Abstract
Corrosion can affect water taste, color, and odor, making it crucial to monitor and control corrosion in the water distribution network to maintain water quality standards. This study used machine learning approaches such as MARS, GMDH, and MPMR to model the corrosion rate in water distribution networks. An experimental setup was established in the running water distribution network for data collection, where several test coupons were inserted into the pipeline. A coupon weight loss method was employed to calculate the corrosion rate. The selected study site was continuously monitored for 315 days to observe the water distribution network (WDN). The physicochemical parameters were regularly tested at the Environmental Engineering Laboratory at NIT Patna. Machine learning analyses, including multivariate adaptive regression splines (MARS), the group method of data handling (GMDH), and multivariate polynomial regression (MPMR), consider 13 features, including pH, temperature, conductivity, total dissolved solids, alkalinity, hardness, calcium hardness, magnesium hardness, chloride, sulfate, nitrate, dissolved oxygen, and time, as input parameters, with the corrosion rate as the output parameter. Energy dispersive X-ray (EDX) analysis revealed changes in composition before and after exposure: the carbon content decreased from 4 to 3%, the oxygen content increased from 20 to 31%, the iron content increased from 21 to 60%, the sulfur content decreased from 3 to 2%, the manganese content decreased from 3 to 1%, and the zinc content decreased from 49 to 1% by weight. The performance of the developed model was assessed via several performance metrics, regression error characteristic (REC) curves, comprehensive measurement (COM), and ranking techniques. On the basis of the performance of the developed models, the proposed MARS model is the most accurate model, with R(2) = 0.9872 for training and R(2) = 0.9741 for the testing phase, followed by the GMDH and MPMR models. The REC curve also demonstrates the superiority of MARS, with lower area-over-the-curve (AOC) values (training: 0.010, testing: 0.015), followed by the GMDH (training: 0.028, testing: 0.024) and MPMR (training: 0.054, testing = 0.074) models. With the lowest COM value (0.172), the MARS model outperforms the GMDH and MPMR models, indicating its superior predictive capability and generalizability.