Conjunction of cluster ensemble-model ensemble techniques for spatiotemporal assessment of groundwater depletion in semi-arid plains

Abstract

• The cluster ensemble technique improved the final cluster structures. • The spatiotemporal ensemble clustering revealed four different patterns of GWL. • LSTM was not appropriate for monthly GWL forecasting compared to ensemble methods. • The neural ensemble has the lowest disorder and a large reduction in uncertainty. • The simultaneous clustering-forecasting ensemble techniques enhanced the outcomes. In this study, first to identify the patterns of groundwater level (GWL) over the Ghorveh-Dehgolan plain (GDP) located in western Iran, as a data pre-processing scheme, three different types of clustering algorithms were applied to monthly GWL data sets of the piezometers. Then, the best structures of all clustering methods were integrated by Combining Multiple Clusterings via Similarity Graph (COMUSA) method to obtain the most homogenous patterns of GWL. The final results of the clustering step indicated that applying COMUSA could enhance the homogeneity of the clusters up to 25%. After dividing the GWL of GDP into four patterns, three single artificial intelligence (AI)-based models were applied to forecast multi-step-ahead GWL of centroid piezometer of each cluster. To benefit from the advantages of the single models, the outcomes were then combined with a neural averaging ensemble (NAE) technique as a post-processing step. Additionally, the assessment of the deep learning (DL) -based long-short-term memory (LSTM) application in multi-step ahead forecasting of GWL showed this method is not an appropriate choice compared to ensemble techniques for modeling the process with limited observed data. The comparison of the proposed GWL forecasting models of this study revealed the superiority of the NAE technique that enhanced the accuracy of the single models up to 23% in the testing phase. It could be concluded that the combination of cluster ensemble and model ensemble techniques could improve the performance of the individual method in reliable forecasting of the future GWL condition and the methodology of this study can be applied to the GWL of other plains.