Long-run forecasting surface and groundwater dynamics from intermittent observation data: An evaluation for 50 years

The accurate prediction of water dynamics is critical for operational water resource management. In this study, we propose a novel approach to perform long-term forecasts of daily water dynamics, including river levels, river discharges, and groundwater levels, with a lead time of 7-30 days. The approach is based on the state-of-the-art neural network, bidirectional long short-term memory (BiLSTM), to enhance the accuracy and consistency of dynamic predictions. The operation of this forecasting system relies on an in-situ database observed for over 50 years with records gauging in 19 rivers, the karst aquifer, the English Channel, and the meteorological network in Normandy, France. To address the problem of missing measurements and gauge installations over time, we developed an adaptive scheme in which the neural network is regularly adjusted and re-trained in response to changing inputs during a long operation. Advances in BiLSTM with extensive learning past-to-future and future-to-past further help to avoid time-lag calibration that simplifies data processing. The proposed approach provides high accuracy and consistent prediction for the three water dynamics within a similar accuracy range as an on-site observation, with approximately 3 % error in the measurement range for the 7 day-ahead predictions and 6 % error for the 30 d-ahead predictions. The system also effectively fills the gap in actual measurements and detects anomalies at gauges that can last for years. Working with multiple dynamics not only proves that the data-driven model is a unified approach but also reveals the impact of the physical background of the dynamics on the performance of their predictions. Groundwater undergoes a slow filtration process following a low-frequency fluctuation, favoring long-term prediction, which differs from other higher-frequency river dynamics. The physical nature drives the predictive performance even when using a data-driven model.