Impact of data processing and robust machine learning process on accurate estimation of specific heat capacity property in energy storage applications

Abstract

The specific heat capacity propery is a critical thermal property because it is directly related to heat storage and transfer. The need for better and novel thermal energy storage materials have led to the discovery of high temperature nanofluids with molten salts (nanosalts). The utilization of novel machine learning tools in accurately estimating the specific heat capacity of nanofluids, would further pave the way for more innovative studies on nanosalt, hence the prediction of specific heat capacity of hybrid nanofluids constitute a significant research area. Despite the few researches conducted in the area of prediction of these fluids, priority with regards to the robust analysis of input variables for different machine learning algorithms have not been given due consideration. Therefore, this present study aims to develop and validate a robust package of machine learning algorithms (twenty machine learning algorithms) in estimating the specific heat capacity of hybrid nanofluids, across different feature spaces. The different algorithm hyperparameters were adjusted to give optimum prediction. A total of 600 data points were retrieved and split into training, testing, and validation data set, for which the validation dataset was not used in the training and testing process. The input variables used in the training of the machine learning models are temperature, volume fraction, the specific heat capacity of the base fluid, the specific heat capacity of individual nanoparticles, and the mixture ratio of individual nanoparticles. The Model A, which constituted the seven (7) input variables gave the best prediction accuracy with validation score of 0.999703, while the Model C, which contained the feature variables of temperature, volume fraction and specific heat capacity of the base fluid gave the least prediction accuracy. The optimal machine algorithms retrieved for the Model A, B, C and D was the Ridge, Gradient Boost regressor, Passive Aggressive regressor, and Kernel ridge regressor. The combination of ensemble tree regressors and online learning algorithms can be coded and used in future prediction analysis of SHC, as they have both shown accurate prediction quality.