Coupling intelligent optimization method

Abstract

The invention relates to a coupling intelligent optimization method, and the method comprises the steps: carrying out the multi-time random sampling of a two-dimensional image of a cooler tube nest through a Latin hypercube sampling method, obtaining one piece of sample data after each sampling, obtaining a plurality of pieces of sample data, calculating the Nusselt number and frictional resistance coefficient of each piece of sample data, and carrying out the calculation of the Nusselt number and frictional resistance coefficient of each piece of sample data; the multiple pieces of sample data and the Nusselt numbers and the frictional resistance coefficients of the multiple pieces of sample data are all included in a database; then, taking a part of the database as a training sample to adjust the convolutional neural network model, and obtaining a predicted Nusselt number and a predicted frictional resistance coefficient of a two-dimensional image of the cooler tube nest, which are adjusted and output by the adjusted convolutional neural network model; and performing prediction result verification on the adjusted convolutional neural network model by using the other part of the database as verification samples when the error of each verification sample is smaller than an error threshold.

Description

Technical Field

[0001] The present application relates to the technical field of air compressors, and in particular to a coupled intelligent optimization method. Background Art

[0002] Air compressors are widely used in industrial production, and the design of the intercooler has an important influence on the efficiency and stability of the compressor.

[0003] However, the internal tubes of the existing intercooler all adopt a hollow cylindrical and straight structure design. Although this design method has a certain convenience in production and processing, this design has the problem of insufficient cooling efficiency. Summary of the invention

[0004] Based on this, it is necessary to provide a coupled intelligent optimization method to address the problem that the internal tubes of the traditional existing intercooler all adopt a hollow cylindrical and straight structural design, resulting in the intercooler's internal cooling effect being not efficient enough.

[0005] The present application provides a coupled intelligent optimization method, comprising:

[0006] Acquire a two-dimensional image of a cooler tube, wherein the two-dimensional image of the cooler tube includes a plurality of tubes;

[0007] The two-dimensional image of the cooler tubes is coupled with the two-dimensional rectangular coordinate system to obtain the coordinate parameters of each tube in the two-dimensional image of the cooler tubes;

[0008] Latin hypercube sampling is performed based on the coordinate parameters of each tube to reconstruct the two-dimensional image of the cooler tube, and the reconstructed two-dimensional image of the cooler tube is used as sample data. The Latin hypercube sampling is performed based on the coordinate parameters of each tube to reconstruct the two-dimensional image of the cooler tube, and the reconstructed two-dimensional image of the cooler tube is used as sample data, and multiple sample data are obtained;

[0009] Based on the lattice Boltzmann method module, the numerical simulation is performed to calculate each sample data, and the Nusselt number and friction coefficient of each sample data are obtained;

[0010] Each obtained sample data, each Nusselt number of the sample data, and each friction resistance coefficient of the sample data are normalized, and each obtained normalized sample data, each normalized Nusselt number of the sample data, and each normalized friction resistance coefficient of the sample data are included in the database as a group of data until there are N groups of data in the database; N is a positive integer and N is greater than 1;

[0011] The M groups of data in the database are used as training samples to adjust the convolutional neural network model to obtain the adjusted convolutional neural network model; M is a positive integer and M is less than N;

[0012] Inputting the two-dimensional image of the cooler tubes into the adjusted convolutional neural network model, starting the adjusted convolutional neural network model, and obtaining the prediction results output by the adjusted convolutional neural network model, wherein the prediction results include the predicted Nusselt number and the predicted friction resistance coefficient;

[0013] The NM group data in the database is used as a verification sample to perform error calculation on the prediction results output by the convolutional neural network model, and the error results of the prediction results of multiple adjusted convolutional neural network models are obtained;

[0014] Determine whether multiple error results are all less than an error threshold;

[0015] If multiple error results are all less than the error threshold, then build an environment for the reinforcement learning algorithm, integrate the environment of the reinforcement learning algorithm into the adjusted convolutional neural network model, and obtain the optimal convolutional neural network model;

[0016] Inputting the two-dimensional image of the cooler tubes into the optimal convolutional neural network model, starting the optimal convolutional neural network model, and obtaining the optimal tube parameters, optimal Nusselt number, and optimal friction resistance coefficient output by the optimal convolutional neural network model;

[0017] If any error result is greater than or equal to the threshold, the method returns to execute the method of adjusting the convolutional neural network model using a part of the database as a training sample to obtain an adjusted convolutional neural network model.

[0018] Furthermore, the two-dimensional image of the cooler tubes is coupled with the two-dimensional rectangular coordinate system to obtain the coordinate parameters of each tube in the two-dimensional image of the cooler tubes, including:

[0019] A random pixel point in the two-dimensional image of the cooler tubes is used as the origin of the coordinates in the two-dimensional rectangular coordinate system;

[0020] Taking the coordinate origin as the anchor point, the coordinate parameters of each tube array in the two-dimensional plane rectangular coordinate system are obtained, and the coordinate parameters of the tube array include the long axis parameter a of the tube array, the long axis parameter b of the tube array, the inclination angle parameter θ of the tube array, and the coordinate parameters (x, y) of the tube array.

[0021] Further, Latin hypercube sampling is performed based on the coordinate parameters of each tube to reconstruct the two-dimensional image of the cooler tube, and the reconstructed two-dimensional image of the cooler tube is used as sample data. The Latin hypercube sampling is performed based on the coordinate parameters of each tube to reconstruct the two-dimensional image of the cooler tube, and the reconstructed two-dimensional image of the cooler tube is used as sample data, and multiple sample data are obtained, including:

[0022] Get the coordinate parameters of each tube column;

[0023] The coordinate parameters of each tube array are adjusted by a Latin hypercube sampling method, and a two-dimensional image of the cooler tube array after the tube coordinate parameters are adjusted is obtained based on the tube arrays after the tube coordinate parameters are adjusted, and the two-dimensional image of the cooler tube array after the tube coordinate parameters are adjusted is recorded as a sample data;

[0024] Repeat the method of Latin hypercube sampling to adjust the coordinate parameters of each tube array, obtain a two-dimensional image of the cooler tube array after the tube coordinate parameters are adjusted based on all the tube arrays after the tube coordinate parameters are adjusted, record the two-dimensional image of the cooler tube array after the tube coordinate parameters are adjusted as a sample data, and obtain at least K sample data.

[0025] Furthermore, the numerical simulation based on the lattice Boltzmann method module calculates each sample data to obtain the Nusselt number and friction coefficient of each sample data, including:

[0026] Select a sample data;

[0027] The sample data is input into the lattice Boltzmann method module, the lattice Boltzmann method module is started, and the temperature field and velocity field of the sample data output by the lattice Boltzmann method module are obtained;

[0028] The friction resistance coefficient of the sample data is obtained by calculating the velocity field of the sample data;

[0029] The Nusselt number of the sample data is obtained by calculating the temperature field of the sample data;

[0030] Return to the step of selecting a sample data until each sample data has been selected once.

[0031] Furthermore, the friction resistance coefficient of the sample data is calculated based on the velocity field of the sample data, including:

[0032] Analyze the velocity field of the sample data to obtain parameters in the velocity field of the sample data, wherein the parameters in the velocity field of the sample data include an inlet average velocity parameter, a flow resistance parameter, a fluid density parameter, a fluid grid number parameter, an air kinematic viscosity parameter, and a Reynolds number;

[0033] The obtained inlet average velocity parameter, air kinematic viscosity parameter and Reynolds number are input into Formula 1 to calculate the equivalent diameter in the velocity field of the sample data;

[0034]

[0035] Where Re is the Reynolds number constant and Re is 150; U is the average inlet velocity; D is the equivalent diameter; υ is the kinematic viscosity of air;

[0036] The obtained equivalent diameter, flow resistance parameter, fluid density parameter, fluid grid number parameter, and inlet average velocity parameter are input into Formula 2 to calculate the friction resistance coefficient in the velocity field of the sample data;

[0037]

[0038] Where f is the friction factor; F D is the flow resistance; n is the number of grids of the fluid; ρ is the density of the fluid; U is; D is the equivalent diameter.

[0039] Furthermore, the Nusselt number of the sample data is calculated based on the temperature field of the sample data, including:

[0040] Analyze the temperature field of the sample data to obtain parameters in the temperature field of the sample data, the sample data including heat flux parameters, surface area parameters, thermal conductivity, fluid average temperature parameters, wall temperature parameters and equivalent diameter;

[0041] The obtained heat flux parameter, surface area parameter, thermal conductivity, fluid average temperature parameter, wall temperature parameter and equivalent diameter are input into Formula 3 to calculate the Nusselt number of the temperature field of the adjusted two-dimensional image of the cooler tube column;

[0042]

[0043] Where Nu is the value of the Nusselt number; Φ is the heat flux; S is the surface area; κ is the thermal conductivity; is the average temperature of the fluid; T w is the wall temperature.

[0044] Further, each sample data, each Nusselt number of each sample data and each friction resistance coefficient of each sample data are normalized, and each normalized sample data, each normalized Nusselt number of each sample data and each normalized friction resistance coefficient of each sample data are included in the database as a group of data until there are N groups of data in the database; N is a positive integer and N is greater than 1, and the above also includes:

[0045] Select a sample data;

[0046] Analyze the sample data to obtain the Nusselt number and friction coefficient of the sample data;

[0047] Determine whether the Nusselt number of the sample data is within the Nusselt number threshold range;

[0048] If the Nusselt number of the sample data is not within the Nusselt number threshold range, the sample data is deleted;

[0049] If the Nusselt number of the sample data is within the Nusselt number threshold range, determining whether the friction resistance coefficient of the sample data is within the friction resistance coefficient threshold range;

[0050] If the friction resistance coefficient of the sample data is not within the friction resistance coefficient threshold range, the sample data is deleted;

[0051] Return to the step of selecting a sample data until each sample data has been selected once.

[0052] Furthermore, the M groups of data in the database are used as training samples to adjust the convolutional neural network model to obtain the adjusted convolutional neural network model; M is a positive integer and M is less than N, including:

[0053] Create training and validation databases;

[0054] Select M groups of data in the database to be included in the training database, and select N groups of data in the database to be included in the verification database;

[0055] Select a set of data from the training database;

[0056] Analyze the group of data to obtain the normalized sample data, the Nusselt number of the normalized sample data, and the friction resistance coefficient of the normalized sample data;

[0057] The normalized sample data is used as the input of the convolutional neural network model, the Nusselt number of the normalized sample data and the friction coefficient of the normalized sample data are used as the output of the convolutional neural network model, and the convolutional neural network model is adjusted to obtain an adjusted convolutional neural network model;

[0058] Returning a set of data in the selected training database until each set of data in the training database has been selected once;

[0059] Get the adjusted convolutional neural network model.

[0060] Furthermore, the NM group data in the database is used as a verification sample to perform error calculation on the prediction results output by the convolutional neural network model to obtain error results of multiple adjusted convolutional neural network model prediction results, including:

[0061] Select a set of data in the verification database;

[0062] Analyze the set of data to obtain the normalized Nusselt number of the sample data and the normalized friction coefficient of the sample data;

[0063] Calculate the error between the normalized Nusselt number of the sample data and the predicted Nusselt number to obtain a first error result;

[0064] Calculating the error between the friction resistance coefficient of the normalized sample data and the predicted friction resistance coefficient to obtain a second error result;

[0065] A set of data in the selected verification database is returned until each set of data in the verification database has been selected once.

[0066] Furthermore, the environment of the reinforcement learning algorithm is constructed, and the environment of the reinforcement learning algorithm is integrated into the adjusted convolutional neural network model to obtain the optimal convolutional neural network model, including:

[0067] The database is used as a training sample for the reinforcement learning algorithm to optimize the adjusted convolutional neural network model and obtain the optimal convolutional neural network model.

[0068] The present application relates to a coupled intelligent optimization method, which uses a Latin hypercube sampling method to perform multiple random sampling of a two-dimensional image of a cooler tube, obtains a sample data after each sampling, and then obtains multiple sample data, then calculates the Nusselt number and friction resistance coefficient of each sample data, and incorporates the multiple sample data and the Nusselt numbers and friction resistance coefficients of the multiple sample data into a database; then uses a part of the database as a training sample to adjust the convolutional neural network model to obtain the predicted Nusselt number and predicted friction resistance coefficient of the two-dimensional image of the cooler tube that are adjusted and output by the adjusted convolutional neural network model, and then uses another part of the database as a verification sample to verify the prediction result of the adjusted convolutional neural network model. When the error of each verification sample is less than the error threshold, the adjusted convolutional neural network model is optimized by a reinforcement learning algorithm to obtain an optimal convolutional neural network model, and then the optimal convolutional neural network model outputs the optimal tube parameters, the optimal Nusselt number and the optimal friction resistance coefficient. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 A flowchart of a coupled intelligent optimization method provided in one embodiment of the present application.

[0070] Figure 2 This is a schematic diagram of the arrangement of the lower tubes, the velocity field, and the temperature field based on the original parameters in the coupled intelligent optimization method provided in one embodiment of the present application.

[0071] Figure 3 A schematic diagram of the two-dimensional shape and arrangement of the tube array after reinforcement learning optimization in the coupled intelligent optimization method provided in one embodiment of the present application. DETAILED DESCRIPTION

[0072] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0073] like Figure 1 As shown, in one embodiment of the present application, the coupled intelligent optimization method includes the following S001 to S012:

[0074] S001, acquiring a two-dimensional image of a cooler tube array, wherein the two-dimensional image of the cooler tube array includes a plurality of tube arrays.

[0075] Specifically, the two-dimensional image of the cooler tube refers to a two-dimensional image of the cross section of the tube of the air compressor intercooler, and the two-dimensional plane is composed of 300×300 grids, and one tube diameter is 50 grids, and there are 9 tubes in a two-dimensional plane image.

[0076] S002, coupling the two-dimensional image of the cooler tubes and the two-dimensional rectangular coordinate system to obtain coordinate parameters of each tube in the two-dimensional image of the cooler tubes.

[0077] Specifically, by coupling the two-dimensional image of the cooler tubes and the two-dimensional rectangular coordinate system, the parameters of each tube in the two-dimensional image of the cooler tubes can be expressed in the form of coordinate parameters, which facilitates the calculation of the parameters of each tube.

[0078] S003, Latin hypercube sampling is performed based on the coordinate parameters of each tube to reconstruct the two-dimensional image of the cooler tube, and the reconstructed two-dimensional image of the cooler tube is used as sample data. The Latin hypercube sampling is performed based on the coordinate parameters of each tube to reconstruct the two-dimensional image of the cooler tube, and the reconstructed two-dimensional image of the cooler tube is used as sample data, and multiple sample data are obtained.

[0079] Specifically, the coordinate parameters of each tube array are subjected to Latin hypercube sampling, and the coordinate parameters of each tube array have their own sampling variation range.

[0080] S004, numerically simulate and calculate each sample data based on the lattice Boltzmann method module to obtain the Nusselt number and friction resistance coefficient of each sample data.

[0081] S005, normalize each sample data, each Nusselt number of the sample data, and each friction resistance coefficient of the sample data, and include each normalized sample data, each normalized Nusselt number of the sample data, and each normalized friction resistance coefficient of the sample data as a group of data into the database until there are N groups of data in the database; N is a positive integer and N is greater than 1.

[0082] S006, using M groups of data in the database as training samples to adjust the convolutional neural network model to obtain an adjusted convolutional neural network model; M is a positive integer and M is less than N.

[0083] Specifically, taking the M groups of data in the database as training samples means randomly taking the M groups of data in the database as training samples, and the M groups of data in the database account for eighty percent of the total data in the database.

[0084] S007, inputting the two-dimensional image of the cooler tubes into the adjusted convolutional neural network model, starting the adjusted convolutional neural network model, and obtaining the prediction results output by the adjusted convolutional neural network model, wherein the prediction results include the predicted Nusselt number and the predicted friction resistance coefficient.

[0085] S008, using the NM group data in the database as a verification sample to perform error calculation on the prediction results output by the convolutional neural network model, and obtaining error results of multiple adjusted convolutional neural network model prediction results.

[0086] Specifically, the NM group data in the database as the verification sample refers to the remaining part of the database after removing the training samples, and the verification sample has no overlapping part with the training sample.

[0087] S009, determining whether the multiple error results are all smaller than the error threshold.

[0088] S010, if multiple error results are all smaller than the error threshold, then build an environment for the reinforcement learning algorithm, integrate the environment for the reinforcement learning algorithm into the adjusted convolutional neural network model, and obtain the optimal convolutional neural network model.

[0089] Specifically, the error threshold is 5 percent.

[0090] S011, inputting the two-dimensional image of the cooler tube column into the optimal convolutional neural network model, starting the optimal convolutional neural network model, and obtaining the optimal tube column parameters, optimal Nusselt number, and optimal friction resistance coefficient output by the optimal convolutional neural network model.

[0091] S012: If any error result is greater than or equal to the threshold, return to execute the step of adjusting the convolutional neural network model using a part of the database as a training sample to obtain an adjusted convolutional neural network model.

[0092] In this embodiment, the two-dimensional image of the cooler tube is randomly sampled multiple times by the Latin hypercube sampling method, and a sample data is obtained after each sampling, and then multiple sample data are obtained, and then the Nusselt number and friction resistance coefficient of each sample data are calculated, and the multiple sample data and the Nusselt numbers and friction resistance coefficients of the multiple sample data are included in the database; then a part of the database is used as a training sample to adjust the convolutional neural network model to obtain the predicted Nusselt number and predicted friction resistance coefficient of the two-dimensional image of the cooler tube adjusted and output by the adjusted convolutional neural network model, and then another part of the database is used as a verification sample to verify the prediction result of the adjusted convolutional neural network model. When the error of each verification sample is less than the error threshold, the adjusted convolutional neural network model is optimized by the reinforcement learning algorithm to obtain the optimal convolutional neural network model, and then the optimal tube parameters, the optimal Nusselt number and the optimal friction resistance coefficient are output by the optimal convolutional neural network model.

[0093] In one embodiment of the present application, the two-dimensional image of the cooler tubes is coupled with the two-dimensional rectangular coordinate system to obtain the coordinate parameters of each tube in the two-dimensional image of the cooler tubes, including the following S002a to S002b:

[0094] S002a, taking a random pixel point in the two-dimensional image of the cooler tubes as the coordinate origin in the two-dimensional plane rectangular coordinate system.

[0095] S002b, taking the coordinate origin as an anchor point, obtaining the coordinate parameters of each tube array in the two-dimensional plane rectangular coordinate system, wherein the coordinate parameters of the tube array include the long axis parameter a of the tube array, the long axis parameter b of the tube array, the inclination angle parameter θ of the tube array, and the coordinate parameters (x, y) of the tube array.

[0096] Specifically, the inclination angle parameter θ of the tube array refers to the angle between the long axis and the y-axis in the two-dimensional rectangular coordinate system; the coordinate parameters (x, y) of the tube array refer to the coordinates of the center point of the tube array in the two-dimensional rectangular coordinate system, and the center point of the tube array refers to the center point of a circle or an ellipse; preferably, an inflection point of the two-dimensional image of the cooler tube array is used as the origin of the two-dimensional rectangular coordinate system.

[0097] In this embodiment, by fusing the two-dimensional image of the cooler tubes with the two-dimensional rectangular coordinate system, it is convenient to express the parameters of multiple tubes in the two-dimensional image of the cooler tubes and to calculate the data.

[0098] In one embodiment of the present application, Latin hypercube sampling is performed based on the coordinate parameters of each tube to reconstruct the two-dimensional image of the cooler tube, and the reconstructed two-dimensional image of the cooler tube is used as sample data. The Latin hypercube sampling is performed based on the coordinate parameters of each tube to reconstruct the two-dimensional image of the cooler tube, and the reconstructed two-dimensional image of the cooler tube is used as sample data, and multiple sample data are obtained, including the following S003a to S003c:

[0099] S003a, obtaining the coordinate parameters of each tube array.

[0100] S003b, adjusting the coordinate parameters of each tube array by Latin hypercube sampling method, obtaining a two-dimensional image of the cooler tube array after the tube coordinate parameters are adjusted based on the tube arrays after all the tube coordinate parameters are adjusted, and recording the two-dimensional image of the cooler tube array after the tube coordinate parameters are adjusted as a sample data.

[0101] S003c, repeatedly executing the method of Latin hypercube sampling to adjust the coordinate parameters of each tube array, obtaining a two-dimensional image of the cooler tube array after the tube coordinate parameters are adjusted based on the tube arrays after the tube coordinate parameters are adjusted, recording the two-dimensional image of the cooler tube array after the tube coordinate parameters are adjusted as a sample data, and obtaining at least K sample data.

[0102] Specifically, adjusting the coordinate parameters of each tube array refers to adjusting each parameter within a certain range. Specifically, the long axis parameter a of the tube array is adjusted within the range of 30 to 60, the long axis parameter b of the tube array is adjusted within the range of 30 to 60, the inclination angle parameter θ of the tube array is adjusted within the range of 0 to 2π, and the y in the coordinate parameters (x, y) of the tube array is adjusted within the range of 0 to 300.

[0103] In this embodiment, by limiting the adjustment range of each tube parameter, the tube parameter adjustment will not produce excessively extreme random data, which is more in line with practical applications.

[0104] In one embodiment of the present application, the numerical simulation based on the lattice Boltzmann method module calculates each sample data to obtain the Nusselt number and friction resistance coefficient of each sample data, including the following S004a to S004e:

[0105] S004a, select a sample data.

[0106] S004b, input the sample data into the lattice Boltzmann method module, start the lattice Boltzmann method module, and obtain the temperature field and velocity field of the sample data output by the lattice Boltzmann method module.

[0107] Specifically, Figure 2 Shown are the structural diagram of the original tube-in-tube arrangement, the temperature field schematic diagram of the original tube-in-tube arrangement, and the velocity field schematic diagram of the original tube-in-tube arrangement, as well as a more specific single-tube arrangement diagram, a single-tube velocity field schematic diagram, and a single-tube temperature field schematic diagram.

[0108] S004c, calculating the friction resistance coefficient of the sample data based on the velocity field of the sample data.

[0109] S004d, calculating the Nusselt number of the sample data based on the temperature field of the sample data.

[0110] S004e, returning to the step of selecting a sample data until each sample data has been selected once.

[0111] Specifically, the lattice Boltzmann method is a computational fluid dynamics method based on the mesoscopic simulation scale. It is between the microscopic molecular dynamics model and the macroscopic continuous model, so it has the advantages of simple description of fluid interactions, easy setting of complex boundaries, easy parallel calculation, and easy implementation of the program.

[0112] In this embodiment, the temperature field and the velocity field of the sample data are calculated respectively by the lattice Boltzmann method, so as to obtain the Nusselt number and the friction resistance coefficient of the sample data.

[0113] In one embodiment of the present application, the friction resistance coefficient of the sample data is obtained by calculating the velocity field of the sample data, including the following steps S005c to S007c:

[0114] S005c, analyzing the velocity field of the sample data to obtain parameters in the velocity field of the sample data, wherein the parameters in the velocity field of the sample data include an inlet average velocity parameter, a flow resistance parameter, a fluid density parameter, a fluid grid number parameter, an air kinematic viscosity parameter, and a Reynolds number.

[0115] S006c, inputting the obtained inlet average velocity parameter, air kinematic viscosity parameter and Reynolds number into Formula 1 to calculate the equivalent diameter in the velocity field of the sample data;

[0116]

[0117] Wherein, Re is the Reynolds number constant and Re is 150; U is the average inlet velocity; D is the equivalent diameter; υ is the kinematic viscosity of air.

[0118] S007c, input the obtained equivalent diameter, flow resistance parameter, fluid density parameter, fluid grid number parameter, and inlet average velocity parameter into Formula 2 to calculate the friction resistance coefficient in the velocity field of the sample data;

[0119]

[0120] Where f is the friction factor; F D is the flow resistance; n is the number of grids of the fluid; ρ is the density of the fluid; U is; D is the equivalent diameter.

[0121] In this embodiment, the friction factor of the sample data is obtained by calculating the velocity field of the sample data. The friction factor is an indicator for evaluating the heat exchange efficiency and can represent the flow resistance inside the heat exchanger.

[0122] In one embodiment of the present application, the Nusselt number of the adjusted two-dimensional image of the cooler tubes is obtained by calculating the temperature field of the adjusted two-dimensional image of the cooler tubes, including the following S005d to S006d:

[0123] S005d, analyzing the temperature field of the sample data to obtain parameters in the temperature field of the sample data, the sample data including heat flux parameters, surface area parameters, thermal conductivity, fluid average temperature parameters, wall temperature parameters and equivalent diameter.

[0124] S006d, input the obtained heat flux parameter, surface area parameter, thermal conductivity, fluid average temperature parameter, wall temperature parameter and equivalent diameter into Formula 3, and calculate the Nusselt number of the temperature field of the adjusted two-dimensional image of the cooler tube;

[0125]

[0126] Where Nu is the value of the Nusselt number; Φ is the heat flux; S is the surface area; κ is the thermal conductivity; is the average temperature of the fluid; T w is the wall temperature.

[0127] In this embodiment, the Nusselt number of the sample data is obtained by calculating the temperature field of the sample data. The Nusselt number is another indicator for evaluating the heat exchange efficiency. The Nusselt number can represent the heat exchange intensity.

[0128] In one embodiment of the present application, each sample data, each Nusselt number of each sample data and each friction resistance coefficient of each sample data are normalized, and each normalized sample data, each normalized Nusselt number of each sample data and each normalized friction resistance coefficient of each sample data are included in the database as a group of data until there are N groups of data in the database; N is a positive integer and N is greater than 1, and the following S005a to S005g are also included before:

[0129] S005a, select a sample data.

[0130] S005b, analyzing the sample data to obtain the Nusselt number and friction resistance coefficient of the sample data.

[0131] S005c, determining whether the Nusselt number of the sample data is within a Nusselt number threshold range.

[0132] S005d: If the Nusselt number of the sample data is not within the Nusselt number threshold range, the sample data is deleted.

[0133] S005e: If the Nusselt number of the sample data is within the Nusselt number threshold range, determine whether the friction resistance coefficient of the sample data is within the friction resistance coefficient threshold range.

[0134] S005f, if the friction resistance coefficient of the sample data is not within the friction resistance coefficient threshold range, the sample data is deleted.

[0135] S005g, returning to the step of selecting a sample data until each sample data has been selected once.

[0136] Specifically, the Nusselt number threshold range refers to the heat exchange intensity being within a set range. If it is lower than the Nusselt number threshold range, the heat exchange intensity is too weak. If it is higher than the Nusselt number threshold range, the heat exchange intensity is too strong. Both too weak and too strong heat exchange intensity are manifestations of inefficiency.

[0137] The friction resistance coefficient threshold range refers to the resistance to the flow inside the heat exchanger. If it is lower than the friction resistance coefficient threshold, the resistance to the flow inside the heat exchanger is too small, resulting in too fast a flow rate and reduced heat transfer efficiency. If it is higher than the friction resistance coefficient threshold, the resistance to the flow inside the heat exchanger is too large, resulting in too high a flow rate and reduced heat transfer efficiency.

[0138] In this embodiment, the Nusselt number and the friction coefficient of the sample data are screened simultaneously, and the sample data is retained only when both the Nusselt number and the friction coefficient meet the standards, so that the sample data has stability and reliability.

[0139] In one embodiment of the present application, the M groups of data in the database are used as training samples to adjust the convolutional neural network model to obtain the adjusted convolutional neural network model; M is a positive integer and M is less than N, including the following S006a to S006g:

[0140] S006a, create a training database and a verification database.

[0141] S006b, select M groups of data in the database to be included in the training database, and select N groups of data in the database to be included in the verification database.

[0142] S006c, select a set of data in the training database.

[0143] S006d, analyzing the group of data to obtain the normalized sample data, the Nusselt number of the normalized sample data, and the friction resistance coefficient of the normalized sample data.

[0144] S006e, using the normalized sample data as the input of the convolutional neural network model, using the Nusselt number of the normalized sample data and the friction resistance coefficient of the normalized sample data as the output of the convolutional neural network model, adjusting the convolutional neural network model to obtain an adjusted convolutional neural network model.

[0145] S006f, returning a set of data in the selected training database until each set of data in the training database has been selected once.

[0146] S006g, obtain the adjusted convolutional neural network model.

[0147] Specifically, the training database and the verification database are independent databases, and there is no overlap or duplication of data between the two.

[0148] In this embodiment, the convolutional neural network model is carefully trained by using the data in the training database so that the convolutional neural network model has the function of predicting the Nusselt number and friction coefficient of the sample data after inputting a sample data.

[0149] In one embodiment of the present application, the NM group data in the database is used as a verification sample to perform error calculation on the prediction results output by the convolutional neural network model to obtain multiple adjusted error results of the convolutional neural network model prediction results, including the following S008a to S006g:

[0150] S008a, selecting a set of data in the verification database.

[0151] S008b, analyzing the group of data to obtain the normalized Nusselt number of the sample data and the normalized friction resistance coefficient of the sample data.

[0152] S008c, calculating the error between the normalized Nusselt number of the sample data and the predicted Nusselt number to obtain a first error result.

[0153] S008d, calculating the error between the friction resistance coefficient of the normalized sample data and the predicted friction resistance coefficient to obtain a second error result.

[0154] S008e, returning a set of data in the selected verification database until each set of data in the verification database has been selected once.

[0155] Specifically, multiple sets of known data in the verification database are used to analyze the errors of the predicted results obtained by the convolutional neural network model, and then the errors of the Nusselt number and the friction resistance coefficient are obtained respectively.

[0156] In this embodiment, a group of data in a verification database is obtained, and sample data in the group of data is used as input and input into a convolutional neural network model to obtain a predicted Nusselt number and a predicted friction resistance coefficient output by the convolutional neural network model. A first error result is obtained by performing an error calculation between the predicted Nusselt number and the measured Nusselt number of the sample data, and a second error result is obtained by performing an error calculation between the predicted friction resistance coefficient and the sample data of the sample data.

[0157] In one embodiment of the present application, the environment of the reinforcement learning algorithm is constructed, and the environment of the reinforcement learning algorithm is integrated into the adjusted convolutional neural network model to obtain the optimal convolutional neural network model, including:

[0158] The database is used as a training sample for the reinforcement learning algorithm to optimize the adjusted convolutional neural network model and obtain the optimal convolutional neural network model.

[0159] Specifically, in reinforcement learning, intelligent agents need a proper environment to learn and interact. However, using the Lattice Boltzmann Method (LBM) as the environment for interaction is a time-consuming process, and each interaction requires the intelligent agent to wait for the LBM calculation to complete, which takes an average of 20 minutes. Therefore, it is crucial to use a proxy model to replace the LBM calculation.

[0160] Deterministic Policy Gradient (DDPG) can solve this problem about continuous action space. It represents the policy as a policy function, mapping the state to a deterministic action. When the policy is a deterministic policy, the action value function obtained by calculating Formula 4 using the Bellman equation is as follows.

[0161]

[0162] Among them, Q π (S t ,a t ) refers to the state S t and action a t The action-value function in the case of Refers to the possible next state S t1 and action a t1 Expectation; r(s t ,a t ) refers to the state S t Execute action a t The immediate reward obtained after the event; γ refers to the discount factor; E π [Q π (s t1 ,a t1 )] refers to the strategy E π For the next state St1 and action a t1 The value function of .

[0163] Therefore, the calculation of internal expectations is avoided, and only external expectations need to be calculated based on the environment. In other words, the action-state value function Q is only related to the environment, but not to the solution of the action.

[0164] Using the Q value, we can iteratively solve the optimal strategy through the following formula 5;

[0165]

[0166] When training a neural network, if the same neural network is used to represent the target network and the current update network (online), the learning process will be very unstable. This is because the same network parameters need to be used to calculate the network gradient while frequently updating the gradient. The solution of DDPG is to create two copies of the neural network for the critics, namely the "critic" Q neural network Q(s,a|θ) and the "critic" Q neural network Q(s,a|θ). Q ) and the “actor” neural network μ θ (s|θ μ ), that is, two independent target networks Q'(s,a|θ are created respectively. Q' ) and μ' θ (s|θ μ' )

[0167] “Actor” Neural Network:

[0168]

[0169] “Commentator” Q-neural network:

[0170]

[0171] In this embodiment, the process of reinforcement learning is as follows:

[0172] 1. Initialize memory D with capacity N.

[0173] 2. Initialize the Q network using randomly generated weights ω.

[0174] 3. Initialize the target Q network using weights ω-=ω.

[0175] 4. Loop through episodes = 1, 2, ..., M

[0176] 4.1. Initialize the initial state S1.

[0177] 4.2. Loop through steps = 1, 2, ..., T

[0178] 4.2.1. Generate actions using the -greedy strategy: select a random action with probability , or select at = maxaQ(St,a;ω).

[0179] 4.2.2. Execute the action, receive the reward and the new state St+1.

[0180] 4.2.3. Store the transformed samples (St, at, rt, St+1) in D.

[0181] 4.2.4. Randomly draw a mini-batch of transformations (Sj,aj,rj,Sj+1) from D.

[0182] 4.2.5. If step j+1 is terminated, set yj=rj, otherwise, set yj=rj+γmaxa′Q(St+1,a′;ω-).

[0183] 4.2.6. Update ω(yj-Q(St,aj;ω))2 using gradient descent.

[0184] 4.3. Update the target Q network every C steps, ω-=ω.

[0185] Specifically, the parameters listed in the original sample data are shown in Table 1;

[0186]

[0187] The parameters listed in the optimized sample data are shown in Table 2;

[0188]

[0189]

[0190] In this embodiment, the tube design parameters of the original cooler model are shown in Table 1. The Nusselt number obtained by numerical calculation is 19.39, the friction coefficient is 0.89, and the tube design parameters of the optimized cooler model are shown in Table 2. The two-dimensional shape and arrangement of the tubes in the sample data after reinforcement learning optimization are shown in Table 2. Figure 3 As shown in the figure, the optimized Nusselt number is 24.47 and the friction coefficient is 2.4. The optimized Nusselt number is 26.2% higher than the original one.

[0191] The technical features of the above-described embodiments may be arbitrarily combined, and the execution order of the method steps is not limited. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0192] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A coupled intelligent optimization method, characterized in that: The coupled intelligent optimization method comprises: Acquire a two-dimensional image of a cooler tube, wherein the two-dimensional image of the cooler tube includes a plurality of tubes; The two-dimensional image of the cooler tubes is coupled with the two-dimensional rectangular coordinate system to obtain the coordinate parameters of each tube in the two-dimensional image of the cooler tubes; Latin hypercube sampling is performed based on the coordinate parameters of each tube to reconstruct the two-dimensional image of the cooler tube, and the reconstructed two-dimensional image of the cooler tube is used as sample data. The Latin hypercube sampling is performed based on the coordinate parameters of each tube to reconstruct the two-dimensional image of the cooler tube, and the reconstructed two-dimensional image of the cooler tube is used as sample data, and multiple sample data are obtained; Based on the lattice Boltzmann method module, the numerical simulation is performed to calculate each sample data, and the Nusselt number and friction coefficient of each sample data are obtained; Each obtained sample data, each Nusselt number of the sample data, and each friction resistance coefficient of the sample data are normalized, and each obtained normalized sample data, each normalized Nusselt number of the sample data, and each normalized friction resistance coefficient of the sample data are included in the database as a group of data until there are N groups of data in the database; N is a positive integer and N is greater than 1; The M groups of data in the database are used as training samples to adjust the convolutional neural network model to obtain the adjusted convolutional neural network model; M is a positive integer and M is less than N; Inputting the two-dimensional image of the cooler tubes into the adjusted convolutional neural network model, starting the adjusted convolutional neural network model, and obtaining the prediction results output by the adjusted convolutional neural network model, wherein the prediction results include the predicted Nusselt number and the predicted friction resistance coefficient; The NM group data in the database is used as a verification sample to perform error calculation on the prediction results output by the convolutional neural network model, and the error results of the prediction results of multiple adjusted convolutional neural network models are obtained; Determine whether multiple error results are all less than an error threshold; If multiple error results are all less than the error threshold, then build an environment for the reinforcement learning algorithm, integrate the environment of the reinforcement learning algorithm into the adjusted convolutional neural network model, and obtain the optimal convolutional neural network model; Inputting the two-dimensional image of the cooler tubes into the optimal convolutional neural network model, starting the optimal convolutional neural network model, and obtaining the optimal tube parameters, optimal Nusselt number, and optimal friction resistance coefficient output by the optimal convolutional neural network model; If any error result is greater than or equal to the threshold, the method returns to execute the method of adjusting the convolutional neural network model using a part of the database as a training sample to obtain an adjusted convolutional neural network model.

2. The coupled intelligent optimization method according to claim 1, characterized in that: The two-dimensional image of the cooler tubes is coupled with the two-dimensional rectangular coordinate system to obtain the coordinate parameters of each tube in the two-dimensional image of the cooler tubes, including: A random pixel point in the two-dimensional image of the cooler tubes is used as the origin of the coordinates in the two-dimensional rectangular coordinate system; Taking the coordinate origin as the anchor point, the coordinate parameters of each tube array in the two-dimensional plane rectangular coordinate system are obtained, and the coordinate parameters of the tube array include the long axis parameter a of the tube array, the long axis parameter b of the tube array, the inclination angle parameter θ of the tube array, and the coordinate parameters (x, y) of the tube array.

3. The coupled intelligent optimization method according to claim 2, characterized in that: The Latin hypercube sampling is performed based on the coordinate parameters of each tube to reconstruct the two-dimensional image of the cooler tube, and the reconstructed two-dimensional image of the cooler tube is used as sample data. The Latin hypercube sampling is performed based on the coordinate parameters of each tube to reconstruct the two-dimensional image of the cooler tube, and the reconstructed two-dimensional image of the cooler tube is used as sample data, and multiple sample data are obtained, including: Get the coordinate parameters of each tube column; The coordinate parameters of each tube array are adjusted by a Latin hypercube sampling method, and a two-dimensional image of the cooler tube array after the tube coordinate parameters are adjusted is obtained based on the tube arrays after the tube coordinate parameters are adjusted, and the two-dimensional image of the cooler tube array after the tube coordinate parameters are adjusted is recorded as a sample data; Repeat the method of Latin hypercube sampling to adjust the coordinate parameters of each tube array, obtain a two-dimensional image of the cooler tube array after the tube coordinate parameters are adjusted based on all the tube arrays after the tube coordinate parameters are adjusted, record the two-dimensional image of the cooler tube array after the tube coordinate parameters are adjusted as a sample data, and obtain at least K sample data.

4. The coupled intelligent optimization method according to claim 3, characterized in that: The numerical simulation based on the lattice Boltzmann method module calculates each sample data to obtain the Nusselt number and friction resistance coefficient of each sample data, including: Select a sample data; The sample data is input into the lattice Boltzmann method module, the lattice Boltzmann method module is started, and the temperature field and velocity field of the sample data output by the lattice Boltzmann method module are obtained; The friction resistance coefficient of the sample data is obtained by calculating the velocity field of the sample data; The Nusselt number of the sample data is obtained by calculating the temperature field of the sample data; Return to the step of selecting a sample data until each sample data has been selected once.

5. The coupled intelligent optimization method according to claim 4, characterized in that: The calculating of the friction resistance coefficient of the sample data based on the velocity field of the sample data comprises: Analyze the velocity field of the sample data to obtain parameters in the velocity field of the sample data, wherein the parameters in the velocity field of the sample data include an inlet average velocity parameter, a flow resistance parameter, a fluid density parameter, a fluid grid number parameter, an air kinematic viscosity parameter, and a Reynolds number; The obtained inlet average velocity parameter, air kinematic viscosity parameter and Reynolds number are input into Formula 1 to calculate the equivalent diameter in the velocity field of the sample data; Where Re is the Reynolds number constant and Re is 150; U is the average inlet velocity; D is the equivalent diameter; υ is the kinematic viscosity of air; The obtained equivalent diameter, flow resistance parameter, fluid density parameter, fluid grid number parameter, and inlet average velocity parameter are input into Formula 2 to calculate the friction resistance coefficient in the velocity field of the sample data; Where f is the friction factor; F D is the flow resistance; n is the number of grids of the fluid; ρ is the density of the fluid; U is; D is the equivalent diameter.

6. The coupled intelligent optimization method according to claim 5, characterized in that: The Nusselt number of the sample data is obtained by calculating the temperature field of the sample data, including: Analyze the temperature field of the sample data to obtain parameters in the temperature field of the sample data, the sample data including heat flux parameters, surface area parameters, thermal conductivity, fluid average temperature parameters, wall temperature parameters and equivalent diameter; The obtained heat flux parameter, surface area parameter, thermal conductivity, fluid average temperature parameter, wall temperature parameter and equivalent diameter are input into Formula 3 to calculate the Nusselt number of the temperature field of the adjusted two-dimensional image of the cooler tube column; Where Nu is the value of the Nusselt number; Φ is the heat flux; S is the surface area; κ is the thermal conductivity; is the average temperature of the fluid; T w is the wall temperature.

7. The coupled intelligent optimization method according to claim 6, characterized in that: Each obtained sample data, each Nusselt number of the sample data and each friction resistance coefficient of the sample data are normalized, and each obtained normalized sample data, each normalized Nusselt number of the sample data and each normalized friction resistance coefficient of the sample data are included in the database as a group of data until there are N groups of data in the database; N is a positive integer and N is greater than 1, and the previous ones also include: Select a sample data; Analyze the sample data to obtain the Nusselt number and friction coefficient of the sample data; Determine whether the Nusselt number of the sample data is within the Nusselt number threshold range; If the Nusselt number of the sample data is not within the Nusselt number threshold range, the sample data is deleted; If the Nusselt number of the sample data is within the Nusselt number threshold range, determining whether the friction resistance coefficient of the sample data is within the friction resistance coefficient threshold range; If the friction resistance coefficient of the sample data is not within the friction resistance coefficient threshold range, the sample data is deleted; Return to the step of selecting a sample data until each sample data has been selected once.

8. The coupled intelligent optimization method according to claim 7, characterized in that: The convolutional neural network model is adjusted using the M groups of data in the database as training samples to obtain the adjusted convolutional neural network model; M is a positive integer and M is less than N, including: Create training and validation databases; Select M groups of data in the database to be included in the training database, and select N groups of data in the database to be included in the verification database; Select a set of data from the training database; Analyze the group of data to obtain the normalized sample data, the Nusselt number of the normalized sample data, and the friction resistance coefficient of the normalized sample data; The normalized sample data is used as the input of the convolutional neural network model, the Nusselt number of the normalized sample data and the friction coefficient of the normalized sample data are used as the output of the convolutional neural network model, and the convolutional neural network model is adjusted to obtain an adjusted convolutional neural network model; Returning a set of data in the selected training database until each set of data in the training database has been selected once; Get the adjusted convolutional neural network model.

9. The coupled intelligent optimization method according to claim 8, characterized in that: The NM group data in the database is used as a verification sample to perform error calculation on the prediction results output by the convolutional neural network model to obtain error results of multiple adjusted convolutional neural network model prediction results, including: Select a set of data in the verification database; Analyze the set of data to obtain the normalized Nusselt number of the sample data and the normalized friction coefficient of the sample data; Calculate the error between the normalized Nusselt number of the sample data and the predicted Nusselt number to obtain a first error result; Calculating the error between the friction resistance coefficient of the normalized sample data and the predicted friction resistance coefficient to obtain a second error result; A set of data in the selected verification database is returned until each set of data in the verification database has been selected once.

10. The coupled intelligent optimization method according to claim 9, characterized in that: The step of building an environment for a reinforcement learning algorithm and integrating the environment for the reinforcement learning algorithm into the adjusted convolutional neural network model to obtain an optimal convolutional neural network model includes: The database is used as a training sample for the reinforcement learning algorithm to optimize the adjusted convolutional neural network model and obtain the optimal convolutional neural network model.

Images