Air separation plant variable load control method and system based on mechanism-data fusion learning

Abstract

The invention belongs to the technical field of air separation plant intelligent control, and discloses an air separation plant intelligent variable load control method and system based on mechanism-data fusion deep learning. Comprising the steps of constructing a space division process dynamic model; training and generating a control sequence based on the operation strategy and the physical constraint; the control sequence is applied to the space division process dynamic model, and a model prediction value is generated; acquiring operation data of an actual system, comparing the operation data with a model predicted value, calculating deviation, importing historical data to carry out feedback correction and rolling optimization on the space division process dynamic model, and continuously correcting the space division process dynamic model until the deviation reaches a set threshold range; and inputting the corrected space division process dynamic model and historical data into a neural network model, training to obtain a mechanism-data fusion learning model, and connecting a control system through a data stream. The load of the air separation equipment is controlled based on a mechanism-data fusion learning model, and the problems of more energy waste and unstable product quality of variable load operation can be solved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent control technology for air separation equipment, specifically a mechanism-data fusion method and system for variable load control of air separation equipment. Background Technology

[0002] The air separation process involves multiple parts, including pressurization, purification, heat exchange, and separation. An air separation unit comprises several components such as compressors, molecular sieves, cooling towers, heat exchangers, and expanders. The entire system is quite complex, making full-process automation crucial for the air separation industry. Currently, domestic air separation plants mainly use distributed control systems (DCS) to monitor process parameters, perform basic control, and provide interlocking protection for equipment. However, the following problems exist in actual production: (1) Due to the strong coupling relationship between the front and rear units of the air separation unit, conventional control cannot effectively solve the coordinated control between the variables of the unit, which affects the stable operation of the unit and thus affects the purity and yield of the product.

[0003] (2) The production load of the air separation unit is often adjusted due to changes in the oxygen demand of downstream users. At present, the industry lacks an automatic load change control system for the entire air separation process, and the load change and operating condition operation are mainly performed manually by the operators. Because the load change operation lasts for a long time, oxygen release or supply shortages often occur, resulting in a lot of energy waste, large fluctuations in operating conditions, and unstable product quality; and it also requires a high level of experience from the operators and is very intensive. Summary of the Invention

[0004] This invention provides a mechanism-data fusion method and system for variable load control of air separation equipment, which solves the problems of long duration of variable load operation, frequent oxygen release or supply shortage, resulting in energy waste, large fluctuations in operating conditions, and unstable product quality.

[0005] To achieve the above objectives, the present invention provides the following technical solution: Mechanism-data fusion-based variable load control methods for air separation equipment include: Based on the parameters and operating conditions under varying loads, calculate and set parameters to construct a dynamic model of the air separation process; The fluctuation range is set based on the operating strategy and physical constraints, and the control sequence is generated by collecting data through the program for training. The control sequence is applied to the dynamic model of the air separation process, different variable load conditions are set, and model prediction values ​​are generated. Obtain the actual system's operating data, compare the operating data with the model's predicted values, calculate the deviation, and import historical data to perform feedback correction and rolling optimization on the air separation process dynamic model, obtaining the corrected air separation process dynamic model, until the deviation reaches the set threshold range. The modified dynamic model of the air separation process and historical data are input into the neural network model for training to obtain the mechanism-data fusion learning model. The load of the air separation equipment is controlled based on the mechanism-data fusion learning model.

[0006] Preferably, the steps of setting the fluctuation range based on the operating strategy and physical constraints, collecting data through a program for training, and generating the control sequence are as follows: A nonlinear model predictive control algorithm is adopted, based on the dynamic model of the space separation process. The fluctuation range is set according to the operation strategy and physical constraints. In each control cycle, an optimization problem including equipment constraints, quality indicators and energy consumption targets is solved. The global optimal solution is searched through particle swarm optimization to generate a control sequence.

[0007] Preferably, the nonlinear model predictive control algorithm is as follows: min

[0008] st

[0009]

[0010]

[0011]

[0012] Where N represents the prediction time domain length, w y The error weights of the predicted output are represented by y. k Let r(t) represent the predicted output, and r(t) represent the expected value of the predicted output. w u The error weight of the input quantity, u k This represents the predicted value of the input quantity, which is adjusted as rolling optimization and feedback correction occur. r (t) represents the expected value of the input, f and g correspond to the equations of the input and output respectively, and correspond to the control sequence and prediction result, u min u max y min y max The upper and lower limits of the input and output quantities are respectively used as constraints for the input and output. x0=x(t) means that the current actual measurement value is used as the initial value for prediction.

[0013] Preferably, the operation strategy and physical constraints are as follows: Based on the load increase or decrease requirements of the air separation equipment, slowly adjust the total intake air volume and monitor the changes in air compressor parameters, set upper and lower limits for load variation and air compressor parameter range, and avoid air compressor surge. Adjust the reflux ratio of the high-pressure and low-pressure fractionation towers and the pressure of the low-pressure fractionation tower; Adjust the oxygen output to change proportionally with the total intake, set a lower limit for product purity, and ensure stable parameters within the fractionation tower.

[0014] Preferably, the method for correcting the dynamic model of the air separation process is to use adaptive PID tuning technology to automatically adjust the PID parameters according to external factors and correct the actuator.

[0015] Preferably, the neural network model is composed of an input layer, multiple one-dimensional convolutional layers, pooling layers, fully connected layers, and an output layer in sequence.

[0016] Preferably, each convolutional layer uses the ReLU activation function and combines Batch Normalization and Dropout techniques to prevent overfitting. The pooling layer is responsible for downsampling and feature compression. The model hyperparameters are tuned through grid search or Bayesian optimization algorithms, and the mean absolute error and root mean square error on the validation set are used as the main evaluation metrics.

[0017] Preferably, during the training phase, the neural network model uses the Adam optimizer to dynamically adjust the learning rate. The initial learning rate is set to 1e-3, and during the training process, the learning rate decay or early stopping strategy is triggered according to the validation error curve. The entire network learns through multiple iterations until the validation set performance converges or the preset stopping condition is met, thus obtaining the mechanism-data fusion learning model. The final model is exported as a lightweight service format and deployed in the control and scheduling platform of the actual system.

[0018] The mechanism-data fusion-based intelligent control model for variable load air separation equipment includes: Physical model module: used to calculate and set parameters and build a dynamic model of the air separation process based on parameters and operating conditions under varying loads; Multivariable coupling and integration module: Based on operating strategies and physical constraints, it is used to set the fluctuation range, collect data through the program for training, and generate control sequences; Model prediction module: used to apply control sequences to the dynamic model of the air separation process, set different variable load conditions, and generate model prediction values; Feedback correction and rolling optimization module: Used to acquire the actual system's operating data, compare the operating data with the model's predicted values, calculate the deviation, and import historical data to perform feedback correction and rolling optimization on the air separation process dynamic model to obtain the corrected air separation process dynamic model until the deviation reaches the set threshold range. Model control module: It is used to input the corrected dynamic model of the air separation process and historical data into the neural network model, train it to obtain the mechanism-data fusion learning model, and control the load of the air separation equipment based on the mechanism-data fusion learning model.

[0019] Preferably, it further includes: Data Flow Module: Used to connect the data flow between the data fusion learning model and the DCS operating system of the spatial separation device.

[0020] Compared with existing technologies, the present invention has the following beneficial effects: The present invention provides a mechanism-data fusion method for variable load control of air separation equipment. By establishing a dynamic model of the air separation process, the variable with the greatest impact on the current load or efficiency is selected from the controlled and monitored variables in each major optimization control unit based on the dynamic model of the air separation process. The global optimal solution is searched based on the model prediction results, and a control sequence for a period of time is generated. On this basis, rolling optimization and updates of subsequent predictions are performed to ensure that the control strategy adapts to load changes in real time, thereby achieving effective control of the entire air separation system with a small amount of computation. Based on the control data, a mechanism-data fusion learning model is established to shorten the online calculation status feedback time and achieve effective control of variable load of air separation equipment. Attached Figure Description

[0021] Figure 1 This is a flowchart of the variable load control method for air separation equipment based on the mechanism and data fusion of this invention; Figure 2 This is a schematic diagram illustrating the working principle of the variable load control method for air separation equipment based on data fusion, as described in this embodiment of the invention. Figure 3 This is a schematic diagram illustrating the working principle of the neural network in an embodiment of the present invention. Figure 4 This is a block diagram of the variable load control system for air separation equipment based on the mechanism and data fusion of this invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0023] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0024] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0025] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0026] like Figure 1 As shown, this invention provides a mechanism-data fusion-based variable load control method for air separation equipment, comprising: S1 calculates and sets parameters based on the parameters and operating conditions under varying loads, and constructs a dynamic model of the air separation process. S2 sets the fluctuation range based on the operating strategy and physical constraints, collects data through the program for training, and generates control sequences. S3 applies the control sequence to the dynamic model of the air separation process, sets different variable load conditions, and generates model prediction values; S4 acquires the actual system's operating data, compares the operating data with the model's predicted values, calculates the deviation, and imports historical data to perform feedback correction and rolling optimization on the air separation process dynamic model, obtaining the corrected air separation process dynamic model until the deviation reaches the set threshold range. S5 inputs the corrected dynamic model of the air separation process and historical data into the neural network model, trains it to obtain a mechanism-data fusion learning model, and controls the load of the air separation equipment based on the mechanism-data fusion learning model.

[0027] Among them, the mechanism-data fusion learning model can connect with the data flow of space separation related equipment and operating system, realizing a rapid response from real-time data to the model and from the model to the control system interface.

[0028] The neural network model consists of an input layer, multiple one-dimensional convolutional layers, pooling layers, fully connected layers, and an output layer. Each convolutional layer uses the ReLU activation function and combines Batch Normalization and Dropout techniques to prevent overfitting. The pooling layers are responsible for downsampling and feature compression. The model hyperparameters are tuned using grid search or Bayesian optimization algorithms, and the mean absolute error and root mean square error on the validation set are used as the main evaluation metrics.

[0029] By establishing a dynamic model of the air separation process, the variable with the greatest impact on the current load or efficiency is selected from the controlled and monitored variables in each major optimization control unit based on the dynamic model. The global optimal solution is searched based on the model prediction results, and a control sequence for a period of time is generated. On this basis, rolling optimization and updates of subsequent predictions are performed to ensure that the control strategy adapts to load changes in real time, thereby achieving effective control of the entire air separation system with a small amount of computation. Based on the control data, a mechanism-data fusion learning model is established to shorten the online calculation status feedback time and achieve effective control of air separation equipment under varying loads.

[0030] like Figure 2 As shown, the detailed steps are as follows: The core working principle of this system revolves around three stages: "model prediction - rolling optimization - feedback correction," combining mechanistic models and data-driven technology to achieve fully automated variable load control. Model prediction phase: A physical model of the air separation process was built based on Aspen Plus to accurately simulate the thermodynamic behavior and fluid dynamics characteristics of key equipment such as compressors, heat exchangers, and fractionation towers. The model was driven by a Python program, inputting parameters under varying load conditions (such as flow rate, pressure, and temperature) and collecting output data (such as oxygen production, purity, and energy consumption) to construct a dynamic model of the air separation process that includes multivariate coupling relationships. This model not only reflects the physical mechanism of the air separation unit but is also calibrated using data from actual plants such as Hebei Aosen Steel Plant to ensure that the simulation results closely approximate engineering realities.

[0031] The rolling optimization stage employs a nonlinear model predictive control (MPC) algorithm. Based on the model prediction results, it solves an optimization problem encompassing equipment constraints, quality indicators, and energy consumption targets within each control cycle. Intelligent algorithms such as particle swarm optimization (PSO) are used to search for the global optimum, generating a control sequence for a future period (e.g., valve opening and compressor speed setpoints). Rolling optimization features "closed-loop rolling and finite-time domain" characteristics; each step executes only the optimal control action for the current moment, while simultaneously updating subsequent predictions based on the latest operating conditions, ensuring the control strategy adapts to load changes in real time.

[0032] The logic of the MPC algorithm is as follows: min

[0033] st Predictive Model

[0034] Model constraints

[0035]

[0036] Where N represents the prediction time domain length, i.e., the number of prediction steps. y This represents the error weights in the predicted output. k This represents the predicted output. r(t) represents the expected value of the predicted output. u This represents the error weight of the input quantity. k This represents the predicted value of the input quantity. This value is adjusted as rolling optimization and feedback correction occur. r (t) represents the expected value of the input, which is generally a constant. f and g correspond to the equations for the input and output, respectively, and represent the control sequence and the prediction result. min u max y min y max These correspond to the upper and lower limits of the input and output quantities, respectively, serving as constraints for the input and output. x0=x(t) represents using the current actual measured value as the initial value for prediction, which is the core of feedback correction and is used to achieve closed-loop control.

[0037] Feedback and correction stage: Real-time sensor measurements (such as oxygen flow rate and fractionation tower pressure) are compared with model predictions to calculate the deviation, which is then imported into historical data for dynamic compensation. Adaptive PID tuning technology is employed to automatically adjust PID parameters based on external factors such as ambient temperature and unit efficiency, fine-tuning actuators such as compressor inlet guide vanes, anti-surge valves, and throttle valves. The corrected control signal is fed back to the model prediction module, forming a "prediction-control-correction" closed loop to suppress multivariate coupling interference and model errors.

[0038] Data integration and lightweight deployment: By training Aspen simulation data and historical factory operating data through neural networks, a "mechanism-data fusion model" is constructed, shortening the online calculation and status feedback time of the MPC algorithm. Finally, the control strategy is embedded into the factory's existing DCS system through data flow connectivity, realizing integrated operation interface and automated control logic, supporting remote monitoring and real-time parameter optimization.

[0039] Neural network as attached Figure 3As shown, the neural network structure is suitable for temporal feature extraction using a convolutional neural network architecture. The network consists of an input layer, multiple one-dimensional convolutional layers, pooling layers, fully connected layers, and an output layer. Each convolutional layer uses the ReLU activation function and incorporates Batch Normalization and Dropout techniques to prevent overfitting; the pooling layers are responsible for downsampling and feature compression. Model hyperparameters (such as the number of convolutional kernels, kernel width, learning rate, batch size, etc.) are tuned using grid search or Bayesian optimization algorithms, with the mean absolute error (MAE) and root mean square error (RMSE) on the validation set as the main evaluation metrics.

[0040] During the training phase, the Adam optimizer is used to dynamically adjust the learning rate, with an initial learning rate set to 1e-3. During training, learning rate decay or early stopping strategies are triggered based on the validation error curve. The entire network learns through multiple iterations until the validation set performance converges or a preset stopping condition is met. After training, the final model is exported as a lightweight service format for rapid deployment and invocation in the control and scheduling platform. This allows for model replacement of the Aspen simulator, accelerating the search for the global optimum and effectively avoiding interface blocking and lag issues.

[0041] To address the problem that traditional DCS control systems cannot effectively handle the coupling and correlation between variables of various devices, this invention selects the variable with the greatest impact on the current load or efficiency from the controlled and monitored variables in each major optimization control unit. It introduces advanced MPC and PSO algorithms to search for the global optimum based on model prediction results and generates a control sequence over a period of time. Rolling optimization updates subsequent predictions are then performed to ensure that the control strategy adapts to load changes in real time, thereby achieving effective control of the entire air separation system with relatively small computational load. To address the problems of oxygen release or supply shortages caused by traditional manual regulation, which lead to energy waste, large fluctuations in operating conditions, and unstable product quality, this invention realizes automatic load and operating condition control of air separation units. That is, based on changes in gas production demand, it automatically and rationally adjusts the set values ​​of relevant control loops and regulating valves to achieve rapid and accurate load and operating condition adjustments, reduce human interference, and lower labor intensity.

[0042] like Figure 4 As shown, the present invention also provides a mechanism-data fusion-based intelligent control system for variable load air separation equipment, comprising: Physical model module: used to calculate and set parameters and build a dynamic model of the air separation process based on parameters and operating conditions under varying loads; Multivariable coupling and integration module: Based on operating strategies and physical constraints, it is used to set the fluctuation range, collect data through the program for training, and generate control sequences; Model prediction module: used to apply control sequences to the dynamic model of the air separation process, set different variable load conditions, and generate model prediction values; Feedback correction and rolling optimization module: Used to acquire the actual system's operating data, compare the operating data with the model's predicted values, calculate the deviation, and import historical data to perform feedback correction and rolling optimization on the air separation process dynamic model to obtain the corrected air separation process dynamic model until the deviation reaches the set threshold range. Model control module: It is used to input the corrected dynamic model of the air separation process and historical data into the neural network model, train it to obtain the mechanism-data fusion learning model, and control the load of the air separation equipment based on the mechanism-data fusion learning model.

[0043] Also includes: Data Flow Module: Used to connect the data flow of the mechanism-data fusion learning model with the spatial separation related equipment and operating system, enabling rapid response from real-time data to the model and from the model to the control system interface.

[0044] Although embodiments of the present invention have been described above in conjunction with the accompanying drawings, the present invention is not limited to the specific embodiments and application fields described above. The specific embodiments described above are merely illustrative and instructive, and not restrictive. Those skilled in the art, guided by the specification, can make many other modifications without departing from the scope of the claims of the present invention, and all of these modifications are within the scope of protection of the present invention.

Claims

1. A variable load control method for air separation equipment based on mechanism-data fusion deep learning, characterized in that, include: Based on the parameters and operating conditions under varying loads, calculate and set parameters to construct a dynamic model of the air separation process; The fluctuation range is set based on the operating strategy and physical constraints, and the control sequence is generated by collecting data through the program for training. The control sequence is applied to the dynamic model of the air separation process, different variable load conditions are set, and model prediction values ​​are generated. Obtain the actual system's operating data, compare the operating data with the model's predicted values, calculate the deviation, and import historical data to perform feedback correction and rolling optimization on the air separation process dynamic model, obtaining the corrected air separation process dynamic model, until the deviation reaches the set threshold range. The modified dynamic model of the air separation process and historical data are input into the neural network model for training to obtain the mechanism-data fusion learning model. The load of the air separation equipment is controlled based on the mechanism-data fusion learning model.

2. The mechanism-data fusion-based variable load control method for air separation equipment according to claim 1, characterized in that, The specific steps for generating control sequences, based on setting fluctuation ranges according to operational strategies and physical constraints, collecting data through a program for training, and then generating control sequences are as follows: A nonlinear model predictive control algorithm is adopted, based on the dynamic model of the space separation process. The fluctuation range is set according to the operation strategy and physical constraints. In each control cycle, an optimization problem including equipment constraints, quality indicators and energy consumption targets is solved. The global optimal solution is searched through particle swarm optimization to generate a control sequence.

3. The mechanism-data fusion-based variable load control method for air separation equipment according to claim 2, characterized in that, The nonlinear model predictive control algorithm is as follows: min s.t Where N represents the prediction time domain length, w y The error weights of the predicted output are represented by y. k Let r(t) represent the predicted output, and r(t) represent the expected value of the predicted output. w u The error weight of the input quantity, u k This represents the predicted value of the input quantity, which is adjusted as rolling optimization and feedback correction occur. r (t) represents the expected value of the input, f and g correspond to the equations of the input and output respectively, and correspond to the control sequence and prediction result, u min u max y min y max The upper and lower limits of the input and output quantities are respectively used as constraints for the input and output. x0=x(t) means that the current actual measurement value is used as the initial value for prediction.

4. The mechanism-data fusion-based variable load control method for air separation equipment according to claim 1, characterized in that, The operational strategy and physical constraints are as follows: Based on the load increase or decrease requirements of the air separation equipment, slowly adjust the total intake air volume and monitor the changes in air compressor parameters, set upper and lower limits for load variation and air compressor parameter range, and avoid air compressor surge. Adjust the reflux ratio of the high-pressure and low-pressure fractionation towers and the pressure of the low-pressure fractionation tower; Adjust the oxygen output to change proportionally with the total intake, set a lower limit for product purity, and ensure stable parameters within the fractionation tower.

5. The mechanism-data fusion-based variable load control method for air separation equipment according to claim 1, characterized in that, The method for correcting the dynamic model of the air separation process is to use adaptive PID tuning technology to automatically adjust the PID parameters according to external factors and correct the actuator.

6. The mechanism-data fusion-based variable load control method for air separation equipment according to claim 1, characterized in that, The neural network model consists of an input layer, multiple one-dimensional convolutional layers, pooling layers, fully connected layers, and an output layer, in sequence.

7. The mechanism-data fusion method for variable load control of air separation equipment according to claim 5, characterized in that, Each convolutional layer employs the ReLU activation function and combines Batch Normalization and Dropout techniques to prevent overfitting. Pooling layers are responsible for downsampling and feature compression. Model hyperparameters are tuned using grid search or Bayesian optimization algorithms, with mean absolute error and root mean square error on the validation set as the main evaluation metrics.

8. The mechanism-data fusion method for variable load control of air separation equipment according to claim 1, characterized in that, During the training phase, the neural network model uses the Adam optimizer to dynamically adjust the learning rate. The initial learning rate is set to 1e-3. During the training process, the learning rate decay or early stopping strategy is triggered according to the validation error curve. The entire network learns through multiple iterations until the validation set performance converges or the preset stopping condition is met, thus obtaining the mechanism-data fusion learning model. The final model is exported as a lightweight service format and deployed in the control and scheduling platform of the actual space division system.

9. A mechanism-data fusion-based intelligent control system for variable load air separation equipment, characterized in that, include: Physical model module: used to calculate and set parameters and build a dynamic model of the air separation process based on parameters and operating conditions under varying loads; Multivariable coupling and integration module: Based on operating strategies and physical constraints, it is used to set the fluctuation range, collect data through the program for training, and generate control sequences; Model prediction module: used to apply control sequences to the dynamic model of the air separation process, set different variable load conditions, and generate model prediction values; Feedback correction and rolling optimization module: Used to acquire the actual system's operating data, compare the operating data with the model's predicted values, calculate the deviation, and import historical data to perform feedback correction and rolling optimization on the air separation process dynamic model to obtain the corrected air separation process dynamic model until the deviation reaches the set threshold range. Model control module: It is used to input the corrected dynamic model of the air separation process and historical data into the neural network model, train it to obtain the mechanism-data fusion learning model, and control the load of the air separation equipment based on the mechanism-data fusion learning model.

10. The mechanism-data fusion intelligent control system for air separation equipment with variable load according to claim 9, characterized in that, Also includes: Data Flow Module: Used to connect the data flow between the data fusion learning model and the DCS operating system of the spatial separation device.

Images