CN117150896A - Supercritical fluid heat transfer coefficient prediction method based on interpretable machine learning - Google Patents
Supercritical fluid heat transfer coefficient prediction method based on interpretable machine learning Download PDFInfo
- Publication number
- CN117150896A CN117150896A CN202311096466.1A CN202311096466A CN117150896A CN 117150896 A CN117150896 A CN 117150896A CN 202311096466 A CN202311096466 A CN 202311096466A CN 117150896 A CN117150896 A CN 117150896A
- Authority
- CN
- China
- Prior art keywords
- heat transfer
- model
- supercritical
- particle
- neural network
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000012546 transfer Methods 0.000 title claims abstract description 112
- 239000012530 fluid Substances 0.000 title claims abstract description 37
- 238000000034 method Methods 0.000 title claims abstract description 24
- 238000010801 machine learning Methods 0.000 title claims abstract description 19
- 239000002245 particle Substances 0.000 claims abstract description 50
- 238000012360 testing method Methods 0.000 claims abstract description 27
- 230000000704 physical effect Effects 0.000 claims abstract description 24
- 238000012549 training Methods 0.000 claims abstract description 24
- 238000003062 neural network model Methods 0.000 claims abstract description 18
- 238000005457 optimization Methods 0.000 claims abstract description 18
- 230000007246 mechanism Effects 0.000 claims abstract description 17
- 238000013528 artificial neural network Methods 0.000 claims abstract description 14
- 230000000694 effects Effects 0.000 claims description 43
- 230000006399 behavior Effects 0.000 claims description 24
- 230000008859 change Effects 0.000 claims description 23
- 230000001133 acceleration Effects 0.000 claims description 15
- 230000006870 function Effects 0.000 claims description 14
- 101001095088 Homo sapiens Melanoma antigen preferentially expressed in tumors Proteins 0.000 claims description 11
- 102100037020 Melanoma antigen preferentially expressed in tumors Human genes 0.000 claims description 11
- 238000004364 calculation method Methods 0.000 claims description 8
- 238000004458 analytical method Methods 0.000 claims description 4
- 238000011156 evaluation Methods 0.000 claims description 4
- 235000003805 Musa ABB Group Nutrition 0.000 claims description 3
- 241000013557 Plantaginaceae Species 0.000 claims description 3
- 235000015266 Plantago major Nutrition 0.000 claims description 3
- 238000002790 cross-validation Methods 0.000 claims description 3
- 230000008569 process Effects 0.000 claims description 3
- 238000012545 processing Methods 0.000 claims description 3
- 238000012795 verification Methods 0.000 claims description 3
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 claims description 3
- 230000002159 abnormal effect Effects 0.000 claims description 2
- 230000004913 activation Effects 0.000 claims description 2
- 230000006978 adaptation Effects 0.000 claims description 2
- 238000004140 cleaning Methods 0.000 claims description 2
- 230000001149 cognitive effect Effects 0.000 claims description 2
- 238000010438 heat treatment Methods 0.000 claims description 2
- 210000002569 neuron Anatomy 0.000 claims description 2
- 238000010606 normalization Methods 0.000 claims description 2
- 230000000630 rising effect Effects 0.000 claims description 2
- 238000010200 validation analysis Methods 0.000 claims description 2
- 238000010276 construction Methods 0.000 claims 1
- 238000007781 pre-processing Methods 0.000 claims 1
- 230000006866 deterioration Effects 0.000 description 13
- 230000004907 flux Effects 0.000 description 10
- 230000001965 increasing effect Effects 0.000 description 6
- 230000002542 deteriorative effect Effects 0.000 description 4
- 238000010586 diagram Methods 0.000 description 4
- 238000005728 strengthening Methods 0.000 description 4
- 230000007423 decrease Effects 0.000 description 3
- 230000002708 enhancing effect Effects 0.000 description 3
- 238000009826 distribution Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000008092 positive effect Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000003090 exacerbative effect Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000013213 extrapolation Methods 0.000 description 1
- 230000002068 genetic effect Effects 0.000 description 1
- 230000000644 propagated effect Effects 0.000 description 1
- 238000005057 refrigeration Methods 0.000 description 1
- 239000000126 substance Substances 0.000 description 1
- 230000007704 transition Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/004—Artificial life, i.e. computing arrangements simulating life
- G06N3/006—Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0499—Feedforward networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/086—Learning methods using evolutionary algorithms, e.g. genetic algorithms or genetic programming
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2113/00—Details relating to the application field
- G06F2113/08—Fluids
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
- G06F2119/08—Thermal analysis or thermal optimisation
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Evolutionary Computation (AREA)
- Health & Medical Sciences (AREA)
- General Physics & Mathematics (AREA)
- Software Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Biophysics (AREA)
- General Health & Medical Sciences (AREA)
- Mathematical Physics (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Medical Informatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Computer Hardware Design (AREA)
- Geometry (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Physiology (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
A supercritical fluid heat transfer coefficient prediction method based on interpretable machine learning is characterized in that a feedforward-back propagation fully-connected neural network model is built in an offline stage, physical data are obtained through REFPROP physical packages, a dimensionless number related to supercritical heat transfer is calculated and built to obtain a data set, and a particle swarm optimization algorithm is adopted to optimize the neural network super-parameters; and in the online stage, the generalization capability is tested on a test set through the trained fully-connected neural network model, and an interpretable algorithm is added to analyze the feature importance degree, so that the supercritical fluid heat transfer coefficient prediction is realized. According to the invention, fluid physical properties under different experimental conditions are obtained through REFPORP physical property packages, dimensionless numbers related to supercritical heat transfer are calculated, and a data set is constructed; building a feedforward-back propagation fully-connected neural network, further adopting a particle swarm optimization algorithm to optimize a model and training and evaluating; finally, an interpretable algorithm is used to interpret the model predictive mechanism in combination with the supercritical heat transfer mechanism.
Description
Technical Field
The invention relates to a technology in the field of supercritical fluid, in particular to a supercritical fluid heat transfer coefficient prediction method based on interpretable machine learning.
Background
Supercritical fluids are considered to be working fluids that improve the thermal efficiency of power cycles and energy conversion due to their unique thermodynamic and transport properties, and have been used in the engineering fields of aerospace, nuclear power, solar energy, refrigeration, geothermal energy, chemical technology, and the like. Physical properties of supercritical fluids undergo drastic nonlinear changes around pseudo-critical temperatures, including drastic increases in specific heat, drastic decreases in dynamic viscosity and density, and the like. While non-uniformity in density results in the generation of both buoyancy and acceleration effects. Ultimately leading to unpredictable heat transfer behavior, including heat transfer enhancement and heat transfer deterioration. The special heat transfer behavior of the supercritical fluid, especially the heat transfer deterioration, brings great challenges to the design and safety of an application scene system, and is therefore important to the prediction of the heat transfer coefficient of the supercritical fluid.
The most common method for predicting the supercritical heat transfer coefficient is currently empirical correlation prediction. Most of the correlations are based on the Dittus-Boelter formula, and different parameters are introduced for describing special effects in supercritical heat transfer according to different theories and interpretations. However, empirical correlation relies on experimental initial conditions including pressure, heat flux density, mass flow, etc., and extrapolation is not adequate and accurate, and it is difficult to predict heat transfer deterioration behavior. In addition, in recent years, due to the rapid development of artificial intelligence, part of research uses an artificial neural network to predict the supercritical heat transfer coefficient, and although higher precision is achieved, a prediction mechanism of a model is not clear, and the interpretation is poor.
Disclosure of Invention
Aiming at the problems of low experience-associated prediction precision and poor interpretability of an artificial neural network model in the prior art, the invention provides a supercritical fluid heat transfer coefficient prediction method based on interpretable machine learning, fluid physical properties under different experimental conditions are obtained through REFPORP physical property packages, dimensionless numbers related to supercritical heat transfer are calculated, and a data set is constructed; building a feedforward-back propagation fully-connected neural network, further adopting a particle swarm optimization algorithm to optimize a model and training and evaluating; and finally, explaining the model by using an interpretable algorithm to obtain the feature importance change of the model prediction in different working conditions and different heat transfer behaviors, and explaining a model prediction mechanism by combining a supercritical heat transfer mechanism.
The invention is realized by the following technical scheme:
the invention relates to a supercritical fluid heat transfer coefficient prediction method based on interpretable machine learning, which comprises the steps of constructing a feedforward-back propagation fully-connected neural network model in an off-line stage, acquiring physical data through REFPROP physical packages, calculating dimensionless numbers related to supercritical heat transfer to obtain a data set, and optimizing neural network super-parameters by adopting a particle swarm optimization algorithm; and in the online stage, the generalization capability is tested on a test set through the trained fully-connected neural network model, and an interpretable algorithm is added to analyze the feature importance degree, so that the supercritical fluid heat transfer coefficient prediction is realized.
The data set is obtained through REFPORP physical property bags, fluid physical properties under different experimental conditions are obtained, dimensionless numbers related to supercritical heat transfer are calculated, and a reliable data set is obtained through pretreatment.
The invention relates to a system for realizing the method, which comprises the following steps: the system comprises a supercritical heat transfer related dimensionless number calculation and arrangement module, a neural network model optimization and training evaluation module and a neural network model interpretation module, wherein: the supercritical heat transfer related dimensionless number calculation and arrangement module obtains fluid physical properties under different experimental conditions through a REFPROP physical property bag, calculates the supercritical heat transfer related dimensionless number, and performs pretreatment to obtain a reliable data set for model training and testing; the neural network model optimization and training evaluation module optimizes the super parameters such as the hidden layer number, the hidden layer node number, the learning rate and the like of the model by using a particle swarm optimization algorithm, substitutes the optimized super parameters into the neural network for training and testing, and carries out error calculation on a prediction result and an experimental value; the neural network model interpretation module obtains the changes of model characteristic parameter importance degree of different working conditions and different heat transfer behaviors according to the characteristic importance degree analysis result of the SHAP interpretable algorithm on the model, and interprets a model prediction mechanism by combining a supercritical heat transfer mechanism.
Technical effects
The invention uses particle swarm optimization algorithm to optimize feedforward-counter propagation fully connected neural network, and builds a supercritical fluid heat transfer coefficient prediction model with higher precision; an interpretable machine learning research method is used, an interpretable algorithm is added to a model for predicting the supercritical fluid heat transfer coefficient, and based on the physical meaning of the input characteristic parameters, the interpretation result is compared with the supercritical heat transfer mechanism, so that the understanding of the model prediction mechanism and decision process can be improved, the 'black box effect' of the model is lightened, and the interpretability of the model is enhanced.
Drawings
FIG. 1 is a flow chart of the present invention;
FIG. 2 is a flow chart of a particle swarm optimization algorithm;
FIG. 3 is a schematic diagram showing a comparison of experimental values and predicted values on a test set;
FIG. 4 is a graph showing the variation of Nu with temperature and the variation of feature importance according to different mass flows;
in the figure: a is the change of Nu along with temperature under different mass flow working conditions; b is 800kg/m 2 s the feature importance of the model under the mass flow working condition; c is 1000kg/m 2 s the feature importance of the model under the mass flow working condition;
FIG. 5 is a graph showing the variation of Nu with temperature and the variation of feature importance according to different heat flux densities;
in the figure: a is under the working conditions of different heat flux densitiesNu varies with temperature; b is 0.2MW/m 2 Feature importance of the model under the working condition of heat flux density; c is 0.4MW/m 2 Feature importance of the model under the working condition of heat flux density;
FIG. 6 is a graph showing the variation of Nu with temperature and the variation of feature importance for different pressures;
in the figure: a is the change of Nu along with temperature under different pressure working conditions; b is the feature importance of the model under the pressure working condition of 23 MPa; c is the feature importance of the model under the working condition of 25 MPa;
fig. 7 is a schematic diagram showing the change of Nu with temperature and the change of feature importance in the heat transfer enhancement behavior;
in the figure: a is the change of Nu with temperature; b is the feature importance of the model of the normal heat transfer behavior region; c is the feature importance of the model of the heat transfer enhancement behavior region;
fig. 8 is a diagram showing the change of Nu with temperature and the change of feature importance in the heat transfer deterioration behavior;
in the figure: a is the change of Nu with temperature; b is the feature importance of the model of the heat transfer enhancement behavior region; c is the feature importance of the model of the heat transfer deterioration behavior zone.
Detailed Description
As shown in fig. 1, this embodiment relates to a supercritical fluid heat transfer coefficient prediction method based on interpretable machine learning, which specifically includes:
and 1, selecting dimensionless numbers related to supercritical heat transfer, and respectively representing physical property change influence, buoyancy effect and acceleration effect in supercritical fluid heat transfer. Based on the experimental data of supercritical water in the existing vertical rising heating pipe, physical property data are obtained through REFPROP physical property packages, and the selected dimensionless number is calculated to construct a data set.
The dimensionless number includes: the Reynolds numbers and Plantains of different qualitative temperatures are used for representing the comprehensive effect; density ratio, dynamic viscosity ratio, thermal conductivity ratio, kinematic viscosity ratio of different qualitative temperatures are dimensionless; a dimensionless number characterizing the buoyancy effect; dimensionless numbers characterizing acceleration effects.
The said solidThe test data comprises 1598 samples, and the coverage range of working condition parameters is as follows: the pressure P is 23-30MPa, the mass flow G is 600-1500kg/m2s, and the heat flow density q w 200-1400kw/m2, pipe diameter of 7.6-26mm, and main flow temperature T b 210.84-449.29 ℃.
And 2, cleaning repeated or abnormal data points in the data set obtained in the step 1, dividing a training set and a testing set, and carrying out normalization processing to obtain a reliable data set for model training and testing.
Step 3, constructing a feedforward-back propagation fully-connected neural network model comprising an input layer, an output layer and a hidden layer, wherein the model transmits input data from the input layer to the output layer, each neuron of the hidden layer weights the input and processes the input through a ReLU activation function, and the result is transmitted to the next layer; after forward propagation, the error between the predicted and actual values is calculated using a mean square error loss function and the error is back propagated.
The parameters of the input layer are 38 dimensionless numbers describing the supercritical heat transfer effect, and the parameters of the output layer are Nu number related to the heat transfer coefficient.
The neural network model updates the weight parameters of the network by adopting an Adam optimizer so as to minimize the loss function, and iterates for a plurality of times until the loss function converges to a minimum value.
And 4, optimizing the hidden layer number, the hidden node number, the initial learning rate, the batch number and the iteration number of the model by using a particle swarm algorithm, namely, representing a super-parameter combination by using the position of each particle, searching and optimizing by adopting five-fold cross validation in a super-parameter space which is set in advance, using the average loss of the five validation as an adaptability function of the particle swarm optimization algorithm, evaluating the effect of the super-parameter combination searched by the particles, and using the minimized adaptability function as an optimizing direction.
The five-fold cross validation means that: the training set is divided into five parts, each part is used as a verification set in turn to verify the super-parameter performance, and the other 4 parts are used as training sets.
As shown in fig. 2, the particle swarm algorithm specifically includes:
4.1 Initializing the position of the particle swarm, i.e. the initial value of each super-parameter in the particle position; calculating the value of the fitness function corresponding to each particle position, namely training a model by using the hyper-parameter combination represented by each particle, and taking the loss on the verification set as the value of the fitness function;
4.2 Updating the particle velocity after updating the optimum value of the individual particles and the optimum value of the whole particle population based on the particle fitness valueAnd position X k (i+1)＝X k (i)+V k (i+1)，X k ＝{X k1 ,X k2 ,X k3 ,…,X kD Continuously calculating a particle adaptation value, and updating an individual extremum and a global extremum until reaching the termination iteration times, wherein: omega is the inertial weight, c 1 For acceleration factor, p best For the individual historical optimal position c 2 G is the acceleration coefficient best X is the historical optimal position of the group k For the position of the kth particle, D represents that there are D hyper-parameters to be optimized, V k (i+1) represents the velocity of the kth particle in iteration (i+1) th generation.
The speed of the particles comprises: an inertial part representing the effect of the current motion state of the particles; a cognitive part representing the distance and direction between the current position of the particle and the self-history optimal position; and a social part representing the distance and direction between the current position of the particles and the optimal position of the group history.
The super-parameter search space is set as follows: the number of hidden layers is 1-4, the number of hidden layer nodes is 1-500, the initial learning rate is set to 0.00009-0.009, the batch number is set to 32-128, and the number of model training iterations is set to 500-2000. The super-parameter optimization result is that the number of hidden layers is 2, the number of hidden nodes is 353 and 417 respectively, the initial learning rate is 1.7052e-03, the batch number is 115, and the iteration number is 1470.
Step 5, substituting the super parameters obtained in the optimizing step 4 into the model obtained in the step 3 for training, and after training, adding MAPE and MAPE on the test setR 2 And testing the model precision for the index.
The test specifically comprises the following steps:the mean absolute error MAPE finally obtained on the test set was 0.7%, R 2 0.99961. And (3) selecting two correlation formulas to carry out precision comparison, wherein the MAPE of the correlation formula of X.Cheng on a test set is 35%, and R2 is 0.2775. The MAPE of the correlation of the watts on the test set is 31%, R 2 0.348.
Through specific practical experiments, nu is subjected to regression prediction by 38 input characteristic parameters based on 1598 flowing heat transfer experimental data of supercritical water in a vertical riser. Comparison of Nu predictions and experimental values over the test set as shown in fig. 4, the distribution of the scatter is close to the y=x distribution. By calculation, the mean absolute value error MAPE obtained on the test set was 0.7%, R 2 0.99961.
In the existing supercritical heat transfer coefficient prediction model, average absolute value errors MAPE of BP neural network (BPNN) and genetic algorithm optimized BP neural network (GA-BP) are respectively 1.46% and 2.13%, and R of packet data processing neural network (GMDH-NN) 2 0.984.
Based on the physical meaning of the input dimensionless number, the SHAP interpretable algorithm is used for analyzing the feature importance of the model, and the interpretation result is combined with the supercritical heat transfer mechanism to explain the prediction mechanism of the model. On one hand, the importance degree analysis of the characteristic parameters is carried out on the influence of different working condition parameters on the model prediction result, wherein the importance degree analysis comprises pressure, heat flow density and mass flow, and the influence of the working condition parameters on the characteristic parameters is analyzed, so that the heat transfer behaviors under different working conditions are different; on the other hand, aiming at the importance degree of the characteristic parameters in different heat transfer behaviors under the same working condition, including normal heat transfer, heat transfer enhancement and heat transfer deterioration, the change of which characteristic parameters is analyzed to cause the change of the heat transfer behavior.
The SHAP interpretable algorithm measures the influence of each feature by calculating the marginal contribution of each feature in the model, thereby further aiming at the black box mouldThe model is explained and this marginal contribution is called shapley value. For a single sample, the expression form of the interpretation model isWherein: m is the number of features, phi 0 Is the average of all sample predictors, also called base value, phi i Shape value for each feature, ++> Representing the feature x under different feature combinations i Participation in and x i Model result change condition when not participating, M represents feature complete set, S represents removal of x i The values of S are various, and the values correspond to different feature combinations respectively. f (x) SU(i) ) And f (x) S ) Respectively represent x under various characteristic combinations i Model results with and without participation, +.>Representing probabilities corresponding to various combinations of parameters.
As shown in fig. 4-8, the model interpretation results include: the scatter diagram represents the change of Nu along with temperature under corresponding working conditions, and the bar chart represents the change of the feature importance. The horizontal axis of the histogram is the shape value of each feature, the larger the value is to represent the larger influence of the feature on the model prediction result, the vertical axis is the input parameter of the model, namely the dimensionless number related to supercritical heat transfer, and the histogram is divided into four types according to the physical meaning of the feature parameter, including: the Reynolds numbers and Plantains of different qualitative temperatures are used for representing the comprehensive effect; density ratio, dynamic viscosity ratio, thermal conductivity ratio, kinematic viscosity ratio of different qualitative temperatures are dimensionless; a dimensionless number characterizing the buoyancy effect; dimensionless numbers characterizing acceleration effects.
The effect of different mass flows on feature importance is first analyzed as shown in fig. 4. As shown in FIG. 4 (a), the diameter of the pipe is 12mm, pressure 23MPa, heat flow density 0.4MW/m 2 Under the working condition, the mass flow rate is from 800kg/m 2 s is increased to 1000kg/m 2 s, nu is unchanged along with the overall trend of temperature change, the peak value at the quasi-critical area is increased, and the heat transfer is enhanced. As shown in fig. 4 (b) and 4 (c), as the mass flow increases, the buoyancy factor and acceleration factor positions that inhibit heat transfer move downward, reducing the deterioration effect on heat transfer, while the Re and Pr positions that have a positive effect on heat transfer move upward, occupying the main positions, increasing the strength of forced convection heat transfer, so that Nu increases, ultimately resulting in enhanced heat transfer.
Further, the effect of different heat flux densities on feature importance was analyzed as shown in fig. 5. As shown in FIG. 5 (a), at a pipe diameter of 26mm, a pressure of 23MPa and a mass flow rate of 600kg/m 2 Under s working condition, the heat flow density is 0.2MW/m 2 To 0.4MW/m 2 The change trend of Nu near the quasi-critical region is changed from increasing to decreasing, and the heat transfer strengthening behavior is converted into heat transfer deterioration. As shown in fig. 5 (b) and 5 (c), the heat flux density increases, resulting in a large temperature gradient at the wall surface, thereby exacerbating the non-uniformity of physical properties, and therefore, the part of the feature map showing the change of physical properties is dominant, especially the density ratio, thereby causing the overall floating force and acceleration factor to move upward, increasing the importance, and improving the deteriorating effect; in addition, the important degree of Re, pr and the like with positive effect on heat exchange is reduced, and finally heat transfer deterioration occurs.
Further, the effect of pressure on feature importance was analyzed as shown in fig. 6. As shown in FIG. 6 (a), at a pipe diameter of 12mm, the mass flow rate was 800kg/m 2 s, heat flux density 0.4MW/m 2 In the working condition, nu of the two groups of working conditions are relatively attached at a position far away from the critical region, and as the pressure is increased from 27MPa to 30MPa, the peak value of Nu at the critical region is reduced, and the heat transfer strengthening effect is weakened. As shown in FIGS. 6 (b) and 6 (c), since the pressure increases and the physical properties change gradually, the Re and Pr number enhanced heat transfer effect at the pseudo-critical region decreases, the importance degree in the feature map decreases, the floating force and acceleration effect factor position in the feature map increases, the deterioration effect on heat transfer increases, and Nu is finally caused at the pseudo-critical regionThe peak at the zone is reduced and the heat transfer enhancement effect is reduced.
Further, the change in the feature importance in the heat transfer strengthening behavior was analyzed, as shown in fig. 7. A group of working conditions for heat transfer enhancement are selected, the pipe diameter is 12mm, the pressure is 23MPa, and the mass flow is 1000kg/m 2 s, heat flux density of 0.2MW/m 2 . As shown in fig. 7 (a), the left side of the broken line shows normal heat transfer behavior, and the right side of the broken line shows heat transfer enhancing behavior. As shown in fig. 7 (b) and (c), from the normal heat transfer to the heat transfer enhancement, re, pr, etc. having a forward effect dominate, the enhancement heat transfer effect is improved, and the position of the whole of the acceleration effect and buoyancy effect factor having a worsening effect is moved downward, the importance degree is reduced, and the effect of deteriorating the heat transfer is reduced, thereby finally enhancing the heat transfer.
Further, the change in the feature importance in the heat transfer strengthening behavior was analyzed, as shown in fig. 8. Selecting a group of working conditions with heat transfer deterioration, wherein the pipe diameter is 7.6mm, the pressure is 25MPa, and the mass flow is 1500kg/m 2 s, heat flux density of 0.4MW/m 2 . As shown in fig. 8 (a), the left side of the broken line shows the heat transfer enhancing behavior, and the right side of the broken line shows the heat transfer deteriorating behavior. As shown in fig. 8 (b) and 8 (c), from the enhancement of heat transfer to the deterioration of heat transfer, the position of the whole acceleration effect factor PIA or the like having the deterioration effect is moved downward, the importance degree is reduced, the effect of deteriorating the heat transfer is improved, the importance degree of Re, pr or the like having the forward effect of heat transfer is reduced, the enhancement of the heat transfer effect is reduced, and finally the transition of the heat transfer behavior is promoted.
In summary, compared with the prior art, the method uses the particle swarm optimization algorithm to optimize the feedforward-back propagation fully-connected neural network, builds the supercritical fluid heat transfer coefficient prediction model, and has higher prediction precision than the empirical correlation type and general neural network model; by adopting an interpretable machine learning method, an interpretable algorithm is combined with the BP neural network, and based on the physical meaning of input characteristic parameters, the change of model prediction characteristic importance in different working conditions and different heat transfer behaviors is obtained, a model prediction mechanism is interpreted, the transparency of the model is improved, and the interpretability of the model is enhanced.
The foregoing embodiments may be partially modified in numerous ways by those skilled in the art without departing from the principles and spirit of the invention, the scope of which is defined in the claims and not by the foregoing embodiments, and all such implementations are within the scope of the invention.
Claims (10)
1. A supercritical fluid heat transfer coefficient prediction method based on interpretable machine learning is characterized in that a feedforward-back propagation fully-connected neural network model is built in an offline stage, physical property data are obtained through REFPROP physical property packages, a data set obtained by calculation and construction of dimensionless numbers related to supercritical heat transfer is optimized by adopting a particle swarm optimization algorithm; and in the online stage, the generalization capability is tested on a test set through the trained fully-connected neural network model, and an interpretable algorithm is added to analyze the feature importance degree, explain a model prediction mechanism and realize the supercritical fluid heat transfer coefficient prediction.
2. The method for predicting the heat transfer coefficient of the supercritical fluid based on the interpretable machine learning according to claim 1, wherein the data set is obtained by acquiring fluid physical properties under different experimental conditions through a REFPORP physical property bag, calculating dimensionless numbers related to supercritical heat transfer, and obtaining a reliable data set through preprocessing.
3. The method for predicting the heat transfer coefficient of a supercritical fluid based on interpretable machine learning of claim 1, comprising the steps of:
step 1, selecting dimensionless numbers related to supercritical heat transfer, wherein the dimensionless numbers are used for representing physical property change influence, buoyancy effect and acceleration effect in supercritical fluid heat transfer; based on the experimental data of supercritical water in the existing vertical rising heating pipe, acquiring physical property data through a REFPROP physical property package, calculating a selected dimensionless number, and constructing a data set;
step 2, cleaning repeated or abnormal data points in the data set obtained in the step 1, dividing a training set and a testing set, and carrying out normalization processing to obtain a reliable data set for model training and testing;
step 3, constructing a feedforward-back propagation fully-connected neural network model comprising an input layer, an output layer and a hidden layer, wherein the model transmits input data from the input layer to the output layer, each neuron of the hidden layer weights the input and processes the input through a ReLU activation function, and the result is transmitted to the next layer; after forward propagation, calculating an error between the predicted value and the true value using a mean square error loss function, and back-propagating the error;
step 4, optimizing the hidden layer number, the hidden node number, the initial learning rate, the batch number and the iteration number of the model by using a particle swarm algorithm, namely, representing a super-parameter combination by using the position of each particle, searching and optimizing by adopting five-fold cross validation in a super-parameter space which is set in advance, using the average loss of the five times validation as an adaptability function of the particle swarm optimization algorithm, evaluating the effect of the super-parameter combination searched by particles, and using the minimized adaptability function as an optimizing direction;
step 5, substituting the super parameters obtained in the optimizing step 4 into the model obtained in the step 3 for training, and after training, performing MAPE and R on the test set 2 Testing the model precision for the index;
and step 6, adding an interpretable algorithm to the model trained in the step 5, and interpreting a model prediction mechanism by combining a supercritical heat transfer mechanism.
4. The method for predicting heat transfer coefficients of a supercritical fluid based on interpretable machine learning of claim 3, wherein the dimensionless number comprises: the Reynolds numbers and Plantains of different qualitative temperatures are used for representing the comprehensive effect; density ratio, dynamic viscosity ratio, thermal conductivity ratio, kinematic viscosity ratio of different qualitative temperatures are dimensionless; a dimensionless number characterizing the buoyancy effect; a dimensionless number characterizing acceleration effects;
the experimental data comprise 1598 samples, and the coverage range of working condition parameters is as follows: the pressure P is 23-30MPa, the mass flow G is 600-1500kg/m2s, and the heat flow density q w Is 200-1400kw/m2,the pipe diameter is 7.6-26mm, and the main flow temperature T b 210.84-449.29 ℃.
5. The method for predicting heat transfer coefficient of supercritical fluid based on interpretable machine learning of claim 3, wherein the parameters of the input layer are 38 dimensionless numbers describing the supercritical heat transfer effect, and the parameters of the output layer are Nu-ser numbers related to the heat transfer coefficients.
6. The method for predicting heat transfer coefficient of supercritical fluid based on interpretable machine learning according to claim 1 or 2, wherein the particle swarm algorithm specifically comprises:
6.1 Initializing the position of the particle swarm, i.e. the initial value of each super-parameter in the particle position; calculating the value of the fitness function corresponding to each particle position, namely training a model by using the hyper-parameter combination represented by each particle, and taking the loss on the verification set as the value of the fitness function;
6.2 Updating the particle velocity after updating the optimum value of the individual particles and the optimum value of the whole particle population based on the particle fitness valueAnd position X k (i+1)＝X k (i)+V k (i+1)，X k ＝{X k1 ，X k2 ，X k3 ，...，X kD Continuously calculating a particle adaptation value, and updating an individual extremum and a global extremum until reaching the termination iteration times, wherein: omega is inertia weight, c1 is acceleration coefficient, p best For the individual historical optimal position c 2 G is the acceleration coefficient best X is the historical optimal position of the group k For the position of the kth particle, D represents that there are D hyper-parameters to be optimized, V k (i+1) represents the velocity of the kth particle in iteration (i+1) th generation.
7. The method for predicting heat transfer coefficients of a supercritical fluid based on interpretable machine learning of claim 6, wherein the particle velocity comprises: an inertial part representing the effect of the current motion state of the particles; a cognitive part representing the distance and direction between the current position of the particle and the self-history optimal position; and a social part representing the distance and direction between the current position of the particles and the optimal position of the group history.
8. The method for predicting heat transfer coefficients of supercritical fluid based on interpretable machine learning of claim 2, wherein the search space of the super-parameters is set as follows: the number of hidden layers is 1-4, the number of hidden layer nodes is 1-500, the initial learning rate is set to 0.00009-0.009, the batch number is set to 32-128, and the number of model training iterations is set to 500-2000; the super-parameter optimization result is that the number of hidden layers is 2, the number of hidden nodes is 353 and 417 respectively, the initial learning rate is 1.7052e-03, the batch number is 115, and the iteration number is 1470.
9. The method for predicting heat transfer coefficient of supercritical fluid based on interpretable machine learning according to claim 1 or 2, wherein the testing specifically comprises:the mean absolute error MAPE finally obtained on the test set was 0.7%, R 2 0.99961; two correlation formulas are selected for precision comparison, wherein the MAPE of the correlation formula of X.Cheng on a test set is 35%, and R2 is 0.2775; the MAPE of the correlation of the watts on the test set is 31%, R 2 0.348.
10. A system for implementing the interpretable machine learning based supercritical fluid heat transfer coefficient prediction method recited in any one of claims 1-9, comprising: the system comprises a supercritical heat transfer related dimensionless number calculation and arrangement module, a neural network model optimization and training evaluation module and a neural network model interpretation module, wherein: the supercritical heat transfer related dimensionless number calculation and arrangement module obtains fluid physical properties under different experimental conditions through a REFPROP physical property bag, calculates the supercritical heat transfer related dimensionless number, and performs pretreatment to obtain a reliable data set for model training and testing; the neural network model optimization and training evaluation module optimizes the super parameters such as the hidden layer number, the hidden layer node number, the learning rate and the like of the model by using a particle swarm optimization algorithm, substitutes the optimized super parameters into the neural network for training and testing, and carries out error calculation on a prediction result and an experimental value; the neural network model interpretation module analyzes the changes of model characteristic parameter importance degree of different working conditions and different heat transfer behaviors according to the characteristic importance degree analysis result of the SHAP interpretable algorithm on the model, and interprets a model prediction mechanism by combining a supercritical heat transfer mechanism.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311096466.1A CN117150896B (en) | 2023-08-29 | Supercritical fluid heat transfer coefficient prediction method based on interpretable machine learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311096466.1A CN117150896B (en) | 2023-08-29 | Supercritical fluid heat transfer coefficient prediction method based on interpretable machine learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN117150896A true CN117150896A (en) | 2023-12-01 |
| CN117150896B CN117150896B (en) | 2024-05-14 |
Family
ID=
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104122291A (en) * | 2014-07-22 | 2014-10-29 | 上海交通大学 | Method for identifying rate of heat transfer from water-cooled wall to working medium of ultra-supercritical thermal power generating unit in real time |
| CN109799533A (en) * | 2018-12-28 | 2019-05-24 | 中国石油化工股份有限公司 | A kind of method for predicting reservoir based on bidirectional circulating neural network |
| CN113919243A (en) * | 2021-09-11 | 2022-01-11 | 西北工业大学 | Supercritical fluid heat transfer associated agent model construction method based on machine learning |
| US20220172050A1 (en) * | 2020-11-16 | 2022-06-02 | UMNAI Limited | Method for an explainable autoencoder and an explainable generative adversarial network |
| CN115618732A (en) * | 2022-10-24 | 2023-01-17 | 上海交通大学 | Nuclear reactor digital twin key parameter autonomous optimization data inversion method |
| CN116542382A (en) * | 2023-05-10 | 2023-08-04 | 长春理工大学 | Sewage treatment dissolved oxygen concentration prediction method based on mixed optimization algorithm |
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104122291A (en) * | 2014-07-22 | 2014-10-29 | 上海交通大学 | Method for identifying rate of heat transfer from water-cooled wall to working medium of ultra-supercritical thermal power generating unit in real time |
| CN109799533A (en) * | 2018-12-28 | 2019-05-24 | 中国石油化工股份有限公司 | A kind of method for predicting reservoir based on bidirectional circulating neural network |
| US20220172050A1 (en) * | 2020-11-16 | 2022-06-02 | UMNAI Limited | Method for an explainable autoencoder and an explainable generative adversarial network |
| CN113919243A (en) * | 2021-09-11 | 2022-01-11 | 西北工业大学 | Supercritical fluid heat transfer associated agent model construction method based on machine learning |
| CN115618732A (en) * | 2022-10-24 | 2023-01-17 | 上海交通大学 | Nuclear reactor digital twin key parameter autonomous optimization data inversion method |
| CN116542382A (en) * | 2023-05-10 | 2023-08-04 | 长春理工大学 | Sewage treatment dissolved oxygen concentration prediction method based on mixed optimization algorithm |
Non-Patent Citations (2)
| Title |
|---|
| NOEMIE CHAGNON-LESSARD .ETAL: "optimal design of geothermal power plants:a comparison of single-pressure and dual-pressure organic rankine cycles", GEOTHERMICS, vol. 86, 31 July 2020 (2020-07-31) * |
| 刘生晖 等: "基于预测-矫正技术的超临界流体强迫对流传热能力预测方法", 中国科学:技术科学, no. 06, 30 June 2017 (2017-06-30) * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109828532B (en) | GA-GBRT-based surface roughness prediction method and technological parameter optimization method | |
| Sun et al. | Elman neural network soft-sensor model of conversion velocity in polymerization process optimized by chaos whale optimization algorithm | |
| CN105608295A (en) | Multi-objective evolutionary algorithm (MOEA) and radial basis function (RBF) neural network optimization modeling method of coking furnace pressure | |
| CN103177291B (en) | The change search volume RNA genetic algorithm modeling method of continuous stirred tank reactor | |
| CN114741961A (en) | Method and system for optimizing wing type fin arrangement structure of printed circuit board type heat exchanger | |
| CN105389442A (en) | Reverse design method for coupling genetic algorithm, neural network and numerical simulation | |
| Velásquez et al. | Surrogate modeling method for multi-objective optimization of the inlet channel and the basin of a gravitational water vortex hydraulic turbine | |
| CN108256688B (en) | Steam flooding development effect prediction method for optimizing GRNN by improving firefly algorithm | |
| CN117150896B (en) | Supercritical fluid heat transfer coefficient prediction method based on interpretable machine learning | |
| Lee et al. | Machine learning-based inverse design methods considering data characteristics and design space size in materials design and manufacturing: a review | |
| CN111708388B (en) | Boiler pressure regulation prediction control method based on GRU-PID | |
| CN117150896A (en) | Supercritical fluid heat transfer coefficient prediction method based on interpretable machine learning | |
| CN116956744A (en) | Multi-loop groove cable steady-state temperature rise prediction method based on improved particle swarm optimization | |
| CN111310974A (en) | Short-term water demand prediction method based on GA-ELM | |
| Xie et al. | Self-adaptively commensal learning-based Jaya algorithm with multi-populations and its application | |
| CN109165792B (en) | Photovoltaic short-term output power prediction method based on SOA-WNN | |
| Glänzel et al. | Parameterization of environmental influences by automated characteristic diagrams for the decoupled fluid and structural-mechanical simulations | |
| Xu et al. | Performance prediction and optimization of annular thermoelectric generators based on a comprehensive surrogate model | |
| CN114519299A (en) | Multi-objective optimization method for double-wall structure of integrated support plate flame stabilizer | |
| Zhang et al. | LS-SVM based on chaotic particle swarm optimization with simulated annealing and application | |
| Jia et al. | Prediction of heat transfer performance of vacuum glass based on extreme gradient boosting algorithm | |
| Cheng et al. | Modeling of main steam temperature using an improved fuzzy genetic algorithm | |
| CN117272244B (en) | Soft measurement modeling method integrating feature extraction and self-adaptive composition | |
| Gao et al. | Soft-sensor modeling of polyvinyl chloride polymerizing process | |
| Ling et al. | An improved PSO algorithm for constrained multiobjective optimization problems |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant |