Novel and robust machine learning approach for estimating the fouling factor in heat exchangers

Abstract

The fouling factor (Rf) is an operating index for measuring an undesirable effect of solids’ deposition on the heat transfer ability of heat exchangers. Accurate prediction of the fouling factor helps appropriate scheduling of the cleaning cycles. Since diverse factors affect this operating feature, it is sometimes hard to estimate the fouling factor accurately using simple empirical or traditional intelligent methods. Therefore, this study employs four up-to-date machine-learning algorithms (Gaussian Process Regression, Decision Trees, Bagged Trees, Support Vector Regression) and a traditional model (Linear Regression) to estimate the fouling factor as a function of operating and constructing variables. The 5-fold cross-validation using 9268 data samples determines the structure of the considered estimators, and 2358 external datasets have been utilized for models’ testing. The relevancy analysis confirms that the most accurate predictions are achieved when the square root of the fouling factor (√Rf) is simulated. The Gaussian Process Regression (GPR) shows the highest level of agreement with the experimental samples in both the model construction and testing stages. The trained GPR model scored an R2 value of 0.98770 and 0.99857 on the internal and external datasets, respectively. The model predicts the overall 11626 experimental samples (Davoudi and Vaferi, 2018) with the MAPE = 13.89%, MSE = 7.02 × 10−4, and R2=0.98999. The proposed GPR model outperforms the previously suggested artificial neural network for estimating the fouling factor in heat exchangers.

Keywords

• Heat exchange
• Fouling
• Gaussian Process Regression
• Decision trees
• Support Vector Regression
• Comparison study