MACH-Flow

Motivation

Accurately predicting river discharge, in particular at ungauged locations, is crucial for water resource management as well as drought and flood monitoring. Traditional hydrological models are generally calibrated on data from a single river catchment. They perform quite well when forecasting at this particular location (satisfactory temporal prediction) but they often generalize poorly at other locations (unsatisfactory spatial prediction). The main goal of this project was to develop models which can predict in space as well as in time. To this end we made use of data from many river gauging stations within a river network and modeled them jointly.

Proposed Approach / Solution

We developed two models in parallel. The first one is a recurrent neural network with long short-term memory (LSTM) cells, following the hydrological state-of-the-art architectures. Here, all gauging stations are considered independent and they share parameters (Figure 1). An extension we implemented in a follow-up paper involved a loss function based on a generalization of the continuous ranked probability score (negative energy score, tailored for multivariate probabilistic forecasts). This allowed to build ensemble predictions which quantify predictive uncertainty in a reliable way. The second model builds on first principles with an additive model with ridge-penalized cubic B-splines. This model includes engineered features for both short- and long-term effects of precipitation and temperature. We developed an additional independent module which can incorporated to this additive model. This module routes water through a river network, by encoding the latter as a directed graph. We have illustrated with a benchmark hydrological model that such spatio-temporal routing can only improve predictions, and this mainly for more downstream locations (Figure 2). A future research avenue, beyond MACH-Flow, is to combine this water routing module (as graph-based convolutions) with a recurrent neural network.

Impact

First, our LSTM neural network is on par with state-of-the-art hydrological models for Switzerland, albeit with much less inputs. This allowed us to reconstruct river discharge at a national scale back to the early 1960s, which is the earliest time for which precipitation and temperature data products are available. This reconstruction is important for the hydrological community as its low computation cost allows one to test various climate change scenarios. Second, the routing model, released as an independent module, is useful for the community as it can be easily added to any local prediction model to improve downstream predictions. Compared to other routing schemes in the hydrological literature, ours require minimal data inputs and accommodates any space discretization (regular grids or others). Also, it can improve predictions even with discharge data only available at a single station. Finally, a version of our additive model has been added to the ensemble of models used operationally on the drought monitoring platform drought.ch maintained by our partners at WSL.