Predicting the thermal performance of heat pipes applying various machine learning methods and a proposed correlation

Abstract

The key purpose of this study is to predict the thermal performance of heat pipes by means of artificial intelligence methods. In this regard, several different prediction models have been employed. These models are proposed with totally nine different parameters as inputs such as dimensions of heat pipes, number of turns, working fluids, inclination angle, filling ratio, and heat input. The output of these models is the thermal performance of the heat pipe. For training the prediction models, a dataset of totally 1196 experimental data was collected from previous studies which includes several different working fluids. The employed prediction methods contain ten machine learning regression methods, seven ANN models with different structures and a correlation model. At first, the machine learning regression methods have been employed and based on the results, the best performance belongs to random forest regressor method with the R2 score and MSE value of 0.90 and 0.003, respectively. Then inspiring physics of the problem, several feed forward back-propagation neural network models have been designed for simulating and estimating the thermal performance of heat pipes. Results verified that the proposed models could be used on a wide range of working fluids with an acceptable accuracy. The values of R2 and MSE of the optimal ANN model are 0.95 and 0.03, respectively which are pretty acceptable for a predicting model working on such a wide range of working fluids. Moreover, a new correlation is proposed to estimate the thermal performance of heat pipes on the basis of the same inputs. This correlation can estimate the thermal performance of the whole dataset with R2 score of 0.60 and MSE value of 0.27, while the minimum and maximum of R2 scores are 0.47 and 0.88 for water and ethanol, respectively. Finally, the results obtained from the mentioned machine learning methods, the ANN models and the proposed correlation were compared. It was found that due to complexity of the problem and wide variety of inputs, the results of using methods based on artificial intelligence are far more accurate than those of the proposed correlation. In fact, choosing more accurate and complicated methods (e.g., the optimal ANN model with structure of 9–32–40-50-64-1), the R2 score of prediction significantly increased from 0.60 to 0.95. Since there is a trade-off between the accuracy and simplicity of the methods, selecting the best method should be considered as a choice problem.