Pressure Drop Prediction in Fluidized Dense Phase Pneumatic Conveying using Machine Learning Algorithms

Abstract

Modeling of pressure drop in fluidized dense phase conveying (FDP) of powders is a tough work as the flow comprises of various interactions among solid, gas and pipe wall. It is difficult to incorporate these interactions into a model. The pressure drop depends on flow, material and geometrical parameters. The existing models show high error when applied to other pipeline configurations of varying pipeline lengths or diameters. The current study investigates the capability of machine learning (ML) techniques to estimate the drop in pressure in FDP conveying of powders. Pneumatic conveying experimental data were used for training the network and then for predicting the pressure drop. For estimating the pressure drop, four distinct ML algorithms light gradient boosting machine (LighGBM)), multilayer perception (MLP), K-nearerst neighbors (KNN), extreme gradient boosting (XGBoost), and were selected. XGBoost model performed better than other models chosen for the study with ±5% error margin while training and testing the data, and ±10% error margin in validating the data. MLP, XGBoost, KNN, and LightGBM models predicted the data of pressure drop with MAE of 5.05, 1.19, 5.72, and 2.85, respectively, for training as well as testing data. Among the four models considered, the model using XGBoost algorithm performed the best, whereas the model using KNN algorithm performed poorly in predicting the FDP conveying pressure drop.