Formulating GA-SOM as a Multivariate Clustering Tool for Managing Heterogeneity of Aquifers in Prediction of Groundwater Level Fluctuation by SVM Model

Multivariate cluster analysis suffers from subjectivity associated with selecting input parameters and the related weights known as “Priority Rates” (PRs). This paper aims to formulate an algorithm referred to as “GA-SOM” to reduce this subjectivity by a hybrid of genetic algorithm (GA) and self-organizing map (SOM), in which GA optimizes the PRs of the input parameters by maximizing silhouette index (SI) along with interrogating SOM to calculate SI. GA-SOM was first applied to clustering groundwater level (GWL) data to identify less heterogeneous clusters in a large and heterogeneous aquifer. Then, GWL fluctuations were predicted by support vector machine (SVM) in 12 representative observation wells (OWs) by GA-SOM from 48 OWs. The results show a significant improvement in the quality of clustering so that the SI value equals 0.61 by using optimized PRs, while it did not exceed 0.38 without using PRs. Performances of SVMs are evaluated using RMSE and NSE, where for all OWs, RMSE and NSE vary in the range of 0.32–0.71 m and 0.91–0.97, respectively. Therefore, the models are found to be fit-for-purpose for predicting GWL in the aquifer.