Neural network prediction of residence time distribution for quasi-2D pebble flow. (15th March 2022)

• Record Type: Journal Article
• Authors: Liu, Yujia; Marquardt, Jeremy; Peng, Sifan; Ge, Liang; Gui, Nan; Yang, X.T.; Tu, J.Y.; Jiang, S.Y.; Kim, Seungjin
• Is Part Of: Chemical engineering science. Volume 250(2022)
• Publication Date: 2022-03-15
• Subjects: Pebble flow -- Residence time -- Quasi-static flow -- Image identification -- PTV algorithm -- Neural network; Chemical engineering -- Periodicals; Génie chimique -- Périodiques; Chemical engineering; Periodicals; Electronic journals; 660
• DOI: 10.1016/j.ces.2021.117363
• Languages: English
• ISSNs: 0009-2509

Abstract

Highlights: A quasi-2D pebble bed is built to get quantitative residence time distribution (RTD). Image identification for large pebbles is achieved by a high-precision PTV algorithm. Neural network method is used to predict RTD of 2D pebble flow with good accuracy. Arc-shaped RTD is explained by radial flow model and linear one is by boundary layer. Prediction result verifies pebble flow follows trends of radial flow model in hopper area. Abstract: An important parameter in silo discharging is residence time distribution (RTD), which describes the time each pebble stays within the bed. RTD can be applied as a key indicator to estimate reaction processes, and a key analytical metric to assess the inherent safety of reactors. However, due to the difficulty in obtaining pebble trajectories, few studies on discharging processes have been carried out. In this research, a pilot experimental study is conducted to get quantitative data and prediction model on RTD with precise predictions. The feasibility of applying neural network methods to RTD prediction is also checked. The accuracy of neural network prediction is confirmed by comparing the error between predicted and experimental results. It is discovered that RTD isoline changes in different regions and characterized by intuitive physical models. Compared to DEM, the neural network has a significant advantage in computing speed while retaining acceptable prediction accuracy.