Machine Learning Aided Numerical and Experimental Investigation of Hydrodynamic Performance in the Circulating Fluidized Bed Boiler

Abstract

This study attempts to illustrate the benefits of integrating the concepts of machine learning algorithms with the field of thermo-fluidic applications. The current work is aimed at identifying effective or significant hydrodynamic input parameters, which are capable of deriving full benefit of fluidization that could yield a better circulating fluidized bed (CFB) furnace design using the Apriori algorithm. For this, historical datasets from the literature are collected and pretreated based on the design under observation. Association analysis performed by this Apriori algorithm measures the comparative strength of parameters under consideration. Also, this algorithm is capable of identifying the right combinations of parameters that can produce maximum fluidization performance. The end results suggested by this Apriori algorithm are validated using the computational fluid dynamics package. For this, the transient behavior of a scaled-down (1:20) reactor model of a real-time industrial CFB boiler is simulated using ansys fluent 18.0. In specific, the effects of fluidizing velocities, inventory heights of the bed, and particle sizes recommended by the Apriori algorithm are investigated. Here, the effects are assessed in terms of volume fraction distribution and axial velocity profile distribution profiles. From the results of simulations, it was clearly found that 2 m/s inlet velocity produced good circulating fluidized bed patterns on a bed inventory height of 0.5 m for a mean particle size of 200 µm. The results obtained from the simulations are once again validated visually against snapshots obtained during real-time laboratory fluidization experimental runs. Among all the cases of comparisons, the best agreement is demonstrated by Apriori algorithm compared to the numerical and the experimentally obtained results. Also, it is found that the manual time taken to identify the right combinations of parameters is drastically reduced by this method compared to conventional optimization algorithm and trial error methods.