Forecasting Groundwater Levels using Nonlinear Autoregressive Networks with Exogenous Input Models: A Case Study from Lake Tuz and Beysehir Lake, Turkey

The water demand is proportionally increasing with the world growing population, increasing urbanization, and changing living standards. Turkey is a country that suffers from water scarcity. Although Central Anatolia of Turkey has important basins such as Lake Tuz and Beysehir Lake, these basins suffer from drying up. Therefore, monitoring the sustainability of water has become a critical issue and require an innovative and accurate management system On the search for accurate techniques and tools to collect all parameters and the ever-increasing data affecting the groundwater levels in one model, we came across the idea of analyzing the GRACE data with the Nonlinear Autoregressive with Exogenous Input (NARX) models, which has been extensively used for nonlinear time-series models. GRACE data provides a cost-effective alternative to locally filed measurements and can be used to monitor the sustainability of Turkey's groundwater resources. GRACE is a mission launched in March 2002. It was designed to show the mass distribution around the planet and monitor changes over time by measuring the gravitational anomalies. In October 2017, the GRACE mission was ended, and in May 2018, the GRACE-FO was launched as GRACE mission successor. However, the GRACE and GRACE-FO missions have short-term temporal gaps caused by battery performance issues and long-term temporal gaps between GRACE and GRACE-FO missions. These gaps distort the temporal variability analysis and statistical study results. In this study, it is aimed to collect all parameters affecting the groundwater level in one model for generating continuous and uninterrupted temporal GRACE-derived terrestrial water storage (TWS) records based on the knowledge of relevant datasets such as temperature, rainfall, evapotranspiration, vegetation indexes, and forecasting TWS 5 months in advance, utilizing Nonlinear Autoregressive with Exogenous Input (NARX) models, an important Artificial Neural Network (ANN) tool. The study area is selected as Lake Tuz and Beysehir Lake, located in central Turkey. The performance of the NARX model was evaluated using standard statistical models, and results are found to be good according to preliminary results. Thus, our study will better understand natural and anthropogenic factors affecting Turkey's groundwater resources.