CN105389442A - Reverse design method for coupling genetic algorithm, neural network and numerical simulation - Google Patents

Abstract

The present invention discloses a reverse design method for coupling a genetic algorithm, a neural network and numerical simulation. The reverse design method comprises: determining a design variable and a design target according to a design object; using a sampling method to generate a neural network training sample, and training the neural network; using crossing and variation processes of the genetic algorithm to simultaneously obtain a plurality of sets of design variable values meeting design requirements, and searching an individual meeting the design requirements from the design variable values; using the neural network and a CFD method to calculate fitness of a new individual; using the neural network to predict a design target value of the new individual, and if the design target value reaches a set threshold value, using a computational fluid mechanics method to calculate a real design target value of the new individual; and using a selective operation to generate a new population: if the new population accords with a convergent standard, ending the design process, otherwise, continuously performing the crossing and variation processes to generate a new population. Compared with a method only using combination of the genetic algorithm and the numerical simulation, the reverse design method disclosed by the present invention reduces 42.1% of the calculation amount of reverse design while ensuring the convergence of the reverse design.

Description

The reverse design method of a kind of genetic algorithm that is coupled, neural network and numerical simulation

Technical field

The present invention relates to the Indoor Environmental Design field of building, particularly the indoor environment reverse design method of a kind of genetic algorithm that is coupled, neural network and numerical simulation.

Background technology

Genetic algorithm is a kind of typical optimization method, it regards the evolutionary process of biotic population as optimizing process, the possible solution of each variable is expressed as a chromosomal fragment, the possible solution of all variablees is arranged in a complete chromosome in order, and every bar chromosome represents the body one by one in evolutionary process.Compared with other optimized algorithm, genetic algorithm more easily explores whole design space, and finds globally optimal solution.Numerical simulation technology can obtain detailed indoor environment data, compared with experimental measurement method, simple to operate, with low cost, and Fluid Mechanics Computation (CFD) model is the computation model be most widely used in indoor environment field.

Use genetic algorithm to combine with CFD, when carrying out reverse design to indoor environment, by intersection, the new individuality of mutation process generation of genetic algorithm, utilize CFD to obtain the design object value of new individuality.But more new individuality may be produced in design process, therefore can produce more CFD computation process, cause the calculated amount of this reverse design method large.

For reducing reverse design calculated amount, can end user's artificial neural networks as the alternative model of CFD model, be applied to indoor environment reverse design.When using genetic algorithm to combine with neural network to carry out reverse design, utilizing the intersection of genetic algorithm, mutation process to produce new individual, utilizing the design object value of the new individuality of neural computing.Due to neural computing time much smaller than CFD computing time, the method can reach the object reducing reverse design calculated amount.But the method, by the impact of neural network prediction error, may cause design process not restrain.Therefore, be necessary to study new reverse design method, ensureing that method for designing is constringent while, reduce reverse design calculated amount.

Summary of the invention

For above-mentioned prior art and Problems existing, the present invention proposes the reverse design method of a kind of genetic algorithm that is coupled, neural network and numerical simulation, on the basis ensureing reverse design procedure converges, by coupling genetic algorithm, neural network and method for numerical simulation, reduce the calculated amount of reverse design method.

The present invention proposes the reverse design method of a kind of genetic algorithm that is coupled, neural network and numerical simulation, the method comprises the following steps:

Step 1, according to design object determination design variable and design object;

Step 2, the use methods of sampling, produce train samples, train neural network;

Step 3, utilize the intersection of genetic algorithm, mutation process simultaneously to obtain design variable value that many groups meet designing requirement, therefrom search for the individuality met design requirement; Use neural network and CFD method to calculate new individual fitness, i.e. design object value, concrete computation process is as follows: the design object value first using the new individuality of neural network prediction, neural network be input as design variable value, export as design object value; If design object value reaches the threshold value of setting, Fluid Mechanics Computation method is used to calculate new individual actual design desired value, according to the boundary condition of design variable value determination design object, hydrodynamic methods is used to calculate the distribution of the speed, temperature etc. of design object inside, and then design object value can be obtained, be individual fitness.

Step 4, by non-dominated ranking method, the individuality in population to be sorted, and use tournament algorithm to select, utilize and select operation to produce new population: if new population meets convergence, design process terminates, otherwise proceed cross and variation process, produce new population.

Compared with the method only using genetic algorithm and numerical simulation to combine, coupling genetic algorithm, neural network and method for numerical simulation, ensureing that reverse design is constringent while, reduce reverse design calculated amount 42.1%.

Accompanying drawing explanation

Fig. 1 is the algorithm model figure of the reverse design method of a kind of genetic algorithm that is coupled of the present invention, neural network and numerical simulation;

Fig. 2 is the Blay model schematic of the specific embodiment of the invention; The geometry of (a) Blay model, (b) stress and strain model;

Fig. 3 is the change curve of target function value with genetic algebra;

Fig. 4 is the calculated amount growth curve figure in reverse design process;

When Fig. 5 is for using genetic algorithm to combine with neural network, objective function is with genetic algebra change curve;

When Fig. 6 is for using genetic algorithm to combine with neural network, CFD, objective function change curve;

When Fig. 7 is for using genetic algorithm to combine with neural network, CFD, calculated amount is with genetic algebra change curve.

Embodiment

Below in conjunction with the drawings and the specific embodiments, be described in further detail technical scheme of the present invention.

The object of reverse design is respectively Blay model and passenger plane Cabin model.As shown in Figure 2, be geometry and the stress and strain model of Blay model.The external structure of Blay model is square, and be of a size of 1.04 × 1.04m, comprise an entrance and an outlet, entrance opening dimension is 0.018m, and outlet size is 0.022m.Inlet velocity is 0.57m/s, and temperature in is 15 DEG C, and wind direction is horizontal direction, and surrounding uses the wall boundary condition without slippage, and wall surface temperature is all set to constant temperature, and monitoring site is near outlet.

The specific embodiment of the invention is described below:

1, the reverse design that combines with CFD of genetic algorithm

Use genetic algorithm to combine with CFD and reverse design is carried out to Blay model.Design variable is inlet velocity and temperature, and design object is monitoring point place speed and temperature.According to numerical result, the velocity magnitude obtaining monitoring point place is 0.1325m/s, and temperature level is 17.67 DEG C, and then obtains the objective function of Blay model reverse design:

$F_{Blay}=\sqrt{{\left(\frac{V-0.1325}{0.1325}\right)}^2+{\left(\frac{T-17.67}{17.67}\right)}^2}\backslash *MERGEFORMAT---\left(1\right)$

The crossover probability of genetic algorithm is 0.8, and mutation probability is 0.1, and maximum genetic algebra was 100 generations, and the condition of convergence of reverse design is FBlay=0.As shown in Figure 3, for target function value is with the change curve of genetic algebra, when calculating for the 33rd generation, reverse design procedure converges, the solution obtained is (0.57,15), and namely inlet velocity is 0.57m/s, and temperature in is 15 DEG C.

Calculated amount growth curve in reverse design process as shown in Figure 4, is calculated amount growth curve when using genetic algorithm to carry out reverse design, uses the case quantity of CFD to represent calculated amount size.When obtaining optimum solution, calculating 235 case altogether, namely having performed 235 CFD and calculating.

2, the reverse design that combines with neural network of genetic algorithm

The method using genetic algorithm to combine with neural network carries out reverse design to Blay model, the data of end user's artificial neural networks to Blay model monitoring point place are predicted, the speed being input as entrance of neural network and temperature value, export the speed into monitoring point place and temperature value.Use Latin hypercube sampling (LHS) method, produce the combination of 80 groups of inlet velocities and temperature as input parameter, these input parameters are used to carry out CFD calculating as boundary condition to Blay model, often organized the output parameter that input parameter is corresponding, the i.e. speed at monitoring point place and temperature value, the output parameter often organizing input parameter and correspondence thereof forms a complete training sample, all training samples are used to train neural network, the neural network trained is combined with genetic algorithm, is applied to the reverse design of Blay model.

As shown in Figure 5, in reverse design process, when using genetic algorithm to combine with neural network, objective function is with genetic algebra change curve; Calculate the incipient stage, objective function curve declines very fast, but remains unchanged after calculating for the 20th generation.When genetic algorithm calculated for 100 generation, target function value does not reach 0.Therefore, when using the method, reverse design does not restrain.

3, the reverse design that combines with neural network, CFD of genetic algorithm

To use and the method that combines with neural network, CFD of genetic algorithm carries out reverse design to Blay model.First the neural network that trains is used to predict new individual desired value, and according to the predictor calculation F of design object _blayvalue, if this value is lower than 0.1, uses CFD to calculate this individuality, obtains the real F of this individuality _blayvalue.

Fig. 6 be target function value with genetic algebra change curve, when calculating for the 34th generation, reverse design procedure converges, the solution obtained is (0.57,15).

As shown in Figure 7, when combining with neural network, CFD for using genetic algorithm, calculated amount is with genetic algebra change curve, and reverse design process need calculate 136 case, and compare with the method that CFD combines with use genetic algorithm, calculated amount decreases 42.1%.

Claims (1)

1. a reverse design method for genetic algorithm, neural network and the numerical simulation of being coupled, it is characterized in that, the method comprises the following steps:

Step 1, according to design object determination design variable and design object;

Step 2, the use methods of sampling, produce train samples, train neural network;

Step 3, utilize the intersection of genetic algorithm, mutation process simultaneously to obtain design variable value that many groups meet designing requirement, therefrom search for the individuality met design requirement; Neural network and CFD method is used to calculate new individual fitness, i.e. design object value; Concrete computation process is as follows: first use the design object value that neural network prediction is new individual, neural network be input as design variable value, export as design object value; If design object value reaches the threshold value of setting, Fluid Mechanics Computation method is used to calculate new individual actual design desired value, according to the boundary condition of design variable value determination design object, hydrodynamic methods is used to calculate the distribution of the speed, temperature etc. of design object inside, and then design object value can be obtained, be individual fitness;

Step 4, by non-dominated ranking method, the individuality in population to be sorted, and use tournament algorithm to select, utilize and select operation to produce new population: if new population meets convergence, design process terminates, otherwise proceed cross and variation process, produce new population.