Short-term and Long-term Groundwater level forecast by applying conceptual and neural networks approaches.

Aquifers are crucial resources for mitigating the impact of droughts, which are expected to exacerbate in the future. Nevertheless, in many cases, there is not enough monitoring data to define distributed models to forecast future potential groundwater (GW) levels. In this work, we apply and compare different approaches for short-term predictions of GW levels. We use conceptual models (e.g., AQUIMOD, EMM, etc) and machine learning techniques (e.g., NAR, NARX, ELMAN, LSTM and/or GRU). We follow the next steps: calibration/training, validation and testing of the behaviour of conceptual models and neural networks for short-term prediction. We consider exogenous variables (precipitation, temperature, recharge, etc.) for short-term prediction. Multiple series of exogenous variables have been generated by using a stochastic weather generator, which will be used to perform a stochastic forecast. The results will be analysed for dry, medium and wet seasonal horizons. The predictions made with "machine learning" are compared with those generated by the conceptual models. In order to assess potential impacts of Climate Change on GW levels, we simulated some simulated some future local climate scenarios within the conceptual models. We analyzed the robustness of the results and their uncertainty. The risk of droughts has also been studied by evaluating the severity of droughts from the series generated by applying the “SPI” indices to the generated series. The method has been applied in two aquifers, namely Campo de Montiel (Center Spain) and Vega de Granada (Southern Spain).

 

Acknowledgments: This research has been partially supported by the projects: STAGES-IPCC (TED2021-130744B-C21) and SIGLO-PRO (PID2021-128021OB-I00), from the Spanish Ministry of Science, Innovation and Universities.