Prediction of Water Quality in the Lower Amudarya Delta Using ANFIS
Abstract
Using soft computing techniques is vital when modeling surface water quality for effective water management and environmental protection. Developing a precise predictive model, especially with numerous source settings and statistics that aren’t consistent, continues to be a challenge. Therefore, more study is needed to increase the efficacy of the prediction models. This paper proposes a framework for pre-processing datasets and optimizing inputs to reduce modeling complexity. This study achieved its goal by using a two-sided detection methodology for outlier elimination and an exhaustive search method to pick critical modeling inputs. Next, the adaptive neuro-fuzzy inference system was used to simulate electrical conductivity and total dissolved solids in the lower Amudarya delta. The modeling technique made use of a broader dataset spanning a 30 -year historical period and monitored weekly. The created models’ prediction capacity was assessed using statistical assessment metrics. Furthermore, the ten times crossvalidations approach was used to solve the modeling overfitting problem. The input optimization findings show that Ca2+, $\mathrm{Na}^{++}$, and $\mathrm{Cl}^{--}$are the most important inputs for electrochemical conversion. Mg2+, HCO3 -, and SO4 2- were chosen to represent total dissolved solids levels. The optimal adaptive neuro-fuzzy inference system models for All solids diffused data exhibited R values of 0.81 and 0.94, with root mean squared error outcomes of $20.6 \mu \mathrm{S} / \mathrm{cm}$ and 11.7 ppm, respectively. The optimal adaptive neuro-fuzzy inference system structure includes a hybrid training procedure with 29 fuzzy rules.