. From the continuously collected data, use the sliding window method to construct the input sequence and the target output.

[0015] Input sequence: Each sample is 2D data of the past 60 time steps, denoted as ; Target output: Predict the pressure in the pre-storage chamber 1 minute later, denoted as .

[0016] Normalize each feature according to the following formula: .

[0017] Sub-step 3.3.2 Initialize the model parameters: Randomly initialize the weight matrices and bias vectors of the LSTM layer and the fully connected layer. For the fully connected layer, initialize the parameters: ; For the LSTM layer, initialize the weight matrix W f 、W i 、W C 、W o and the corresponding bias vectors b f 、b i 、b C 、b o .

[0018] Sub-step 3.3.3 Start Training Epoch: Divide the entire training process into multiple Epochs, and each Epoch traverses the entire training set once.

[0019] Sub-step 3.3.4 Load Training Data: Load the training data in batches, and the current batch of data is .

[0020] Sub-step 3.3.5 Forward Propagation to Calculate Predicted Values: Input the current batch of data into the LSTM network, and after multiple layers of calculations, obtain the predicted output: ; Among them, f(X;θ) represents the forward propagation process of the network, and θ represents all model parameters.

[0021] Sub-step 3.3.6 Calculate the Loss Function: Use the mean squared error to measure the error between the predicted value and the true value: ; Among them, N is the number of samples, y i is the true value, is the predicted value of the model.

[0022] Sub-step 3.3.7 Backward Propagation to Calculate Gradients: According to the loss function L, calculate the gradients of each parameter through the backward propagation algorithm: ; Among them, g represents the gradient value; is the gradient vector, representing the gradient of the loss function L with respect to the parameter θ, which describes the rate of change of the loss function at the current parameter value.

[0023] Sub-step 3.3.8 Adam Optimizer Updates Parameters: The Adam optimizer updates the parameters according to the following formula: ; Among them, g t is the current gradient, β1 and β2 are momentum parameters, α is the learning rate, is a small constant to prevent division by zero, and θ t represents the model parameters.

[0024] Sub-step 3.3.9 Evaluate the Performance of the Validation Set: After each Epoch, use the validation set data for forward propagation, and calculate the validation loss and evaluation metrics.

[0025] Sub-step 3.3.10 Determine Whether to Meet the Stopping Condition: Determine whether to meet the stopping condition according to the validation set metrics. If the validation loss no longer decreases or reaches the preset number of training epochs, end the training; otherwise, continue with the next Epoch.

[0026] Sub-step 3.4 Model Validation.

[0027] Step 4: The intelligent control system adjusts the relief valve in advance according to the coke charging state signal and the disturbance prediction result. The intelligent control system controls the pressure in the pre-storage chamber, which is divided into pressure control when the furnace lid is open for coke charging and pressure control when the furnace lid is closed without coke charging.

[0028] Furthermore, in Step 2, the specific process of adjusting the opening of the relief valve when the furnace lid is open includes the following sub-steps:

[0029] Sub-step 2.1.1 Initial adjustment: Directly increase the corresponding percentage on the basis of the original opening of the relief valve according to experience, and set the initial opening of the relief valve as D initial , and the percentage increase on the basis of the original opening of the relief valve is , then the adjusted opening of the relief valve is expressed as: .

[0030] Sub-step 2.1.2 Deviation correction: Further adjust the opening of the relief valve according to the deviation between the pressure in the pre-storage chamber and the set value. If the deviation is positive, increase the opening of the relief valve; if the deviation is negative, decrease the opening of the relief valve. Set the target pressure of the pre-storage chamber as P target , the current pressure of the pre-storage chamber is P current , and the adjustment coefficient of the relief valve is K adjust , then the opening of the relief valve that needs to be adjusted is expressed as: ; The final opening of the relief valve is expressed as: .

[0031] Furthermore, in Step 2, when the furnace lid is closed without coke charging, the specific process of adjusting the circulating gas relief valve includes the following sub-steps:

[0032] Sub-step 2.2.1 Calculate the pressure deviation value of the pre-storage chamber: The deviation value between the current pressure value P current of the pre-storage chamber and the target pressure value P target is expressed as: .

[0033] Sub-step 2.2.2 According to the deviation value, the PID controller calculates the control quantity u(t) for adjusting the opening of the relief valve. The control quantity u(t) is calculated by the following formula: ; where K p is the proportional gain, K i is the integral gain, and K d is the derivative gain.

[0034] Further, in sub-step 3.1, for each time step t, construct the input-output pair (X_t, y_{t + 12}), where X_t = {x_{t - 59}, x_{t - 1}, …, x_t}; the total length of the historical data is T, and for each valid time point t satisfying t ≥ 60 and t + 12 ≤ T, construct the sample pair: ; Among them, ; y t+12 is the pressure in the pre-storage chamber measured at time t + 12.

[0035] Further, in sub-step 3.2, the network structure of the LSTM prediction model includes:

[0036] Input layer: Receive an input sequence with a shape of (60, 2). The input data is the data of the past 60 time steps, and each time step contains 2 features, then ; Among them, x t,1 represents the circulating air volume at the t-th time step, and x t,2 represents the air guiding volume at the t-th time step.

[0037] LSTM layer: Stack 2 LSTM layers to capture the long-term dependencies in the input sequence, and set 50 hidden units for each layer. For each time step t, the calculation process of the LSTM layer includes the following sub-steps:

[0038] Sub-step 3.2.1 Forget gate layer, calculate how much information in the previous cell state C t−1 needs to be forgotten: ; Among them, W f is the forget gate weight matrix, b f is the forget gate bias vector, is the previous hidden state, is the current input, and σ(·) represents the Sigmoid function.

[0039] Sub-step 3.2.2 Input gate layer, calculate which information in the current input needs to be written into the cell state: ; Generate a new candidate cell state: ; Among them, W i 、W c are the weight matrices of the candidate cell state, b i 、b c are the bias vectors of the candidate cell state.

[0040] Sub-step 3.2.3 Combine the results of the forget gate and the input gate to update the cell state: ; Among them, ⊙ represents element-wise multiplication.

[0041] Sub-step 3.2.4 Output gate layer, calculate which parts of the cell state are used as the current output: ; Among them, W o is the weight matrix of the output gate, and b o is the bias vector of the output gate.

[0042] Sub-step 3.2.5 Finally update the hidden state: ; The dimension of the hidden state h t is 50, that is, each LSTM layer outputs a 50-dimensional vector.

[0043] Dropout layer: Add a Dropout layer after each LSTM layer to prevent the model from overfitting. During the training process, the Dropout layer randomly sets the outputs of some neurons in the previous layer to 0 with a set probability of 0.2, that is, 20%; assuming h represents the vector output by the LSTM layer, the calculation process of the Dropout layer is expressed as: ; Among them: is a mask vector, and each element mi is generated according to the Bernoulli distribution , where p = 0.8 is the probability that a neuron is retained; ☉ represents element-wise multiplication.

[0044] Fully connected layer: Input the hidden state at the last time step into the fully connected layer and map it to a single output, representing the predicted pre-chamber pressure value: ; Among them, is the weight matrix of the fully connected layer, is the bias vector of the fully connected layer.

[0045] Furthermore, in sub-step 3.4, the verification process of the LSTM prediction model includes the following sub-steps:

[0046] Sub-step 3.4.1 Performance evaluation metrics: During the verification process, the following two metrics are mainly used to measure the prediction error:

[0047] The root mean square error (RMSE) measures the standard deviation between the predicted value and the true value: .

[0048] Mean Absolute Error (MAE) measures the average absolute difference between the predicted values ​​and the true values ; Among them, N is the number of validation samples, y i is the actual pre-storage chamber pressure value of the i-th sample, is the pre-storage chamber pressure value predicted by the model.

[0049] Sub-step 3.4.2 Parameter tuning Adjust the model parameters according to the RMSE and MAE indicators, and repeat the training and validation until a model with generalization ability that meets the control requirements is obtained.

[0050] Sub-step 3.4.3 Feedback to control strategy: After confirming that the verification results have achieved the predetermined target, the prediction model is embedded in the control system, and the predicted future pressure value is used to adjust the relief valve opening through the feedforward control strategy to ensure that the pre-storage chamber pressure is maintained within the set range.

[0051] Furthermore, in step 4, the pressure control when the furnace cover is opened for coke loading specifically includes the following sub-steps:

[0052] Step 4.1.1 Detect the furnace cover opening signal and immediately enter the expert system control mode. First, make an initial adjustment. According to the rule of thumb, the opening of the relief valve is directly increased by a fixed ratio, and then the deviation adjustment is made. The deviation adjustment rules are as follows:

[0053] When the deviation ΔP exceeds the set threshold When the pressure reaches the target range, immediately increase the relief valve opening by 10% to accelerate the pressure release. If the pressure still does not drop to the target range, increase it further.

[0054] When the deviation is within a reasonable range When the pressure is too low, maintain the current relief valve opening.

[0055] When the deviation is lower than the set threshold When the pressure is too low, gradually reduce the relief valve opening by 5% to prevent the pressure from being too low.

[0056] The adjustment period is set to 5-10 seconds.

[0057] Step 4.1.2 The LSTM prediction system starts working at this time and predicts the pressure trend in the next 1 minute based on past data. If it is predicted that the pressure will rise, the relief valve opening will be increased in advance. If it is predicted that the pressure will drop, the relief valve opening will be reduced in advance.

[0058] The LSTM prediction model inputs the circulating air volume and air guiding air volume data of the past 5 minutes, outputs the pressure of the pre-storage chamber after 1 minute, and cooperates with the expert system for early compensation during the coke charging stage.

[0059] Sub-step 4.1.3 Real-time pressure feedback: Based on the real-time sensor feedback data, the expert system performs secondary fine-tuning.

[0060] Furthermore, in step 4, for the pressure control when the furnace lid is closed and no coke is charged, it specifically includes the following sub-steps:

[0061] Sub-step 4.2.1 The furnace lid is closed, the pressure inside the pre-storage chamber tends to be stable, and it enters the PID control mode.

[0062] Sub-step 4.2.2 The LSTM prediction system continues to monitor disturbances, predicts external air volume disturbances, fine-tunes the circulating fan and the relief valve, and maintains the pressure of the pre-storage chamber within the target pressure range.

[0063] The LSTM prediction model inputs the circulating air volume and air guiding air volume data of the past 5 minutes, outputs the pressure of the pre-storage chamber after 1 minute, and the prediction result is used for the feed-forward compensation of the PID controller.

[0064] Sub-step 4.2.3 Only retain the PID and LSTM prediction adjustments, monitor the system fluctuations in real time, and make adjustments at any time.

[0065] The beneficial effects of the present invention are:

[0066] 1) Realize automatic operation: Reduce the dependence on manual operation and reduce the labor intensity of operators.

[0067] 2) Ensure the stable operation of the coke dry quenching furnace: Through dynamic switching, avoid the instability problems that may be brought about by the unified control of all working conditions by a single control model, thereby accurately predicting and real-time adjusting the pressure of the pre-storage chamber, reducing pressure fluctuations, and stabilizing the pressure.

[0068] 3) Improve the safety of the system operation: Build control models for different working conditions during the coke charging process of opening and closing the furnace lid respectively, and automatically switch the control model according to the working conditions to adapt to the periodic working condition changes, improve the robustness of the system, and prevent the phenomenon of coke charging with fire and smoke.

[0069] 4) Prediction of pre-storage chamber pressure disturbance based on neural network: Predict in advance the pressure changes caused by the fluctuations of the circulating fan and the air guiding air volume, and then reduce the parameter fluctuations brought about by the system lag through feed-forward control. Description of the Drawings

[0070] Figure 1 It is the flow block diagram of the pre-storage chamber pressure control of the present invention.

[0071] Figure 2It is the process control model operation flowchart during the coke charging in the coke dry quenching furnace of the present invention.

[0072] Figure 3 It is the process control model operation flowchart when no coke is charged in the coke dry quenching furnace of the present invention.

[0073] Figure 4 It is the flowchart of the training process of the disturbance prediction model of the present invention.

[0074] Figure 5 It is the flowchart of the operation of the intelligent control model of the present invention. Detailed implementation manners

[0075] The present invention will be further described in detail below in conjunction with the implementation examples. By describing these implementation examples in sufficient detail, those skilled in the art can understand and practice the present invention. Logical, implementation and other changes can be made to the implementation without departing from the gist and scope of the present invention. Therefore, the following detailed description should not be construed in a limiting sense, and the scope of the present invention is only defined by the claims.

[0076] Refer to Figures 1 to 5 , the present invention proposes a pressure control method for the pre-storage chamber of a coke dry quenching furnace based on artificial intelligence, which specifically includes the following steps.

[0077] Step 1: Raw data collection and data preprocessing.

[0078] The pressure signal in the pre-storage chamber of the coke dry quenching furnace usually has strong volatility and high-order harmonic components, which will increase system interference and cause the control system to be unstable. To effectively suppress these high-frequency noises and improve the smoothness of the signal, a first-order inertial filter is used to process the raw data of the pressure in the pre-storage chamber of the coke dry quenching furnace.

[0079] The first-order inertial filter is a low-pass filter, and its output signal is the weighted average of the input signal and the output signal of the previous moment. Its mathematical expression is:

[0080] Among them, y(t) is the output at the current moment, x(t) is the input at the current moment, y(t−1) is the output of the previous moment, and α is the filtering coefficient, and its value range is between 0 and 1.

[0081] In the processing of the pressure signal in the pre-storage chamber of the coke dry quenching furnace, using a first-order inertial filter can effectively smooth the raw signal and reduce the influence of high-frequency noise. Set the initial output value y(0) of the filter and take the first input signal value. The value of the filtering coefficient α is selected as 0.85 according to actual needs. For each new input x(t), calculate the current output y(t) and update the output value of the previous moment.

[0082] Step 2: Switching the process control model of the pre-storage chamber of the coke dry quenching furnace.

[0083] During the coke charging process of the coke dry quenching furnace, the operation of opening and closing the furnace lid needs to be carried out periodically, and the process model of the system will change every time the furnace lid is opened. To ensure the stability and accuracy of the control system, it is necessary to establish independent process models for different working conditions of charging with the furnace lid open and not charging with the furnace lid closed, and automatically switch the corresponding models according to the coke charging signal to achieve precise control.

[0084] When the furnace lid is opened, significant changes will occur in parameters such as the gas flow and pressure inside the coke dry quenching furnace. At this time, it is necessary to control the pressure of the pre-storage chamber for this process to prevent the phenomenon of positive pressure and smoke.

[0085] During the coke charging process of the coke dry quenching furnace, the negative pressure of the pre-storage chamber is maintained by adjusting the discharge amount through the discharge valve. According to the current pressure of the pre-storage chamber, the opening degree of the discharge valve is adjusted in sections, which specifically includes the following sub-steps:

[0086] Sub-step 2.1.1 Initial adjustment: Directly increase the corresponding percentage on the basis of the original opening degree of the discharge valve according to experience to quickly reduce the pressure of the pre-storage chamber. Set the initial opening degree of the discharge valve as D initial , and the percentage increase on the basis of the original opening degree of the discharge valve is , then the adjusted opening degree of the discharge valve is expressed as: .

[0087] Sub-step 2.1.2 Deviation correction: Further adjust the opening degree of the discharge valve according to the deviation between the pressure of the pre-storage chamber and the set value. If the deviation is positive (that is, the pressure of the pre-storage chamber is higher than the set value), increase the opening degree of the discharge valve; if the deviation is negative (that is, the pressure of the pre-storage chamber is lower than the set value), reduce the opening degree of the discharge valve. Set the target pressure of the pre-storage chamber as P target , the current pressure of the pre-storage chamber is P current , and the adjustment coefficient of the discharge valve is K adjust , then the opening degree that the discharge valve needs to be adjusted is expressed as: ; The final opening degree of the discharge valve is expressed as: .

[0088] When not charging with the furnace lid closed, the coke dry quenching furnace operates normally, and the gas flow and pressure conditions inside it will return to a relatively stable state. At this time, the pressure of the pre-storage chamber is controlled by adjusting the circulating gas discharge valve to maintain the negative pressure of the pre-storage chamber, which specifically includes the following sub-steps.

[0089] Sub-step 2.2.1 Calculate the pressure deviation value of the pre-storage chamber: The deviation value between the current pressure value P current of the pre-storage chamber and the target pressure value P target is expressed as: ;

[0090] Sub-step 2.2.2: According to the deviation value, the PID controller calculates the control variable u(t) to adjust the opening degree of the bleed valve, change the bleed volume, and thus regulate the pressure in the pre-storage chamber. The control variable u(t) is calculated by the following formula: ; where K p is the proportional gain, K i is the integral gain, and K d is the derivative gain.

[0091] Step 3: Prediction of the pressure disturbance in the pre-storage chamber.

[0092] During the non-coke charging process of the coke dry quenching furnace, the pressure in the pre-storage chamber is affected by multiple disturbance factors, and the main disturbances include the air volume of the circulation fan and the air volume of the air guide. The circulation fan adjusts the air flow in the pre-storage chamber by changing the circulation speed of the circulating gas in the closed system, which has an important impact on the pressure. The air guide system introduces external air, changes the gas volume and flow rate in the pre-storage chamber, and also affects the pressure stability.

[0093] The present invention uses an LSTM neural network to construct a prediction model to predict the pressure in the pre-storage chamber. By predicting the future pressure change in advance, it provides feedforward information for the controller to achieve more accurate pressure regulation.

[0094] The prediction of the pressure disturbance in the pre-storage chamber specifically includes the following sub-steps:

[0095] Sub-step 3.1 Data preprocessing: Data is collected once every 5 seconds, and a total of 60 data points are collected within 5 minutes. The length of the input sequence is 60. For each input sequence, the data of the past 5 minutes (length of 60) is used as the input, and the corresponding output is the pressure in the pre-storage chamber 1 minute later.

[0096] Assume that the time series data is X = {x_1, x_2, …, x_T}, where each x_t contains two characteristic quantities, namely the circulating air volume and the air volume of the air guide. The corresponding pressure sequence in the pre-storage chamber is Y = {y_1, y_2, …, y_T}.

[0097] For each time step t, an input-output pair (X_t, y_{t + 12}) is constructed, where X_t = {x_{t - 59}, x_{t - 1}, …, x_t}.

[0098] The total length of the historical data is T (the number of sampled data points). For each valid time point t satisfying t ≥ 60 and t + 12 ≤ T, sample pairs are constructed: ; Among them, ; y t+12 is the pressure in the pre-storage chamber measured at time t + 12.

[0099] Sub-step 3.2: Design of the LSTM prediction model. The network structure of the LSTM prediction model includes:

[0100] Input layer: Receives an input sequence with a shape of (60, 2). The input data is the data of the past 60 time steps, and each time step contains 2 features (recirculation air volume and air guide volume). Then ; Among them, x t,1 represents the recirculation air volume at the t-th time step, and x t,2 represents the air guide volume at the t-th time step.

[0101] LSTM layer: Stack 2 LSTM layers to capture the long-term dependencies in the input sequence, and set 50 hidden units for each layer. For each time step t, the calculation process of the LSTM layer includes the following sub-steps:

[0102] Sub-step 3.2.1 Forget gate layer, calculate how much information in the previous cell state C t-1 needs to be forgotten: ; Among them, W f is the forget gate weight matrix, b f is the forget gate bias vector, is the previous hidden state, is the current input, and σ(·) represents the Sigmoid function.

[0103] Sub-step 3.2.2 Input gate layer, calculate which information in the current input needs to be written into the cell state: ; Generate a new candidate cell state: ; Among them, W i , W c are the weight matrices of the candidate cell state, and b i , b c are the bias vectors of the candidate cell state.

[0104] Sub-step 3.2.3 Combine the results of the forget gate and the input gate to update the cell state: ; Among them, ⊙ represents element-wise multiplication.

[0105] Sub-step 3.2.4 Output gate layer, calculating which parts of the cell state are used as the current output: ; where W o is the weight matrix of the output gate, and b o is the bias vector of the output gate.

[0106] Sub-step 3.2.5 Final update of the hidden state: ; The dimension of the hidden state h t is 50, that is, each LSTM layer outputs a 50-dimensional vector.

[0107] Dropout layer: A Dropout layer is added after each LSTM layer to prevent the model from overfitting. During training, the Dropout layer randomly sets the outputs of some neurons in the previous layer to 0 with a set probability (0.2, i.e., 20%), so that the model does not rely on a specific combination of neurons during each training.

[0108] Assuming h represents the vector output by the LSTM layer, the calculation process of the Dropout layer is expressed as: ; where: is a mask vector, and each element mi is generated according to the Bernoulli distribution , where p = 0.8 is the probability that a neuron is retained (so there is a 20% probability of being disconnected); ☉ represents element-wise multiplication.

[0109] During testing, no random masking is performed, but the output is multiplied by the retention rate p to maintain a consistent expectation.

[0110] Fully connected layer: The hidden state at the last time step (the 60th time step) is input into the fully connected layer and mapped to a single output, representing the predicted prechamber pressure value: ; where, is the weight matrix of the fully connected layer, is the bias vector of the fully connected layer.

[0111] Sub-step 3.3: Model training. The training process of the LSTM prediction model includes the following sub-steps:

[0112] Sub-step 3.3.1 Data preprocessing and sample construction: Normalize the recirculation air volume and the air guide volume to ensure consistent numerical ranges.

[0113] From the continuously collected data, the input sequence and the target output are constructed using the sliding window method.

[0114] Input sequence: Each sample is 2D data for the past 60 time steps, denoted as ; Target output: Predict the pressure in the pre-storage room 1 minute later (i.e., 12 time steps later), denoted as .

[0115] Normalize each feature according to the following formula: .

[0116] Divide the samples into a training set and a validation set, with a common ratio of 80% for training and 20% for validation.

[0117] Sub-step 3.3.2 Initialize model parameters: Randomly initialize the weight matrices and bias vectors of the LSTM layer and the fully connected layer. For the fully connected layer, initialize the parameters: . For the LSTM layer, initialize the weight matrices W f , W i , W C , W o and the corresponding bias vectors b f , b i , b C , b o .

[0118] Sub-step 3.3.3 Start training Epoch: Divide the entire training process into multiple Epochs, and each Epoch traverses the entire training set once.

[0119] Sub-step 3.3.4 Load training data (Batch): Load the training data in batches (each batch contains 64 samples). The current batch of data is .

[0120] Sub-step 3.3.5 Forward propagation to calculate the predicted value: Input the current batch of data into the LSTM network, and after calculation through multiple layers (including Dropout), obtain the predicted output: ; where f(X;θ) represents the forward propagation process of the network, and θ represents all model parameters.

[0121] Sub-step 3.3.6 Calculate the loss function: Use the mean squared error (MSE) to measure the error between the predicted value and the true value: ; where N is the number of samples, y iis the true value, is the model predicted value.

[0122] Sub-step 3.3.7 Backpropagation to calculate gradients: According to the loss function L, calculate the gradients of each parameter through the backpropagation algorithm: ; where g represents the gradient value; is the gradient vector, representing the gradient of the loss function L with respect to the parameter θ, which describes the rate of change of the loss function at the current parameter values.

[0123] Sub-step 3.3.8 Adam optimizer to update parameters: The Adam optimizer updates the parameters according to the following formula: ; where g t is the current gradient, β1 and β2 are momentum parameters, α is the learning rate, is a small constant to prevent division by zero, and θ t represents the model parameters.

[0124] Sub-step 3.3.9 Evaluate the performance of the validation set: After each Epoch, use the validation set data for forward propagation to calculate the validation loss and evaluation metrics (RMSE, MAE).

[0125] .

[0126] Sub-step 3.3.10 Determine whether the stopping condition is met: Determine whether the stopping condition is met according to the validation set metrics, such as the validation loss no longer decreasing or reaching the preset number of training epochs. If the stopping condition is met, end the training; otherwise, continue with the next Epoch.

[0127] Sub-step 3.4 Model validation: The LSTM prediction model validation process includes the following sub-steps:

[0128] Sub-step 3.4.1 Performance evaluation metrics: During the validation process, the following two metrics are mainly used to measure the prediction error:

[0129] Root Mean Square Error (RMSE) measures the standard deviation between the predicted value and the true value: .

[0130] Mean Absolute Error (MAE) measures the average absolute difference between the predicted value and the true value ; where N is the number of validation samples, and y i is the true pre-storage room pressure value of the i-th sample, The pre-stored chamber pressure value for model prediction.

[0131] Since the prediction task of this model is used for real-time control of the pre-stored chamber pressure, the prediction accuracy does not need to be too high. The target value of the pre-stored chamber pressure is -100~-50Pa, the RMSE is controlled within 5~15Pa, and the MAE is less than 10Pa, which can effectively provide feedforward information for the control strategy.

[0132] Sub-step 3.4.2 Parameter tuning: According to the RMSE and MAE indicators, adjust the model parameters (including learning rate, batch size, Dropout rate, etc.). Repeat training and validation until a model with generalization ability meeting the control requirements is obtained.

[0133] Sub-step 3.4.3 Feedback to the control strategy: After confirming that the verification result reaches the predetermined target, embed the prediction model into the control system, and adjust the opening of the bleed valve with the predicted future pressure value through the feedforward control strategy to ensure that the pre-stored chamber pressure is maintained within the set range.

[0134] Step 4: The intelligent control system adjusts the bleed valve in advance according to the coking state signal and the disturbance prediction result to prevent fluctuations in the CDQ system caused after feedback.

[0135] The intelligent control system of the present invention adopts a strategy of coordinated control by multiple controllers, combines PID control, expert system to simulate manual adjustment, and LSTM predictive control to ensure the stability and control accuracy of the pre-stored chamber pressure of the CDQ. The intelligent control system controls the pre-stored chamber pressure in two cases: pressure control when the furnace lid is open for coking and pressure control when the furnace lid is closed without coking.

[0136] For the pressure control when the furnace lid is open for coking, it specifically includes the following sub-steps:

[0137] Step 4.1.1 Detect the furnace lid opening signal and immediately enter the expert system control mode. Since the furnace lid is open during the coking process, the pre-stored chamber is connected to the outside world, and the system pressure fluctuates violently. At this time, the traditional PID control cannot respond in time, so the expert system is used to simulate manual adjustment.

[0138] First, perform initial adjustment. According to the empirical rule, directly increase the opening of the bleed valve by a fixed ratio (10%~30%). Then perform deviation adjustment, and the deviation adjustment rules are as follows:

[0139] When the deviation ΔP far exceeds the set threshold immediately increase the opening of the bleed valve by 10% to accelerate pressure release. If it still does not drop to the target range, further increase it.

[0140] When the deviation is within a reasonable range maintain the current opening of the bleed valve.

[0141] When the deviation is lower than the set threshold the opening of the relief valve is gradually reduced by 5% to prevent too low pressure.

[0142] The adjustment period is set to 5 - 10 seconds to adapt to the dynamic changes of the system.

[0143] Step 4.1.2 The LSTM prediction system starts to work at this time, predicts the pressure trend in the next 1 minute based on past data, and adjusts the relief valve in advance to prevent fire and smoke caused by too high pressure. If it is predicted that the pressure will rise, the opening of the relief valve is increased in advance. If it is predicted that the pressure will drop, the opening of the relief valve is reduced in advance.

[0144] The LSTM prediction model inputs the data of the circulating air volume and the air volume of the empty guide in the past 5 minutes, outputs the pressure of the pre - storage chamber after 1 minute, and cooperates with the expert system for early compensation during the coke charging stage.

[0145] Sub - step 4.1.3 Real - time pressure feedback: Based on the real - time sensor feedback data, the expert system makes secondary fine - tuning.

[0146] For the pressure control when the furnace lid is closed and no coke is charged, it specifically includes the following sub - steps:

[0147] Sub - step 4.2.1 The furnace lid is closed, and the pressure inside the pre - storage chamber tends to be stable, entering the PID control mode to ensure the stability of the system. The PID control process is as follows:

[0148] Set the target pressure value P exttarget and calculate the deviation between the current pressure value P extcurrent and the target value: ΔP = P extcurrent- P exttarget .

[0149] The PID control formula: ; After calculating the control quantity u(t), adjust the opening of the relief valve to stabilize the system pressure.

[0150] Sub - step 4.2.2 The LSTM prediction system continues to monitor the disturbance, predicts the external air volume disturbance, makes fine - tuning to the circulating fan and the relief valve, and maintains the pressure of the pre - storage chamber within the target pressure range (2.5 - 3.5 kPa) to reduce the system lag.

[0151] The LSTM prediction model inputs the data of the circulating air volume and the air volume of the empty guide in the past 5 minutes, outputs the pressure of the pre - storage chamber after 1 minute, and the prediction result is used for the feed - forward compensation of the PID controller.

[0152] Sub-step 4.2.3 Only retain the PID and LSTM prediction adjustments, monitor the system fluctuations in real time, and make adjustments at any time. The intelligent control system enters the normal operation state and waits for the next round of coke charging signal.

[0153] The prechamber pressure control method of the present invention avoids the instability problems that may be brought about by a single control model for unified control of all working conditions through dynamic model switching, thereby achieving more refined and accurate control. The prechamber pressure disturbance prediction based on the neural network can predict in advance the pressure changes caused by the fluctuations of the recycle fan and the air guide air volume, and then reduce the parameter fluctuations caused by the system lag through feedforward control.

[0154] The present invention has been described in detail above. The above description is only the preferred embodiment of the present invention, and it cannot limit the scope of the present invention. That is, all equal changes and modifications made according to the scope of this application should still fall within the scope covered by the present invention.

Claims

1. A pressure control method for the pre-storage chamber of a coke dry quenching furnace based on artificial intelligence, characterized in that, It includes the following steps: Step 1: Raw data collection and data preprocessing; Step 2: Switching of the process control model in the pre-storage chamber of the coke dry quenching furnace; During the coke charging process of the coke dry quenching furnace, independent process models are established for two working conditions: coke charging with the furnace lid open and no coke charging with the furnace lid closed; When the furnace lid is open, the negative pressure in the pre-storage chamber is maintained by adjusting the opening degree of the relief valve; When no coke is charged with the furnace lid closed, the pressure in the pre-storage chamber is controlled by adjusting the circulating gas relief valve to maintain the negative pressure in the pre-storage chamber; Step 3: Prediction of pre-storage chamber pressure disturbance; An LSTM neural network is used to construct a prediction model, and by predicting future pressure changes in advance, feedforward information is provided for the controller; The prediction of pre-storage chamber pressure disturbance specifically includes the following sub-steps: Sub-step 3.1 Data preprocessing; Data is collected once every 5 seconds, and a total of 60 data points are collected within 5 minutes. The length of the input sequence is 60; For each input sequence, the data of the past 5 minutes is used as the input, and the corresponding output is the pressure in the pre-storage chamber 1 minute later; Set the time series data as X = {x_1, x_2, …, x_T}, where each x_t contains two characteristic quantities: circulating air volume and empty guide air volume, and the corresponding pre-storage chamber pressure sequence is Y = {y_1, y_2, …, y_T}; Sub-step 3.2 Design of the LSTM prediction model; Sub-step 3.3 Model training: The training process of the LSTM prediction model includes the following sub-steps: Sub-step 3.3.1 Data preprocessing and sample construction: Normalize the circulating air volume and empty guide air volume; From the continuously collected data, the input sequence and the target output are constructed using the sliding window method; Input sequence: Each sample is 2D data of the past 60 time steps, denoted as ； Target output: Predict the pressure in the pre-storage chamber 1 minute later, denoted as ； Each feature is normalized according to the following formula: ； Sub-step 3.3.2 Initialize model parameters: randomly initialize the weight matrices and bias vectors of the LSTM layer and the fully connected layer; for the fully connected layer, initialize the parameters: ; For the LSTM layer, initialize the weight matrices W f , W i , W C , W o and the corresponding bias vectors b f , b i , b C , b o ; Sub-step 3.3.3 Start training Epoch: Divide the entire training process into multiple Epochs, and each Epoch traverses the entire training set once; Sub-step 3.3.4 Load training data: Load the training data batch by batch, and the current batch of data is ; Sub-step 3.3.5 Forward propagation to calculate the predicted value: Input the current batch of data into the LSTM network, and after multiple-layer calculations, obtain the predicted output: ; Among them, f(X;θ) represents the forward propagation process of the network, and θ represents all model parameters; Sub-step 3.3.6 Calculate the loss function: Use the mean square error to measure the error between the predicted value and the true value: ； where N is the number of samples, and y i is the true value, and is the predicted value of the model; Sub-step 3.3.7 Backpropagation to calculate the gradient: According to the loss function L, calculate the gradients of each parameter through the backpropagation algorithm: ； Among them, g represents the gradient value; is the gradient vector, representing the gradient of the loss function L with respect to the parameter θ, which describes the rate of change of the loss function at the current parameter value; Sub-step 3.3.8 Update parameters using the Adam optimizer: The Adam optimizer updates the parameters according to the following formula: ； Among them, g t is the current gradient, β1 and β2 are momentum parameters, α is the learning rate, is a small constant to prevent division by zero, and θ t represents the model parameters; Sub-step 3.3.9 Evaluate the performance of the validation set: After each Epoch, use the validation set data for forward propagation, calculate the validation loss and evaluation metrics; Sub-step 3.3.10 Determine whether the stopping condition is met: Determine whether the stopping condition is met according to the validation set metrics. If the validation loss no longer decreases or reaches the preset number of training epochs, end the training; Otherwise, continue with the next Epoch; Sub-step 3.4 Model validation; Step 4: The intelligent control system adjusts the bleeder valve in advance according to the coking state signal and the disturbance prediction result; the intelligent control system controls the pressure in the pre-storage chamber, which is divided into pressure control when the furnace cover is open for coking and pressure control when the furnace cover is closed and no coking is carried out.

2. The method according to claim 1, wherein In Step 2, the specific process of adjusting the opening of the bleeder valve when the furnace cover is open includes the following sub-steps: Sub-step 2.1.1 Initial adjustment: Directly increase the corresponding percentage based on the original opening of the relief valve according to experience, and set the initial opening of the relief valve as D initial , and the percentage increase based on the original opening of the relief valve is , then the adjusted opening of the relief valve is expressed as: ； Sub-step 2.1.2 Deviation correction: Further adjust the opening of the bleeder valve according to the deviation between the pressure in the pre-storage chamber and the set value; if the deviation is positive, increase the opening of the bleeder valve; if the deviation is negative, decrease the opening of the bleeder valve. Set the target pressure of the pre-storage chamber to P target , and the current pressure of the pre-storage chamber is P current , and the adjustment coefficient of the bleed valve is K adjust , then the opening degree that the bleed valve needs to be adjusted is expressed as: ； The final opening of the bleeder valve is expressed as: 。 3. The method according to claim 1, characterized in that, In Step 2, when the furnace cover is closed and no coking is carried out, the specific process of adjusting the circulating gas bleeder valve includes the following sub-steps: Sub-step 2.2.1 Calculate the pressure deviation value of the pre-storage chamber: The current pressure value P of the pre-storage chamber current and the target pressure value P target The deviation value between them is expressed as: ； Sub-step 2.2.2 According to the deviation value, the PID controller calculates the control quantity u(t) for adjusting the opening of the bleeder valve, and the control quantity u(t) is calculated by the following formula: ； where K p is the proportional gain, K i is the integral gain, and K d is the derivative gain.

4. The method according to claim 1, wherein In sub-step 3.1, for each time step t, construct the input-output pair (X_t, y_{t + 12}), where X_t = {x_{t - 59}, x_{t - 1}, …, x_t}; the total length of the historical data is T, and for each valid time point t satisfying t ≥ 60 and t + 12 ≤ T, construct the sample pair: ; Among them, ; y t+12 is the pressure in the pre-storage chamber measured at time t + 12.

5. The method according to claim 1, wherein In Sub-step 3.2, the network structure of the LSTM prediction model includes: Input layer: Receive an input sequence with a shape of (60, 2). The input data is the data of the past 60 time steps, and each time step contains 2 features. Then ； where x t,1 represents the circulating air volume at the t-th time step, and x t,2 represents the air guiding volume at the t-th time step; LSTM layer: Stack 2 LSTM layers to capture the long-term dependencies in the input sequence, and set 50 hidden units for each layer; for each time step t, the calculation process of the LSTM layer includes the following sub-steps: Sub-step 3.2.1 Forget gate layer, calculate how much information in the previous cell state C t−1 needs to be forgotten: ； Among them, W f is the forgetting gate weight matrix, b f is the forgetting gate bias vector, is the hidden state at the previous moment, is the current input, and σ(·) represents the Sigmoid function; Sub-step 3.2.2 Input gate layer, calculate which information in the current input needs to be written into the cell state: ； Generate a new candidate cell state: ； Among them, W i and W c are the weight matrices of the candidate cell states, and b i and b c are the bias vectors of the candidate cell states; Sub-step 3.2.3 Combine the results of the forget gate and the input gate to update the cell state: ； where, ⊙ represents element-wise multiplication; Sub-step 3.2.4 Output gate layer, calculate which part of the cell state is used as the current output: ； Among them, W o is the weight matrix of the output gate, and b o is the bias vector of the output gate; Sub-step 3.2.5 Finally update the hidden state: ； Hidden state h t has a dimension of 50, that is, each LSTM layer outputs a 50-dimensional vector; Dropout layer: Add a Dropout layer after each LSTM layer to prevent the model from overfitting; during the training process, the Dropout layer randomly sets the outputs of some neurons in the previous layer to 0 with a set probability of 0.2, that is, 20%; set h to represent the vector output by the LSTM layer, then the calculation process of the Dropout layer is expressed as: ； where: is a masking vector, and each element mi is generated according to a Bernoulli distribution , where p = 0.8 is the probability that the neuron is retained; ☉ represents element-wise multiplication; Fully connected layer: Input the hidden state of the last time step into the fully connected layer and map it to a single output, representing the predicted pressure value of the pre-storage chamber: ； Among them, is the weight matrix of the fully connected layer, is the bias vector of the fully connected layer.

6. The method according to claim 1, wherein In Sub-step 3.4, the verification process of the LSTM prediction model includes the following sub-steps: Sub-step 3.4.1 Performance evaluation indicators: During the verification process, the following two indicators are mainly used to measure the prediction error: Root Mean Square Error (RMSE) measures the standard deviation between the predicted value and the true value: ； Mean Absolute Error (MAE) measures the average absolute difference between the predicted value and the true value ； where N is the number of verification samples, and y i is the true pre-storage chamber pressure value of the i-th sample, and is the predicted pre-storage chamber pressure value by the model; Sub-step 3.4.2 Parameter tuning: According to the RMSE and MAE indicators, adjust the model parameters, repeat training and verification until a model with generalization ability meeting the control requirements is obtained; Sub-step 3.4.3 Feedback to the control strategy: After confirming that the verification result reaches the predetermined target, embed the prediction model into the control system, and adjust the opening of the bleeder valve with the predicted future pressure value through the feedforward control strategy to ensure that the pressure in the pre-storage chamber is maintained within the set range.

7. The method according to claim 1, wherein In Step 4, for the pressure control when the furnace cover is open for coking, it specifically includes the following sub-steps: Step 4.1.1 Detect the signal of the furnace lid opening, and immediately enter the expert system control mode; first, perform initial adjustment. According to the empirical rule, directly increase the opening of the bleed valve by a fixed ratio, and then perform deviation adjustment. The deviation adjustment rules are as follows: When the deviation ΔP far exceeds the set threshold immediately increase the opening of the relief valve by 10% to accelerate the pressure release. If the pressure still fails to drop to the target range, further increment it; When the deviation is within a reasonable range maintain the current opening degree of the relief valve; When the deviation is lower than the set threshold the opening of the relief valve is gradually reduced by 5% to prevent the pressure from being too low; The adjustment period is set to 5 - 10 seconds; Step 4.1.2 The LSTM prediction system starts to work at this time. According to the past data, it predicts the pressure trend in the next 1 minute. If it predicts that the pressure will rise, increase the opening of the bleed valve in advance. If it predicts that the pressure will fall, reduce the opening of the bleed valve in advance; The LSTM prediction model inputs the circulating air volume and blanking air volume data in the past 5 minutes, and outputs the pressure in the prechamber after 1 minute, which is used to cooperate with the expert system for early compensation during the coke charging stage; Sub-step 4.1.3 Real-time pressure feedback: Based on the real-time sensor feedback data, the expert system performs secondary fine-tuning.

8. The method according to claim 7, wherein In Step 4, for the pressure control when the furnace lid is closed and no coke is charged, it specifically includes the following sub-steps: Sub-step 4.2.1 The furnace lid is closed, and the pressure inside the prechamber tends to be stable, and enter the PID control mode; Sub-step 4.2.2 The LSTM prediction system continues to monitor the disturbance, predicts the external air volume disturbance, and fine-tunes the circulating fan and the bleed valve to maintain the pressure in the prechamber within the target pressure range; The LSTM prediction model inputs the circulating air volume and blanking air volume data in the past 5 minutes, and outputs the pressure in the prechamber after 1 minute. The prediction result is used for the feed-forward compensation of the PID controller; Sub-step 4.2.3 Only retain the PID and LSTM prediction adjustments, monitor the system fluctuations in real time, and make adjustments at any time.

Images