Machine learning optimization simulated moving bed multi-tower temperature swing adsorption process and system for C4F6 purification

Abstract

The invention discloses a machine learning optimization simulated moving bed multi-tower temperature swing adsorption process and system for C4F6 purification. The simulated moving bed is used for adsorbing various halogenated hydrocarbon impurities in different regions, so that the removal efficiency of trace impurity gas in C4F6 is remarkably improved. Furthermore, through collection of preliminary operation data, a machine learning model is established, and optimized process parameters are generated in combination with an NSGA-II optimization algorithm. The process is continuously optimized through iteration, so that the prediction precision of the machine learning model is gradually improved, and meanwhile, better process parameters are obtained, so that the performance of the simulated moving bed temperature swing adsorption process is improved, the impurity content in the product gas is effectively reduced, and stable production of high-purity C4F6 is realized. According to the invention, physical adsorption enhancement and an intelligent optimization algorithm are deeply fused, and the novel C4F6 purification process and the process optimization method thereof can effectively reduce the energy consumption in the electronic special gas purification process and improve the product yield.

Description

Technical Field

[0001] The present application belongs to the field of chemical separation technology, and specifically relates to a hexafluorobutadiene machine learning optimized simulated moving bed multi-tower temperature swing adsorption process system and method. Background Art

[0002] High-purity electronic specialty gases (ESGs) are key upstream raw materials for national strategic industries such as semiconductor integrated circuits and are known as the "blood" of the electronics industry. Hexafluorobutadiene (C4F6) ranks among the top three in the global ESG market, primarily used in etching processes. Currently, the production process for C4F6 primarily uses halogenated olefins. The process is lengthy and complex, with a wide variety of impurities. Subsequent conventional distillation cannot meet the required performance requirements for ESGs.

[0003] Simulated Moving Bed (SMB) technology overcomes the long non-adsorption cycle, low throughput per unit time, and low productivity of fixed beds, and is widely used in separation fields such as petrochemicals, the food industry, chiral pharmaceuticals, and natural products. Patent CN111100683A discloses a method for separating C10-C15 alkenes and alkanes using simulated moving bed technology. The desorbent is a mixture of hexane and hexene, but the system product produced by this process has a low purity of only 90% to 95%. CN115721965A discloses a simulated moving bed technology for simultaneously separating three components from a feedstock, but the desorbent must be removed by distillation, which consumes a lot of energy.

[0004] Simulated moving bed (SMB) is a complex continuous preparative chromatography technique. Due to the coupled influences between its operating parameters (including flow rates and switching times between different zones), process optimization is challenging. Therefore, it is necessary to simulate and optimize the SMB process to determine reasonable and efficient operating conditions. Traditional SMB parameter optimization requires solving the mechanistic model described by partial differential equations thousands of times, which is time-consuming. Hu Rong et al. (Hu Rong, Yang Minglei, and Qian Feng. "Application of a Multi-Objective Teaching Optimization Algorithm in the Optimization of Xylene Adsorption Separation Process." Journal of Chemical Industry and Engineering (China) 66.1 (2015): 326-332) applied a multi-objective teaching optimization algorithm to the separation of p-xylene, with the objective functions of maximizing p-xylene recovery, minimizing eluent consumption, and simultaneously maximizing recovery and purity. The algorithm demonstrated good convergence and distribution. However, the number of SMB mathematical model solutions did not significantly decrease, and the time consumption remained long. Li et al. (Li, Yan, et al. "Multi-objective optimization of sequential simulated movingbed for the purification of xylo-oligosaccharides." Chemical Engineering Science 211(2020):115279.) used NSGA-II to solve the optimization problems of maximizing purity and yield, maximizing yield and minimizing eluent consumption in the process of separating and purifying oligoxylose from industrial syrup using sequential SMB. Similarly, since the NSGA-II optimization algorithm often involves more than a hundred iterations in the optimization process, the calculation of each iteration still returns to the solution of the mathematical model, and the computational pressure has not been reduced. Many researchers have also tried various optimization algorithms, but none of them have escaped the difficulty and time-consuming problems involved in solving strict mathematical models.

[0005] Yao et al. (Yao, Chuanyi, et al. "Using a machine learning model for the optimal design of simulated moving bed processes and its application toseparate rebaudioside A and stevioside." Journal of Chemical Technology&Biotechnology 96.9(2021):2558-2568.) used machine learning algorithms to establish random forest models and deep neural network models for the separation processes of steviol glycoside A and steviol glycoside, respectively. The results showed that the average absolute errors of the two models were less than 0.19% and 0.08%, respectively, and both models were able to predict the purity of the product well. It proves that the machine learning model is a good alternative model, which greatly improves the computing time required for the optimization process. The construction of its data set still requires a lot of time to solve the mathematical model.

[0006] The method of purifying C4F6 in the prior art includes:

[0007] Patent US6544319B1 discloses a method for purifying hexafluorobutadiene using 5A molecular sieve. After treatment by this method, the purity of hexafluorobutadiene can be increased from 99.90% to 99.96%. However, during the adsorption purification process, exothermic conditions occur, and the temperature rises rapidly, resulting in the production of approximately 0.1 wt% of hexafluoro-2-butyne, which affects production yield and purity.

[0008] Patent JP2004339187A discloses a method for purifying hexafluorobutadiene using molecular sieves and activated carbon. This method involves passing the hexafluorobutadiene feedstock through the molecular sieve and activated carbon sequentially, reducing the H2O and HF contents to below 1 ppm. While this method can reduce the H2O and HF contents, it does not remove other fluorocarbon impurities. Furthermore, the activated carbon adsorbs a large amount of the feedstock, reducing the hexafluorobutadiene yield, making it unsuitable for industrial production.

[0009] Patent CN118812323A discloses the use of an extractive distillation process to remove alcohols, water, and hexafluorobutene impurities from C4F6. The extractant is selected from acetonitrile, butanone, and dimethylformamide. After purification, the water content is less than 10ppm, the hexafluorobutene impurity content is less than 10ppm, and the alcohol impurity content is less than 10ppm. However, the process is relatively complex. Patents CN111138240B and CN115340437A also use extractive distillation processes, using different extractants for different impurities. However, the extraction and distillation are coupled, which is relatively complex and has the same problem of high distillation energy consumption.

[0010] Patent CN118831340A discloses an ultra-gravity distillation purification device and method for C4F6 purification. Using centrifugal force generated by a motor rotating a rotating disk, instead of gravity, the device rapidly achieves equilibrium between the gas and liquid phases. The resulting C4F6 product has a purity exceeding 99.99%. However, this method is still essentially a distillation method, resulting in high energy consumption.

[0011] Patent CN117599443A discloses a C4F6 purification device and method. The crude gas can be purified by deacidification treatment, first adsorption treatment, second adsorption treatment, first distillation treatment and second distillation treatment in sequence to achieve a purity of more than 99.99%. However, due to the multi-stage adsorption and distillation involved, the process is long, the loss is high, and the yield is low.

[0012] In view of the problems mentioned above with the simulated moving bed and the fact that the existing technology has not yet addressed the removal of structurally similar halogenated hydrocarbon impurities in C4F6 crude gas, it is necessary to seek a method for purifying C4F6 crude gas containing multiple structurally similar halogenated hydrocarbon impurities and to propose a machine learning framework to quickly optimize the simulated moving bed. Summary of the Invention

[0013] The present invention aims to provide a machine learning-optimized simulated moving bed (SMB) multi-tower temperature swing adsorption (TSA) process and system for C₄F₆ purification. This process can increase the C₄F₆ concentration in crude gas from over 96% to over 99.99%. This method utilizes molecular sieve materials and activated carbon as adsorbents to design and implement a multi-tower SMBTTA process for electronic specialty gas purification. This process is applicable to crude C₄F₆ gas containing structurally similar halogenated hydrocarbon impurities.

[0014] The specific technical solutions of the present invention are as follows:

[0015] A machine learning optimized simulated moving bed multi-tower temperature swing adsorption system for C4F6 purification, comprising a simulated moving bed system, a machine learning algorithm model, and an NSGA-II optimization algorithm;

[0016] The simulated moving bed system includes six adsorption towers, each of which is equipped with a vacuum pump, a crude gas inlet valve, an adsorption tower drain valve, a discharge valve, a steam inlet valve, a steam outlet valve, a cooling water inlet valve, and a cooling water outlet valve; the crude gas inlet valve is connected to the C4F6 crude gas; the adsorption tower drain valve is connected to the vacuum pump and then input into the tail gas collection device; the discharge valve is connected to the product gas storage device; the steam inlet valve is connected to the high-temperature steam; and the cooling water inlet valve is connected to the cooling water;

[0017] The simulated moving bed system records the process parameters of each operation, the time of each valve switching, and the purity, recovery rate, and content of various impurities in the product gas of the C4F6 process parameters and the valve switching sequence, which together constitute a set of samples; multiple sets of samples are organized into a data set; the process parameters include: a specified pressure P1 in the adsorption tower performing a depressurization operation, a specified desorption temperature T1 in the adsorption tower performing a heating operation, a specified adsorption pressure P2 in the adsorption tower performing a pressurization operation, and the switching sequence ti (i=1, 2, 3, 4, 5, 6) of each valve;

[0018] The data set is used to train a machine learning algorithm model, and the trained machine learning algorithm model is combined with the NSGA-II optimization algorithm to infer multiple sets of better process parameters and valve switching timings; the process parameters and timings are applied to the simulated moving bed multi-tower temperature swing adsorption process system to obtain more samples, which are merged with the original data set to form a new data set; this is repeated multiple times until the performance of the simulated moving bed multi-tower temperature swing adsorption process system no longer improves further.

[0019] In the above system, each single tower in the simulated moving bed multi-tower temperature swing adsorption process system is filled with two adsorbents at the same time; the two adsorbents are molecular sieve and activated carbon; the filling ratio of the molecular sieve and activated carbon is in the range of 1:3 to 3:1.

[0020] In the above system, the adsorption tower is a shell and tube device, steam and cooling water flow through the tube layer, and the shell layer is filled with adsorbent.

[0021] In the above system, the C4F6 content in the C4F6 crude gas is 96%.

[0022] A machine learning optimized simulated moving bed multi-tower temperature swing adsorption process for C4F6 purification using the above system comprises the following steps:

[0023] (1) Two adsorption towers perform adsorption operations. The feed valves connecting the bottoms of all adsorption towers performing adsorption operations and the crude gas storage device are opened, the discharge valves connecting the tops of all adsorption towers performing adsorption operations and the product gas storage device are opened, the other valves connected to all adsorption towers performing adsorption operations are closed, and the vacuum pumps of the adsorption towers are turned off. During this process, C4F6 crude gas enters the adsorption tower from the bottom of the tower, impurities are adsorbed by the adsorbent bed, and the resulting C4F6 pure gas is discharged from the top of the tower and directed to the C4F6 gas storage device.

[0024] (2) At the same time, an adsorption tower performs a pressure reduction operation. The drain valve connecting the bottom of the adsorption tower and the vacuum pump is opened, the exhaust valve between the vacuum pump of the adsorption tower and the tail gas storage device is opened, the vacuum pump of the adsorption tower is opened, and other valves connected to the adsorption tower are closed. During this process, the pressure of the adsorption tower gradually decreases from the adsorption set pressure to the specified pressure, and the gas in the adsorption tower is discharged from the bottom of the tower together with a small amount of halogenated hydrocarbon impurities desorbed from the bed and directed to the tail gas storage device.

[0025] (3) At the same time, an adsorption tower performs a heating operation, opens a drain valve connecting the bottom of the adsorption tower performing the heating operation and the vacuum pump, opens an exhaust valve between the vacuum pump of the adsorption tower performing the heating operation and the tail gas storage device, opens the vacuum pump of the adsorption tower performing the heating operation, opens the steam inlet valve between the top of the adsorption tower performing the heating operation and the steam inlet line, opens the steam outlet valve between the bottom of the adsorption tower performing the heating operation and the steam outlet line, and closes other valves connected to the adsorption tower performing the heating operation. During this process, the temperature of the adsorption tower gradually rises to the specified desorption temperature and is maintained there, and a large amount of halogenated hydrocarbon impurities are desorbed from the bed of the adsorption tower, discharged from the bottom of the tower and directed to the tail gas storage device.

[0026] (4) At the same time, an adsorption tower performs a cooling operation. The drain valve connecting the bottom of the adsorption tower performing the cooling operation and the vacuum pump is opened, the exhaust valve between the vacuum pump of the adsorption tower performing the cooling operation and the tail gas storage device is opened, the vacuum pump of the adsorption tower performing the cooling operation is opened, the inlet valve between the bottom of the adsorption tower performing the cooling operation and the cooling water inlet line is opened, the outlet valve between the top of the adsorption tower performing the cooling operation and the cooling water outlet line is opened, and the other valves connected to the adsorption tower performing the cooling operation are closed. During this process, the temperature of the adsorption towers gradually drops to room temperature, and a small amount of residual gas in the tower is discharged from the bottom of the tower and directed to the tail gas storage device.

[0027] (5) Simultaneously, a pressurized adsorption tower performs a pressurization operation. The feed valve connecting the bottom of the adsorption tower performing the pressurized adsorption operation and the crude gas storage device is opened. The discharge valves connecting the tops of all adsorption towers performing the pressurized adsorption operation and the product gas storage device are opened. Other valves connected to the adsorption tower performing the pressurized adsorption operation are closed, and the vacuum pump of the adsorption tower is turned off. During this process, the crude C4F6 gas enters the adsorption tower from the bottom of the tower. The pressure in the adsorption tower gradually rises to the specified adsorption pressure. Impurities are adsorbed by the adsorbent bed, and the resulting pure C4F6 gas is discharged from the top of the tower and directed to the C4F6 gas storage device.

[0028] (6) Record the process parameters of each operation, the time of each valve switching, and the C4F6 purity, recovery rate, and the content of various impurities in the product gas when the process parameters and the valve switching timing match, which together constitute a set of samples. Organize multiple sets of these samples into a data set.

[0029] (7) The dataset obtained in step (6) is used to train a machine learning model. The trained machine learning model is combined with an optimization algorithm to infer multiple sets of more optimal process parameters and valve switching timings. The process parameters and timings are applied to the simulated moving bed multi-tower temperature swing adsorption process system, and step (6) is repeated to obtain more samples, which are merged with the original dataset to form a new dataset.

[0030] (8) Repeat steps (6) and (7) several times until the performance of the simulated moving bed multi-tower temperature swing adsorption process system is no longer improved.

[0031] Preferably, each single tower in the simulated moving bed multi-tower temperature swing adsorption process system is filled with two adsorbents at the same time.

[0032] Preferably, the two adsorbents are ZSM-5 molecular sieve and activated carbon.

[0033] Preferably, the filling ratio of ZSM-5 molecular sieve and activated carbon in the single tower is in the range of 1:3 to 3:1.

[0034] Preferably, the adsorption tower is a shell and tube device, with steam and cooling water flowing through the tube layer and the shell layer filled with adsorbent.

[0035] Preferably, the C4F6 content in the C4F6 crude gas is 96%.

[0036] Preferably, the halogenated hydrocarbon impurities are 1,1,2,3,4,4,4-heptafluoro-1-butene, 1,1,1,3,4,4,4-heptafluoro-1-butene, 1,1,3,4,4-pentafluoro-1,3-butadiene, and 1,1,4,4-tetrafluoro-1,3-butadiene.

[0037] Preferably, in step (2), the pressure is reduced to a specified pressure of 0.1 bar to 0.2 bar.

[0038] Preferably, in step (3), the heating to a specified temperature is 100°C to 150°C.

[0039] Preferably, in step (5), the pressurization is performed to a specified pressure of 5 bar to 7 bar.

[0040] Preferably, the process parameters in step (6) include the specified pressure P1 in step (2), the specified temperature T1 in step (3), the specified pressure P2 in step (5), the valve switching timing t in step (6), and the specified pressure P1 in step (2). i (i=1,2,3,4,5,6).

[0041] Preferably, the multiple groups of samples in step (6) are 10 to 50 groups, preferably 20 to 30 groups.

[0042] Preferably, the machine learning model in step (7) is a neural network model, the number of its hidden layers is 1, the number of neurons in the hidden layer is 10 to 30, the activation function is the ReLU function, the optimizer is the Adam optimizer, and the loss function is the mean square error (MSE).

[0043] Preferably, the optimization algorithm in step (7) is a non-dominated sorting genetic algorithm (NSGA-II).

[0044] Preferably, the number of groups of samples in step (7) is 20 to 30 groups.

[0045] Preferably, the performance in step (8) is no longer further improved, and the content changes of 1,1,2,3,4,4,4-heptafluoro-1-butene, 1,1,1,3,4,4,4-heptafluoro-1-butene, 1,1,3,4,4-pentafluoro-1,3-butadiene, and 1,1,4,4-tetrafluoro-1,3-butadiene in the product gas before and after one cycle are all less than 1 ppm, and the recovery rate changes are less than 1%.

[0046] Compared with the prior art, the present invention has the following advantages:

[0047] One advantage of the present invention is that different adsorbents are filled in each single tower at the same time, and multiple impurities in the C4F6 crude gas can be adsorbed simultaneously, and multiple halogenated hydrocarbon impurities with similar structures can be removed at the same time.

[0048] One advantage of the present invention is that it provides a multi-tower temperature swing adsorption process for purifying C4F6 by simulating moving bed operation. In the adsorption operation, the present invention adopts a multi-tower series connection mode and simulates the moving bed operation mode, so that the adsorbent phase and the gas mobile phase in the tower form a countercurrent flow. Compared with the conventional pressure swing adsorption process, the utilization rate of the adsorbent is greatly improved.

[0049] One advantage of the present invention is that it uses a machine learning algorithm to optimize process parameters to overcome the obstacles to process optimization caused by multi-tower processes. In addition, this iterative training method has low data requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 Schematic diagram of machine learning optimized simulated moving bed multi-tower temperature swing adsorption process system.

[0051] Figure 2 This is a flow chart of the simulated moving bed multi-tower temperature swing adsorption process described in the present invention.

[0052] Description of symbols in the figure: P, vacuum pump; VH i , steam inlet valve; VHb i , steam outlet valve; VC i , cooling water inlet valve; VCb i , cooling water outlet valve; VF i , crude gas inlet valve; VP i , product gas outlet valve; VW i , adsorption tower drain valve; VWb i , exhaust valve.

[0053] Figure 3 Schematic diagram of the cross-sectional structure of the adsorption tower.

[0054] Figure 4 is the average regression coefficient R of the neural network model of Example 1 for the process performance index 2 As the number of iterations increases.

[0055] Figure 5 is the average regression coefficient R of the neural network model of Example 2 for the process performance index 2 As the number of iterations increases.

[0056] Figure 6 is the average regression coefficient R of the neural network model of Example 3 for the process performance index 2 As the number of iterations increases.

[0057] Figure 7 is the average regression coefficient R of the neural network model of Example 4 for the process performance index 2 As the number of iterations increases.

[0058] Figure 8 is the average regression coefficient R of the proportional neural network model on the process performance index 2 As the number of iterations increases.

[0059] Figure 9is the average regression coefficient R of the proportional neural network model on the process performance index 2 As the number of iterations increases.

[0060] Figure 10 is the average regression coefficient R of the proportional neural network model on the process performance index 2 As the number of iterations increases.

[0061] Figure 11 is the average regression coefficient R of the proportional neural network model on the process performance index 2 As the number of iterations increases. DETAILED DESCRIPTION

[0062] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but the present invention is not limited thereto.

[0063] The following examples use the following process steps:

[0064] (1) Two adsorption towers perform adsorption operation, open the feed valves connecting the bottom of all adsorption towers performing adsorption operation and the crude gas storage device, open the discharge valves connecting the top of all adsorption towers performing adsorption operation and the product gas storage device, close other valves connected to all adsorption towers performing adsorption operation, and turn off the vacuum pumps of the adsorption towers; during this process, C4F6 crude gas enters the adsorption tower from the bottom of the tower, impurities are adsorbed by the adsorbent bed, and the obtained C4F6 pure gas is discharged from the top of the tower and directed to the C4F6 gas storage device;

[0065] (2) At the same time, an adsorption tower performs a depressurization operation, opens a drain valve connected between the bottom of the adsorption tower performing the depressurization operation and the vacuum pump, opens an exhaust valve between the vacuum pump of the adsorption tower performing the depressurization operation and the tail gas storage device, opens the vacuum pump of the adsorption tower performing the depressurization operation, and closes other valves connected to the adsorption tower performing the depressurization operation; during this process, the pressure of the adsorption tower gradually decreases from the adsorption set pressure to the specified pressure, and the gas in the adsorption tower is discharged from the bottom of the tower together with a small amount of halogenated hydrocarbon impurities desorbed from the bed layer and directed to the tail gas storage device;

[0066] (3) At the same time, an adsorption tower performs a heating operation, opens a drain valve connected between the bottom of the adsorption tower performing the heating operation and the vacuum pump, opens an exhaust valve between the vacuum pump of the adsorption tower performing the heating operation and the tail gas storage device, opens the vacuum pump of the adsorption tower performing the heating operation, opens a steam inlet valve between the top of the adsorption tower performing the heating operation and the steam inlet line, opens a steam outlet valve between the bottom of the adsorption tower performing the heating operation and the steam outlet line, and closes other valves connected to the adsorption tower performing the heating operation; during this process, the temperature of the adsorption tower gradually rises to the specified desorption temperature and is maintained, and a large amount of halogenated hydrocarbon impurities are desorbed from the bed of the adsorption tower, discharged from the bottom of the tower and directed to the tail gas storage device;

[0067] (4) At the same time, an adsorption tower performs a cooling operation, opens a drain valve connected between the bottom of the adsorption tower performing the cooling operation and the vacuum pump, opens an exhaust valve between the vacuum pump of the adsorption tower performing the cooling operation and the tail gas storage device, opens the vacuum pump of the adsorption tower performing the cooling operation, opens the inlet valve between the bottom of the adsorption tower performing the cooling operation and the cooling water inlet line, opens the outlet valve between the top of the adsorption tower performing the cooling operation and the cooling water outlet line, and closes other valves connected to the adsorption tower performing the cooling operation. During this process, the temperature of the adsorption tower gradually drops to room temperature, and a small amount of residual gas in the tower is discharged from the bottom of the tower and directed to the tail gas storage device;

[0068] (5) At the same time, a pressurized adsorption tower performs a pressurization operation, opens the feed valve connected between the bottom of the adsorption tower performing the pressurized adsorption operation and the crude gas storage device, opens all the discharge valves connected to the top of the adsorption tower performing the pressurized adsorption operation and the product gas storage device, closes the other valves connected to the adsorption tower performing the pressurization operation, and turns off the vacuum pump of the adsorption tower. During this process, the C4F6 crude gas enters the adsorption tower from the bottom of the tower, the pressure in the adsorption tower gradually rises to the specified adsorption pressure, the impurities are adsorbed by the adsorbent bed, and the obtained C4F6 pure gas is discharged from the top of the tower and directed to the C4F6 gas storage device;

[0069] (6) Recording the process parameters of each operation, the time of each valve switching, and the purity of C4F6, the recovery rate, and the content of various impurities in the product gas that match the process parameters and the valve switching timing, together forming a set of samples; arranging multiple sets of the samples into a data set;

[0070] (7) using the data set obtained in step (6) to train a machine learning model, and combining the trained machine learning model with an optimization algorithm to infer multiple sets of more optimal process parameters and valve switching timings; applying the process parameters and timings to the simulated moving bed multi-tower temperature swing adsorption process system, repeating step (6) to obtain more samples, and merging them with the original data set to form a new data set;

[0071] (8) Repeat steps (6) and (7) several times until the performance of the simulated moving bed multi-tower temperature swing adsorption process system is no longer improved.

[0072] Example 1

[0073] according to Figure 1 To build this system, a simulated moving bed multi-tower temperature swing adsorption process system, a machine learning model, and an NSGA-II optimization algorithm are used. Figure 2 As shown, it includes six adsorption towers, each tower is connected to a vacuum pump P; a steam inlet valve VH i ; Steam outlet valve VHb i ; Cooling water inlet valve VC i ; Cooling water outlet valve VCb i ; Crude gas inlet valve VF i ; Product gas outlet valve VP i ; Adsorption tower drain valve VW i ; Exhaust valve VWb i The screenshot structure of the adsorption tower is as follows: Figure 3 As shown, steam and cooling water flow through the tube side, and the adsorbent is filled in the shell side; the above i = 1, 2, 3, 4, 5, 6.

[0074] The crude gas contains 1% of the four impurities: 1,1,2,3,4,4,4-heptafluoro-1-butene, 1,1,1,3,4,4,4-heptafluoro-1-butene, 1,1,3,4,4-pentafluoro-1,3-butadiene, and 1,1,4,4-tetrafluoro-1,3-butadiene. The desorption pressure is 0.2 bar, the heating temperature is 125°C, and the adsorption pressure is 5 bar. The tower is filled with ZSM-5 molecular sieve and activated carbon in a 1:1 ratio.

[0075] The operating sequence of each adsorption tower was switched according to Table 1, and the switching timing of each valve was cycled according to Table 2. Twenty-five sets of parameters were initially collected for training the neural network model, with 15 hidden layer neurons. After initial training, NSGA-II was used to optimize and predict the 25 sets of process parameters and transmit them back to the simulated moving bed multi-tower temperature swing adsorption process system, thus entering the iterative cycle.

[0076] Figure 4The regression coefficient R of the neural network model for each process performance index increases with the number of iterations. 2 The values continued to increase, approaching stability after 20 rounds, reaching a regression coefficient of 0.990 for the training set and 0.987 for the test set. Simultaneously, the recovery rate of the simulated moving bed multi-tower temperature swing adsorption process system reached 90%. The contents of 1,1,2,3,4,4,4-heptafluoro-1-butene, 1,1,1,3,4,4,4-heptafluoro-1-butene, 1,1,3,4,4-pentafluoro-1,3-butadiene, and 1,1,4,4-tetrafluoro-1,3-butadiene in the product gas were 8 ppm, 8 ppm, 5 ppm, and 7 ppm, respectively, and the purity of C4F6 in the product gas reached 99.998%.

[0077] Example 2

[0078] according to Figure 1 To build this system, a simulated moving bed multi-tower temperature swing adsorption process system, a machine learning model, and an NSGA-II optimization algorithm are used. Figure 2 As shown, each tower is connected to a vacuum pump P; a steam inlet valve VH i ; Steam outlet valve VHb i ; Cooling water inlet valve VC i ; Cooling water outlet valve VCb i ; Crude gas inlet valve VF i ; Product gas outlet valve VP i ; Adsorption tower drain valve VW i ; Exhaust valve VWb i The screenshot structure of the adsorption tower is as follows: Figure 3 As shown, steam and cooling water go through the tube side, and the adsorbent is filled in the shell side.

[0079] The crude gas contains 1% of the four impurities: 1,1,2,3,4,4,4-heptafluoro-1-butene, 1,1,1,3,4,4,4-heptafluoro-1-butene, 1,1,3,4,4-pentafluoro-1,3-butadiene, and 1,1,4,4-tetrafluoro-1,3-butadiene. The desorption pressure is 0.1 bar, the heating temperature is 130°C, and the adsorption pressure is 7 bar. The tower is filled with ZSM-5 molecular sieve and activated carbon in a ratio of 1:2.

[0080] The operating sequence of each adsorption tower was switched according to Table 1, and the switching timing of each valve was cycled according to Table 2. Twenty-five sets of parameters were initially collected for training the neural network model, with 15 hidden layer neurons. After initial training, NSGA-II was used to optimize and predict the 25 sets of process parameters and transmit them back to the simulated moving bed multi-tower temperature swing adsorption process system, thus entering the iterative cycle.

[0081] Figure 5 The regression coefficient R of the neural network model for each process performance index increases with the number of iterations. 2 The values continued to increase, approaching stability after 20 rounds, reaching a regression coefficient of 0.989 for the training set and 0.983 for the test set. Simultaneously, the recovery rate of the simulated moving bed multi-tower temperature swing adsorption process system reached 85%. The contents of 1,1,2,3,4,4,4-heptafluoro-1-butene, 1,1,1,3,4,4,4-heptafluoro-1-butene, 1,1,3,4,4-pentafluoro-1,3-butadiene, and 1,1,4,4-tetrafluoro-1,3-butadiene in the product gas were 1 ppm, 3 ppm, 7 ppm, and 8 ppm, respectively, and the purity of C4F6 in the product gas reached 99.998%.

[0082] Example 3

[0083] according to Figure 1 To build this system, a simulated moving bed multi-tower temperature swing adsorption process system, a machine learning model, and an NSGA-II optimization algorithm are used. Figure 2 As shown, each tower is connected to a vacuum pump P; a steam inlet valve VH i ; Steam outlet valve VHb i ; Cooling water inlet valve VC i ; Cooling water outlet valve VCb i ; Crude gas inlet valve VF i ; Product gas outlet valve VP i ; Adsorption tower drain valve VW i ; Exhaust valve VWb i The screenshot structure of the adsorption tower is as follows: Figure 3 As shown, steam and cooling water go through the tube side, and the adsorbent is filled in the shell side.

[0084] The crude gas contains 1% of the four impurities: 1,1,2,3,4,4,4-heptafluoro-1-butene, 1,1,1,3,4,4,4-heptafluoro-1-butene, 1,1,3,4,4-pentafluoro-1,3-butadiene, and 1,1,4,4-tetrafluoro-1,3-butadiene. The desorption pressure is 0.1 bar, the heating temperature is 100°C, and the adsorption pressure is 5 bar. The tower is filled with ZSM-5 molecular sieve and activated carbon in a 1:1 ratio.

[0085] The operating sequence of each adsorption tower was switched according to Table 1, and the switching timing of each valve was cycled according to Table 2. Twenty-five sets of parameters were initially collected for training the neural network model, with 15 hidden layer neurons. After initial training, NSGA-II was used to optimize and predict the 25 sets of process parameters and transmit them back to the simulated moving bed multi-tower temperature swing adsorption process system, thus entering the iterative cycle.

[0086] Figure 6 The regression coefficient R of the neural network model for each process performance index increases with the number of iterations. 2 The values continued to increase, approaching stability after 15 rounds, reaching a regression coefficient of 0.980 for the training set and 0.977 for the test set. Simultaneously, the recovery rate of the simulated moving bed multi-tower temperature swing adsorption process system reached 89%. The contents of 1,1,2,3,4,4,4-heptafluoro-1-butene, 1,1,1,3,4,4,4-heptafluoro-1-butene, 1,1,3,4,4-pentafluoro-1,3-butadiene, and 1,1,4,4-tetrafluoro-1,3-butadiene in the product gas were 8 ppm, 9 ppm, 7 ppm, and 8 ppm, respectively, and the purity of C4F6 in the product gas reached 99.996%.

[0087] Example 4

[0088] according to Figure 1 To build this system, a simulated moving bed multi-tower temperature swing adsorption process system, a machine learning model, and an NSGA-II optimization algorithm are used. Figure 2 As shown, each tower is connected to a vacuum pump P; a steam inlet valve VH i ; Steam outlet valve VHb i ; Cooling water inlet valve VC i ; Cooling water outlet valve VCb i ; Crude gas inlet valve VF i ; Product gas outlet valve VP i ; Adsorption tower drain valve VW i ; Exhaust valve VWb i The screenshot structure of the adsorption tower is as follows: Figure 3 As shown, steam and cooling water go through the tube side, and the adsorbent is filled in the shell side.

[0089] The crude gas contains 1% of the four impurities: 1,1,2,3,4,4,4-heptafluoro-1-butene, 1,1,1,3,4,4,4-heptafluoro-1-butene, 1,1,3,4,4-pentafluoro-1,3-butadiene, and 1,1,4,4-tetrafluoro-1,3-butadiene. The desorption pressure is 0.1 bar, the heating temperature is 150°C, and the adsorption pressure is 5 bar. The tower is filled with ZSM-5 molecular sieve and activated carbon in a 2:1 ratio.

[0090] The operating sequence of each adsorption tower was switched according to Table 1, and the switching timing of each valve was cycled according to Table 2. Twenty-five sets of parameters were initially collected for training the neural network model, with 100 hidden layer neurons. After initial training, NSGA-II was used to optimize and predict the 25 sets of process parameters and transmit them back to the simulated moving bed multi-tower temperature swing adsorption process system, thus entering the iterative cycle.

[0091] Figure 7 The regression coefficient R of the neural network model for each process performance index increases with the number of iterations. 2 The regression coefficients for the training set and the test set reached 0.997 and 0.968 respectively, indicating some overfitting due to the large number of neurons. The recovery rate of the multi-tower temperature swing adsorption process system simulated by moving bed operation reached 87%. The contents of 1,1,2,3,4,4,4-heptafluoro-1-butene, 1,1,1,3,4,4,4-heptafluoro-1-butene, 1,1,3,4,4-pentafluoro-1,3-butadiene, and 1,1,4,4-tetrafluoro-1,3-butadiene in the product gas were 3 ppm, 1 ppm, 2 ppm, and 2 ppm, respectively. The purity of C4F6 in the product gas reached 99.999%.

[0092] In order to further illustrate the superiority of the present application, the present application also provides the following comparative examples.

[0093] Comparative Example 1

[0094] The system model, feed gas setting, desorption pressure, adsorption pressure and heating temperature were the same as those in Example 1. The tower was filled with ZSM-5 molecular sieve and activated carbon in a ratio of 10:1.

[0095] The operating sequence of each adsorption tower was switched according to Table 1, and the switching timing of each valve was cycled according to Table 2. Twenty-five sets of parameters were initially collected for training the neural network model, with 15 hidden layer neurons. After initial training, NSGA-II was used to optimize and predict the 25 sets of process parameters and transmit them back to the simulated moving bed multi-tower temperature swing adsorption process system, thus entering the iterative cycle.

[0096] Figure 8 The regression coefficient R of the neural network model for each process performance index increases with the number of iterations. 2The values continued to increase, approaching stability after 20 iterations, reaching a regression coefficient of 0.990 for the training set and 0.988 for the test set. However, after 80 iterations, the concentrations of 1,1,2,3,4,4,4-heptafluoro-1-butene and 1,1,1,3,4,4,4-heptafluoro-1-butene fluctuated within 50 ppm, and the concentrations before and after a single cycle could never be less than 1 ppm. The final recovery rate of the simulated moving bed multi-tower temperature swing adsorption process system was 80%. The contents of 1,1,2,3,4,4,4-heptafluoro-1-butene, 1,1,1,3,4,4,4-heptafluoro-1-butene, 1,1,3,4,4-pentafluoro-1,3-butadiene, and 1,1,4,4-tetrafluoro-1,3-butadiene in the product gas were 1560ppm, 1635ppm, 1ppm, and 2ppm, respectively. The purity of C4F6 in the product gas only reached 99.68%.

[0097] Comparative Example 2

[0098] The system model, feed gas setting, desorption pressure, adsorption pressure and heating temperature are the same as those in Example 2. The tower is filled with ZSM-5 molecular sieve and Cu-BTC in a ratio of 1:2.

[0099] The operating sequence of each adsorption tower was switched according to Table 1, and the switching timing of each valve was cycled according to Table 2. Twenty-five sets of parameters were initially collected for training the neural network model, with 15 hidden layer neurons. After initial training, NSGA-II was used to optimize and predict the 25 sets of process parameters and transmit them back to the simulated moving bed multi-tower temperature swing adsorption process system, thus entering the iterative cycle.

[0100] Figure 9 The regression coefficient R of the neural network model for each process performance index increases with the number of iterations. 2 The values continued to increase, approaching stability after 40 rounds. The final regression coefficient for the training set reached 0.985, and the regression coefficient for the test set reached 0.952, indicating some overfitting. The recovery rate of the simulated moving bed multi-tower temperature swing adsorption process system was low, at 75%. The contents of 1,1,2,3,4,4,4-heptafluoro-1-butene, 1,1,1,3,4,4,4-heptafluoro-1-butene, 1,1,3,4,4-pentafluoro-1,3-butadiene, and 1,1,4,4-tetrafluoro-1,3-butadiene in the product gas were 105 ppm, 206 ppm, 20 ppm, and 28 ppm, respectively. The purity of C4F6 in the product gas was 99.96%.

[0101] Comparative Example 3

[0102] The system model, feed gas setting, desorption pressure, adsorption pressure, heating temperature and adsorbent filling are the same as those in Example 3.

[0103] The operating sequence of each adsorption tower was switched according to Table 1, and the switching timing of each valve was cycled according to Table 2. The machine learning model used a random forest algorithm, and 25 sets of parameters were initially collected for model training. After initial training, NSGA-II was used to optimize and predict the 25 sets of process parameters and transmit them back to the simulated moving bed multi-tower temperature swing adsorption process system, thus entering the iterative cycle.

[0104] Figure 10 The regression coefficient R of the random forest model for each process performance index is shown as the number of iterations increases. 2 The overall increase is, but the R 2 There were obvious wild fluctuations, and convergence was impossible. The machine learning model performed poorly, resulting in suboptimal predicted process parameters. After 100 iterations, the recovery rate of the simulated moving bed multi-tower temperature swing adsorption process system was 67%. The contents of 1,1,2,3,4,4,4-heptafluoro-1-butene, 1,1,1,3,4,4,4-heptafluoro-1-butene, 1,1,3,4,4-pentafluoro-1,3-butadiene, and 1,1,4,4-tetrafluoro-1,3-butadiene in the product gas were 2058 ppm, 3079 ppm, 3543 ppm, and 5748 ppm, respectively. The purity of C4F6 in the product gas was only 98.56%.

[0105] Comparative Example 4

[0106] The system model, feed gas setting, adsorption pressure and adsorbent filling were the same as those in Example 4. The desorption pressure was 0.5 bar and the heating temperature was 50°C.

[0107] The operating sequence of each adsorption tower was switched according to Table 1, and the switching timing of each valve was cycled according to Table 2. Twenty-five sets of parameters were initially collected for training the neural network model, with 100 hidden layer neurons. After initial training, NSGA-II was used to optimize and predict the 25 sets of process parameters and transmit them back to the simulated moving bed multi-tower temperature swing adsorption process system, thus entering the iterative cycle.

[0108] Figure 11 The regression coefficient R of the neural network model for each process performance index increases with the number of iterations. 2The results continued to increase, approaching stability after 10 rounds. The final regression coefficient for the training set reached 0.980, and the regression coefficient for the test set reached 0.954, indicating significant overfitting due to the large number of neurons. Furthermore, the performance of the simulated moving bed declined due to insufficient vacuum during regeneration, low temperatures, and insufficient bed regeneration. The recovery rate was 78%, and the contents of 1,1,2,3,4,4,4-heptafluoro-1-butene, 1,1,1,3,4,4,4-heptafluoro-1-butene, 1,1,3,4,4-pentafluoro-1,3-butadiene, and 1,1,4,4-tetrafluoro-1,3-butadiene in the product gas were 1078 ppm, 2034 ppm, 190 ppm, and 325 ppm, respectively. The purity of C4F6 in the product gas was 99.63%.

[0109] Table 1 Operating status and timing of each adsorption tower during the temperature swing adsorption-simulated moving bed process for purification of C4F6

[0110]

[0111] Table 2 Switching status and function description of all valves in each cycle when purifying C4F6 by temperature swing adsorption-simulated moving bed process

[0112]

[0113]

[0114]

Claims

1. A machine learning optimized simulated moving bed multi-tower temperature swing adsorption system for C4F6 purification, characterized in that: Includes simulated moving bed system, machine learning algorithm model, and NSGA-II optimization algorithm; The simulated moving bed system includes an adsorption tower, each adsorption tower is equipped with a vacuum pump (P), a crude gas inlet valve (VF i ), adsorption tower drain valve (VW i ), discharge valve (VP i )、Steam inlet valve (VH i )、Steam outlet valve (VHb i )、Cooling water inlet valve (VC i ) and cooling water outlet valve (VCb i ); the crude gas inlet valve (VF i ) Connect C4F6 crude gas; adsorption tower drain valve (VW i ) is connected to the vacuum pump (P) and then input to the tail gas collection device; the discharge valve (VP i ) is connected to the product gas storage device; the steam inlet valve (VH i ) Connect high temperature steam; cooling water inlet valve (VC i ) cooling water connection; The simulated moving bed system records the process parameters of each operation, the time of each valve switching, and the C4F6 purity, recovery rate and content of various impurities in the product gas that match the process parameters and the valve switching timing, which together constitute a set of samples; Arrange the multiple groups of samples into a data set; the process parameters include: a specified pressure P1 in the adsorption tower performing a depressurization operation, a specified desorption temperature T1 in the adsorption tower performing a heating operation, a specified adsorption pressure P2 in the adsorption tower performing a pressurization operation, and a switching sequence ti (i=1, 2, 3, 4, 5, 6) of each valve; The data set is used to train a machine learning algorithm model, and the trained machine learning algorithm model is combined with the NSGA-II optimization algorithm to infer multiple sets of more optimal process parameters and valve switching timings; the process parameters and timings are applied to the simulated moving bed multi-tower temperature swing adsorption process system to obtain more samples, which are merged with the original data set to form a new data set; Repeat several times until the performance of the simulated moving bed multi-tower temperature swing adsorption process system is no longer improved.

2. The machine learning optimized simulated moving bed multi-tower temperature swing adsorption process system for C4F6 purification according to claim 1, characterized in that: Each single tower in the simulated moving bed multi-tower temperature swing adsorption process system is filled with two adsorbents at the same time; the two adsorbents are molecular sieve and activated carbon; and the filling ratio of the molecular sieve and activated carbon is in the range of 1:3 to 3:

1.

3. The machine learning optimized simulated moving bed multi-tower temperature swing adsorption process system for C4F6 purification according to claim 1, characterized in that: The adsorption tower is a shell and tube type device, steam and cooling water flow through the tube layer, and the shell layer is filled with adsorbent.

4. The machine learning optimized simulated moving bed multi-tower temperature swing adsorption process system for C4F6 purification according to claim 1, characterized in that: The C4F6 content in the C4F6 crude gas is 96%.

5. A machine learning optimized simulated moving bed multi-tower temperature swing adsorption process for C4F6 purification using the system according to any one of claims 1 to 4, characterized in that: The steps include: (1) Two adsorption towers perform adsorption operation, open the feed valves connecting the bottom of all adsorption towers performing adsorption operation and the crude gas storage device, open the discharge valves connecting the top of all adsorption towers performing adsorption operation and the product gas storage device, close other valves connected to all adsorption towers performing adsorption operation, and turn off the vacuum pumps of the adsorption towers; during this process, C4F6 crude gas enters the adsorption tower from the bottom of the tower, impurities are adsorbed by the adsorbent bed, and the obtained C4F6 pure gas is discharged from the top of the tower and directed to the C4F6 gas storage device; (2) At the same time, an adsorption tower performs a depressurization operation, opens a drain valve connected between the bottom of the adsorption tower performing the depressurization operation and the vacuum pump, opens an exhaust valve between the vacuum pump of the adsorption tower performing the depressurization operation and the tail gas storage device, opens the vacuum pump of the adsorption tower performing the depressurization operation, and closes other valves connected to the adsorption tower performing the depressurization operation; during this process, the pressure of the adsorption tower gradually decreases from the adsorption set pressure to the specified pressure, and the gas in the adsorption tower is discharged from the bottom of the tower together with a small amount of halogenated hydrocarbon impurities desorbed from the bed layer and directed to the tail gas storage device; (3) At the same time, an adsorption tower performs a heating operation, opens a drain valve connected between the bottom of the adsorption tower performing the heating operation and the vacuum pump, opens an exhaust valve between the vacuum pump of the adsorption tower performing the heating operation and the tail gas storage device, opens the vacuum pump of the adsorption tower performing the heating operation, opens a steam inlet valve between the top of the adsorption tower performing the heating operation and the steam inlet line, opens a steam outlet valve between the bottom of the adsorption tower performing the heating operation and the steam outlet line, and closes other valves connected to the adsorption tower performing the heating operation; during this process, the temperature of the adsorption tower gradually rises to the specified desorption temperature and is maintained, and a large amount of halogenated hydrocarbon impurities are desorbed from the bed of the adsorption tower, discharged from the bottom of the tower and directed to the tail gas storage device; (4) At the same time, an adsorption tower performs a cooling operation, opens a drain valve connected between the bottom of the adsorption tower performing the cooling operation and the vacuum pump, opens an exhaust valve between the vacuum pump of the adsorption tower performing the cooling operation and the tail gas storage device, opens the vacuum pump of the adsorption tower performing the cooling operation, opens the inlet valve between the bottom of the adsorption tower performing the cooling operation and the cooling water inlet line, opens the outlet valve between the top of the adsorption tower performing the cooling operation and the cooling water outlet line, and closes other valves connected to the adsorption tower performing the cooling operation. During this process, the temperature of the adsorption tower gradually drops to room temperature, and a small amount of residual gas in the tower is discharged from the bottom of the tower and directed to the tail gas storage device; (5) At the same time, a pressurized adsorption tower performs a pressurization operation, opens the feed valve connected between the bottom of the adsorption tower performing the pressurized adsorption operation and the crude gas storage device, opens all the discharge valves connected to the top of the adsorption tower performing the pressurized adsorption operation and the product gas storage device, closes the other valves connected to the adsorption tower performing the pressurization operation, and turns off the vacuum pump of the adsorption tower. During this process, the C4F6 crude gas enters the adsorption tower from the bottom of the tower, the pressure in the adsorption tower gradually rises to the specified adsorption pressure, the impurities are adsorbed by the adsorbent bed, and the obtained C4F6 pure gas is discharged from the top of the tower and directed to the C4F6 gas storage device; (6) Recording the process parameters of each operation, the time of each valve switching, and the purity of C4F6, the recovery rate, and the content of various impurities in the product gas that match the process parameters and the valve switching timing, together forming a set of samples; arranging multiple sets of the samples into a data set; (7) using the data set obtained in step (6) to train a machine learning model, and combining the trained machine learning model with an optimization algorithm to infer multiple sets of more optimal process parameters and valve switching timings; applying the process parameters and timings to the simulated moving bed multi-tower temperature swing adsorption process system, repeating step (6) to obtain more samples, and merging them with the original data set to form a new data set; (8) Repeat steps (6) and (7) several times until the performance of the simulated moving bed multi-tower temperature swing adsorption process system is no longer improved.

6. The process according to claim 1, characterized in that The halogenated hydrocarbon impurities are 1,1,2,3,4,4,4-heptafluoro-1-butene, 1,1,1,3,4,4,4-heptafluoro-1-butene, 1,1,3,4,4-pentafluoro-1,3-butadiene or 1,1,4,4-tetrafluoro-1,3-butadiene.

7. The process according to claim 1, characterized in that In step (2), the pressure is reduced to a specified pressure of 0.1 bar to 0.2 bar; in step (3), the pressure is heated to a specified desorption temperature of 100° C. to 200° C.; and in step (5), the pressure is increased to a specified adsorption pressure of 3 bar to 10 bar.

8. The process according to claim 1, characterized in that The process parameters in step (6) include the specified pressure P1 in step (2), the specified desorption temperature T1 in step (3), the specified adsorption pressure P2 in step (5), the valve switching timing t in step (6), and the specified adsorption pressure P2 in step (5). i (i=1,2,3,4,5,6); The number of groups of samples in step (6) is 10 to 50 groups.

9. The process according to claim 1, characterized in that In step (7), the machine learning model is a neural network model; the optimization algorithm is a non-dominated sorting genetic algorithm (NSGA-II), and the number of groups of samples is 20 to 30 groups.

10. The process according to claim 1, characterized in that: In step (8), the performance is no longer further improved. Before and after one cycle, the content changes of 1,1,2,3,4,4,4-heptafluoro-1-butene, 1,1,1,3,4,4,4-heptafluoro-1-butene, 1,1,3,4,4-pentafluoro-1,3-butadiene, and 1,1,4,4-tetrafluoro-1,3-butadiene in the product gas are all less than 1 ppm, and the recovery rate changes are less than 1%.

Images