Dry quenching circulating air system state prediction method based on deep learning

Abstract

The invention belongs to the technical field of dry quenching recirculating air systems, and particularly relates to a dry quenching recirculating air system state prediction method based on deep learning, which comprises the following steps: firstly, collecting data and preprocessing data, then constructing a deep neural network model with a core of a selective state space block and a Kalman filter map neural network algorithm, and finally, predicting the state of the dry quenching recirculating air system. And after the model construction is completed, performing model training, optimization and verification, developing a self-adaptive control strategy according to a prediction result, and finally integrating the deep neural network model and the self-adaptive control strategy into a control system of a dry quenching production line to predict the state of the dry quenching circulating air system. By accurately capturing the dynamic relationship among key parameters such as the circulating air volume, the coke discharge temperature and the coke discharge amplitude, accurate prediction of the system state is realized, the production efficiency and the product quality of dry quenching can be improved, the operation cost can be reduced, the stability of the system is enhanced, and the economic benefit is increased. And powerful support is provided for intelligent and automatic development of the coking industry.

Description

Technical Field

[0001] The present invention belongs to the technical field of dry coke quenching circulating air systems, and in particular relates to a method for predicting the state of a dry coke quenching circulating air system based on deep learning. Background Art

[0002] The CDQ circulating air system is the core component of the CDQ process. Its main function is to send the cooled inert gas into the CDQ furnace through the circulating fan to exchange heat with the red coke to achieve the cooling of the red coke. The state of the circulating air system directly affects the efficiency of CDQ and the product quality. In actual production, the fluctuation of parameters such as circulating air volume, coke discharge temperature and coke discharge amplitude will directly affect the thermal efficiency of CDQ and the cooling quality of coke. Therefore, accurate prediction and control of these parameters are of great significance to improve the production efficiency of CDQ, reduce energy consumption and ensure product quality.

[0003] Traditional CDQ circulating air system control methods mainly rely on empirical rules and simple mathematical models, which have certain limitations when dealing with complex industrial processes. With the advancement of industrial intelligent manufacturing, it has become an inevitable trend to optimize and upgrade traditional industrial processes using advanced information technology and intelligent algorithms. As a cutting-edge artificial intelligence technology, deep learning has been successfully applied in image recognition, natural language processing, speech recognition and other fields, providing new ideas for state prediction and intelligent control of CDQ circulating air systems. Summary of the invention

[0004] In view of the various shortcomings of the existing technology, the inventors have researched and designed a CDQ circulating air system state prediction method based on deep learning in long-term practice. By accurately capturing the dynamic relationship between key parameters such as circulating air volume, coke discharge temperature and coke discharge amplitude, accurate prediction of the system state can be achieved. This can not only improve the production efficiency and product quality of CDQ, but also reduce operating costs and enhance the stability of the system, providing strong support for the intelligent and automated development of the coking industry.

[0005] A method for predicting the state of a CDQ circulating air system based on deep learning of the present invention comprises the following steps.

[0006] Step 1: Data collection and preprocessing.

[0007] Step 2: Build a deep neural network model; build the STG-Mamba model, capture the state evolution of the system in the time dimension through the selective state space model, and set the selective state space block and Kalman filter graph neural network algorithm in the STG-Mamba model; the selective state space block is used to accurately describe the dynamic evolution of the space-time graph network, and capture the complex heterogeneous dependencies in the space-time graph network through input-dependent edge construction and dynamic node feature selection; the Kalman filter graph neural network algorithm is used for adaptive space-time graph structure generation and upgrading, which smoothly adapts to the context of selective state space evolution while maintaining linear complexity.

[0008] Step 3: Embed the causal inference algorithm into the deep neural network model.

[0009] Step 4: Integrate anomaly detection and fault diagnosis modules into the deep neural network model, and use pattern recognition technology to monitor the system status in real time. When abnormal trends or potential faults are detected, the system will immediately issue an early warning.

[0010] Step 5: Introduce environmental parameters into the deep neural network model and use the environmental parameters as model input so that the model can adapt to different environmental conditions, including temperature, humidity, and wind speed.

[0011] Step 6: Embed a user-customized module in the deep neural network model and provide a user-customized interface. Users can adjust the model parameters and structure according to their specific needs to achieve modular design.

[0012] Step 7: Model training and validation; Use the preprocessed data to train the constructed deep neural network model; During the training process, use the cross-validation method to evaluate the performance of the model to ensure that the model can accurately capture the dynamic relationship between parameters and make accurate predictions.

[0013] Step 8: Introduce explainable artificial intelligence technology into model training, retain the feature representation of the intermediate layer when training the deep learning model, and then use LIME or Shapley value algorithm to analyze the model's prediction decision, so that operators can understand the model's prediction logic and results.

[0014] Step 9: Perform real-time performance optimization and cloud-edge collaborative optimization on the deep neural network model.

[0015] Step 10: Based on the model prediction results, develop an adaptive control strategy to dynamically adjust the circulation air volume and other parameters.

[0016] Step 11: System integration and testing: Integrate the deep neural network model and adaptive control strategy into the control system of the CDQ production line, predict the state of the CDQ circulating air system, and obtain the prediction results.

[0017] Furthermore, the specific method of data collection and preprocessing in step 1 is to first collect time series data of circulating air volume, coke discharge temperature, coke discharge amplitude, and steam flow from the dry quenching cycle control system, including coke discharge set amount, calibrated coke discharge amount and system error, and these data are used as the basis for model training and prediction; the collected data is then preprocessed, and the preprocessing steps include eliminating abnormal data, clearing outliers, and retaining normal data; the preprocessed data is arranged in chronological order and normalized so that its value range is between 0 and 1.

[0018] Furthermore, clearing outliers in preprocessing specifically includes the following sub-steps.

[0019] Sub-step 1.1.1 Calculate quartiles: For each parameter’s time series data, calculate its first quartile Q1 and third quartile Q. Quartiles are three points that divide the data into four equal parts. The data below Q1 is 25%, and the data above Q3 is 25%. Sort the data from small to large. The position of Q1 is (n+1) / 4, and the position of Q3 is 3(n+1) / 4. If Q1 and Q3 are not integers, round them up.

[0020] Sub-step 1.1.2 Calculate the interquartile range (IQR): IQR = Q3-Q, which represents the range of the middle 50% of the data.

[0021] Sub-step 1.1.3: Set the outlier range: Q1-1.5×IQR and Q3+1.5×IQR are used as the boundaries of outliers. Data points outside the set range are considered outliers.

[0022] Sub-step 1.1.4 Data screening: Check whether each data point is within the set outlier range. If the value of a data point is less than Q1-1.5×IQR or greater than Q3+1.5×IQ, mark the data point as an outlier and remove it.

[0023] Furthermore, the design method of the selective state space block in step 2 is as follows: State update equation: h(t k+1 )=A - h(t k )+B - x(t k )h(t k+1 ) Among them, h(t) represents the state vector, A - is the state transfer matrix, B - is the control input matrix, x(t) is the input vector; Output equation: y(t k+1 )=Ch(t k+1 )y(tk+1 ) Where y(t) is the output vector and C is the output matrix.

[0024] The calculation method of the Kalman filter graph neural network is as follows.

[0025] Sub-step 2.1, first perform the prediction step and use the following formula to predict the state: H - (t k+1 )=A - H(t k )H - (t k+1 )=A - H(t k ) Among them, H(t k ) means that at time step t k The state estimate, H - (t k+1 ) means that at time step t k+1 state estimation; The error covariance prediction is performed using the following formula: P - (t k+1 )=A - P(t k )A -T +QP - (t k+1 ) Among them, P(t k ) is at time step t k The updated error covariance matrix, P - (t k+1 ) is at time step t k+1 The prediction error covariance matrix, A -T is the transpose of matrix A, and Q is the process noise covariance matrix.

[0026] Sub-step 2.2, perform the update step and calculate the Kalman gain using the following formula: K(t k+1 )=P - (t k+1 )C T (CP - (t k+1 )C T +R) -1 The following formula is used to calculate the status update: H(t k+1 )=H - (t k+1 )+K(tk+1 )(y(t k+1 )-CH - (t k+1 )) The error covariance is updated using the following formula: P(t k+1 )=(IK(t k+1 )C)P - (t k+1 )P(t k+1 ) Where K(t) is the Kalman gain, C T is the transpose of matrix C, I is the identity matrix, and R is the observation noise covariance matrix.

[0027] Furthermore, the specific method of embedding the causal inference algorithm into the deep neural network model in step 3 includes the following sub-steps.

[0028] Sub-step 3.1: Use Do-calculus or causal graph model to identify and infer the causal relationship between circulation air volume, coke discharge temperature, coke discharge amplitude, and steam flow rate.

[0029] Sub-step 3.2: Embed a causal inference module in the deep learning model to ensure that the prediction results can reflect the true causal effect. The causal effect can be expressed as: E[Yi(do(X=x))]-E[Yi(X)] Where Yi(do(X=x) represents the expected value of the result Y when X is set to x; Yi(X) represents the expected value of X to Y in the natural state.

[0030] In step 4, key features are extracted from the state data of the circulating air system, and an anomaly detection model is constructed using the One-Class SVM or Isolation Forest algorithm. Combined with the expert system, the cause of the detected anomaly is analyzed and diagnosed.

[0031] Furthermore, in step 5, environmental parameters are first monitored, and then the environmental parameters are used as model inputs to dynamically adjust the model parameters according to environmental changes and perform environmental adaptability adjustments to maintain the accuracy of the prediction. The environmental adaptability adjustment can be expressed as: θnew=θold+βΔE Where θnew is the updated model parameter, θold is the old model parameter, β is the adjustment coefficient, and ΔE is the impact of environmental changes.

[0032] Furthermore, in step 6, a user customized module is embedded in the deep neural network model, and the specific method includes the following sub-steps.

[0033] Sub-step 6.1, collect user requirements and determine customized functions.

[0034] Sub-step 6.2, conduct modular design, design the model into a modular structure, so as to facilitate adjustment according to user needs; the interface definition of modular design can be expressed by the following formula: f(input,θuser)=output Where f is the prediction function of the model, θuser is the model parameter defined by the user, input is the model input, and output is the model output.

[0035] Furthermore, in step 7, a k-fold cross validation method is used to evaluate the performance of the model, which specifically includes the following sub-steps.

[0036] Sub-step 7.1: Divide the dataset: Randomly divide the preprocessed dataset into k subsets of equal size.

[0037] Sub-step 7.2: Test in turns: For each subset, use it as the test set and the remaining k-1 subsets as the training set, train the deep neural network model, and evaluate the performance of the model on the test set.

[0038] Sub-step 7.3 Calculate the average performance: average the performance indicators of k tests to obtain the final performance evaluation of the model.

[0039] Furthermore, the prediction decision of the model is analyzed using LIME or Shapley value algorithm, which specifically includes the following sub-steps.

[0040] Sub-step 8.1 Model training: Preserve the feature representation of the intermediate layers when training the deep learning model.

[0041] Substep 8.2 Select a data point: Select a data point x that needs to explain the prediction result.

[0042] Sub-step 8.3 Generate neighborhood data LIME: Select a data point x that needs to explain the prediction result, generate neighborhood data of x, these data points are close to x in the feature space; then calculate the weight of each neighborhood data point based on the distance between the neighborhood data and x, and use the Gaussian kernel function to calculate the weight: wi=exp(-d(x,xi)2 / σ2 where d(x,xi) is the distance between x and the neighboring data point xi, and σ is the bandwidth parameter of the Gaussian kernel.

[0043] Sub-step 8.4: Train the proxy model LIME: Select an interpretable linear model as the proxy model, use the generated neighborhood data and its corresponding model prediction results to train the proxy model. During the training process, use the calculated weights to weight each neighborhood data point so that the proxy model pays more attention to data points close to x.

[0044] Sub-step 8.5 Calculate Shapley value: Calculate the Shapley value of each feature; list all possible feature combinations, and for each feature combination, calculate the prediction result of the model under this combination. For each feature, calculate its marginal contribution under different feature combinations.

[0045] Substep 8.6 extracts the importance of features from the surrogate model LIME or Shapley value and generates an explanation report that explains the contribution of each feature to the prediction results.

[0046] Furthermore, in step 9, the method for real-time performance optimization includes the following sub-steps.

[0047] Sub-step 9.1.1 Model simplification: Use knowledge distillation technology to transfer the knowledge of the large model to the small model. In knowledge distillation, the loss function of the student model can be expressed by the following formula: Ldistill=LCE+λLKL Where LCE is the cross entropy loss, LKL is the KL divergence loss, and λ is the weight to balance the two losses.

[0048] Sub-step 9.1.2 Algorithm optimization: Embed a lightweight neural network architecture MobileNet or ShuffleNet into the model.

[0049] Sub-step 9.1.3 Real-time feedback: The model performs online learning and updates model parameters in real time to respond to environmental changes.

[0050] Furthermore, the specific method of cloud-edge collaborative optimization in step 9 includes the following sub-steps.

[0051] Sub-step 9.2.1 Edge computing: Deploy lightweight models at the edge for real-time data processing and preliminary analysis.

[0052] Sub-step 9.2.2 Cloud collaboration: Upload edge data to the cloud for deep learning model training and optimization.

[0053] Sub-step 9.2.3 Collaborative optimization: Through the collaborative work of the cloud and the edge, the rapid iteration of the model is achieved. The objective function of the collaborative optimization of the cloud and the edge can be expressed as: minθLcloud(θ)+λLedge(θ) Where Lcloud(θ) is the loss function on the cloud side, Ledge(θ) is the loss function on the edge side, and λ is the weight to balance the two losses.

[0054] Furthermore, the specific method for developing the adaptive control strategy in step 10 is as follows.

[0055] Sub-step 10.1 Predict model output: Use deep neural networks to predict the parameters of the circulating air system.

[0056] Sub-step 10.2 Control strategy design: Design an adaptive PID control algorithm based on the prediction results. The parameter update of the adaptive PID control algorithm can be expressed by the following formula: Ki(k+1)=Ki(k)+α[e(k)-e(k-1)] Where Ki is the parameter of the PID controller, e(k) is the error at the current moment, and α is the learning rate.

[0057] Sub-step 10.3 Dynamic adjustment: Adjust the control parameters according to the calculated update parameters.

[0058] The beneficial effects of the present invention are:

[0059] 1) Deep integration of deep learning and industrial processes: Apply deep learning technology to the state prediction of the CDQ circulating air system to realize the intelligence and automation of industrial processes.

[0060] 2) Multi-parameter collaborative prediction: Develop models that can simultaneously predict multiple key parameters to improve prediction accuracy and system response speed.

[0061] 3) Adaptive control strategy: According to the model prediction results, the circulating air volume is dynamically adjusted to maintain the exhaust temperature within the optimal range and reduce energy waste.

[0062] 4) Anomaly detection and early warning system: Integrate anomaly detection algorithms to monitor forecast data in real time. Once an abnormal trend is found, an early warning will be issued immediately to provide support for system maintenance and fault prevention. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is a comparison chart of circulating air volume prediction of the implementation case of the present invention.

[0064] Figure 2 It is a comparison chart of the coke discharge temperature prediction of the implementation case of the present invention.

[0065] Figure 3 It is a comparison diagram of the prediction of the defocusing amplitude of the implementation case of the present invention.

[0066] Figure 4 It is a comparison chart of steam flow prediction of the implementation case of the present invention. DETAILED DESCRIPTION

[0067] The present invention will be further described in detail below in conjunction with implementation examples. By describing these implementation examples in sufficient detail, it is possible for those skilled in the art to understand and practice the present invention. Without departing from the spirit and scope of the present invention, logical, achievable and other changes may be made to implementation. Therefore, the following detailed description should not be construed as limiting, and the scope of the present invention is limited only by the claims.

[0068] The present invention proposes a method for predicting the state of a CDQ circulating air system based on deep learning, which specifically includes the following steps.

[0069] Step 1: Data collection and preprocessing. First, it is necessary to collect time series data of circulating air volume, coke discharge temperature, coke discharge amplitude, and steam flow from the dry quenching cycle control system, including coke discharge set amount, calibrated coke discharge amount, and system error. These data will serve as the basis for model training and prediction. The collected data is preprocessed to ensure the quality and availability of the data. The preprocessing steps include removing abnormal data, removing outliers, and retaining normal data. In addition, the data will be arranged in chronological order and normalized so that its numerical range is limited to between 0 and 1 to meet the input requirements of the model.

[0070] The circulating air volume is a very important parameter in the CDQ circulating air system, which directly affects the cooling efficiency of red coke. The circulating air volume determines the ability of inert gas to exchange heat with red coke, thus affecting the cooling rate and final temperature of coke.

[0071] The coke discharge temperature refers to the temperature of the coke when it is discharged after being cooled in the dry quenching furnace. This parameter is crucial to ensure the quality of the coke and the smooth progress of the subsequent process. If the coke discharge temperature is too high, it may cause the coke to spontaneously combust during the subsequent transportation and storage process; if the coke discharge temperature is too low, it may mean that the circulating air volume is too large, resulting in energy waste.

[0072] The coke discharge amplitude reflects the vibration of coke when it is discharged from the dry quenching furnace. Appropriate vibration helps to discharge coke evenly and avoid coke accumulation or blockage in the furnace. The amplitude of coke discharge needs to be precisely controlled according to actual working conditions to ensure the stability and continuity of the dry quenching process.

[0073] During the CDQ process, a certain amount of steam will be generated after the circulating air exchanges heat with the red coke. The size of the steam flow rate can reflect the efficiency and degree of heat exchange. By predicting the steam flow rate, the circulating air volume and other related parameters can be adjusted in time to optimize the thermal energy utilization in the CDQ process and improve energy utilization efficiency.

[0074] When removing abnormal data, some reasonable data range rules are set according to the process knowledge and historical experience of the coke dry quenching circulating air system. For example, the normal range of the circulating air volume may be between Qmin and Qmax, and the normal range of the coke discharging temperature may be between Tmin and Tmax, etc. Traverse the collected data, and for each parameter, check whether its value is within the set normal range. If the value of a certain data point exceeds the normal range, that is, x < Qmin or x > Qmax (taking the circulating air volume as an example), then mark this data point as abnormal data and remove it.

[0075] In preprocessing, removing outliers specifically includes the following sub-steps.

[0076] Sub-step 1.1.1 Calculate quartiles: For the time series data of each parameter, calculate its first quartile Q1 (i.e., the 25th percentile) and the third quartile Q3 (i.e., the 75th percentile). Quartiles are three points that divide the data into four equal parts. Below Q1 is 25% of the data, and above Q3 is 25% of the data.

[0077] Sort the data from smallest to largest. The position of Q1 is (n + 1) / 4 (round up if not an integer), and the position of Q3 is 3(n + 1) / 4 (round up if not an integer).

[0078] Sub-step 1.1.2 Calculate the interquartile range (IQR): IQR = Q3 - Q1, representing the range of the middle 50% of the data.

[0079] Sub-step 1.1.3 Set the outlier range: Usually, Q1 - 1.5×IQR and Q3 + 1.5×IQR are used as the boundaries of outliers. Data points outside this range are considered outliers.

[0080] Sub-step 1.1.4 Data screening: Check whether each data point is within the set outlier range. If the value of a certain data point is less than Q1 - 1.5×IQR or greater than Q3 + 1.5×IQR, then mark this data point as an outlier and remove it.

[0081] The preprocessed data is normalized according to the following sub-steps.

[0082] Sub-step 1.2.1 Data sorting: Arrange the time series data of each parameter collected in chronological order.

[0083] Sub-step 1.2.2 For each parameter’s time series data, calculate its minimum value xmin and maximum value xmax. The minimum value xmin is obtained by the formula xmin=min(x1,x2,…,xn), and the maximum value xmax is obtained by the formula xmax=max(x1,x2,…,xn). Where xi is the value of the i-th data point and n is the number of data points.

[0084] Sub-step 1.2.3 Convert each data point xi to a normalized value xi′ using the formula xi′=(xi-xmin) / (xmax-xmin). After this processing, the data will be scaled to the interval [0,1]. If the data needs to be scaled to other intervals, such as [a,b], it is further converted according to the following formula: xi′=a+(ba)(xi-xmin) / (xmax-xmin).

[0085] Sub-step 1.2.4 replaces the original data with the normalized values ​​to obtain a normalized data set for subsequent model training and prediction.

[0086] Step 2: Build a deep neural network model. Build the STG-Mamba model. The STG-Mamba model regards the CDQ furnace circulating air system as a dynamic spatial-temporal graph (STG) system, and uses the selective state space model (SSSM) to capture the state evolution of the system in the time dimension. The selective state space block (GS3B) and Kalman filter graph neural network (KFGN) algorithm are set in the STG-Mamba model.

[0087] The STG-Mamba model adopts an encoder-decoder architecture and uses GS3B as the basic module to achieve efficient sequence data modeling.

[0088] GS3B is the basic module of the STG-Mamba model, which is used to accurately describe the dynamic evolution of the STG network. It captures the complex heterogeneous dependencies in the STG network through input-dependent edge construction and dynamic node feature selection. The design method of GS3B is as follows: State update equation: h(t k+1 )=A - h(t k )+B - x(t k )h(t k+1 ) Among them, h(t) represents the state vector, A - is the state transfer matrix, B - is the control input matrix, x(t) is the input vector; Output equation: y(t k+1 )=Ch(t k+1 )y(t k+1 ) Where y(t) is the output vector and C is the output matrix.

[0089] KFGN is used for adaptive spatiotemporal graph structure generation and upgrading, smoothly adapting to the context of selective state space evolution while maintaining linear complexity. The calculation method of KFGN is as follows.

[0090] Sub-step 2.1, first perform the prediction step and use the following formula to predict the state: H - (t k+1 )=A - H(t k )H - (t k+1 )=A - H(t k ) Among them, H(t k ) means that at time step t k The state estimate, H - (t k+1 ) means that at time step t k+1 state estimation; The error covariance prediction is performed using the following formula: P - (t k+1 )=A - P(t k )A -T +QP - (t k+1 ) Among them, P(t k ) is at time step t k The updated error covariance matrix, P - (t k+1 ) is at time step t k+1 The prediction error covariance matrix, A -T is the transpose of matrix A, and Q is the process noise covariance matrix.

[0091] Sub-step 2.2, perform an update step and calculate the Kalman gain using the following formula: K(t k+1 )=P - (t k+1 )C T (CP - (t k+1 )C T +R) -1 The following formula is used to calculate the status update: H(t k+1 )=H - (t k+1 )+K(t k+1 )(y(t k+1 )-CH - (t k+1 )) The error covariance is updated using the following formula: P(t k+1 )=(IK(t k+1 )C)P - (t k+1 )P(t k+1 ) Where K(t) is the Kalman gain, C T is the transpose of matrix C, I is the identity matrix, and R is the observation noise covariance matrix.

[0092] Step 3: Embed the causal inference algorithm into the deep neural network model.

[0093] In order to improve the prediction accuracy and interpretability of the model, the present invention introduces causal inference technology to clarify the causal relationship between various parameters in the CDQ circulating air system. By using the causal inference algorithm, the deep neural network model can not only identify the correlation between parameters, but also infer the causal effect between them, thereby providing operators with deeper insights and enhancing the credibility of the model. The specific method includes the following sub-steps.

[0094] Sub-step 3.1: Use causal inference algorithms, such as Do-calculus or causal graph models, to identify and infer the causal relationship between circulation air volume, coke discharge temperature, coke discharge amplitude, and steam flow rate.

[0095] Sub-step 3.2: Embed a causal inference module in the deep learning model to ensure that the prediction results can reflect the true causal effect. The causal effect can be expressed as: E[Yi(do(X=x))]-E[Yi(X)] Where Yi(do(X=x) represents the expected value of the result Y when X is set to x; Yi(X) represents the expected value of X to Y in the natural state.

[0096] Step 4: Integrate anomaly detection and fault diagnosis modules into the deep neural network model, and use pattern recognition technology to monitor the system status in real time. Once an abnormal trend or potential failure is detected, the system will immediately issue an early warning to provide support for timely maintenance and fault prevention.

[0097] The specific method is as follows: first, key features are extracted from the state data of the circulating air system, and an anomaly detection model is constructed using the One-Class SVM or Isolation Forest algorithm. Combined with the expert system, the cause of the detected anomaly is analyzed and diagnosed.

[0098] Step 5: Introduce environmental parameters into the deep neural network model and use them as model input. The model can adapt to different environmental conditions such as temperature and humidity, which improves the generalization ability of the model.

[0099] The specific method is: first, monitor the environmental parameters, monitor the temperature, humidity and other environmental parameters in real time, then use the environmental parameters as model input to adapt to different environmental conditions, and finally dynamically adjust the model parameters according to environmental changes to make environmental adaptability adjustments to maintain the accuracy of the prediction. Environmental adaptability adjustments can be expressed as: θnew=θold+βΔE Where θnew is the updated model parameter, θold is the old model parameter, β is the adjustment coefficient, and ΔE is the impact of environmental changes.

[0100] The STG-Mamba model achieves linear time complexity O(n) through a carefully designed algorithm, significantly reducing the computational overhead compared to Transformer-based methods, including the number of floating-point operations (FLOPs) and test inference time.

[0101] Step 6: Embed a user-customized module in the deep neural network model, provide a user-customized interface, and allow users to adjust model parameters and structure according to their specific needs to achieve modular design. This flexibility enables the model to adapt to the needs of different users and different working conditions, improving the practicality and market competitiveness of the model. The specific method includes the following sub-steps.

[0102] Sub-step 6.1, collect user requirements and determine customized functions.

[0103] Sub-step 6.2, conduct modular design and design the model into a modular structure so that it can be adjusted according to user needs. The interface definition of modular design can be expressed as: f(input,θuser)=output Where f is the prediction function of the model, θuser is the model parameter defined by the user, input is the model input, and output is the model output.

[0104] Step 7: Model training and validation. Use the preprocessed data to train the constructed deep neural network model. During the training process, methods such as cross-validation are used to evaluate the performance of the model to ensure that the model can accurately capture the dynamic relationship between parameters and make accurate predictions.

[0105] The present invention adopts the k-fold cross-validation method to evaluate the performance of the model, which specifically includes the following sub-steps.

[0106] Sub-step 7.1: Divide the data set: Randomly divide the preprocessed data set into k subsets (folds) of equal size, and try to keep the data distribution consistent in each subset.

[0107] Sub-step 7.2: Test in turns: For each subset, use it as a test set and the remaining k-1 subsets as training sets, train the deep neural network model, and evaluate the performance of the model on the test set. In this way, each subset has the opportunity to be used as a test set once.

[0108] Sub-step 7.3 Calculate the average performance: average the performance indicators (such as accuracy, mean square error, etc.) of k tests to obtain the final performance evaluation of the model.

[0109] Step 8: Introduce explainable artificial intelligence (XAI) technology into model training to enable operators to understand the model's prediction logic and results, display the model's decision-making process through visualization technology, enhance users' trust in the model, improve the model's interpretability, and promote the widespread application of the model. The specific method includes the following sub-steps.

[0110] Sub-step 8.1 Model training: Preserve the feature representation of the intermediate layers when training the deep learning model.

[0111] Substep 8.2 Select a data point: Select a data point x that needs to explain the prediction result.

[0112] Sub-step 8.3 Generate neighborhood data (LIME): Generate neighborhood data of x and calculate the weight of each neighborhood data point. LIME (Local Interpretable Model-agnostic Explanations) is a local interpretable model-agnostic explanation method that explains a single prediction by learning an interpretable proxy model in the local neighborhood of the model.

[0113] First, select a data point x that needs to explain the prediction result, and generate neighborhood data of x, which are close to x in the feature space. You can use perturbation methods (such as adding noise, permuting eigenvalues, etc.) to generate neighborhood data. Then, calculate the weight of each neighborhood data point based on the distance between the neighborhood data and x. Common distance measurement methods include Euclidean distance, Manhattan distance, etc. The weight can be calculated using the Gaussian kernel function: wi=exp(-d(x,xi)2 / σ2 where d(x,xi) is the distance between x and the neighboring data point xi, and σ is the bandwidth parameter of the Gaussian kernel.

[0114] Substep 8.4 Train surrogate model (LIME): Use the neighborhood data and its weights to train an interpretable linear surrogate model. First, choose an interpretable linear model as the surrogate model, such as linear regression or logistic regression. Then, use the generated neighborhood data and its corresponding model prediction results to train the surrogate model. During the training process, use the calculated weights to weight each neighborhood data point so that the surrogate model pays more attention to data points close to x.

[0115] Substep 8.5 Calculate Shapley value (Shapley value algorithm): Calculate the Shapley value of each feature. Select a data point x for which you want to explain the prediction result, consider all possible feature combinations (coalitions), and for each feature combination, calculate the prediction result of the model under that combination. For each feature i, calculate its marginal contribution under different feature combinations. The marginal contribution refers to the change in the model prediction result after adding feature i.

[0116] Substep 8.6 extracts the importance of features from the surrogate model LIME or Shapley value and generates an explanation report that explains the contribution of each feature to the prediction results.

[0117] Extract the importance of features from the trained surrogate model. For linear models, the importance of features can be measured by the coefficients (weights) of the model. The larger the absolute value of the coefficient, the greater the influence of the feature on the prediction result. Based on the importance of the features, generate an explanation report to explain the contribution of each feature to the prediction result. For example, you can list the top features that have the greatest impact on the prediction result and their contribution direction (positive contribution or negative contribution).

[0118] Extract the importance of the feature from the calculated Shapley value. The larger the absolute value of the Shapley value, the greater the influence of the feature on the prediction result. Based on the Shapley value of the feature, generate an explanation report to explain the contribution of each feature to the prediction result. For example, you can list the top features that have the greatest impact on the prediction result and their contribution direction (positive contribution or negative contribution).

[0119] Step 9: Perform real-time performance optimization and cloud-edge collaborative optimization on the deep neural network model.

[0120] Sub-step 9.1, in view of the real-time requirements in industrial applications, the present invention optimizes the model structure, reduces the computational complexity, and improves the response speed of the model. By adopting a lightweight neural network architecture and an efficient algorithm, it is ensured that the model can run quickly under limited computing resources to meet the requirements of online prediction and real-time control. The specific method includes the following sub-steps.

[0121] Sub-step 9.1.1 Model simplification: Use knowledge distillation technology to transfer the knowledge of the large model to the small model to reduce the computational complexity. In knowledge distillation, the loss function of the student model can be expressed as: Ldistill=LCE+λLKL Where LCE is the cross entropy loss, LKL is the KL divergence loss, and λ is the weight to balance the two losses.

[0122] Sub-step 9.1.2 Algorithm optimization: Select or design a lightweight neural network architecture, such as MobileNet or ShuffleNet.

[0123] Sub-step 9.1.3 Real-time feedback: Implement the model's online learning mechanism to update model parameters in real time in response to environmental changes.

[0124] Sub-step 9.2 combines cloud computing and edge computing to achieve cloud-edge collaborative optimization of the model. By performing real-time data processing and preliminary analysis on the edge, and deep learning and model optimization on the cloud, the data processing efficiency and model running speed are improved, which specifically includes the following sub-steps.

[0125] Sub-step 9.2.1 Edge computing: Deploy lightweight models at the edge to perform real-time data processing and preliminary analysis.

[0126] Sub-step 9.2.2 Cloud collaboration: Upload edge data to the cloud for training and optimization of deep learning models.

[0127] Sub-step 9.2.3 Collaborative optimization: Through the collaborative work of the cloud and the edge, the rapid iteration and optimization of the model can be achieved. The objective function of the collaborative optimization of the cloud and the edge can be expressed as: minθLcloud(θ)+λLedge(θ) Where Lcloud(θ) is the loss function on the cloud side, Ledge(θ) is the loss function on the edge side, and λ is the weight to balance the two losses.

[0128] Step 10: Based on the model prediction results, develop an adaptive control strategy to dynamically adjust the circulation air volume and other key parameters to maintain the system in the best working state. This adaptive control strategy can automatically adjust the control parameters according to real-time data and prediction results to improve the energy efficiency and stability of the system. The specific method includes the following sub-steps.

[0129] Sub-step 10.1 Predict model output: Use deep neural network to predict the parameters of the circulating air system, such as circulating air volume, coke discharge temperature, coke discharge amplitude, and steam flow rate.

[0130] Sub-step 10.2 Control strategy design: Based on the prediction results, design an adaptive PID control algorithm. The parameter update of the adaptive PID control algorithm can be expressed by the following formula: Ki(k+1)=Ki(k)+α[e(k)-e(k-1)] Where Ki is the parameter of the PID controller, e(k) is the error at the current moment, and α is the learning rate.

[0131] Sub-step 10.3 Dynamic adjustment: Adjust control parameters in real time to maintain the system in the best working state.

[0132] Step 11: System integration and testing: Integrate the prediction model and control strategy into the control system of the CDQ production line, and conduct field testing and optimization to ensure the accuracy and reliability of the model in the actual production environment.

[0133] The following is a specific implementation case - the status prediction of the CDQ circulating air system in a large coking plant.

[0134] 1) Background: In the production process of the CDQ circulating air system of a large coking plant, the circulating air volume, coke discharge temperature, and coke discharge amplitude and other parameters fluctuate greatly, resulting in low production efficiency and unstable product quality. Traditional control methods mainly rely on empirical rules and simple mathematical models, which cannot accurately predict and control these parameters. Therefore, it is decided to introduce the prediction method based on deep learning of the present invention.

[0135] 2) Data collection and preprocessing

[0136] Data collection: Time series data of circulating air volume, coke discharge temperature, coke discharge amplitude, and steam flow rate from January 2024 to June 2024 were collected from the CDQ cycle control system, including the coke discharge set amount, calibrated coke discharge amount, and system error.

[0137] Data preprocessing: Eliminate abnormal data: Use minimum-maximum normalization and Z-score method to eliminate abnormal data.

[0138] Remove outliers: Use box plot and Z-score method to remove outliers.

[0139] Normalization: Normalize the data to the interval [0,1].

[0140] 3) Model construction and training

[0141] Build a deep neural network model: Build the STG-Mamba model and set up the selective state space block and Kalman filter graph neural network algorithm.

[0142] Model training: The model was trained using the preprocessed data, and the model performance was evaluated using the 5-fold cross-validation method.

[0143] Introducing causal inference: Use Do-calculus to identify the causal relationship between parameters and embed a causal inference module.

[0144] Integrated anomaly detection and fault diagnosis module: Use One-Class SVM to build an anomaly detection model and combine it with an expert system for fault diagnosis.

[0145] Introduce environmental parameters: monitor environmental parameters such as temperature, humidity, wind speed, etc., and dynamically adjust model parameters.

[0146] User customized module: Provide customized interface and adjust model parameters and structure according to user needs.

[0147] 4) Model verification and testing

[0148] Model validation: The model is validated using data from July 2024 to evaluate the model’s predictive accuracy and stability.

[0149] Real-time performance optimization: Use knowledge distillation technology and lightweight neural network architecture (such as MobileNet) to simplify the model and implement online learning mechanism.

[0150] Cloud-edge collaborative optimization: Deploy lightweight models on the edge for real-time data processing, and train and optimize deep learning models on the cloud.

[0151] 5) Prediction results and comparison

[0152] See also Figure 1~Figure 4 ,In each figure, the red line is the actual value, the blue line is the ,prediction result of the method of the present invention, and the green line is the ,prediction result of the traditional method.

[0153] The prediction results of various parameters are as follows: Circulation air volume prediction: mean absolute error (MAE) is 2.3%, mean square error (MSE) is 3.1%; De-coking temperature prediction: MAE is 1.8℃, MSE is 2.5℃; Focus discharge amplitude prediction: MAE is 1.2%, MSE is 1.5%; Steam flow prediction: MAE is 2.1%, MSE is 2.8%.

[0154] Traditional method prediction results: Circulation air volume prediction: MAE is 4.5%, MSE is 5.8%; De-coking temperature prediction: MAE is 3.2℃, MSE is 4.1℃; Focus discharge amplitude prediction: MAE is 2.5%, MSE is 3.0%; Steam flow prediction: MAE is 3.8%, MSE is 4.5%.

[0155] 6) Effect comparison

[0156] Prediction accuracy: In the prediction of all key parameters, the MAE and MSE of the method of the present invention are significantly lower than those of the traditional method, indicating that the method of the present invention can predict the system state more accurately.

[0157] System stability: Through real-time performance optimization and cloud-edge collaborative optimization, the method of the present invention shows higher stability and response speed in actual production, and can adjust control parameters in time to maintain the system in the best working state.

[0158] Energy utilization efficiency: Based on the adaptive control strategy of the prediction results, the method of the present invention can dynamically adjust the circulating air volume, reduce energy waste, improve energy utilization efficiency, and reduce operating costs.

[0159] 7) Conclusion

[0160] Through the test results of actual application cases, the deep learning-based CDQ circulating air system state prediction method of the present invention is significantly superior to the traditional method in terms of prediction accuracy, system stability and energy utilization efficiency. The method of the present invention can not only improve the production efficiency and product quality of CDQ, but also reduce operating costs, enhance system stability, and provide strong support for the intelligent and automated development of the coking industry.

[0161] The present invention has been described in detail above. The above description is only a preferred embodiment of the present invention and should not limit the scope of implementation of the present invention. That is, all equivalent changes and modifications made within the scope of this application should still fall within the scope of the present invention.

Claims

1. A method for predicting the state of a CDQ circulating air system based on deep learning, characterized in that: The following steps are involved: Step 1: Data collection and preprocessing; Step 2: Build a deep neural network model; build the STG-Mamba model, set the selective state space block and Kalman filter graph neural network algorithm in the STG-Mamba model; the selective state space block is used to describe the dynamic evolution of the spatiotemporal graph network, and capture the complex heterogeneous dependencies in the spatiotemporal graph network through input-dependent edge construction and dynamic node feature selection; the Kalman filter graph neural network algorithm is used for adaptive spatiotemporal graph structure generation and upgrading, smoothly adapting to the context of selective state space evolution while maintaining linear complexity; Step 3: Embed the causal inference algorithm into the deep neural network model; Step 4: Integrate anomaly detection and fault diagnosis modules into the deep neural network model, and use pattern recognition technology to monitor the system status in real time. When abnormal trends or potential faults are detected, the system will immediately issue an early warning. Step 5: Introduce environmental parameters into the deep neural network model and use the environmental parameters as model input to enable the model to adapt to different environmental conditions, including temperature, humidity, and wind speed; Step 6: Embed a user-customized module in the deep neural network model and provide a user-customized interface. Users can adjust model parameters and structures according to their specific needs to achieve modular design. Step 7: Model training and verification: Use the preprocessed data to train the constructed deep neural network model; During the training process, a cross-validation method is used to evaluate the performance of the model to ensure that the model can accurately capture the dynamic relationship between parameters and make accurate predictions; Step 8: Introduce explainable AI technology into model training, retain the feature representation of the intermediate layer when training the deep learning model, and then use LIME or Shapley value algorithm to analyze the model's prediction decision, so that operators can understand the model's prediction logic and results; Step 9: Perform real-time performance optimization and cloud-edge collaborative optimization on the deep neural network model; Step 10: Based on the model prediction results, develop an adaptive control strategy to dynamically adjust the circulation air volume and other parameters; Step 11: System integration and testing: Integrate the deep neural network model and adaptive control strategy into the control system of the CDQ production line, predict the state of the CDQ circulating air system, and obtain the prediction results.

2. The method according to claim 1, characterized in that In step 1, the specific method of data collection and preprocessing is to first collect the time series data of circulating air volume, coke discharge temperature, coke discharge amplitude, and steam flow from the dry coke quenching cycle control system, including the coke discharge setting amount, the calibrated coke discharge amount, and the system error. These data are used as the basis for model training and prediction; the collected data is then preprocessed, and the preprocessing steps include eliminating abnormal data, clearing outliers, and retaining normal data; The preprocessed data are arranged in chronological order and normalized so that their values ​​range from 0 to 1.

3. The method according to claim 2, characterized in that The removal of outliers in preprocessing includes the following sub-steps: Sub-step 1.1.1 Calculate quartiles: For each parameter's time series data, calculate its first quartile Q1 and third quartile Q. Quartiles are three points that divide the data into four equal parts. The data below Q1 accounts for 25% of the data, and the data above Q3 accounts for 25% of the data. Sort the data from small to large, with Q1 at (n+1) / 4 and Q3 at 3(n+1) / 4. If Q1 and Q3 are not integers, round them up. Sub-step 1.1.2 Calculate the interquartile range (IQR): IQR = Q3-Q, which represents the range of the middle 50% of the data; Sub-step 1.1.3: Set the outlier range: Q1-1.5×IQR and Q3+1.5×IQR are used as the boundaries of outliers. Data points outside the set range are considered outliers. Sub-step 1.1.4 Data screening: Check whether each data point is within the set outlier range. If the value of a data point is less than Q1-1.5×IQR or greater than Q3+1.5×IQ, mark the data point as an outlier and remove it.

4. The method according to claim 1, characterized in that: In step 2, the design method of the selective state space block is: State update equation: h(t k+1 )=A - h(t k )+B - x(t k )h(t k+1 ); Among them, h(t) represents the state vector, A - is the state transfer matrix, B - is the control input matrix, x(t) is the input vector; Output equation: y(t k+1 )=Ch(t k+1 )y(t k+1 ); Among them, y(t) is the output vector and C is the output matrix; The calculation method of the Kalman filter graph neural network is: Sub-step 2.1, first perform the prediction step and use the following formula to predict the state: H - (t k+1 )=A - H(t k )H - (t k+1 )=A - H(t k ); Among them, H(t k ) means that at time step t k The state estimate, H - (t k+1 ) means that at time step t k+1 state estimation; The error covariance prediction is performed using the following formula: P - (t k+1 )=A - P(t k )A -T +QP - (t k+1 ) ; Among them, P(t k ) is at time step t k The updated error covariance matrix, P - (t k+1 ) is at time step t k+1 The prediction error covariance matrix, A -T is the transpose of matrix A, Q is the process noise covariance matrix; Sub-step 2.2, perform an update step and calculate the Kalman gain using the following formula: K(t k+1 )=P - (t k+1 )C T (CP - (t k+1 )C T +R) -1 ; The following formula is used to calculate the status update: H(t k+1 )=H - (t k+1 )+K(t k+1 )(y(t k+1 )-CH - (t k+1 )); The error covariance is updated using the following formula: P(t k+1 )=(I-K(t k+1 )C)P - (t k+1 )P(t k+1 ); Where K(t) is the Kalman gain, C T is the transpose of matrix C, I is the identity matrix, and R is the observation noise covariance matrix.

5. The method according to claim 1, characterized in that In step 3, the specific method of embedding the causal inference algorithm into the deep neural network model includes the following sub-steps: Sub-step 3.1: Use Do-calculus or causal graph model to identify and infer the causal relationship between circulation air volume, coke discharge temperature, coke discharge amplitude, and steam flow rate; Sub-step 3.2: Embed a causal inference module in the deep learning model to ensure that the prediction results can reflect the true causal effect; the causal effect can be expressed as: E[Yi(do(X=x))]-E[Yi(X)]; Where Yi(do(X=x) represents the expected value of the result Y when X is set to x; Yi(X) represents the expected value of X to Y in the natural state; In step 4, key features are extracted from the state data of the circulating air system, and an anomaly detection model is constructed using the One-Class SVM or IsolationForest algorithm. Combined with the expert system, the cause of the detected anomaly is analyzed and diagnosed.

6. The method according to claim 1, characterized in that In step 5, environmental parameters are first monitored, and then the environmental parameters are used as model inputs. Model parameters are dynamically adjusted according to environmental changes to make environmental adaptability adjustments to maintain the accuracy of the prediction. Environmental adaptability adjustments can be expressed by the following formula: θnew=θold+βΔE; Where θnew is the updated model parameter, θold is the old model parameter, β is the adjustment coefficient, and ΔE is the impact of environmental changes; In step 6, a user customized module is embedded in the deep neural network model. The specific method includes the following sub-steps: Sub-step 6.1, collect user requirements and determine customized functions; Sub-step 6.2, conduct modular design, design the model into a modular structure, so as to facilitate adjustment according to user needs; the interface definition of modular design can be expressed by the following formula: f(input,θuser)=output; Where f is the prediction function of the model, θuser is the model parameter defined by the user, input is the model input, and output is the model output.

7. The method according to claim 1, characterized in that In step 7, the k-fold cross validation method is used to evaluate the performance of the model, which includes the following sub-steps: Sub-step 7.1: Divide the data set: randomly divide the preprocessed data set into k subsets of equal size; Sub-step 7.2: Test in turns: For each subset, use it as the test set and the remaining k-1 subsets as the training set, train the deep neural network model, and evaluate the performance of the model on the test set; Sub-step 7.3 Calculate the average performance: average the performance indicators of k tests to obtain the final performance evaluation of the model.

8. The method according to claim 1, characterized in that In step 8, explainable artificial intelligence technology is introduced into model training. The feature representation of the intermediate layer is retained when training the deep learning model, and then the LIME or Shapley value algorithm is used to analyze the prediction decision of the model. Specifically, the following sub-steps are included: Sub-step 8.1 Model training: retain the feature representation of the intermediate layers when training the deep learning model; Sub-step 8.2 Select data point: Select a data point x that needs to explain the prediction result; Sub-step 8.3 Generate neighborhood data LIME: Select a data point x that needs to explain the prediction result, generate the neighborhood data of x, calculate the weight of each neighborhood data point based on the distance between the neighborhood data and x, and use the Gaussian kernel function to calculate the weight: wi=exp(-d(x,xi)2 / σ2; Where d(x,xi) is the distance between x and the neighborhood data point xi, and σ is the bandwidth parameter of the Gaussian kernel; Sub-step 8.4: Train the surrogate model LIME: Select an interpretable linear model as the surrogate model, use the generated neighborhood data and its corresponding model prediction results to train the surrogate model, and during the training process, use the calculated weights to weight each neighborhood data point; Sub-step 8.5 Calculate Shapley value: Calculate the Shapley value of each feature; list all possible feature combinations, and for each feature combination, calculate the prediction result of the model under this combination; for each feature, calculate its marginal contribution under different feature combinations; Substep 8.6 extracts the importance of features from the surrogate model LIME or Shapley value and generates an explanation report that explains the contribution of each feature to the prediction results.

9. The method according to claim 1, characterized in that: In step 9, the real-time performance optimization method includes the following sub-steps: Sub-step 9.1.1 Model simplification: Use knowledge distillation technology to transfer the knowledge of the large model to the small model. In knowledge distillation, the loss function of the student model can be expressed by the following formula: Ldistill = LCE + λLKL; Where LCE is the cross entropy loss, LKL is the KL divergence loss, and λ is the weight to balance the two losses; Sub-step 9.1.2 Algorithm optimization: embed lightweight neural network architecture MobileNet or ShuffleNet into the model; Sub-step 9.1.3 Real-time feedback: The model performs online learning and updates model parameters in real time to respond to environmental changes; In step 9, the specific method of cloud-edge collaborative optimization includes the following sub-steps: Sub-step 9.2.1 Edge computing: Deploy lightweight models at the edge for real-time data processing and preliminary analysis; Sub-step 9.2.2 Cloud collaboration: Upload edge data to the cloud for deep learning model training and optimization; Sub-step 9.2.3 Collaborative optimization: Through the collaborative work of the cloud and the edge, the rapid iteration of the model is achieved; the objective function of the collaborative optimization of the cloud and the edge can be expressed as: minθLcloud(θ)+λLedge(θ); Where Lcloud(θ) is the loss function on the cloud side, Ledge(θ) is the loss function on the edge side, and λ is the weight to balance the two losses.

10. The method according to claim 1, characterized in that In step 10, the specific method of developing the adaptive control strategy includes the following sub-steps: Sub-step 10.1 Predict model output: Use deep neural network to predict the parameters of the circulating air system; Sub-step 10.2 Control strategy design: Design an adaptive PID control algorithm based on the prediction results. The parameter update of the adaptive PID control algorithm can be expressed by the following formula: Ki(k+1)=Ki(k)+α[e(k)-e(k-1)]; Where Ki is the parameter of the PID controller, e(k) is the error at the current moment, and α is the learning rate; Sub-step 10.3 Dynamic adjustment: Adjust the control parameters according to the calculated update parameters.

Images