Previsão do comportamento dos processos de transferência de oxigênio induzidos por plumas de bolhas em tanques de água usando aprendizado de máquina e cinética de ordem superior

Abstract

The process of oxygen transfer to water has several applications in engineering and is referred to in the literature as a reaction with kinetics of order 1 derived from Fick’s diffusion law. The complexity of this phenomenon, however, introduces challenges to its evaluation, especially in artificial aeration systems, where air injection usually results in flows through bubble plumes that transfer gas by different mechanisms. From this, several studies have analyzed the problem proposing sophistications to this approach. This study explores a machine learning (ML) approach to predict the behavior of oxygen transfer by air injection in water tanks by predicting the reaeration coefficient k2 considering higher-order kinetics, an approach that has been little explored in this context. A total of 99 replicates of an experiment were conducted over 300 seconds each in laboratory tanks for different inlet conditions: air injection flow rates (Qair), temperatures (T ), initial dissolved oxygen concentrations (C0) and saturation (Cs), different types of air diffusers (tdi f ), volumes (Volt ) and water heights in the tanks (hw). After adjusting the coefficients k2 from the dissolved oxygen (DO) concentration time series and subsequently constructing the final dataset, eight ML models were studied considering coefficients with orders from 1 to 4 as output variables. The coefficients adjusted to higher orders resulted in the most promising models (in contrast to order 1), with emphasis on order 2, which is justified by the complex mechanism composed of four main gas transfer processes that occur simultaneously. For this order, the tree-based models obtained the best metrics, especially the regression tree (RT), with r2 = 0.668, RMSE = 7.00, MAE = 4.81, MAPE = 0.200 and RMSLE = 0.268 . From this, an analysis of feature importance was made for a better understanding of the interaction of the variables in the dynamics of the phenomenon, with estimates through the tree and linear models. These models agreed on the least important variables for predicting k2 in the context of these experiments, which were the air injection flow rate (Qair), the average plume velocity at the surface (Vplum) and the presence or absence of porous airstones, while for the more accurate models (tree-based models), the most important variables were the volume (Volt ) and the water height in the tanks (hw).