Constraining a lumped rainfall-runoff model with piezometry to improve low-flow simulation

Most lumped hydrological models are focused on the rainfall-runoff relationship, since climatic conditions are the driving force of the hydrological behaviour of a catchment. Many hydrological models, like the ones used by the French national PREMHYCE platform, only take climatic variables as inputs – daily rainfall and potential evaporation – to simulate and forecast low-flows. Yet, a hydrological drought is generally a medium- to long-term phenomenon, which is the consequence of long records of dry climatic conditions. Daily lumped hydrological models often struggle to integrate these records to reproduce catchment memory.

In many French catchments, it was observed that this memory of past hydroclimatic conditions is well represented in piezometric signals that are broadly available over the national territory. Indeed, aquifers, especially the large ones, do store water on the long, feeding rivers during droughts: aquifers are not only water carriers – the etymology for the word aquifer – they are also memory carriers. A dataset of 108 catchments, each of them being associated with one or several piezometers, was used to investigate whether the GR6J daily lumped rainfall-runoff model could be constrained by piezometric time series to improve low-flow simulations. We found that a particular state of the model, the exponential store, is particularly well correlated with piezometry in most studied catchments.

In order to get a univocal relationship between the exponential store and piezometry, a multi-objective calibration approach was implemented, optimising both (i) flow simulation with a criterion focused on low-flows and (ii) affine correspondence between the exponential store level and piezometry. For that purpose, a new parameter was added to the model. The modified calibration was then evaluated through a split-sample test and the performance in simulating particular drought events. The calibrated store-piezometry relationship can now be used for data assimilation to improve low-flow forecasting.