Coke oven gas collector pressure control method based on multivariable constraint optimization

Abstract

The invention relates to the technical field of coking industry, in particular to a coke oven gas collecting pipe pressure control method based on multivariable constraint optimization, which comprises the following steps: collecting, screening and preprocessing historical operation data of a coke oven gas collecting pipe system under different operation conditions; designing a multivariable nonlinear constraint optimization control algorithm; introducing a constraint reflecting the actual operation of a coke oven system into a nonlinear constraint optimization problem, and adding penalty terms for a control input constraint and a gas collector pressure constraint into an optimization objective function; and at the beginning of each control period, acquiring key parameters of the coke oven system in real time, and performing prediction calculation on the key parameters of the coke oven system through the NARX neural network model. The method has the advantages that the generated data set can be ensured to comprehensively cover the states of the coke oven gas collecting pipe system in various actual operation scenes, so that the subsequently trained NARX neural network model can fully learn dynamic behavior modes of the system under different working conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of coking industry, and in particular to a coke oven gas collecting pipe pressure control method based on multivariable constraint optimization. Background Art

[0002] In the coking industry, coke oven is the core equipment for producing coke, and the pressure control of its gas collecting pipe system is extremely critical to improving production efficiency and ensuring stable operation of equipment. x Coal gas containing harmful components such as chlorinated coal, chlorinated coal, and chlorinated coal. Coke oven gas collecting systems are commonly used in industry to recover and process these gases, improving energy efficiency, reducing direct emissions of harmful gases, and contributing to environmental protection and resource recycling.

[0003] With the advancement of the national energy conservation and emission reduction strategy, the coking industry is committed to optimizing control systems and improving the accuracy of coke oven gas collector pressure regulation to meet energy conservation and efficiency requirements. The stability of the gas collector pressure is crucial to improving the quality of the coking process, safe production, and environmental protection. However, in actual applications, the gas collector system faces many complex problems, such as multiple variables, strong coupling, significant nonlinearity, large external disturbances, and dynamic time-varying. Inconsistent operating conditions between the coke oven and the carbonization chamber can lead to uneven pressure distribution; key operating links in the carbonization chamber, such as pushing coke, loading coal, and reversing heating, can cause severe pressure disturbances; and unreasonable pressure settings can cause serious environmental pollution, energy waste, and safety hazards.

[0004] Currently, industrial sites often employ single-variable PID control strategies for complex pressure systems, but this approach has significant limitations. Firstly, the complexity and nonlinear characteristics of coke oven systems make PID control less adaptable to multi-loop coupling, making it difficult to meet the demands of multivariable dynamic regulation. Secondly, existing equipment designs often reserve significant redundancy to ensure safety, resulting in low equipment utilization and significant resource waste. Furthermore, the diverse structures and topologies of different coke ovens and carbonization chambers make it difficult to construct a unified overall system model, further limiting the control system's optimization capabilities. Summary of the Invention

[0005] The purpose of the present invention is to provide a coke oven gas collecting pipe pressure control method based on multivariable constrained optimization, which can significantly improve the control accuracy, system response speed and stability. By applying data-driven neural network modeling and multivariable nonlinear predictive control algorithm, high-precision dynamic regulation of the coke oven gas collecting pipe pressure can be achieved, and the gas collecting pipe pressure control process can be optimized at the same time, thereby reducing energy consumption, improving coke quality, and ultimately ensuring the operating economy and safety of the coke oven gas collecting pipe system.

[0006] To achieve the above object, the present invention is implemented through the following technical solutions:

[0007] A coke oven gas collecting pipe pressure control method based on multivariable constrained optimization, comprising:

[0008] S1. Collect, filter, and preprocess historical operating data of the coke oven gas collector system under different operating conditions to generate a data set for training, validating, and testing a nonlinear autoregressive NARX neural network model with exogenous input. Train and validate the nonlinear autoregressive NARX neural network model with exogenous input to obtain multivariable coupling characteristics describing the coke oven gas collector system.

[0009] S2. Design a multivariable nonlinear constrained optimization control algorithm based on the NARX neural network model;

[0010] S3. Introducing constraints that reflect the actual operation of the coke oven system into the nonlinear constrained optimization problem, including the upper and lower limits of the coke oven gas collector pressure, the safe operating range of valve opening and blower speed, and energy consumption limits. Penalty terms for control input constraints and gas collector pressure constraints are added to the optimization objective function to impose penalties for violations.

[0011] S4. At the beginning of each control cycle, the key parameters of the coke oven system are collected in real time, and the key parameters of the coke oven system are predicted and calculated through the NARX neural network model to obtain the predicted value of the coke oven gas collecting pipe system state in the future prediction time domain. The control parameters are adjusted according to the optimization result of the constrained optimization control output, and the first value of the optimal control input increment sequence is applied to the control input. The control input value is updated to complete the control of the gas collecting pipe pressure.

[0012] In S1, preprocessing includes merging the data with aligned time by taking the intersection of the time axis, and mapping the data to the normalized interval using the maximum and minimum normalization method to eliminate the influence of the magnitude differences of different variables on the training and prediction of the nonlinear autoregressive NARX neural network model with exogenous input;

[0013] The screening adopts the dynamic segmentation method. The dynamic segmentation is carried out according to the identification results of the coke oven operating conditions, and the coke oven gas collector system operation data is divided into multiple subsets;

[0014] The dataset is obtained by merging the data with the aligned time and performing maximum and minimum value normalization preprocessing by taking the intersection of the time axis.

[0015] In S2, the multivariable nonlinear constrained optimization control algorithm is as follows:

[0016] In each control cycle, a nonlinear constrained optimization problem is constructed to predict the dynamic behavior of the coke oven gas collector system at multiple moments in the future. The optimization objective function is to minimize the sum of the squares of the tracking errors between the gas collector pressure and the set value in the future prediction time domain, the weighted sum of the squares of the control input incremental changes, and includes nonlinear system dynamic equation constraints based on the nonlinear autoregressive NARX neural network model with exogenous input. The control variables are dynamically adjusted through a rolling optimization method based on sequential quadratic programming, so that the coke oven gas collector system can follow the gas collector pressure set value while suppressing the pressure fluctuations caused by the carbonization chamber under different operating conditions.

[0017] The multivariable nonlinear constrained optimization control algorithm combines the dynamic changes of coke oven operating conditions with the optimization of control variables. The dynamic changes are predicted by the nonlinear autoregressive NARX neural network model with exogenous input.

[0018] In S3, the upper / lower limit constraints of the coke oven gas collecting pipe pressure are set according to the coke oven production safety requirements and process requirements, and the valve opening and blower speed are introduced into the safety operation constraints.

[0019] Penalty terms for control input constraints and manifold pressure constraints are added to the optimization objective function. This is achieved by introducing slack variables to allow small violations of hard constraints and applying penalty terms to the slack variables in the optimization objective function.

[0020] In S4, the real-time collected gas collecting pipe pressure is fed back to the control system and compared with the set value of the gas collecting pipe pressure. The control parameters are adjusted according to the optimization results output by the multivariable nonlinear constrained optimization control algorithm to form a feedback loop.

[0021] The nonlinear autoregressive NARX neural network model with exogenous input is obtained through offline training. The input of the nonlinear autoregressive NARX neural network model with exogenous input includes current and historical key parameters. The output of the nonlinear autoregressive NARX neural network with exogenous input is the predicted value of the gas collecting pipe pressure in multiple future prediction time domains.

[0022] The rolling optimization method of sequential quadratic programming is used to solve the nonlinear constrained optimization problem by converting it into multiple quadratic programming sub-problems;

[0023] The quadratic programming subproblem is solved using a quadratic programming solver. When solving the quadratic programming subproblem, the Gauss-Newton method is used to approximate the Hessian matrix.

[0024] The optimization objective function includes minimizing the sum of the squares of the tracking errors between the gas collecting pipe pressure and the set value in the future prediction time domain, and the weighted sum of the sum of the squares of the incremental changes in the control input. It also includes restrictions on the control input. The restrictions are dynamically adjusted according to the coke oven gas production and pressure fluctuation limit requirements to ensure the energy optimization operation of the coke oven system while ensuring the stability of the gas collecting pipe pressure.

[0025] Adjusting the control parameters according to the optimization result of the constrained optimization control output includes applying the first value of the optimal control input increment sequence to the control input and updating the control input value by incrementally increasing it.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] 1. It can ensure that the generated data set comprehensively covers the status of the coke oven gas collecting pipe system in various actual operating scenarios. This enables the subsequently trained NARX neural network model to fully learn the dynamic behavior patterns of the system under different working conditions, avoiding inaccurate or unpredictable predictions under certain working conditions due to missing data. The screening process can remove abnormal data, erroneous data, and irrelevant data, retaining the data that is valuable for model training. Preprocessing (such as data alignment and normalization) can eliminate the adverse effects of dimensional differences of different variables on model training, making the data more standardized and neat, in line with the requirements of the neural network model for input data, thereby improving the efficiency and accuracy of model training and laying a good foundation for subsequent model training, verification, and testing.

[0028] 2. The coke oven gas collector system has strong multivariable coupling and nonlinear dynamic characteristics. Traditional linear models cannot accurately describe its complex behavior. As a nonlinear model, the NARX neural network can effectively capture and learn the complex coupling relationships between multiple variables in the system through its nonlinear mapping capabilities. After training and verification, the NARX neural network model can accurately describe the dynamic characteristics of the coke oven gas collector system under different operating conditions, providing high-precision system model support for subsequent pressure control. By training and verifying on data sets containing data from different operating conditions, the NARX neural network model can learn the general laws of the system, rather than being limited to the behavioral patterns under a specific operating condition. When applied to the actual control of the coke oven gas collector system, the model has good adaptability and generalization capabilities to unprecedented operating conditions or system state changes, ensuring the reliability of prediction and control effects under various actual operating conditions.

[0029] 3.3. Since the NARX neural network model can accurately describe the multivariable coupling characteristics of the coke oven gas collecting pipe system, the multivariable nonlinear constrained optimization control algorithm designed based on this model can fully consider the mutual influence and coupling relationship between the variables in the system, which enables the control algorithm to more accurately coordinate and optimize the control of multiple control variables (such as valve opening, blower speed, etc.), avoiding the system imbalance or poor control effect caused by single variable control, thereby improving the performance and stability of the gas collecting pipe pressure control; the operation of the coke oven gas collecting pipe system is subject to multiple constraints, including upper and lower pressure limits, valve opening range, blower speed range, and energy consumption limit, etc. The optimization control algorithm based on the NARX neural network model can incorporate these complex constraints into the control strategy. By constructing and solving nonlinear constrained optimization problems, it ensures that the control results meet the stable operation requirements of the system and meet the actual constraints of industrial production, thereby achieving comprehensive optimization of the system in terms of safety, economy, and efficiency;

[0030] 4. Introducing constraints such as the upper and lower limits of the coke oven gas collecting pipe pressure, the safe operating area of ​​the valve opening and the blower speed can strictly limit the changes of the control variables and system states within the safe range, prevent equipment damage, production accidents or environmental pollution caused by conditions such as excessively high or low pressure, excessively large or small valve opening, and abnormal blower speed, and ensure the safe and stable operation of the coke oven system; considering energy consumption limit constraints and adding penalty terms for control input constraints and gas collecting pipe pressure constraints to the optimization objective function can enable the optimization algorithm to minimize energy waste and unnecessary consumption while meeting the system pressure control requirements. At the same time, the setting of penalty terms can improve the robustness of the optimization problem to measurement noise and model errors, allowing for minor violations of constraints in some cases. However, by imposing penalties on violations, the system can optimize its operation as much as possible within a safe and economical range, reducing production costs and improving the economic benefits of the enterprise;

[0031] 5. Real-time collection of key parameters of the coke oven system and inputting them into the NARX neural network model for prediction can obtain the predicted state value of the coke oven gas collector system within a certain time range in the future (within the prediction time domain) in advance. This enables the control system to understand the possible development trends and changes of the system in advance, such as the rising or falling trend of pressure, the possible pressure fluctuation amplitude, etc., thereby providing a basis for taking corresponding control measures in advance, and enhancing the foresight and initiative of the control system; since the operating conditions of the coke oven system are dynamically changing, real-time prediction can promptly reflect the latest changes in the system status. The control system can quickly adjust the control strategy based on these real-time prediction results, so that the control parameters can adapt to the changes in the operating conditions in a timely manner, improve the system's response speed and adaptability to dynamic disturbances (such as pressure fluctuations caused by operations such as coal loading and coke pushing), ensure that the pressure of the coke oven gas collector is always stable within the set range, and maintain the normal operation of the coking process;

[0032] 6. The first value of the optimal control input increment sequence output by the constrained optimization control is promptly applied to the control input, which can quickly convert the optimization results into actual control actions, such as adjusting the valve opening, changing the blower speed, etc. This timely update of the control parameters can ensure that the system can be controlled according to the current optimal solution in each control cycle, avoiding the problem of control lag and improving the real-time and effectiveness of the gas collecting pipe pressure control; by continuously updating the control input value, the control system can continuously fine-tune and optimize the coke oven gas collecting pipe pressure to keep the system in the best operating state. At the same time, since the optimization algorithm is based on the consideration of system constraints and dynamic characteristics, each adjustment of the control parameters is carried out within a safe and reasonable range, thereby ensuring the stability and continuity of the control process, avoiding system oscillation or instability caused by excessive or inappropriate adjustment of the control parameters, and ensuring the precise control of the coke oven gas collecting pipe pressure and the long-term stable operation of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a structural diagram of the coke oven gas collecting pipe pressure control system based on multivariable constrained optimization.

[0034] Figure 2 3 is a modeling effect diagram of the neural network model established in the embodiment.

[0035] Figure 3 It is a schematic diagram of the constrained optimization control structure based on neural network established by the present invention.

[0036] Figure 4 2 is a diagram showing the gas collecting pipe pressure control effect in the embodiment. DETAILED DESCRIPTION

[0037] The present invention will be described in detail below with reference to the accompanying drawings, but it should be noted that the implementation of the present invention is not limited to the following embodiments.

[0038] The following examples are implemented under the premise of the technical solution of the present invention, and provide detailed implementation methods and specific operating processes, but the scope of protection of the present invention is not limited to the following examples. The methods used in the following examples are conventional methods unless otherwise specified.

[0039] Example 1

[0040] A coke oven gas collector pressure control method based on multivariable constrained optimization, specifically including:

[0041] S1. Collect, filter, and preprocess historical operating data of the coke oven gas collector system under different operating conditions to generate a data set for training, validating, and testing a nonlinear autoregressive NARX neural network model with exogenous input. Train and validate the nonlinear autoregressive NARX neural network model with exogenous input to obtain multivariable coupling characteristics describing the coke oven gas collector system.

[0042] Preprocessing includes merging the data with aligned time by taking the intersection of the time axis and mapping the data to the normalized interval using the maximum and minimum normalization method to eliminate the influence of the magnitude differences of different variables on the training and prediction of the nonlinear autoregressive NARX neural network model with exogenous input;

[0043] The screening process uses a dynamic segmentation approach. Dynamic segmentation is performed based on the identification results of coke oven operating conditions. The coke oven gas collector system operating data is divided into multiple subsets. Coke oven operating conditions include but are not limited to coal loading, coke pushing, and reversing. This allows for more accurate capture of the system's dynamic characteristics under different operating conditions. The identification results are used to construct a representative data set.

[0044] The dataset is obtained by merging the data with aligned time and performing maximum and minimum value normalization preprocessing by taking the intersection of the time axis. The dataset is used to train the NARX neural network model that describes the multivariable strong coupling and nonlinear autoregressive with exogenous input of the coke oven gas collecting pipe system.

[0045] S2. Based on the NARX neural network model, a multivariable nonlinear constrained optimization control algorithm is designed. The content is as follows:

[0046] Within each control cycle, a nonlinear constrained optimization problem is constructed to predict the dynamic behavior of the coke oven gas collector system at multiple future moments. The optimization objective function is to minimize the sum of the squares of the tracking errors between the gas collector pressure and the set value within the future prediction time domain, and the weighted sum of the squares of the incremental changes in the control inputs. This problem also includes nonlinear system dynamic equation constraints based on a nonlinear autoregressive NARX neural network model with exogenous inputs. A rolling optimization approach based on sequential quadratic programming is used to dynamically adjust control variables such as valve opening and blower speed, enabling the coke oven gas collector system to follow the set value of the gas collector pressure while suppressing pressure fluctuations in the carbonization chamber under different operating conditions and controlling the incremental changes in inputs such as valve opening and blower speed. The rolling optimization approach of sequential quadratic programming achieves efficient real-time optimization by transforming the nonlinear constrained optimization problem into multiple quadratic programming subproblems. The quadratic programming subproblems are solved using a quadratic programming solver. When solving these quadratic programming subproblems, the Gauss-Newton method is used to approximate the Hessian matrix to reduce the computational burden of each iteration and accelerate the optimization process.

[0047] The optimization objective function includes minimizing the sum of the squares of the tracking errors between the gas collecting pipe pressure and the set value in the future prediction time domain, and the weighted sum of the sum of the squares of the incremental changes in the control input. It also includes restrictions on the control input. The restrictions are dynamically adjusted according to the coke oven gas production and pressure fluctuation limit requirements to ensure the energy optimization operation of the coke oven system and reduce electricity consumption while ensuring the stability of the gas collecting pipe pressure.

[0048] The nonlinear autoregressive NARX neural network model with exogenous input is obtained through offline training. The input of the nonlinear autoregressive NARX neural network model with exogenous input includes current and historical key parameters. The current and historical key parameters include the gas collecting pipe pressure display value, the gas collecting pipe regulating valve optimization setting, the blower inlet regulating valve optimization setting, the blower return valve optimization setting, the blower speed optimization setting, the blower front / rear suction, and the high-pressure ammonia flow rate. The output of the nonlinear autoregressive NARX neural network with exogenous input is the predicted value of the gas collecting pipe pressure in multiple future prediction time domains.

[0049] The multivariable nonlinear constrained optimization control algorithm combines the dynamic changes of the coke oven operating conditions with the optimization of control variables such as valve opening and blower speed. The dynamic changes are predicted through a nonlinear autoregressive NARX neural network model with exogenous input. This can not only effectively suppress pressure fluctuations, but also reduce the system's energy consumption by optimizing the control variables while ensuring pressure stability.

[0050] S3. Constraints reflecting the actual operation of the coke oven system are introduced into the nonlinear constrained optimization problem, including upper and lower limits on the coke oven gas manifold pressure, safe operating ranges for valve opening and blower speed, and energy consumption limits. This ensures that the control results meet the safety and economic requirements of industrial operation. Penalties for the control input constraints and the gas manifold pressure constraint are incorporated into the optimization objective function. Slack variables are introduced to allow for minor violations of hard constraints, and penalties are applied to the slack variables in the optimization objective function to improve the solvability of the optimization problem and its robustness to measurement noise and model errors. The upper and lower limits on the coke oven gas manifold pressure are set within a safe range based on the safety and process requirements of the coke oven. For example, if the coke oven gas manifold pressure is too low and negative, air will enter the furnace, causing coke combustion and degrading coke quality. Excessive pressure can lead to raw gas leakage, causing environmental pollution and reducing raw gas recovery. Therefore, the upper and lower limits on the coke oven gas manifold pressure should be within a safe range that meets the safety and process requirements of the coke oven. Safety operation constraints are introduced for the valve openings of the gas collecting pipe regulating valve, blower inlet regulating valve, blower return valve, and the blower speed used to generate negative pressure to suck gas. For example, the rate of change limit of control variables such as valve opening and blower speed or the minimum value constraint of operating parameters such as gas collecting pipe pressure and blower speed are set to prevent instability caused by too fast or too slow response of the coke oven gas collecting pipe system, and avoid the decrease in gas transmission efficiency caused by low-speed operation.

[0051] S4. At the beginning of each control cycle, key parameters such as the coke oven gas collecting pipe pressure of the coke oven system are collected in real time, and the key parameters of the coke oven system are predicted and calculated through the NARX neural network model to obtain the predicted value of the coke oven gas collecting pipe system state in the future prediction time domain. According to the optimization results of the constrained optimization control output, the control parameters such as the valve opening of the gas collecting pipe regulating valve, blower inlet regulating valve, blower return valve, etc., the blower operating speed, etc. are adjusted. The first value of the optimal control input increment sequence is applied to the control input, and the control input value is updated to adapt to the operating condition changes, improve the system's robustness to dynamic disturbances, and complete the control of the gas collecting pipe pressure; the real-time collected gas collecting pipe pressure is fed back to the control system and compared with the set value of the gas collecting pipe pressure. The control parameters are adjusted through the optimization results output by the multivariable nonlinear constrained optimization control algorithm to form a feedback loop, ensuring that the pressure regulation system can quickly respond to external disturbances while improving the stability and robustness of the control. Adjusting control parameters according to the optimization results of the constrained optimization control output includes applying the first value of the optimal control input increment sequence to the control input and updating the control input value by incremental increment. The control input includes, for example, a valve opening set value and a booster fan operating speed set value.

[0052] Example 2

[0053] Taking the gas collecting pipe of a coal coking chemical industry site as an example, the coke oven gas collecting pipe pressure control method based on multivariable constrained optimization provided by the present invention is specifically described. The specific steps are as follows:

[0054] Step A: Collect and pre-process historical operating data of the gas collecting pipe system and construct a nonlinear autoregressive neural network prediction model with exogenous input;

[0055] See Figure 1 , collect long-term operation data from the integrated measurement and control system of the coke oven system. In order to build a data set that can accurately capture the dynamic characteristics of the system, the data is filtered to identify key process variables under typical working conditions. Key parameters include but are not limited to: gas collecting pipe pressure (such as the pressure PR of the No. 1 gas collecting pipe) 1a , No. 2 gas collecting pipe pressure PR 1b 、No. 3 gas collecting pipe pressure PR 2a 、No. 4 gas collecting pipe pressure PR 2b ), gas main pressure PR before primary cooler 41 , the valve opening of the gas collecting pipe regulating valve (such as the opening of the No. 1 regulating valve PVI 1a , No. 2 regulating valve opening PVI 1b , No. 3 regulating valve opening PVI 2a , No. 4 regulating valve opening PVI 2b ), valve opening of blower inlet regulating valve (PVI 41 ), valve opening of blower return valve (PVI 42 ), blower speed (SI), blower front / rear pressure (PR 42 PR 43 ), high-pressure ammonia water flow rate (such as No. 1 high-pressure ammonia water flow rate FR1, No. 2 high-pressure ammonia water flow rate FR2).

[0056] Perform multi-step preprocessing on the collected raw data:

[0057] First, data cleaning is performed, including removing outliers and filling missing values;

[0058] Secondly, to address the possible time deviations among different variables, we use the time axis intersection method to merge the aligned data to ensure the temporal synchronization of the data of different variables, which facilitates subsequent data analysis and modeling.

[0059] Thirdly, for different operating conditions such as coal loading, coke pushing, and reversing, the system operation data is divided into multiple subsets using a dynamic segmentation method to more comprehensively cover and reflect the system dynamic characteristics under different operating conditions;

[0060] Finally, the processed data is normalized, preferably using maximum and minimum value normalization, to map the data to the unit interval, effectively eliminating the problem of variable weight imbalance in the modeling process caused by dimensional differences, and improving the training efficiency and prediction accuracy of the model;

[0061] Finally, the preprocessed high-quality dataset is divided into training set, validation set and test set in appropriate proportions for training and evaluating the nonlinear autoregressive neural network model with exogenous input.

[0062] A lightweight nonlinear autoregressive neural network with external input is used to approximate the nonlinear dynamics of the gas collector. The input layer of the model (i.e., the nonlinear autoregressive neural network model with external input) receives the current and historical gas collector pressure (PR 1a PR 1b PR 2a PR 2b ) and the optimal setting of the opening of the gas collecting pipe regulating valve (PV 1a PV 1b PV 2a PV 2b ), the optimal setting of the blower inlet regulating valve opening (PV 41 ), the optimal setting of the blower return valve opening (PV 42 ), the optimized given blower speed (SP) and other control input variables, the output layer predicts the predicted value of the gas collecting pipe pressure in multiple prediction time domains in the future; in the neural network modeling process, a multi-layer neural network model with an appropriate number of layers and neurons can be constructed to capture the complex multi-variable coupling relationship and nonlinear characteristics in the system. The commonly used neural network training algorithm Adam optimizer is used to train the model, and the performance of the model on the validation set is evaluated by the cross-validation method, and the network model with the best prediction accuracy and generalization ability is selected for the subsequent control system design. In this embodiment, the trained nonlinear autoregressive neural network model with exogenous input can achieve high prediction accuracy for each variable, especially the high fitting degree of the gas collecting pipe pressure, see Figure 2 , its actual field measurement data is close to the neural network prediction results, and the high-precision prediction model can be used as a reliable prediction model for the nonlinear model predictive controller.

[0063] Step B: Design a multivariable nonlinear constrained optimization control algorithm based on a neural network model.

[0064] See Figure 3 , constrained optimization control architecture, uses nonlinear autoregressive neural network with exogenous input to build a prediction model, and realizes dynamic and precise control of coke oven gas collecting pipe pressure through rolling optimization.

[0065] The algorithm is divided into two stages:

[0066] Offline stage

[0067] Based on the historical operating data of the coke oven system, a multivariable prediction model is constructed using a neural network. The prediction accuracy of the model is optimized through training. Once the model is trained, it can be used for online control.

[0068] Online stage

[0069] Based on the trained prediction model, a nonlinear constrained optimization problem is constructed at each sampling moment, and an efficient optimization solver is used to solve it in real time and dynamically adjust the control variables.

[0070] Based on the multivariable nonlinear characteristics of the coke oven gas collecting pipe system, a neural network model with nonlinear autoregressive structure is constructed. The prediction formula is as follows:

[0071] y k =f(y k-1 ,u k-1 ) ①

[0072] In formula ①, y k represents the pressure value at time k, y k-1 Indicates the pressure value at the previous moment, u k-1 Indicates the control input at the previous moment, including the optimal setting of the gas collecting pipe regulating valve (PV 1a PV 1b PV 2a PV 2b ), the optimal setting of the blower inlet regulating valve (PV 41 ), the optimization setting of the blower return valve (PV 42 ), the blower speed is optimized (SP), considering the influence of historical data, the model is expressed as:

[0073]

[0074] In formula ②, y k+i|k represents the predicted value of the manifold pressure at the sampling time k for the future time k+i; y k+i-1 The pressure vector of the gas collecting pipe at time k+i-1 and before can be expressed as [y k+i-1 ,y k+i-2 ,…,y k+i-p ], p represents the delay order of pressure history data, u k+i-1 The control input vector representing time k+i-1 and multiple historical moments before can be expressed as [u k+i-1 ,u k+i-2 ,…,u k+i-q ], q is the delay order of the control input historical data; d k+i-1represents the known disturbance input vector at time k+i-1 and previous historical moments, such as high-pressure ammonia spraying, coke pushing, coal loading, environmental factors, etc.; f represents the nonlinear mapping relationship realized by the trained neural network model.

[0075] The neural network adopts a multi-layer perceptron structure, the activation function uses ReLU, the output layer is linearly activated, and the mean square error is used as the loss function. The network parameters are trained using the Adam optimization algorithm, and cross-validation is used to select the optimal network structure and hyperparameters. The resulting neural network model can realize pressure prediction at multiple moments in the future, with the prediction error controlled within ±2%, and has good dynamic prediction performance.

[0076] Based on the obtained lightweight neural network model, a nonlinear model predictive controller is constructed with this model as the prediction subsystem. The controller's goal is to minimize the deviation between the system's predicted output and the target value within a finite prediction horizon, while taking into account the smoothness of the control input and the actual system constraints. The controller optimization objective function is as follows:

[0077]

[0078] The objective function weight matrix Q of the optimization problem N , Q and R are used to weigh the impact of pressure tracking error and control input change on system performance. N represents the preset control time domain, and it is assumed that the prediction time domain is equal to the control time domain; y k+N represents the predicted value at the k+Nth moment; y r Indicates the set value of pressure fluctuation; y k+i It represents the predicted value at the kth moment for the k+ith moment in the future; Q represents the weight matrix of pressure fluctuation; R represents the weight matrix of control input.

[0079] To improve the computational efficiency and solvability of the algorithm, the hard constraints are converted into penalty terms in the objective function, that is, penalties are imposed for minor violations of the hard constraints. By introducing slack variables, the optimization problem is rewritten as follows:

[0080]

[0081] In formula ④, ε is the slack variable, S is the penalty weight matrix, which is used to control the tolerance of constraint violation; Δu k+i It is the incremental form of the control input, which represents the incremental control input at the k+i-th moment predicted at the k-th moment.

[0082] The sequential quadratic programming method is used to solve the above nonlinear constrained optimization problem. The nonlinear objective function and constraints are linearized and quadratically approximated, respectively, and then converted into multiple quadratic programming sub-problems for iterative solution. The derivative information of the neural network is used to construct the gradient of the optimization problem, and the Gauss-Newton method is used to approximate the Hessian matrix. That is, for the form ||J(Δu)|| 2 The Hessian of the term is approximately expressed as The product of Jacobian matrices is used to approximate the Hessian matrix, avoiding the explicit calculation of second-order derivatives, significantly reducing the computational burden of each iteration, and speeding up the optimization solution process.

[0083] Step C: Introduce system constraints during the optimization process to achieve card edge control of constrained optimization.

[0084] A variety of constraints are introduced to limit the system operation to a safe range, optimize energy efficiency, and prevent abnormal behavior caused by parameter settings or operating conditions fluctuations. The pressure of the coke oven gas collecting pipe system is limited to remain within the set range. By introducing upper and lower limit constraints, it is ensured that the pressure control sequence generated during the optimization process always meets the following requirements:

[0085] y min ≤y k+i ≤y max ,i=1,2,…N ⑤

[0086] In formula 4, to protect the stable operation of the control equipment and system, a safe operating range is set for key operating variables such as valve opening and fan speed to avoid system over-response or loss of control caused by full closure or full opening. The constraint is expressed as:

[0087] u min ≤u k+i ≤u max ,i=0,1,…N-1 ⑥

[0088] Δu min ≤Δu k+i ≤Δu max ,i=0,1,…N-1 ⑦

[0089] In formula ⑥, u k+i represents the control input at the k+i-th moment predicted at the k-th moment; u min Indicates the minimum value of the control input; u max Indicates the maximum value of the control input;

[0090] In formula ⑦, Δu k+i represents the increment of the control input at the k+i-th moment predicted at the k-th moment; Δu minIndicates the minimum value of the control input increment; Δu max Indicates the maximum value of the control input increment;

[0091] The system dynamic constraints are the dynamic characteristics predicted by the neural network model. Multi-step prediction is used to simulate the future system dynamics, thereby identifying dynamic interference in advance and enhancing the system's adaptability and robustness to dynamic disturbances.

[0092] Step D: Use the neural network model to predict the pressure, and dynamically adjust the valve opening and fan speed according to the optimization results to close the loop control of the coke oven gas collecting pipe system.

[0093] The key operating parameters of the coke oven system are collected in real time through the sensor network and used as input data for prediction calculations. These parameters include the current gas collecting pipe pressure, valve opening, and fan speed. The real-time collected parameters are input into the offline trained neural network model to predict the pressure change trend y in the next N control time domains. k+1 、y k+2 ,…,y k+N ; The neural network model quickly obtains the prediction results through forward calculation and passes them to the optimization module.

[0094] Based on the prediction results, combined with the dynamic constraints of the system, the optimization results are calculated through the nonlinear constrained optimization control algorithm, and the results are obtained after meeting the accuracy requirements of the KKT conditions. In each control cycle, after solving the optimization problem, the first value of the control sequence is extracted. As the control command for the current sampling moment, the valve opening and blower speed are dynamically adjusted to achieve precise control of the manifold pressure. The remaining parameters are recalculated in the next cycle to adapt to dynamic system changes. The control command is issued to the execution unit in real time, dynamically adjusting the manifold regulating valve position and blower speed, thereby achieving high-precision closed-loop control of the coke oven manifold pressure. Furthermore, during the control process, the system can sense changes in operating status in real time and adaptively respond to possible disturbances, operational switching, or model deviations, thereby improving system robustness and stability.

[0095] The coke oven gas collector pressure control method proposed in this invention, which uses a multivariable constrained optimization method based on a neural network, was verified using actual measured data from a coke oven gas collector system at a domestic coal coke chemical industry site. During the verification, the typical set value for the coke oven gas collector pressure was 150 Pa. The coke oven gas collector system was controlled using the nonlinear model predictive controller based on a neural network designed by this invention, and the control effect achieved was shown in Figure 2. Figure 4. It can be seen that under the control of the multivariable constrained optimization control algorithm of the present invention, the pressure of the four gas collecting pipes can accurately track the set value of 150Pa, reflecting the expected goal of multivariable control. In order to further demonstrate the tracking control performance of the designed controller, the set value of the gas collecting pipe pressure was modified from 150Pa to 149Pa at the 1000th sampling moment of the simulation verification. Figure 4 It can be seen from the response curve that the pressure of the four gas collecting pipes can quickly respond to the change of the set value and quickly and stably maintain around 149 Pa, indicating that the method of the present invention has good set value tracking ability. Figure 4 The local area of ​​the pressure fluctuation response of the gas collecting pipe system under dynamic interference is further magnified. As can be seen from the figure, despite the existence of interference, the control method designed by the present invention can still effectively control the gas collecting pipe pressure and maintain it within a smaller fluctuation range. Statistical analysis of the verified control results shows that near the set value of 150Pa, the proportion of the gas collecting pipe pressure within the fluctuation range of ±20Pa is as high as 99.98%. Compared with the existing technology, the existing coke oven gas collecting pipe pressure control usually adopts a DCS-based PID control system. Due to the dynamic disturbances caused by significant multivariable coupling, nonlinear characteristics and coke oven reversing heating in the coke oven gas collecting pipe system, traditional PID control is difficult to achieve high-precision control. According to the statistical analysis of the measured operating data of the existing DCS PID control system under the same working conditions, the proportion of the gas collecting pipe pressure within the fluctuation range of ±20Pa of 150Pa is only 44.5%, and 24.3% of the data fluctuations exceed the range of ±50Pa. The above comparative analysis demonstrates that, compared to traditional PID control methods, the neural network-based multivariable constrained optimization control method proposed in this paper significantly reduces the control fluctuations of the gas header pressure, significantly improving the system's control accuracy, dynamic performance, and robustness. This higher control accuracy helps improve the operational stability of the coke oven and reduce gas emissions, thereby indirectly helping to reduce energy consumption and positively impacting coke quality.

[0096] In summary, the coke oven gas collecting pipe pressure control method based on multivariable constrained optimization of the present invention provides a novel industrial control solution for the coking industry. The present invention first uses a neural network to construct a multivariable nonlinear prediction model of the coke oven gas collecting pipe system, and designs an efficient constrained optimization control architecture in combination with a sequential quadratic programming algorithm. On the basis of introducing constraints such as upper and lower limits of pressure and safe operating areas, precise dynamic control of pressure is achieved through rolling optimization. This method can significantly improve the accuracy and stability of system pressure tracking, reduce energy consumption fluctuations, and improve coke quality and production efficiency. The method provided by the present invention is suitable for pressure control in complex industrial environments, provides important technical support for the intelligent development of the coking industry, and has broad engineering application value.

[0097] The present invention can ensure that the generated data set comprehensively covers the status of the coke oven gas collecting pipe system in various actual operation scenarios, so that the subsequently trained NARX neural network model can fully learn the dynamic behavior pattern of the system under different working conditions, avoiding the situation where the model makes inaccurate or unpredictable predictions under certain working conditions due to data loss; the screening process can remove abnormal data, erroneous data and irrelevant data, and retain the data parts that are valuable for model training; preprocessing (such as data alignment, normalization and other operations) can eliminate the adverse effects of different variable dimensional differences on model training, making the data more standardized and neat, and meeting the requirements of the neural network model for input data, thereby improving The efficiency and accuracy of model training lay a good foundation for subsequent model training, verification and testing; the coke oven gas collecting pipe system has multi-variable strong coupling and nonlinear dynamic characteristics, and traditional linear models are difficult to accurately describe its complex behavior. As a nonlinear model, the NARX neural network can effectively capture and learn the complex coupling relationship between multiple variables in the system through its own nonlinear mapping ability. The trained and verified NARX neural network model can accurately describe the dynamic characteristics of the coke oven gas collecting pipe system under different working conditions, providing high-precision system model support for subsequent pressure control; by training and verifying on a data set containing data from different working conditions, NAR The NARX neural network model can learn the general laws of the system, rather than being limited to the behavior pattern under a specific working condition. When applied to the actual coke oven gas collector system control, the model can have good adaptability and generalization capabilities to unseen working conditions or system state changes, ensuring the reliability of prediction and control effects under various actual operating conditions. Because the NARX neural network model can accurately describe the multivariable coupling characteristics of the coke oven gas collector system, the multivariable nonlinear constrained optimization control algorithm designed based on this model can fully consider the mutual influence and coupling relationship between the various variables in the system, which enables the control algorithm to more accurately control multiple control variables (such as valve opening, blower Speed, etc.) to avoid system imbalance or poor control effects caused by single-variable control, thereby improving the performance and stability of gas collector pressure control. The operation of the coke oven gas collector system is subject to multiple constraints, including upper and lower pressure limits, valve opening range, blower speed range, and energy consumption limits. The optimization control algorithm based on the NARX neural network model can incorporate these complex constraints into the control strategy. By constructing and solving nonlinear constrained optimization problems, it ensures that the control results meet both the stable operation requirements of the system and the actual constraints of industrial production, achieving comprehensive optimization of the system in terms of safety, economy, and efficiency.Introducing constraints such as the upper and lower limits of the coke oven gas collecting pipe pressure, the safe operating area of ​​the valve opening and the blower speed can strictly limit the changes of the control variables and system states within a safe range, prevent equipment damage, production accidents or environmental pollution caused by conditions such as excessively high or low pressure, excessively large or small valve opening, and abnormal blower speed, and ensure the safe and stable operation of the coke oven system; considering energy consumption limit constraints and adding penalty items for control input constraints and gas collecting pipe pressure constraints to the optimization objective function can prompt the optimization algorithm to minimize energy waste and unnecessary consumption while meeting the system pressure control requirements. At the same time, the setting of penalty items can improve the optimization problem's resistance to measurement noise and model errors. The robustness of the coke oven system allows for minor violations of constraints in certain circumstances, but by imposing penalties on violations, the system can optimize its operation as much as possible within a safe and economical range, reduce production costs, and improve the economic benefits of the enterprise; by collecting the key parameters of the coke oven system in real time and inputting them into the NARX neural network model for prediction, the state prediction value of the coke oven gas collecting pipe system within a certain time range in the future (within the prediction time domain) can be obtained in advance, which enables the control system to understand the possible development trends and changes of the system in advance, such as the rising or falling trend of pressure, the possible pressure fluctuation amplitude, etc., thereby providing a basis for taking corresponding control measures in advance and enhancing the foresight and initiative of the control system; due to the coke oven system, the key parameters of the coke oven system are collected in real time and input into the NARX neural network model for prediction, which can obtain the state prediction value of the coke oven gas collecting pipe system within a certain time range in the future (within the prediction time domain) in advance, which enables the control system to understand the possible development trends and changes of the system in advance, such as the rising or falling trend of pressure, the possible pressure fluctuation amplitude, etc., thereby providing a basis for taking corresponding control measures in advance and enhancing the foresight and initiative of the control system; due to the coke oven system, the key parameters of the coke oven system are collected in real time and input into the NARX neural network model for prediction The operating conditions of the furnace system are dynamically changing. Real-time prediction can reflect the latest changes in the system status in a timely manner. The control system can quickly adjust the control strategy based on these real-time prediction results, so that the control parameters can adapt to the changes in the operating conditions in a timely manner, improve the system's response speed and adaptability to dynamic disturbances (such as pressure fluctuations caused by operations such as coal loading and coke pushing), ensure that the pressure of the coke oven gas collecting pipe is always stable within the set range, and maintain the normal progress of the coking process; the first value of the optimal control input increment sequence of the constraint optimization control output is applied to the control input in a timely manner, and the optimization results can be quickly converted into actual control actions, such as adjusting the valve opening, changing the blower speed, etc. This timely control parameter update can ensure the system During each control cycle, control is performed according to the current optimal solution, avoiding control lag and improving the real-time and effectiveness of gas collector pressure control. By continuously updating the control input value, the control system can continuously fine-tune and optimize the coke oven gas collector pressure, keeping the system in the optimal operating state. At the same time, because the optimization algorithm is based on system constraints and dynamic characteristics, each adjustment of the control parameters is carried out within a safe and reasonable range, thus ensuring the stability and continuity of the control process and avoiding system oscillation or instability caused by excessive or inappropriate control parameter adjustments. This ensures precise control of the coke oven gas collector pressure and the long-term stable operation of the system.

Claims

1. A coke oven gas collecting pipe pressure control method based on multivariable constrained optimization, characterized in that: include: S1. Collect, filter, and preprocess historical operating data of the coke oven gas collector system under different operating conditions to generate a data set for training, validating, and testing a nonlinear autoregressive NARX neural network model with exogenous input. Train and validate the nonlinear autoregressive NARX neural network model with exogenous input to obtain multivariable coupling characteristics describing the coke oven gas collector system. S2. Design a multivariable nonlinear constrained optimization control algorithm based on the NARX neural network model; S3. Introducing constraints that reflect the actual operation of the coke oven system into the nonlinear constrained optimization problem, including the upper and lower limits of the coke oven gas collector pressure, the safe operating range of valve opening and blower speed, and energy consumption limits. Penalty terms for control input constraints and gas collector pressure constraints are added to the optimization objective function to impose penalties for violations. S4. At the beginning of each control cycle, the key parameters of the coke oven system are collected in real time, and the key parameters of the coke oven system are predicted and calculated through the NARX neural network model to obtain the predicted value of the coke oven gas collecting pipe system state in the future prediction time domain. The control parameters are adjusted according to the optimization result of the constrained optimization control output, and the first value of the optimal control input increment sequence is applied to the control input. The control input value is updated to complete the control of the gas collecting pipe pressure.

2. The coke oven gas collecting pipe pressure control method based on multivariable constrained optimization according to claim 1, characterized in that: In S1, the preprocessing includes merging the data of the aligned time by taking the intersection of the time axis, and mapping the data to the normalized interval using the maximum and minimum normalization method to eliminate the influence of the magnitude differences of different variables on the training and prediction of the nonlinear autoregressive neural network model with exogenous input; The screening adopts a dynamic segmentation method, which is performed based on the recognition results of the coke oven operating conditions, and divides the coke oven gas collecting pipe system operation data into multiple subsets; The data set is obtained by merging the data with the aligned time and performing maximum and minimum value normalization preprocessing by taking the intersection of the time axis.

3. The coke oven gas collecting pipe pressure control method based on multivariable constrained optimization according to claim 1, characterized in that: In S2, the multivariable nonlinear constrained optimization control algorithm is as follows: In each control cycle, a nonlinear constrained optimization problem is constructed to predict the dynamic behavior of the coke oven gas collector system at multiple moments in the future. The optimization objective function is to minimize the sum of the squares of the tracking errors between the gas collector pressure and the set value in the future prediction time domain, the weighted sum of the squares of the control input incremental changes, and includes nonlinear system dynamic equation constraints based on the nonlinear autoregressive NARX neural network model with exogenous input. The control variables are dynamically adjusted through a rolling optimization method based on sequential quadratic programming, so that the coke oven gas collector system can follow the gas collector pressure set value while suppressing the pressure fluctuations caused by the carbonization chamber under different operating conditions.

4. The coke oven gas collecting pipe pressure control method based on multivariable constrained optimization according to claim 3, characterized in that: The multivariable nonlinear constrained optimization control algorithm combines the dynamic changes of the coke oven operating conditions with the optimization of the control variables, and the dynamic changes are predicted by a nonlinear autoregressive NARX neural network model with exogenous input.

5. The coke oven gas collecting pipe pressure control method based on multivariable constrained optimization according to claim 1, characterized in that: In S3, the upper / lower limit constraints of the coke oven gas collecting pipe pressure are set according to the coke oven production safety requirements and process requirements, and the valve opening and blower speed are introduced into the safety operation constraints. The addition of penalty items for control input constraints and manifold pressure constraints in the optimization objective function is achieved by introducing slack variables to allow small violations of hard constraints, and applying penalty items to the slack variables in the optimization objective function.

6. The coke oven gas collecting pipe pressure control method based on multivariable constrained optimization according to claim 1, characterized in that: In S4, the real-time collected gas collecting pipe pressure is fed back to the control system and compared with the set value of the gas collecting pipe pressure. The control parameters are adjusted according to the optimization results output by the multivariable nonlinear constrained optimization control algorithm to form a feedback loop.

7. The coke oven gas collecting pipe pressure control method based on multivariable constrained optimization according to claim 1, characterized in that: The nonlinear autoregressive NARX neural network model with exogenous input is obtained through offline training. The input of the nonlinear autoregressive NARX neural network model with exogenous input includes current and historical key parameters. The output of the nonlinear autoregressive NARX neural network model with exogenous input is the predicted value of the gas collecting pipe pressure in multiple future prediction time domains.

8. The coke oven gas collecting pipe pressure control method based on multivariable constrained optimization according to claim 3, characterized in that: The rolling optimization method of sequential quadratic programming is solved by converting the nonlinear constrained optimization problem into multiple quadratic programming sub-problems; The quadratic programming subproblem is solved using a quadratic programming solver. When solving the quadratic programming subproblem, the Gauss-Newton method is used to approximate the Hessian matrix.

9. The coke oven gas collecting pipe pressure control method based on multivariable constrained optimization according to claim 1, characterized in that: The optimization objective function includes minimizing the weighted sum of the square of the tracking error between the gas collecting pipe pressure and the set value in the future prediction time domain and the square of the incremental change of the control input. It also includes a restriction item on the control input. The restriction item is dynamically adjusted according to the coke oven gas production and pressure fluctuation limit demand to ensure the energy optimization operation of the coke oven system under the premise of stable gas collecting pipe pressure.

10. The coke oven gas collecting pipe pressure control method based on multivariable constrained optimization according to claim 1, characterized in that: The adjusting of the control parameters according to the optimization result of the constrained optimization control output includes applying the first value of the optimal control input increment sequence to the control input and updating the control input value by incremental increment.

Images