Machine learning model optimization for removal of steroid hormones from wastewater

Abstract

In the past few decades, there has been a significant focus on detecting steroid hormones in aquatic environments due to their influence on the endocrine system. Most compounds of these pollutants are the natural steroidal estrogens, i.e., estrone (E1), 17β-Estradiol (E2), and the synthetic estrogen 17α-Ethinylestradiol (EE2). The Moving-Bed Biofilm Reactor (MBBR) technique is appropriate for eliminating steroid hormones. This study centers on creating a model to estimate the effectiveness of the MBBR system regarding its ability to eliminate E1, E2, and EE2. The results were modeled with artificial neural networks (ANNs). The Particle Warm Optimization (PSO) and Levenberg Marquardt (LM) algorithms were selected for network training. The models incorporated five input parameters, encompassing the COD loading rate, initial levels of E1, E2, and EE2 steroid hormones, and Hydraulic Retention Time (HRT). The optimum removal conditions (three steroid hormones and COD) were determined using the optimized ANN based on both PSO and LM algorithms. The optimal transfer functions for the hidden and output layers were identified as tan-sigmoid and linear, respectively. The best ANN structures (Neurons in input, hidden, and output layers) and correlation coefficients (R) were 5:9:4, with R = 0.9978, and 5:10:4, with R = 0.9982 for the trained networks with LM and PSO algorithms, respectively. Eventually, the input parameters' importance was ranked using sensitivity analysis (SA) through Pearson correlation and developed ANNs.