Hybrid modelling of piezometric head – a large sample test

This work aims at reliably modeling water table fluctuations, in an operational groundwater management perspective. It is based on a large set of ca 100 piezometers, representative of the hydrogeological diversity of French groundwater exploitations (mostly aquifers presenting dual-porosity dynamics, often located in the phreatic domain).

This study, a component of the Water Resources Forecast SUEZ’s project, partially funded by the French Ecological Transition Agency (ADEME’s innov’eau initiative), compares several approaches for the modelling of daily piezometric head and its fluctuations induced by recharge:

• a conceptual model (derived from an existing rainfall-runoff hydrological model, whose ability to reproduce piezometric time-series — through one of its conceptual reservoirs);
• a classic AI approach (a non-parametric and data-based method using random forest algorithms applied to data-engineered features, e.g. rolling sums of meteorological inputs);
• a few hybrid approaches resulting from various combinations of the two previous solutions. 

All the above mentioned methods use daily meteorological data (precipitation and evapotranspiration time series) as inputs for the modeling chain. Model efficiency is assumed at the daily step using the Nash-Sutcliffe efficiency criterion, over three distinct 5-years periods where observed piezometric time-series are available.

Based on our results, we discuss the potential of using hybrid models for short-, medium- and long-term operational forecasting.