Intelligent Algorithm‐Based Neural Modeling of Two Suspended Growth Bioreactor Systems for Wastewater Treatment

The choice of appropriate training algorithms often impacts the convergence and fitness of ANN models. In this regard, three conventional (Levenberg–Marquardt, Bayesian regularization, scaled conjugate gradient) and three evolutionary (genetic algorithm, particle swarm, reptile search) algorithms were explored for their suitability in the training of feedforward backpropagation (BP) neural networks, modeling experimental datasets from airlift bioreactor (ALB) and inverse fluidized bed bioreactor (IFBB). The Selenite removal efficiency (RE

selenite

) (%) in ALB was modelled as a function of influent selenite concentration (IC

selenite

) (10–60 mg/L) and hydraulic retention time (HRT) (24–72 h), while RE

selenite

(%) and chemical oxygen demand removal efficiency (RE

COD

) (%) in IFBB considered IC

selenite

(12.7–635 mg/L) and HRT (12–48 h) as the input parameters. After training and testing, the reptile search algorithm offers slightly better prediction accuracy compared to the Levenberg–Marquardt algorithm, with the lowest

RMSE

(ALB: 0.946; IFBB: 1.352) and

MAE

(ALB: 0.701; IFBB: 0.927), and reasonably high

R

values (ALB: 0.996, 0.983, 0.995; IFBB: 0.985, 0.952, 0.980), on training, testing, and complete datasets. The optimized neural network topologies were 2‐18‐1 and 2‐15‐2, respectively, for ALB and IFBB.