Development and Evaluation of Hybrid Artificial Neural Network Architectures for Modeling Spatio-Temporal Groundwater Fluctuations in a Complex Aquifer System

Abstract

The proper design, development, and appropriate tuning of the Hybrid Neural Network architecture, mainly for its parsimoniousity and optimal training can help practitioners to generate a robust predictive tool for modeling several important hydrological processes within the water resources sector. In this paper, the Feedforward Artificial Neural Network (FFANN) and the hybrid WANN model have been developed, and later, coupled with the Gamma and M-tests (GT) approach for forecasting spatio-temporal groundwater fluctuations in a complex alluvial aquifer system. The performance of these hybrid models were evaluated using goodness-of-fit criteria. An analysis of the modeling results indicates that the GT coupled with the WANN model was able to provide significantly improved results, with lower values of the root mean square error (RMSE) and higher values of the NSE metric for the 1-week and 3-week lead times. Hence, utilizing this hybrid model, the groundwater level prediction tests were extended for 6-week and 12-week lead times with the GT approach, coupled with the WANN hybrid model only. The results showed that the accuracy of the GT-WANN hybrid model was better for the unconfined aquifer system compared to the leaky confined aquifer system. Furthermore, the present study also examined the interdependence between different model inputs and output variables for the selected study sites by means of the Wavelet Coherent Analysis (WCA). These results indicated that all the model’s input variables have a significant effect on the groundwater level of unconfined aquifers, and confirmed the nature of the aquifers tapped within the present study sites. The study finally concludes that the GT-WANN approach can be a robust predictive tool for modeling spatio-temporal fluctuations of groundwater levels.