Multi-step-ahead forecasting of groundwater levels using a Long Short-Term Memory based Encoder-Decoder (LSTM-ED) model

Abstract

Operational forecasts of groundwater levels provide critical real-time knowledge during extreme events, such as floods and droughts. This study proposes a Long Short-Term Memory based Encoder-Decoder (LSTM-ED) model for multi-step-ahead groundwater level forecasting. The LSTM-ED is a well-suited architecture for sequence-to-sequence modelling tasks but has not yet been applied to forecast groundwater levels. The proposed LSTM-ED model is designed in the context of the Danish online monitoring system grundvandsstanden.dk to serve as operational groundwater level forecasting system. In the encoder LSTM model, sequences of past precipitation, temperature and groundwater levels are processed to initialize the decoder LSTM model which, in addition takes in forecast sequences of precipitation and temperature to output a sequence of groundwater levels. We train LSTM-ED models individually for each well, with all data aggregated to daily timescale. We demonstrate the performance of the LSTM-ED architecture for numerous wells from grundvandsstanden.dk and test varying lead times of up to 30 days. The LSTM-ED model forecasts are contrasted with simple benchmark models as well as with a sequence-to-sequence LSTM model that does not incorporate forecasts of precipitation and temperature for outputting the groundwater sequence. Initial results underpin that integrating forecasts of precipitation and temperature is a crucial component, especially for wells with shallow intakes where surface and sub-surface processes are well connected. The sequence-to-sequence LSTM model yields similar accuracy as the simple benchmark models, whereas accuracy clearly improves for the LSTM-ED model. Overall, this study highlights the potential of LSTM-ED models as an operational tool for multi-step-ahead forecasting of groundwater levels.