Predictive Modelling of a Sequential Anaerobic–Aerobic Biofilm Reactor for Tannery Wastewater Treatment: A Comparative Study of Regression and Artificial Neural Network Approaches

Abstract

Tannery wastewater is characterized by high chemical oxygen demand (COD), tannins, chromium and sulfide content, and its treatment by a single biological process is often insufficient to meet discharge standards. This study develops predictive regression and artificial neural network (ANN) models for a laboratory-scale sequential anaerobic-aerobic biofilm-based reactor treating tannery wastewater, using experimental data collected under varying hydraulic retention time (HRT), organic loading rate (OLR), aeration rate, influent pH, VSS/TS ratio, and influent/effluent COD and volatile fatty acids (VFA). Port-wise linear regression models were first fitted to describe pH stabilization and the volatile fatty acid to alkalinity (VFA/Alkalinity) ratio across three sampling ports, together with the biogas production response to variations in HRT (10-60 h), OLR (1-7 kg/m3/d), and aeration rate (2-12 L/min). A multilayer feed-forward back-propagation neural network (BPNN), with five inputs (OLR, HRT, aeration rate, pH and VFA/Alkalinity), three hidden neurons, and two outputs (biogas production and percentage COD removal), was subsequently trained on 215 data points and tested on 46 unseen data points. The trained network achieved a training performance index of R = 0.93 and testing performance indices of R = 0.99 for both biogas production and percentage COD removal, with an average prediction error of approximately 5% and an overall model accuracy of 95%. The results indicate that the BPNN model substantially outperforms simple linear regression in capturing the nonlinear dynamics of the sequential anaerobic-aerobic biofilm reactor and can serve as a reliable tool for process monitoring and optimization of tannery wastewater treatment.