Mechanistic, neural network, and intelligent hybrid models for a three-phase fluidised-bed biofilm reactor : a thesis presented in partial fulfillment of the requirements for the degree of Doctor of Philosophy in Environmental Engineering at Institute of Technology and Engineering, Massey University
Over the past three decades, considerable amount of research efforts have been undertaken in order to develop a mathematical model for a three-phase fluidised-bed biofilm reactor (TPFBBR). Although biofilm properties such as biofilm thickness and its density are allowed to vary with biofilm growth in the model to simulate the real TPFBBR system, they are assumed to be constant in the majority of models developed for a TPFBBR. The main goal of this thesis is to develop mathematical models incorporating dynamic biofilm growth for a TPFBBR using three different modelling approaches such as a mechanistic model, a neural network model, and an intelligent hybrid model with a neurofuzzy model. This thesis consists of three parts. Firstly, a dynamic biofilm growth model, which reflects the variation of biofilm thickness and its density in time, is developed. This model is derived from a biomass balance equation and is solved by the method of characteristics. The biofilm detachment model is proposed and incorporated within the dynamic biofilm growth model. The dynamic biofilm growth model with detachment is then combined with a reaction-diffusion model and reactor model to form an integrated model of a TPFBBR. Simulation method of integrated model incorporating the dynamic biofilm growth model is developed. It is observed that results predicted are in good agreement with experimental data and the integrated model proposed provides a valuable tool to predict performance of a TPFBBR. Secondly, the sequential neural network model, which is composed of two parts, namely, the neural process estimator and the neural process predictor, is developed to describe the task of process estimation and prediction for a TPFBBR. In order to implement the sequential neural network model, multilayer feedforward neural network (MFNN) with cascaded-correlation (C-C) learning and extended Kalman filtering (EKF) learning, and generalized regression neural network (GRNN) are used. Results shows that the sequential neural network model has the feasibility as intelligent estimators and dynamic predictors and gives considerably good results in process estimation and prediction for a TPFBBR. Finally, this thesis shows how a combination of both mechanistic and empirical modelling approaches, called a hybrid model, can be implemented and utilised for modelling a TPFBBR. The neurofuzzy model as an empirical part of hybrid model is used to estimate the variation of the biofilm thickness and biofilm density, and is combined with mechanistic model-based reaction-diffusion and axial-dispersion models to predict the dynamic behavior and performance of a TPFBBR according to the variation of biofilm density and biofilm thickness. This hybrid modelling approach due to its flexibility shows a unified framework through incorporation of strong points of both mechanistic and empirical models, and provides a new modelling framework with a great potential to be applied to other types of biofilm reactors.