Research On The Application Of Machine Learning In The Field Of Convective Heat Transfer With Single-phase Heat Flow

Transient thermal performance has been a matter of great interest in various thermal and energy engineering.Generally,there are three commonly used methods to obtain transient thermal properties:experiments,theoretical calculations and numerical simulations.However,the use of these three traditional methods to obtain transient thermal properties has varying degrees of limitations.In most cases in practical thermal fluid engineering,theoretical solutions are not feasible;pure experiments do not give researchers universally credible results when the operating conditions change;and numerical simulations face the problem of being extremely inefficient,as it usually takes a significant amount of time to calculate the required transient information for a liquid-solid-thermal coupling phenomenon.This paper first proposes a novel machine-learning based data-driven approach to predict transient thermal performance other operating conditions by using data obtained from certain operating conditions(from the three methods mentioned above).In this paper,a single-phase forced convection system is chosen as a demonstration of transient thermal performance,and results are first obtained numerically for 39 conditions with different inlet velocities and heat flow densities;then,data from 32 conditions are randomly selected as the training data set,while the other seven conditions are used as the test set.Using the training dataset,the neural network model was trained and gave high accuracy predictions for the test set,where an average accuracy of 97.3%was obtained.The processing time for predicting transient performance from 0 to 300 s can be reduced to 1 s,which shows that the trained model has good properties for ultra-real-time prediction.In addition,machine learning modelling,especially in large data high precision regression problems,is more dependent on large amounts of training data.However,such large amounts of data are difficult to obtain in the field of heat transfer,i.e.the training data is not that large,and therefore the data size required for the training set should be kept as small as possible,as data acquisition is very inconvenient to obtain.Therefore,the training set data size needs to be kept as small as possible without sacrificing prediction accuracy.On top of the work mentioned above,a method for selecting smart-sized training set data is proposed based on the idea of active learning.The smart-sized training data set size is reduced by 50%and the prediction accuracy is still improved.In particular,the worst-case prediction accuracy remains above 90%,an improvement of 3.5%over the previous results.Small sample modelling developed using deep learning algorithms provides a powerful information processing tool that can enhance the understanding of non-linear systems in thermofluid mechanics and even transform current lines relevant to and more widely used in thermal fluids,and the proposed approach also promises to accurately model complex thermal fluid systems.