Author(s): Amos Agossou; Jeong_Seok Yang; Jae Beom Lee

Linked Author(s): Amos Agossou

Keywords: GRI; ML; RNN; LSTM; Groundwater drought

Abstract: In recent years climate change has been responsible for many floods and droughts. it is important to take protective measure against these. One of these measures is future potential groundwater drought forecasting system in which groundwater level is predicted multiple time ahead. Although several research have applied machine learning (ML) techniques to predict flood or drought, there is lack of clearance regarding which technique is most reliable in groundwater level prediction. the present study aims to investigate the prediction of future potential groundwater drought from historical groundwater level and rainfall data using ML. From several existing ML techniques, the study adopted Long Short-Term Memory (LSTM) network, which is a type of Recurrent Neural Network (RNN) that readily reflects time-series data. The same network is built in two different works where the first work presents LSTM for groundwater (GW) level forecasting using only previously observed GW level data as the input without resorting to any other type of data and information about a groundwater basin. The second work has relied on a variety of inputs such as pumping rates and precipitation with only GW level in output. The model is trained over 18 years of data, tested, and validated over 5 years of the most recent record to predict multi step time ahead GW level and the accuracy of the model is validated using coefficient of determination and root mean squared error. During the process of calibration, a list of important model’s parameters such as: optimizer, activate function and epochs are modified one after another and the performance of the model is evaluated after a single modification. Both LSTM model predict 4 lags and up to 30 lags ahead GW level with an accuracy (coefficient of determination) about 90% and Root Mean Square Error about 0.012m. A comparison between both works proved that the LSTM with a variety of input gives better result for GW level prediction only when the correlation between GW level and related input parameter exceeds 60%. This paper’s results demonstrate the superiority of the LSTM with multivariate input (input parameters should have a high correlation coefficient) over the LSTM with only one input for groundwater level prediction and the success of the LSTM in reliable groundwater level prediction. From the predicted GW level, potential future groundwater drought is studied using groundwater resource index (GRI). The quality of this work’s results proves the capacity of ML in groundwater prediction, and affirms the importance of gathering high-quality, long-term, precipitation, pumping rates and GW level data for predicting key groundwater characteristics useful in sustainable groundwater management.

DOI: https://doi.org/10.3850/IAHR-39WC2521711920221278

Year: 2022